In a recent report with Stephen Rose, I argued that the narrative of a “shrinking middle class” was based on a kernel of truth, but one that undermines economic pessimism. We showed that while 36 percent of families were part of what we called the “core middle class” in 1979, the share had fallen to 31 percent by 2024. However, the share of families who fell short of the middle class shrank even more. The middle class has not been hollowed out; rather, the overall decline stems from the net movement of families upward into the upper-middle class. That group, with incomes between 5 and 15 times the 2024 federal poverty guidelines, rose from 10 percent of families in 1979 to 31 percent in 2024.

Analyses that find a hollowed-out middle invariably rely on definitions of the middle class that peg thresholds to how the typical family is doing. In that case, even if everyone is better off over time in inflation-adjusted terms, if the middle’s gains are stronger than those of families lower down, more people can fall short of “the middle.” The Pew Research Center, for example, found that the share of families that were “lower-income” rose between 1971 and 2023, even though the purchasing power of those lower-income families rose by 55 percent. The explanation for this seeming paradox is that “middle-income” families saw a 60 percent gain, making it harder to reach the middle-income threshold if income rose more slowly than that.

The point of my paper with Rose was that claims of a “hollowing out” of the middle class wrongly reinterpret widespread gains across the income distribution as rising insecurity and declining living standards. Unbeknownst to us, a perfect example of this misinterpretation appeared a week before we published our report in Chicago magazine. The offending article title blared that “Chicago’s Middle Class Is Disappearing.” My reanalysis of the data behind the piece indicates it would be difficult to articulate a more misleading conclusion. Fewer Chicagoans live in middle-class neighborhoods than in 1970—but only because more live in richer neighborhoods.

Subscribe now

The Chicago article illustrates “the trajectories of the rich and poor in Chicago” by describing homeless persons panhandling around the corner from $800,000 houses. It quotes a demographer asserting that, “The middle class is declining because the top 10 percent is swelling…and the bottom 20 percent is swelling.” The piece concludes, “As America races toward extremes of rich and poor, Chicago is racing even faster.”

The piece included no hard numbers on trends (nor did it explain how the bottom 20 percent can ever “swell” to more than one in five people). But it linked to another article on the local public radio website (WBEZ) cheerily titled, “The Middle Class Is Shrinking Everywhere — In Chicago It’s Almost Gone.” Both included maps created by the University of Illinois at Chicago’s Nathalie P. Voorhees Center for Neighborhood and Community Improvement, which paint an alarming picture:

Source: Nathalie P. Voorhees Center for Neighborhood and Community Improvement, College of Urban Planning and Public Affairs, University of Illinois Chicago, https://voorheescenter.uic.edu/chicago_communities/.

The growth of very high-income census tracts (which you can think of as large neighborhoods) is apparent in the spread of dark blue areas. But what really stands out is the spread of very low-income census tracts (red/orange). The beige middle-class tracts do, indeed, seem to disappear. The Voorhees Center data indicate that the share of Chicago tracts that were middle-income fell from 50 percent in 1970 to just 16 percent in 2017. Meanwhile, the share of lower-income tracts rose from 42 percent to 62 percent, and the share of higher-income tracts rose from 8 percent to 22 percent. Sounds pretty bad!

This bidirectional exodus from the middle is what populists on both the left and the right have in mind when they bemoan the hollowing out of the middle class. But the Voorhees Center methodology has the same shortcoming as Pew’s analyses of the shrinking middle class. Both define middle-class status relative to a benchmark that changes over time and is tied to typical contemporary income. If everyone’s income doubles, the middle class is no larger, yet everyone’s income has doubled.

In our paper, Steve and I argue that, for the purpose of assessing how much better off families are, using absolute thresholds to define classes makes far more sense. Everyone’s income doubling moves people out of lower classes and into higher ones, accurately depicting the rise in living standards.

To assess the claims around the Voorhees Center data and see how the conclusion changes when using absolute thresholds, I turned to two artificial intelligence (AI) assistants—the first time I have relied on AI for data analysis. I primarily used Claude, Anthropic’s assistant, and relied on OpenAI’s ChatGPT to check the code Claude produced.

I first set out to replicate the original analysis. The methodological details for the Voorhees Center analysis are sparse. The thresholds defining the five classes are shown in the graphic above and in other graphics online. The researchers estimated per capita income for each census tract in the city. They then compared each tract’s income to the per capita income for seven counties in the Chicago metropolitan area. (These seven counties only comprise part of today’s metro area, which includes 14 counties, and only six of the seven were in the 1970 definition of the metro area. Planning agencies in Chicago generally refer to the seven-county “metropolitan region.”) The Voorhees Center researchers created five classes, separated by thresholds pegged to percentages of metro income. For our purposes, the key thresholds define the middle class: per capita income of at least 80 percent but no more than 120 percent of the metro per capita income.

Except for the 2017 results, their estimates are based on decennial census data aggregated at the tract level. Census tract boundaries can change over time, but the Brown University Longitudinal Tract Database (LTDB) holds them fixed so that shifts in the tract definitions don’t affect results. The Voorhees Center analysts held the boundaries at their 2010 definitions. The 2017 estimates are from the American Community Survey (ACS), conducted annually, with tract-level totals based on five years of aggregated data. Those tract-level estimates already reflect 2010 boundaries.

Claude obtained the 1970 tract estimates for the Chicago metro area from the LTDB, 2017 estimates from the 2013-2017 ACS data on the Census Bureau website, and 2010 Chicago city and census tract shapefiles (for mapping) from the Census Bureau. It produced Python scripts to analyze the data and create census tract maps for 1970 and 2017. As with fully human analyses, the process involved numerous iterations to correct mistakes and refine the methods, aided by ChatGPT’s review of the Python code and this human’s substantive knowledge. The code is available on request, but I can’t interpret it for you as I don’t know Python!

On to the results. According to a chart attributed to the Voorhees Center, in 1970, 42 percent of tracts in the city of Chicago were “lower income” (combining the “low income” and “very low income” categories), 50 percent were “middle income,” and just 8 percent were “higher income” (combining “high income” and “very high income”). I believe that, rather than these percentages being the share of tracts, they are the percentage of Chicago’s population living in these tracts. My replication attempt estimated these figures at 42 percent, 51 percent, and 7 percent—extremely similar.

Moreover, the map that Claude produced (on the right) also very closely resembled the original Voorhees Center map (on the left):

Chicago census tracts by income class, 1970. Shading ranges from red/orange (very low income, or under 60% of the metro average individual income) to salmon (low income, or 60-80% of the metro average) to beige (middle income, or 80-120% of the metro average) to teal (120-140% of the metro average) to dark blue (above 140% of the metro average). The original is on the left; my replication attempt is on the right. Original is from the WBEZ website, based on Jessica Kursman and Nick Zettel, “Who Can Live in Chicago? Part I,” Voorhees Center for Neighborhood & Community Improvement, June 6, 2018. Unpopulated tracts are omitted.

The replication was also successful for 2017. Whereas the Voorhees Center found the lower-income share rising to 62 percent, the middle-income share falling to 16 percent, and the upper-income share rising to 22 percent, my replication produced estimates of 60 percent, 19 percent, and 22 percent. Here are the maps:

Chicago census tracts by income class, 2017. Shading ranges from red/orange (very low income, or under 60% of the metro average individual income) to salmon (low income, or 60-80% of the metro average) to beige (middle income, or 80-120% of the metro average) to teal (120-140% of the metro average) to dark blue (above 140% of the metro average). The original is on the left; my replication attempt is on the right. Original is from the WBEZ website, based on Jessica Kursman and Nick Zettel, “Who Can Live in Chicago? Part I,” Voorhees Center for Neighborhood & Community Improvement, June 6, 2018. Unpopulated tracts are omitted.

Reassured of Claude’s ability to analyze the data, I next determined how the size of the middle class changed when using absolute rather than relative thresholds. I determined the income levels corresponding to the 1970 thresholds, which were $20,942, $27,923, $41,885, and $48,865 (in 2025 dollars). I used the Personal Consumption Expenditures price index for the inflation adjustment. Note that the 1970 census data actually record income from 1969.

Then I used the same real (inflation-adjusted) thresholds for 2017. That means that instead of “middle income” requiring 80 to 120 percent of the 2017 metro average income, it requires 80 to 120 percent of the 1970 metro average income (adjusted for inflation to keep income in terms of constant purchasing power). Using this approach, the share of people living in middle income tracts fell in half from 1970 to 2017—from 51 percent to 25 percent. The share living in tracts below the middle income was roughly constant—42 percent in 1970 and 43 percent in 2017. In contrast, the share living in tracts above middle income more than quadrupled, rising from 7 percent to 31 percent. The resulting maps (1970 on the left, 2017 on the right) show significant growth in the number of tracts above and below middle income. However, the tracts below the middle income lost population. The number of lower income tracts spread, but fewer people lived in them.

Chicago census tracts by income class, 1970 and 2017, using fixed class thresholds. Shading ranges from red/orange (very low income, or under 60% of the 1970 metro average individual income) to salmon (low income, or 60-80% of the 1970 metro average) to beige (middle income, or 80-120% of the 1970 metro average) to teal (120-140% of the 1970 metro average) to dark blue (above 140% of the 1970 metro average). The 1970 map is on the left; the 2017 map is on the right. Unpopulated tracts are omitted.

Remember that “2017” here actually refers to 2013-2017, which includes the middle of the economic recovery following the Great Recession. The national unemployment rate was 4.9 percent in 1970, while it averaged 5.6 percent from 2013 to 2017. If we instead use 2015-2019 as the end point (average national unemployment rate of 4.4 percent), the middle income share falls from 51 percent to 26 percent, the lower-income share falls from 42 percent to 36 percent, and the higher income share jumps from 7 percent to 38 percent.

Finally, we can use the most recent data available, which are from 2020 to 2024. Over this period, unemployment averaged 4.9 percent—the same as in 1970. The analysis of the ACS data requires converting Chicago’s 2020 census tract boundaries back to the 2010 definitions, which Claude did using a Census Bureau crosswalk file, apportioning income and population from 2020 tracts to 2010 tracts based on their geographic overlap. Only about 8 percent of Chicago tracts had to be split, while another 8 percent were combined.

From 1970 to 2024, the share of Chicagoans who lived in middle income tracts fell from 51 percent to 25 percent. The share living in tracts falling short of middle income dropped from 42 percent to 28 percent. Meanwhile, the share living in upper income tracts rose sevenfold—from 7 percent to a whopping 48 percent. Looking at the top group, very high income tracts were home to just 4 percent of Chicagoans in 1970 but 38 percent in 2024.

Chicago census tracts by income class, 1970 and 2024, using fixed class thresholds. Shading ranges from red/orange (very low income, or under 60% of the 1970 metro average individual income) to salmon (low income, or 60-80% of the 1970 metro average) to beige (middle income, or 80-120% of the 1970 metro average) to teal (120-140% of the 1970 metro average) to dark blue (above 140% of the 1970 metro average). The 1970 map is on the left; the 2024 map is on the right. Unpopulated tracts are omitted.

The commentators wielding the Voorhees Center data are not wrong when they declare that Chicago’s middle class is shrinking. However, it doesn’t follow that everyone now is either struggling or affluent. Fewer Chicagoans live in middle income neighborhoods today because more of them live in upper income neighborhoods. Happily, fewer people also live in lower income neighborhoods than in the past.

Researchers who allow class thresholds to rise as incomes rise get results that reflect not just changes in living standards but also changes in inequality. Surely both trends are important, but they are separate aspects of economic well-being. If the rich were getting richer while the middle class and poor were getting poorer, we might worry that rising inequality was hurting living standards. However, what has really happened is that the entire distribution of income has shifted upward—just more at the top than at the middle, and more at the middle than at the bottom. The rich have gotten richer, and the middle and poor have gotten richer.

To simply characterize the result as a “disappearing middle class” or to think it necessitates asking, “Who can live in Chicago?” is to presume that rising inequality has left the middle class worse off and swelled the ranks of the poor. In reality, per capita income in the median Chicagoan’s census tract rose from $29,600 in 1970 to $39,300 in 2024 (both in 2025 dollars)—an increase of one-third. Using relative thresholds and letting class thresholds increase over time, the average Chicagoan in a lower income census tract lived in a tract with a per capita income of $22,300 in 1970 but $27,800 in 2024 (25 percent higher). For Chicagoans in middle income tracts, the increase was from $32,400 to $49,900 (54 percent). That is hardly the impression conveyed by headlines about a disappearing middle class.

If we ask whether neighborhoods meet the standard of middle class income that prevailed in 1970, Chicago looks like the nation. The middle is shrinking because neighborhoods are moving up, not because they’re being hollowed out. Nearly half of Chicago’s tracts now exceed what would have been the higher income threshold in 1970. Fewer than 30 percent fall below what would have been the middle income threshold.

None of this is to deny that Chicago, like many American cities, faces real challenges related to concentrated poverty, residential segregation, and uneven development. Some neighborhoods have indeed been left behind in both relative and absolute terms. But the aggregate picture is one of broad, if unequal, progress. Decrying a hollowed-out middle is just a gloomy way of acknowledging that the upper middle class has boomed.

One final note related to AI. The analysis here was both easier and more difficult than I expected. Initially, ChatGPT kept changing as it tried to produce my results. With the benefit of experience coaxing ChatGPT, I obtained maps of Chicago in Claude ridiculously easily, though ChatGPT then found a handful of fairly large errors in the Claude code. I recognized issues that neither assistant caught. Extra tokens were purchased and burned through.

Data analysis remains too tricky for AI-assisted amateurs, which should both reassure seasoned researchers about their own value and inspire fear about the damage that overconfident dilettantes can now inflict by telling Claude to answer an empirical question as if it were Alexa setting a timer.

But, man oh man. Having spent quite a lot of time on the Chicago project, it was simply stunning to then watch Claude produce these charts in under seven minutes after telling it,

Now repeat the 1970-to-2024 analyses for New York City (all 5 boroughs) using inflation-adjusted thresholds based on the same 1970 percentages of metro per capita income used in the Chicago analyses, but this time using the New York metro area (across multiple states, as it was defined in 1970). Provide the percentage distributions across the 5 income categories in both years and the 1970 and 2024 maps of New York City. Use all the same analytic and presentation decisions we made for the Chicago analyses. Provide the code.

New York City census tracts by income class, 1970 and 2024, using fixed class thresholds. Shading ranges from red/orange (very low income, or under 60% of the 1970 metro average individual income) to salmon (low income, or 60-80% of the 1970 metro average) to beige (middle income, or 80-120% of the 1970 metro average) to teal (120-140% of the 1970 metro average) to dark blue (above 140% of the 1970 metro average). The 1970 map is on the left; the 2024 map is on the right. Unpopulated tracts are omitted.

And if you’re curious, the share of New Yorkers in lower income, middle income, and higher income tracts went from 43/43/14 to 21/33/46 from 1970 to 2024. Fortunately, it looks like the middle class really is shrinking everywhere.

This week, Michael Brendan Dougherty wrote that he doesn’t think he’s “ever been so depressed” as when he read my recent report with Steve Rose, “The Middle Class is Shrinking Because of a Booming Upper-Middle Class.” It found, well, that the middle class is shrinking because of a booming upper-middle class. Dougherty’s depression was inspired not by that conclusion but by his conviction that it was misleading — masking the lousy state of the middle class. But when we take a close look at the middle class, there are definitely reasons why Dougherty should cheer up.

(I didn’t ask ChatGPT to make it kitsch, but I like it. Look at that floating %.)

Dougherty lays out what, to him, are “the trends behind” my study’s findings. The first of these is that “we have traded some economic security for dynamism.” His evidence for this claim, he says, is “the share of workers jumping directly from one employer to another in a year rising from roughly 6 percent of the labor force in the late 1970s to nearly 9 percent by the late 1990s.”

I’ll confess I couldn’t determine where these numbers came from. But whatever one should make of job-to-job transitions as an indicator of economic insecurity, they have likely fallen. One measure of job-to-job flows is the share of workers who hold two or more jobs over a year’s time. That share fell from about 16 percent in 1979 to 15 percent in 1989 to about 12 percent in 2006 (all relative peaks for their business cycles). The share of workers who were in another job one month later fell from 1997 to 2013. Other research shows a drop in job-to-job transitions from 1975 to 2014, from 1979 to 2015, from 1988 to 2013, from 1990 to 2011, from 2000 to 2010, from 2000 to 2016, and from 2000 to 2020. A couple of studies show a flatter trend from 1992 to 2003 and from 1994 to 2006. The only paper I found that shows an increase in job-to-job transitions also finds that the overall job separation rate fell — that is, transitions from a job to having no job fell by more than the increase in job-to-job transitions. That hardly suggests increased insecurity.

Dougherty’s claim that dynamism has increased goes against what is essentially unanimity among economists that dynamism has been in decline since at least the 1970s. Nearly 15 years ago, many commentators were as sure as Dougherty is now that economic insecurity had increased, but I showed that any changes were small in historical context. Americans rate their own personal finances about as highly as they did 25 years ago.

Subscribe now

The second issue Dougherty raises is his claim that “the overwhelming contributor to the growing upper-middle class is not higher productivity . . . It’s more hours worked, and at a higher wage.” Let me quickly just point out that according to data I assembled for an earlier paper, between 1979 and 2022, net productivity (which excludes depreciation — don’t ask) in the nonfarm business sector rose 97 percent while hourly wages rose 85 percent. So it doesn’t exactly feel like productivity growth was unimportant. More good news!

Furthermore, it’s not the case that only women have seen wage gains while men “saw stagnant or modest growth in their wages” and deterioration in their employment. Rather, the numbers behind Figure 2 in this paper of mine indicate that the median wage of men ages 25 to 54 rose by 16 to 29 percent from 1989 to 2023 (and I’d advocate hard for the 29 percent as the better number). One can wish that increase was stronger, but it nevertheless means that men are better off than ever.

And it’s not at all clear that the economy is the main villain here. Part of the downshift just reflects that men had a monopoly on the best jobs until women’s opportunities opened up. Moreover, the decline in marriage and fertility may have played a strong role. In the mid-20th century, when men were far more likely to be sole breadwinners, there was more pressure on them to take the highest-paying job they could find and to stay in that job through thick and thin (while hoping for a promotion). As fewer men over time were husbands or fathers, that pressure declined. Men could take lower-paying jobs, leave jobs they didn’t like, and take longer to find new jobs than in the days where they were solely responsible for a family. The increase in earnings among wives also alleviated these pressures.

While many populists remember the 1980s fondly, median male wages fell from 1979 to 1989 while rising thereafter. No one wants to admit it, but Boomer men (not Millennials or Gen Z) have had the worst wage trends.

Male employment has been falling for a long time; the decline in male labor force participation goes back at least to the 1950s and probably to the 1930s. Moreover, relatively little of this decline concerns men who tell government surveyors they want a job, and little of it concerns men who cite the state of the economy as the reason for their nonwork. From 1967 to 2019, only 12 percent of the decline in employment for men ages 25 to 54 was due to inability to find a job — slightly higher than the share accounted for by early retirement, but lower than the share accounted for by increased school attendance and by growth in househusbands. Declining work among those saying they are disabled or sick is a big part of the story, though less so for the past 25 years. It’s unclear how to interpret this trend, but it’s fairly clear much of it is tied to policy reforms in the 1970s that made federal disability programs more accessible and generous.

Dougherty is certainly right that women work more hours than in the past. But it’s not clear how important this has been for propping up family income. As Jeremy Horpedahl’s analyses suggest, much more important than rising hours among wives has been their higher hourly pay as their opportunities have increased. And while Dougherty asserts that the “number of personal hours spent in leisure at home or maintaining the home has dramatically gone down as well,” the best work on this question indicates that leisure increased among both married men and married women between 1965 and 2003. Among men and women with children (married or not), leisure was also higher in 2003 than in 1965, though it may have been lower for women than in 1985.

Dougherty writes that there are vast swaths of women who have to “settl[e] for marriages across class or for single life.” But remember that men are doing better than ever in absolute terms. Also remember that men still tend to make more than women. The “problem” here is that if men or women will only marry if the husband makes some multiple of what the wife makes, fewer will marry today than in the past. The economy did not create this “problem,” except insofar as it has encouraged more women to prioritize earnings. Men aren’t going to make the same multiple of women’s earnings as they did 75 years ago without some massive affirmative action program for them or state and employer discrimination against women.

I don’t necessarily disagree with Dougherty about some of the costs of the dual-earner model, though I think the benefits have outweighed them and that it’s not my business to convince wives and husbands otherwise. I don’t necessarily disagree with Dougherty that Americans are chasing consumption of market goods too much relative to investing in family and community life, but I think this has little to do with increased economic insecurity or stagnation. I think the decline in marriage — reflecting the nation’s affluence rather than economic stagnation — is a much more concerning issue than the rise of two-earner families among those who do marry. Populists may not agree with the economic choices that free-thinking women and men have made in recent years, but they have the right to make those decisions — and they’re not being coerced by a failing economy to do so.

This example of Cass withholding information from his readers in order to make a rhetorical point resonated with me because I had just experienced another example of it—the attribution to me by Cass of a finding I do not stand behind. Not only did I tell him multiple times that he was using out-of-date results of mine, I pointed him to updated results, created additional ones, and even provided him with the data for the new numbers. Here’s the story, complete with the email exchange that closed it out.

Subscribe now

Our saga begins with a chart in a paper I wrote and ends with a sentence in a new American Compass report citing me. In late 2022, I wrote Bringing Home the Bacon, which examined whether the evolution of young men’s earnings could explain the sharp decline in sole-breadwinner families or the dramatic increase in single motherhood. Many populists argue that a deterioration in men’s economic standing has led to these changes. My report showed that real median annual compensation among young men was essentially the same in 2019 as in 1969 and that by various “marriageability” thresholds, young men were “at, near, or above historic highs.” That ruled out declining male earnings an explanation for the striking changes in the family that occurred over this period.

In my paper, I made a number of conservative methodological choices because I wanted to show that male marriageability had not declined even using methods that worked against that result. Nevertheless, in public events, podcasts, and even the inaugural post for his “Understanding America” Substack, Cass highlighted that young men’s earnings were lower or no higher than “50 years ago.” He did so again during our 2024 debate on the state of the economy.

After the latter, I took to X to share some updated results that I didn’t get a chance to mention in the debate. I indicated that, using an improved price index that I had developed earlier that month, young men’s real median post-tax compensation rose 20 percent from 1973 to 2019, or $7,200, and rose 24 percent ($8,500) from 1989 to 2019. Optimistically, I wrote, regarding whether young men’s earnings have stagnated over 50 years, “I’ll trust my chart doesn’t get cited anymore in support of that claim!” I also stated unambiguously that, “In case it’s not clear, the chart [showing stagnant earnings] was what I considered the best evidence then, but it is not the best evidence now. You [Cass] can still cite it, obviously, but you should either say why you think it is still the best evidence or clarify that you don’t care.”

No response was forthcoming, but Cass returned to the question one year later, asking whether I’d updated Bringing Home the Bacon yet. I replied it was “on my list,” but noted that in the meantime, I had produced updated estimates of the median pre-tax earnings trend for men ages 25-29 and that they indicated an increase of 21 percent from 1973 to 2023. The increase from 1989 to 2023 was 25 percent. I provided a link to that 2025 paper, Don’t Choose Your Own Adventure, which was published at Civitas Outlook, on my Substack, and on AEI’s site.

Despite the new published results, on Monday, February 23 of this year, Cass reached out to AEI’s communications team to request permission to reproduce for a forthcoming American Compass report the chart from Bringing Home the Bacon showing earnings stagnation. Because I no longer believe that that chart provides the most accurate available depiction of young men’s earnings trends, I indicated to our comms team that while I would not grant permission, I would publish a new version of the chart, using the improved methods from Don’t Choose Your Own Adventure, and he could use that one. My understanding was that he needed a chart by the end of the week.

My new chart differed from the earlier one in three ways. I extended the earnings trends from 2020 to 2024. I refined the inflation adjustment using the “More Accurate Consumer Price Index,” or MACPI, which I developed in 2024. And I switched from including all young men (workers and nonworkers alike) to only year-round workers, based on extensive analyses in Don’t Choose Your Own Adventure. (The latter choice isn’t that consequential relative to other defensible ways of dealing with nonworkers and part-year workers.)

I wrote a piece describing my updated estimates and including a new chart developed for Cass’s use. Notably, the piece also provided results that simply extended the same analyses from the old paper to 2024, finding that while median earnings were little different in 2024 than in 1973, they rose by 15-22 percent from 1989 to 2024 (or $6,800-$9,100). Any decline had ended 30 years ago. At the very least, then, Cass could have cited these estimates, perhaps explaining why he objected to my preferred ones. But even these estimates aren’t fairly described as “stagnation” without noting that a period of decline has been followed by three decades of non-negligible gains. (Of course, that pattern is inconsistent with the China Shock having harmed American workers, a central part of many American Compass narratives.)

My piece went into a fair amount of detail on my improvements to the original Bringing Home the Bacon estimates, citing the 17,500-word 2024 report I wrote justifying my new price index and the 4,400-word Don’t Choose Your Own Adventure. (The latter’s appendix added another 3,100 words.) I concluded that the median earnings of young men had risen 24-40 percent between 1973 and 2024 and 42-46 percent between 1989 and 2024. This does not constitute stagnation.

AEI published my new piece early the afternoon of Thursday, February 26. I also published it to my Substack. At some point later that day, my comms person emailed Cass, sending him a link to the AEI version of the piece. We also gave him an image file of the new chart and an Excel file with the chart and the numbers behind it.

Alas, before the weekend arrived, Cass indicated to my comms person that he would just refer to the numbers in the 2022 report without reproducing any chart or using any of the updated estimates.

After sitting on it for the weekend, I sent the following email to Cass on Monday evening, March 2. It’s a long one, because I wanted to make sure I was clear about my position on the earnings trend. Anything in brackets below has been added here and was not part of the original email.

Hi Oren,

I was happy to hear last week of your interest in reproducing a chart of mine. Unfortunately, I understand that you’ve chosen not to reproduce the chart I’ve just published that improves on the one in which you were interested (Figure 5 from my 2022 report). Could I ask why? It was my understanding that you needed a chart by the end of last week, which led me to accelerate my work updating the 2022 report and publish something publicly by your deadline.

As I say in the new piece, posted in two different places on the AEI website, I view the new chart as a clear improvement on the one in the 2022 report, which I am in the process of updating. I spent weeks (months, actually) researching price indexes, producing the 2024 paper to which I refer and link in the new piece. I went so far as to try to validate the MACPI last year by comparing various estimates of absolute mobility using different price indexes against subjective data on absolute mobility from public opinion surveys (also mentioned in the new piece). [That piece is available here.] Both the evidence on price measurement and the validation exercise point toward the superiority of the MACPI. The MACPI paper is filled with references to academic and government economists who have argued that the CPI family of indexes and the PCEPI are biased in the direction of indicating too much inflation. [“CPI” is the Consumer Price Index for Urban Consumers, and “PCEPI” is the Personal Consumption Expenditures Price Index.]

I also spent weeks researching the question of how to deal with non-earners—I dare say that I’ve gone into more depth on this question than anyone ever has. I refer and link to that research in the piece too.

To be clear, when I am done updating the 2022 numbers and writing them up, AEI will publish a new version of the paper that will constitute an improvement on the 2022 one. I anticipate taking down the old paper and having the old URL redirect to the new version, and the new version will clearly indicate what has changed and why. It will also include a link to the archived 2022 paper for those who want to see the old, inferior numbers. I would be surprised if you want to link to the old version and have the reader who follows the link see estimates that conflict with your claims about what I find. I’d also be surprised if you instead prefer to link to an archived version of the old report if its author is telling everyone that it’s an inferior version of analyses that the author once viewed as the best evidence at the time but has since improved.

To be clear, the figure you have repeatedly cited to date always was presented as a conservative trend. I wrote that, “The trend in men’s lifetime earnings would likely show that men have done somewhat better over time.” I noted that the trends for older men get around some of the issues involved and showed that the trends of older men indicated rising earnings. I also wrote, “There are other reasons to think that the ideal pre-tax earnings measures would show a somewhat better trend than in Figure 5, as discussed in Appendix A.”

In Appendix A, I specifically mention the issue of how to deal with non-earners, saying, “There are good reasons for excluding at least some of these categories of non-earning men. However, the issues that would result from doing so are either unclear or would make earnings trends appear better than shown in the report. To be conservative, I leave them in the data.”

I also link to an earlier piece I wrote at the Manhattan Institute to justify using the PCE deflator to adjust for inflation. That older report notes, “Using the PCE looks like a very conservative choice for a cost-of-living adjustment.” [That piece is available here. See Appendix 2.]

In both instances, I went with conservative choices because I wanted to show that even making such choices, male marriageability has not declined. Having subsequently done more research into these questions (and seen my conservative choices be misconstrued by analysts), I have improved the methods. (Indeed, I think it is the case that the MACPI is a conservative choice in some regards, as my 2024 paper on the index indicates.)

Regardless, it’s surprising to me that having been given permission to use a more accurate version of the chart you wanted to reproduce--a chart published on AEI’s website with a clear explanation for its superiority over the older chart--and having been provided the actual estimates behind the new chart and a graphic file of the chart, you would choose not only NOT to use the more accurate chart but to cite the old numbers that the author has told you and the world are less accurate than the updated ones.

It’s all the more surprising to me that you have a preference for the numbers I have said are inferior because your cost-of-thriving research seems to reject price indexes all together as a way of accounting for the rising cost of living. I would have thought that you’d either reject ALL of my estimates as illegitimate or emphasize that while you disagree with conventional price indexes as a way to adjust for the cost of living, a believer in price indexes finds that there’s a substantial rise in men’s earnings. It seems inconsistent with your past work to say instead that YOU prefer one set of results over the one the author prefers, even though you reject the whole exercise of inflation adjustment.

It’s certainly your prerogative to choose as you’ve done, but it feels like an intellectually dishonest decision. I guess that’s just not what I would expect of the founder of a think tank seeking to help working Americans by aligning policies with the evidence. As ever, thanks for your interest in my work. Happy to answer any questions you have about these estimates. Best, Scott

Cass’s reply, midday on Tuesday, March 3, was brief:

Hi Scott,

I have never rejected the whole exercise of inflation adjustment and of course I regularly publish and cite inflation-adjusted data. My concern here, and the reason I will not use your newly published Substack chart, is that you are substituting your own personal inflation index for those provided by public statistics agencies. I consider that improper as a methodological matter and unwise as a political one. Not that you asked for my advice, but for whatever it’s worth, I genuinely believe it will be a mistake for you and for AEI to publish a revised version of your analysis that materially changes the answer by making an adjustment that you invented yourself.

I’ll note again that even if Cass insisted on citing numbers I deem inferior, he could have cited estimates from either Don’t Choose Your Own Adventure or my latest paper that included more recent years but otherwise used the same methods as the old numbers. I’ll also note Cass’s interjection of political considerations in a conversation about empirical methods, a revealing aside.

I sent a final reply, that afternoon, to which Cass did not respond.:

I mean, I don’t know what “Substack chart” means. I posted the piece to Substack after it was posted to AEI’s website. The links you were sent were to AEI’s website. Anway [sic], I’m not sure why you’d denigrate Substack publications given how, ahem, “Commonplace” it is to publish to that source. [This was a reference to the American Compass newsletter that is published on Substack.]

Your objection to my developing an alternative to official government price indexes would ring truer if you used them consistently and used the ones that government agencies advocated consistently (rather than the CPI-U for long-term trends). If you’re unaware of what those agencies have advocated, my MACPI piece goes into detail about what BLS, CBO, the Census Bureau, and the Federal Reserve Board have said about them. Is there other evidence that you’re going on when selecting which price indexes to use or reject? [“BLS” is the Bureau of Labor Statistics, and “CBO” is the Congressional Budget Office.]

It’s also frustrating as someone who wrote tens of thousands of words critiquing YOUR attempt at an alternative to government price indexes to just have you dismiss my effort without any substantive critique. [This was referring to my paper, with Jeremy Horpedahl, on Cass’s flawed Cost of Thriving Index.] What did I get wrong in the MACPI paper? Did I misrepresent the research of economists who study the topic? Is my validation exercise fatally flawed in some way? Rather than address the evidence, you seem all too willing to just go with the estimates that make the case you want to make—to engage in policy-based evidence making.

A more confident and honest researcher would accurately convey the conclusions of other researchers he cites rather than withhold those conclusions. Or, if he insisted on citing less accurate evidence that the original researcher has updated, he at least would indicate that the original researcher disputes the claim being made and say why.

I used to get more upset about this stuff, but we have each chosen our intended audiences, be they more or less interested in evidence. I doubt I can influence your core audience, and I think mine understands the nature of your project. Do as you will! Best, Scott

I’m saddened to report that American Compass’s new paper, on workforce development, was released on Thursday, March 5, and the foreword, coauthored by Cass, cited my old report:

For young men, especially, the situation has become dire. Their outcomes are worse at every stage of the conventional pipeline, and their earnings have now stagnated for half a century. Research published by the American Enterprise Institute’s Scott Winship in 2022 found that men between the ages of 25 and 29 earned lower compensation in 2020 than in 1970, both pre- and post-tax.

This paragraph didn’t so much as include a citation, let alone any caveat about my own updated views on this question. It included no reference to my just-released updated figures, nor to the updated figures in Don’t Choose Your Own Adventure. Cass simply cited a researcher’s no-longer-operative finding without including any of the original nuance the researcher offered, ignoring two updates to the results in question, and shrugging off the researcher’s privately and publicly stated view that the earlier estimates no longer reflect the best evidence.

Cass calls himself an economist, but he was trained as a lawyer. Economists (and other social scientists) are interested in the best evidence. Cass, like a skilled litigator, is interested in the best evidence that strengthens the case he has already decided to argue. Cass uses evidence not because he is interested in getting closer to the truth, as a strategy for finding effective policy solutions; he uses it to advance his perspective, unmoored though it may be to the best evidence and unhelpful though it may be to the people for whom he claims to advocate.

We’re also debuting our revamped website, which looks much cleaner than before. One new feature of the site is the “Spotlight on Work Incentives” page. There we’ve highlighted three previously published pieces each by Kevin Corinth, Robert Doar, Angela Rachidi, Matt Weidinger, and me on the importance of work incentives in evaluating safety net policies, as well as a few multi-authored pieces. Beyond these highlights, on the webpage you can find many more briefs, op-eds, essays, and reports we’ve written on the importance of work.

I’m (relatively) new to AEI, having arrived mid-2020, and I realized that some of what I’ve written in the past exists off the AEI website. I figured I’d list some of this research here, alongside the pieces highlighted on the COSM page. These pieces cover my evolution from a center-left sociology grad student who believed that the 1996 welfare reform would harm families to a center-right AEI scholar who believes work incentives are important enough to dedicate a Substack post to them. What a long, strange trip, etc…

Subscribe now

“How Did the Social Policy Changes of the 1990s Affect Material Hardship Among Single Mothers? Evidence from the CPS Food Security Supplement,” with Christopher Jencks, Harvard University John F. Kennedy School of Government Faculty Research Working Papers Series, No. RWP04-027 (June, 2004)

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=600601

“Welfare Reform Worked—Don’t Fix It,” with Christopher Jencks, Christian Science Monitor (July 21, 2004)

https://www.csmonitor.com/2004/0721/p09s02-coop.html

“Understanding Welfare Reform,” with Christopher Jencks, Harvard Magazine (November 1, 2004)

https://www.harvardmagazine.com/2004/11/understanding-welfare-re-html

“Welfare Redux,” with Christopher Jencks and Joe Swingle, The American Prospect (February 21, 2006)

https://prospect.org/2006/02/21/welfare-redux/

“Safety-Net Reforms to Protect the Vulnerable and Expand the Middle Class,” Room to Grow: Conservative Reforms for a Limited Government and a Thriving Middle Class, YG Network (May, 2014)

https://web.archive.org/web/20160911165338/http://conservativereform.com/wp-content/uploads/2014/05/Room-To-Grow.pdf

“Would a Block-Granted Safety Net Mean Less Aid to Families?” Forbes.com (July 31, 2014)

https://www.forbes.com/sites/scottwinship/2014/07/31/would-a-block-granted-safety-net-mean-less-aid-to-families/

“Will Welfare Reform Increase Upward Mobility?” Forbes.com (March 26, 2015)

https://www.forbes.com/sites/scottwinship/2015/03/26/will-welfare-reform-increase-upward-mobility/

“Welfare Reform Reduced Poverty And No One Can Contest It,” Forbes.com (January 11, 2016)

https://www.forbes.com/sites/scottwinship/2016/01/11/welfare-reform-reduced-poverty-and-no-one-can-contest-it/

“Welfare Reform’s Success and the War on Immobility,” Law and Liberty Forum (May 6, 2016)

https://lawliberty.org/forum/welfare-reforms-success-and-the-war-on-immobility/

Poverty After Welfare Reform, Manhattan Institute (August 22, 2016)

https://manhattan.institute/article/poverty-after-welfare-reform

“The Personal Responsibility and Work Opportunity Reconciliation Act: Success, Failure, or Incomplete?” presentation at the Welfare Reform Turns 20: Looking Back, Going Forward conference, Cato Institute (August 22, 2016)

https://www.cato.org/multimedia/events/panel-1-personal-responsibility-work-opportunity-reconciliation-act-success

“Why the 1996 Welfare Reform Benefited Poor Children,” National Review (September 1, 2026)

https://www.nationalreview.com/2016/09/welfare-reform-child-poverty-1996-law-poor-children/amp/

“Yes, the ’96 Welfare reform Helped Reduce Child Poverty,” National Review (September 7, 2016)

https://www.nationalreview.com/2016/09/welfare-reform-child-poverty-reduced-1996-prwora/

“The Romney Child Allowance Proposal Is a Move in the Wrong Direction,” AEIdeas, American Enterprise Institute (February 4, 2021)

https://www.aei.org/opportunity-social-mobility/the-romney-child-allowance-proposal-is-a-move-in-the-wrong-direction/

“The Conservative Case Against Child Allowances,” American Enterprise Institute (March 5, 2021)

https://www.aei.org/research-products/report/the-conservative-case-against-child-allowances/

“Do People Know What’s Good for Them? An Argument for Conditioning Cash Transfers,” RealClearPolicy (May 7, 2021)

https://www.aei.org/op-eds/do-people-know-whats-good-for-them-an-argument-for-conditioning-cash-transfers/

“Deprivation Is Not Simply a Material Matter,” Education Next (July 6, 2021)

https://www.aei.org/articles/deprivation-is-not-simply-a-material-matter/

“Reforming Tax Credits to Promote Child Opportunity and Aid Working Families,” American Enterprise Institute (July 29, 2021)

https://www.aei.org/research-products/report/reforming-tax-credits-to-promote-child-opportunity-and-aid-working-families/

“New Evidence on the Benefits and Costs of an Expanded Child Tax Credit,” AEIdeas, American Enterprise Institute (October 7, 2021)

https://www.aei.org/opportunity-social-mobility/new-evidence-on-the-benefits-and-costs-of-an-expanded-child-tax-credit/

“Did a National Academy of Sciences Committee on Poverty Reduction Overstate the Benefits and Understate the Costs of a Child Allowance?” American Enterprise Institute (December 16, 2021)

https://www.aei.org/research-products/working-paper/did-a-national-academy-of-sciences-committee-on-poverty-reduction-overstate-the-benefits-and-understate-the-costs-of-a-child-allowance/

“Second Time’s the Charm?” with Yuval Levin, National Review (June 15, 2022)

https://www.aei.org/op-eds/second-times-the-charm/

“How Much Would Creating a Child Allowance Reduce Work Among Parents?” American Enterprise Institute (December 9, 2022)

https://www.aei.org/research-products/working-paper/how-much-would-creating-a-child-allowance-reduce-work-among-parents/

“The True Cost of Expanding the Child Tax Credit,” New York Times (December 20, 2022)

https://www.aei.org/op-eds/the-true-cost-of-expanding-the-child-tax-credit/

“The Work Incentive and Employment Effects of Eliminating the Child Tax Credit’s Annual Income Requirement,” with Kevin Corinth, Angela Rachidi, and Matt Weidinger, American Enterprise Institute (January 19, 2024)

https://www.aei.org/research-products/working-paper/the-work-incentive-and-employment-effects-ofeliminating-the-child-tax-credits-annual-incomerequirement/

“Per-Child Benefit in Wyden-Smith Child Tax Credit Bill Would Discourage Full-Time Work for Families with Multiple Children,” with Kevin Corinth, AEIdeas, American Enterprise Institute (January 29, 2024)

https://www.aei.org/opportunity-social-mobility/per-child-benefit-in-wyden-smith-child-tax-credit-bill-would-discourage-full-time-work-for-families-with-multiple-children/

“How Sensitive Are Single Mothers’ Work Decisions to a Change in Incentives? Correcting Misperceptions of the Evidence,” with Kevin Corinth, COSM Commentary, American Enterprise Institute (January 30, 2024)

https://www.aei.org/articles/how-sensitive-are-single-mothers-work-decisions-to-a-change-in-incentives-correcting-misperceptions-of-the-evidence/

“The Wyden-Smith Child Tax Credit and Work: Responding to Critics,” with Kevin Corinth, COSM Commentary, American Enterprise Institute (January 31, 2024)

https://www.aei.org/articles/the-wyden-smith-child-tax-credit-and-work-responding-to-critics/

“Research by a Top Biden Administration Economist Reinforces the Importance of Work Incentives in the Child Tax Credit and the Safety Net,” COSM Commentary, American Enterprise Institute (February 6, 2024)

https://www.aei.org/articles/research-by-a-top-biden-administration-economist-reinforces-the-importance-of-work-incentives-in-the-child-tax-credit-and-the-safety-net/

“Another Flawed Analysis Shows That Single Mothers Are Highly Sensitive to Changes in Work Incentives,” COSM Commentary, American Enterprise Institute (February 8, 2024)

https://www.aei.org/articles/another-flawed-analysis-shows-that-single-mothers-are-highly-sensitive-to-changes-in-work-incentives/

“Options for Improving the Child Tax Credit Provisions in H.R. 7024, the Tax Relief for American Families and Workers Act of 2024,” with Kevin Corinth, Leslie Ford, Angela Rachidi, and Matt Weidinger, COSM Commentary (February 27, 2024)

https://www.aei.org/articles/options-for-improving-the-child-tax-credit-provisions-in-h-r-7024-the-tax-relief-for-american-families-and-workers-act-of-2024/

“An Early Look at the Child Tax Credit Changes in the Tax Relief for American Families and Workers Act of 2024,” with Kevin Corinth, American Enterprise Institute (March 28, 2024)

https://www.aei.org/research-products/report/an-early-look-at-the-child-tax-credit-changes-in-the-tax-relief-for-american-families-and-workers-act-of-2024/

“Critiquing Bastian (2022, 2023, 2024, and Forthcoming): On Child Tax Credit Reform and the Sensitivity of Single Mothers to Work Incentives,” American Enterprise Institute (May 16, 2024)

https://www.aei.org/research-products/working-paper/critiquing-bastian-2022-2023-2024-and-forthcomingon-child-tax-credit-reformand-the-sensitivity-of-single-mothers-to-work-incentives/

“Work Requirements For Medicaid could Increase Income and Reduce Poverty,” Civitas Outlook (June 17, 2025)

https://www.aei.org/articles/work-requirements-for-medicaid-could-increase-income-and-reduce-poverty/

Median pre-tax earnings of men ages 25-29 rose 24 percent ($10,800) from 1973 to 2024, and median post-tax compensation rose 40 percent ($15,500). The gains from 1989 to 2024 were 42 percent ($16,300) and 46 percent ($17,000)

A few years ago (2022), I wrote a report arguing that while fewer families are pursuing the traditional sole-breadwinner model than in the past, and while single parenthood has increased over the past 50 years, the problem is not that men have become less marriageable economically. I defined different marriageability thresholds related to the earnings of young sole-breadwinning married fathers in the past. The analyses showed that young men’s marriageability, at worst, was no lower in 2020 than in 1980. It was sometimes modestly lower than its peak in 1969 or 1973, but usually higher than in the early 1960s. I argued that rising affluence was the main reason that more women work today, fewer parents marry or stay married, and men work less and at less remunerative jobs.

Assessing long-run trends in earnings involves a number of important methodological choices. Since publishing the marriageability paper, I have conducted several deep analyses looking more narrowly at some of these options. In 2024, I reviewed how well conventional price indexes measure inflation. Based on research by academic and government economists over several decades, I concluded (like many others) that these price indexes tend to overstate inflation. That means that analyses like my marriageability study that adjust nominal earnings for inflation using a conventional price index tend to understate earnings growth over time.

My 2024 paper used the quantitative evidence from this research on price indexes to create a “More Accurate Consumer Price Index,” or “MACPI.” Relative to much of the past literature, my resulting adjustment was somewhat conservative; using other estimates of bias would have shown an even slower rise in prices than the MACPI indicates. I showed the implications for a variety of long-term economic trends of using the MACPI instead of other popular price indexes. I subsequently validated the MACPI by comparing objective trends in and levels of absolute mobility to subjective perceptions. The objective absolute mobility results were more consistent with subjective perceptions when using the MACPI to adjust for inflation than they were using the PCEPI.

In my marriageability paper, I had used the personal consumption expenditures price index (PCEPI). In the 2024 paper, I found that, relying on the PCEPI, the median earnings of men who worked full-time, year-round rose by 11 percent between 1973 and 2023. However, using the MACPI, earnings rose 38 percent.

Subscribe now

In 2025, I conducted a more extensive analysis of men’s and women’s median earnings over fifty years. This study paid particularly close attention to two methodological issues. First, I explored how to address non-workers and workers who only work part of the year. Much of this non-work or part-year work is voluntary (retirement, e.g.) or involves circumstances not closely related to transitory economic conditions (such as enrollment in school or serious disability). Because median earnings trends can be affected by changes in the prevalence of these circumstances, handling them in an appropriate manner is vital to producing meaningful earnings trends.

In my marriageability paper, which looked only at young men, I had included nearly all non-workers and workers, excluding only those who were in school and worked less than the entire year. The alternative that I considered was to exclude all non-workers. In the 2025 paper, I found that from 1973 to 2023, among men ages 15 and older, median earnings fell by 18 percent ($6,500) if all non-workers were included but rose by 20 percent ($9,100) if only workers were included. (These estimates used the PCEPI.)

To obtain more meaningful results, the 2025 paper estimated trends for a subset of the sample who I could link to earnings records from a year earlier or a year later. I looked at trends after (1) swapping in non-workers’ full-year earnings in an adjacent year (where available) if their non-work was not due to economic conditions, (2) excluding the remaining non-workers and part-year earners except those whose status was due to economic conditions or their health or disability, and (3) excluding a number of disabled non-workers who would have been so even in 1973. A compromise choice that did not rely on linked records was to look at trends for all year-round workers (including part-time workers), which closely tracked the trend using my more complicated subsample. The median for this group rose by 13 percent from 1973 to 2023 ($7,300), again, using the PCEPI.

The second methodological issue I considered in my 2025 paper involved different ways to measure earnings or compensation. I included analyses in which I added three forms of pre-tax earnings to my measure. The first component added was the value of nonwage benefits provided by employers to their workers. (Employers are indifferent to the forms that total compensation to employees takes.) The second addition involved the payroll taxes ostensibly paid by employers but widely assumed by economists to come out of employee pay. (Again, employers are indifferent to the forms that total compensation takes.) The third addition was the share of the corporate income tax that similarly comes out of employee pay. (Because of the tax, both workers and investors receive less than they otherwise would.)

Using the MACPI, I found that median pre-tax compensation, so defined, rose by 46 percent ($23,200) among year-round male workers from 1973 to 2023. The conventional pre-tax earnings measure rose 40 percent ($17,900). In the marriageability paper, I had included measures of pre-tax earnings, pre-tax compensation, post-tax earnings, and post-tax compensation, but I did not include employees’ share of corporate income tax in their pre-tax earnings.

Given this reassessment of key measurement choices since the publication of my marriageability paper, and given that a few more years of data have become available in the interim, I am in the process of updating that paper. Here, I preview the findings by including revised and extended results for the paper’s Figure 5, which showed trends in the median earnings and compensation of 25- to 29-year-old men from 1962 to 2020. (The fully revised paper will be published later this year.)

The original Figure 5 showed that the median pre-tax earnings of young men fell 9 percent from 1969 to 2019 (both business cycle peaks), while median post-tax compensation was flat. I didn’t report estimates from 1973 to 2019, but those ranged from a drop of 3 percent to a decline of 13 percent. As already noted, these estimates used the PCEPI to adjust for inflation, excluded only students who worked less than the full year, and did not include as pre-tax earnings the share of the corporate income tax paid by employees.

What do the updated numbers show? I use 1973 as my starting year because I was unable to reliably extend the MACPI further back in time. My estimates that try to replicate the numbers in the original paper now range from a drop of 12 percent from 1973 to 2019 to an increase of 2 percent. (It’s possible the underlying data from IPUMS have changed slightly since 2022, and the PCEPI values have changed slightly. Small modifications to my code might have also made a difference.) From 1973 to 2024, they range from a 6 percent drop to a 6 percent increase.

Note that even these estimates are not really evidence of 50 years of stagnation. Instead, there was a period of decline that ended 35 years ago, followed by a period of gains. From 1973 to 1989 (both business cycle peaks), median earnings fell 13 to 19 percent, while they rose 15 to 22 percent from 1989 to 2024 ($6,800 to $9,100 in 2024 dollars).

In my original paper, my sample included men who were 26 to 30 years old when they were interviewed, meaning that they were 25 to 29 years old one year earlier, which was the calendar year for which men were to report their earnings. Because men were generally interviewed between February and April, most of them would have been 26 to 30 years old at the end of the previous year. In the intervening years since my original marriageability paper, I decided it made more sense to look at men who were 25 to 29 years old when interviewed, most of whom would have been 25 to 29 at the end of the previous calendar year. The median change in earnings in this case ranges from a drop of 1 percent to an increase of 12 percent. I’ll stick with this group for the remainder of the analyses.

The first major improvement we’ll consider is the switch from the PCEPI to the MACPI. This change makes a big difference. Now pre-tax earnings (not including corporate income taxes, employer payroll taxes, or nonwage benefits) rise 23 percent from 1973 to 2024 ($9,200). Median post-tax compensation rises 36 percent ($12,700).

Second, if we also switch to all year-round workers, the increase from 1973 to 2024 is 24 percent for pre-tax earnings ($10,800) and 40 percent for post-tax compensation ($15,500).

Finally, the addition of the employee’s share of corporate income taxes to pre-tax compensation turns out not to make any meaningful difference. From 1973 to 2024, median pre-tax compensation excluding the employee’s share of corporate income taxes rose by 33 percent, or $16,600. If we add in corporate taxes, the increase is 32 percent, or $16,300.

The figure at the top of this post updates Figure 5 from my original paper. It leaves corporate income taxes out of pre-tax earnings and pre-tax compensation but uses the MACPI instead of the PCEPI, looks at full-year workers instead of everyone but non-working or part-year students, and looks at men 25 to 29 years old in the survey year instead of men 26 to 30 years old.

Note that these estimates continue to show that the median earnings of young men declined from 1973 to 1989 (by 4 to 13 percent). However, that period of decline is 35 years in the past. From 1989 to 2024, median earnings rose between 42 and 46 percent. That amounted to an additional $19,800 in pre-tax compensation and $17,000 in post-tax compensation.

It seems difficult to characterize the past 35 years as one of “stagnation” for young men. What accounts for the perception that they are not doing well? First, many observers have interpreted the decline in men’s labor force participation as a failure of the labor market to provide opportunities to less-skilled men. However, the evidence is more consistent with rising affluence affording men more opportunities to voluntarily take time out of the workforce (including through more accessible and more generous disability benefits).

Second, the decline in marriage (another effect of rising affluence) has hurt the ability of men and women to pay down debt, save, and afford a down payment on a home. Third, falling marriage and fertility and increased work among women have reduced the burdens on men to maximize their earnings in order to bear the responsibilities of breadwinning. This lessened responsibility, in turn, has reduced men’s incentives to invest in their human capital.

Finally, people feel better about their own circumstances than the mood of economic declensionism would suggest. Americans have a view of how others are doing that is inaccurately negative as an empirical matter. Fully 70 to 75 percent of Americans near the age of 40 believe that they have higher real income than their parents did at the same age, and roughly the same share actually do. These misperceptions are likely to do with the negativity of media accounts as well as incentives on the part of elected officials, foundations, researchers, and advocates to emphasize problems in need of fixing.

We do have problems in need of fixing, of course. But we would do well to recognize that they are less about the economy or economic policy failing us and more about how the individual choices we have made as the nation has grown richer have altered the economic, social, and political landscape.

That study’s title, “The Economic Impacts of Tax Expenditures: Evidence from Spatial Variation across the US,” hardly conveyed its importance in launching a project that continues to this day.³ The paper, authored by economists Raj Chetty and Nathaniel Hendren of Harvard University and Patrick Kline and Emmanuel Saez of the University of California, Berkeley, was nominally about the effect of itemized tax deductions and the earned income tax credit on intergenerational mobility. However, the media headlines it garnered rightly ignored the tax angle; the value of the study was in its innovative use of restricted-access data from the Internal Revenue Service (IRS) to produce the most comprehensive study to date on the geography of opportunity in America.

Using that data, constituting millions of records linking parents to their grown children, Chetty and his colleagues were able to demonstrate for the first time where Americans were more likely to escape childhood poverty, and where they tended to remain mired in generational hardship. In subsequent years, their “Equality of Opportunity Project” became Opportunity Insights (OI), and it has produced a string of pathbreaking studies primarily relying on the IRS data. These analyses have focused on residential mobility, college selection, racial disparities, social capital, the differential prospects of girls and boys, and housing policy.⁴

However, one consistent finding in the OI research has remained outside the spotlight: the apparent importance of family structure for child outcomes. It’s not so much that it has been hidden offstage in the wings; from the start family structure has been a visible supporting player. Leonhardt’s article opened with a vignette about a single mother of three living in Atlanta. He noted the study’s finding that, “Income mobility was…higher in areas with more two-parent households,” highlighting it among a small number of other factors correlated with mobility.

But family structure has evaded center stage in public conversations around the OI research. This omission is unfortunate, for the evidence consistently suggests that family structure is at least as important as any other single factor OI has identified. Our hope is that by bringing together this evidence in one place, policymakers, researchers, and journalists will afford the importance of two-parent families the attention it deserves.

Our review summarizes the OI papers in which family structure is mentioned as a potential causal factor affecting social mobility or other outcomes—11 of them in all. We do not organize the review chronologically, choosing instead to order similar studies together and build toward narrower findings. We highlight some interpretive issues as they come up, which complicate a number of the results.

Our main conclusions are as follows:

Growing Up with a Single Parent

• The relationship between family-level single parenthood and adult outcomes is relatively under-examined in the OI research. The effect of family-level single parenthood on child mobility generally seems modest in the OI research, but the methods used probably understate its importance.

• In several papers assessing the relationship between community-level single parenthood and community rates of mobility, the association is nearly as strong for the children of married parents as for the children of single parents. OI interpretations of these results lean toward the conclusion that what is harmful is growing up around single parents rather than growing up with a single parent oneself.

• OI’s evidence suggests that men who grew up with a single parent tend to have employment rates that are lower than women who grew up with a single parent, even when they have the same parental income. In contrast, holding parental income constant, men who were raised by married parents are more likely to be employed than women raised by married parents. This result is consistent with the idea that growing up fatherless is more harmful for boys than for girls.

• OI finds that the male black-white mobility gap (the difference between the adult incomes of white and black men with the same parental income) is as large among boys with married parents as it is among boys with a single parent. That runs contrary to the idea that family-level single parenthood is responsible for the black-white gap.

• In another paper, OI found that the black-white mobility gap for low-income children (combining boys and girls) has narrowed in recent years. But it narrowed by similar amounts for the children of married and single parents. Moreover, the black-white gap in single parenthood among low-income children narrowed only modestly. The widening of the mobility gap between lower-income and upper-income white children also occurred regardless of family structure.

• These results, however, probably understate the importance of growing up with a single parent.

  • Family structure is assessed in a single year of childhood, but some children with married parents in that year experienced single parenthood earlier or later.

  • Moreover, if family structure affects parental income, then the mobility measures used by OI (which control for parental income) will understate how family structure affects adult outcomes.

  • Similarly, showing that, for instance, the black-white gap in men’s income, controlling for parental income, is the same whether or not children grew up with a single parent does not establish that single parenthood doesn’t affect the black-white gap. If most black children have single parents and correspondingly lower incomes, while most white children have married parents and correspondingly higher incomes, then holding constant parental income will miss the importance of family structure.

  • Finally, married parents with lower incomes likely have unobserved disadvantages, since they potentially have two earners but are nonetheless relatively poor. Similarly, single parents with higher incomes likely have unobserved advantages. When comparing children of single and married parents with the same income, unobserved disadvantages or advantages might mask the harmful impact of family structure in correlational analyses.

Growing Up in a Community with High Rates of Single Parenthood

• Places with higher rates of single parenthood tend to have lower upward mobility. Across the OI papers, single parenthood is repeatedly among the strongest predictors of a place’s mobility.

• This relationship recurs whether looking across commuting zones (aggregations of counties), counties, or census tracts.

• When pitted against other predictors of community mobility rates in multivariate analyses, single parenthood rates usually remain strongly associated with mobility, even when other predictors no longer look important.

• In analyses that estimate each commuting zone’s or county’s causal effect on mobility, rates of single parenthood are only somewhat less strongly correlated with the size of the causal effects than they are with the mobility rates.

• However, community rates of single parenthood aren’t consistently strongly correlated with mobility gaps between boys and girls or between blacks and whites.

• In one OI study, communities with higher rates of single motherhood tended to be associated with lower mobility for boys but higher mobility for girls. Higher rates of single fatherhood were associated with higher mobility for black boys but lower mobility for girls and white boys. In non-poor communities, higher rates of single motherhood among low-income black families were associated with lower black male upward mobility but not with lower white male upward mobility or lower black female upward mobility. Higher rates of single motherhood among low-income white families were associated with lower white male upward mobility but not lower black male upward mobility.

• As noted, the association between community single parenthood and upward mobility rates generally was only somewhat lower for children of married parents than for children of single parents. This could indicate that growing up around single parents hurts even children in intact families. However, for the methodological reasons already noted, the OI papers probably understate the differential impact of community single parenthood on children depending on their family structure. The similarity of the correlation for children of married and single parents could also indicate that unobserved community factors are actually driving the associations for both.

Marriage As an Outcome

• Most of the OI papers focus on adulthood household income to assess mobility. However, the OI researchers note that household income partly reflects the number of earners, which in turn is related to whether someone is married. Therefore, family structure in both childhood and adulthood can affect mobility rates.

• One OI paper found that the gap between blacks and whites in upward mobility was two-thirds smaller using adult individual income as the basis for assessing mobility rather than household income. A primary reason is that marriage rates are far lower among blacks than among whites, even holding parental income constant. One potential way that childhood family structure affects adult household income is by affecting adulthood marriage, a question unexplored in the OI research.

The OI studies have not been designed to rigorously estimate the causal effect of family structure on child outcomes, and most of the evidence in the papers is correlational, though (like the OI researchers) we sometimes use causal language to describe the results for convenience. Assessing the causal effect of family structure on child outcomes is an extraordinarily difficult task, one that can be appreciated by thinking through what an experiment would look like that randomly assigned “single parenthood” to children and then measured the effect on some outcome. One of us, Winship, has explored the conceptual and methodological issues involved elsewhere.⁵

We conclude with suggestions for future research on family structure and mobility that OI could explore.

“The Economic Impact of Tax Expenditures: Evidence from Spatial Variation across the US,” Raj Chetty, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez. First released July 2013. Issued as an Internal Revenue Service SOI Working Paper (2015)

The much-trumpeted 2013 paper was finalized two years later with an absence of fanfare.⁶ By that point, the headline findings had been carved off into a new paper (discussed next). But it is worth summarizing the original draft that started it all, as it included results related to family structure that did not reappear in subsequent papers.

Like most subsequent papers from OI, this one defined “upward mobility” as the typical adulthood income percentile rank for someone who grew up in a family at the 25^th percentile of parental income. That is, a mobility value of 40 indicates that children with parents who were better off than 25 percent of parents nationally (and poorer than 75 percent) typically had incomes as adults that were higher than 40 percent of their same-age peers. This mobility measure is available for 741 local areas called “commuting zones” (CZs), which are comprised of contiguous counties.

In the 2013 version of the paper, Chetty and his colleagues considered the correlation across CZs of mobility with 28 community-level predictors.⁷ A correlation indicates, for example, the extent to which the CZs with the highest rates of single parenthood are also the CZs with the lowest upward mobility. Correlations range from -1 to 1, with negative values indicating a “the higher, the lower” relationship and positive ones indicating “the higher, the higher.” A correlation of 0 means that single parenthood rates have no linear relationship with upward mobility, while larger (or more negative) values indicate a stronger relationship. Correlations of -1 or 1 mean that if you know a CZ’s single parenthood rate, and you know the linear relationship between single parenthood rates and mobility across CZs, you can exactly predict a CZ’s mobility rate from its single parenthood rate.

(Of note, a correlation does not tell you the magnitude of the increase or decrease in mobility as single parenthood rates change, only how well mobility is predicted by single parenthood. It’s possible that single parenthood predicts mobility very well, but increases in single parenthood rates have a substantively small impact on mobility.)

The two predictors most strongly correlated with mobility were the share of families in a CZ headed by a single mother (-0.76) and a CZ’s divorce rate (-0.69). CZ teen birth rates ranked ninth (-0.55).⁸ (Note that the team specifically looked at community rates of single motherhood.)

In multivariate analyses—pitting several predictive variables against each other—the share of families headed by a single mother remained strongly correlated with upward mobility (-0.35), even after “controlling for” high school dropout rates, social capital, income inequality, and unobserved factors common to all commuting zones within a given state.⁹ That suggests that the apparent relationship between single motherhood and mobility is not simply due to any of these competing explanations creating a misimpression. The 2015 version of the paper set up single motherhood rates as a rival predictor of mobility as against high school dropout rates, social capital, income inequality, and racial segregation. It found that single motherhood had the strongest association—a correlation of -0.49 even after controlling for those competing factors.¹⁰

When pitted against the share of a CZ’s families that were black and against the generosity of a state’s earned income tax credit, the correlation between single motherhood and mobility remained -0.76 in the 2013 version of the paper.¹¹ Since single motherhood is more common among black families, we might be concerned that the relationship between single motherhood and mobility is simply picking up the fact that, for some other reason, black children are more likely to have single mothers and lower mobility. This result makes that possibility unlikely. In contrast, while the correlation between the share of a CZ’s population that was black and the CZ’s mobility was -0.61, once single motherhood rates and state EITC generosity were accounted for the relationship between race and mobility disappeared entirely.¹²

Finally, the paper presented one other correlation without comment, but it would feature in the team’s next paper. The correlation between having more single mothers in a CZ and the mobility rate among children living with married parents was also very large (-0.65).¹³ That is, in CZs with high rates of single motherhood, even the kids of married parents tended to have lower mobility.

“Where Is the Land of Opportunity? The Geography of Intergenerational Mobility in the United States,” Chetty, Hendren, Kline, and Saez. First released January 2014. Published in the Quarterly Journal of Economics (2014).

The initial paper—sans the focus on tax provisions—evolved into what became the flagship OI study.¹⁴ In “Where Is the Land of Opportunity?,” the same team of authors highlighted five predictors of a CZ’s upward mobility, noting that “The fraction of children living in single-parent households is the single strongest correlate of upward income mobility among all the variables we explored.”¹⁵

Specifically, of 35 variables examined (including most of those in the 2013 paper), the CZ single motherhood rate came out on top, with a correlation of -0.76.¹⁶ The fraction of adults divorced placed 13^th this time (-0.49), while the share of adults married ranked 11^th (0.57). While teen birth rates were considered in the earlier paper, neither the current study nor the final version of the 2013 paper from 2015 included them. (Indeed, in no subsequent paper did OI revisit teen birth rates as a predictor.) Using the spreadsheets OI provides, it is possible to obtain a county-level correlation of teen birth rates and mobility. That is -0.56—very close to the -0.55 given for the CZ-level correlation in the 2013 paper, where teen birth rates were ranked 9^th out of 28 variables examined.¹⁷

When controlling for the share of the CZ population that was African American and for income growth from 2000 to 2010, the single motherhood rate came out on top again (-0.61), while the share divorced had the second-strongest correlation (-0.57).¹⁸ Finally, the single motherhood rate had the strongest correlation with an alternative measure of mobility (the “rank-rank slope”), with only the black share of the population coming close to it among the other 34 variables.¹⁹ (The rank-rank slope indicates how many income percentiles apart the poorest and richest children tend to be in adulthood. It indicates the extent to which poorer and richer children converge when they grow up.)

In another striking set of findings, when the authors conducted multivariate analyses, pitting five top predictors of mobility against each other, single motherhood emerged as the factor appearing most important.²⁰ Holding the other four variables constant, the single motherhood rate had a stronger correlation with mobility (-0.49) than did commuting times (an indicator of sprawl), income inequality, high school dropout rates, and social capital. Single motherhood looked even more important using the rank-rank slope, with the 0.59 correlation being twice as large as the short-commuting-time correlation of -0.29.²¹ (A lower rank-rank slope indicates more mobility, so the signs of the correlations are the opposite of those looking at the upward mobility measure.) In a model that also held constant all the unobserved factors common to CZs within a state, the single motherhood correlation was unchanged.²²

Finally, controlling for the share of a CZ that was African American, in addition to the other top-five variables, strengthened the relationship between single motherhood rates and mobility, lifting the correlation from -0.49 to -0.58.²³ That relationship, in other words, looks even stronger after accounting for the fact that blacks have lower mobility than non-blacks.

One important issue to keep in mind throughout this review is that the importance of growing up with a single parent may be understated by using income mobility as the outcome of interest. Mobility measures involve looking at outcomes after conditioning on parental income—for instance, assessing the typical adult income for someone who grew up at the 25^th percentile of parental income. To see how this might downplay the importance of family structure, imagine that the only way growing up with a single parent is harmful is by reducing the income available to parents for investing in children. (Single parents have less income than two parents, on average.) The OI researchers start with families at the 25^th percentile of income and then assess how their kids turn out. But if the only way single parenthood can affect how kids turn out is by lowering their parent’s income, then holding constant parental income will mean that no variation across CZs in how kids turn out can be explained by family structure.

The problem isn’t solved by using something like the rank-rank slope, which is just the correlation between parental income percentile and adult income percentile. If single parenthood only harms children by determining their parental income percentile, then we should not expect that variation across CZs in single parenthood will be related to variation in how strongly parental income is tied to the income of adult children.

In reality, family structure potentially affects children in other ways besides lowering parental income. But at least part of its impact is likely missed as a consequence of conditioning on parent income. Therefore, when looking across CZs to see how strongly single motherhood rates are correlated with upward mobility, the OI approach will tend to understate the importance of growing up fatherless for kids’ subsequent income. Fatherlessness affects whether children are at the 25^th percentile, not just where they end up among those who also are raised at the 25^th percentile. (This is also an issue for something like the black share of a CZ, where being black in America can affect whether one’s parent is at the 25^th percentile.)

What the OI analyses generally ask when they look at single parenthood rates is, “Controlling for parental income, how strongly does growing up with a single parent correlate with adult income?” But if growing up with a single parent lowers parental income, then controlling for parental income will understate its importance. Ideally, we’d want to see correlations between parental income percentile rank and adult income rank (the rank-rank slope) and between single parenthood and adult income rank, and then both correlations with the other factor controlled.

There is another problem with asking how important single parenthood is conditional on parental income. As noted, single parents have lower incomes than married parents, on average. If single parenthood causes parental income to be lower, then comparing single parents and married parents with the same income likely involves a particularly disadvantaged set of married parents. After all, married parents potentially can have two incomes. If they do not, that may be because one or both of them has attributes that both make them less attractive as employees and that might harm their children. If they do have two incomes and still remain at, for instance, the 25^th percentile of parental income, then they are likely to lack skills that would earn them more and that may be beneficial for their children.

In short, married parents at the 25^th percentile are probably more disadvantaged in unobserved ways relative to the typical married parent, than are single parents at the 25^th percentile relative to the typical single parent. These unobserved disadvantages, if not statistically controlled, will tend to make the outcomes of children raised by lower-income married parents worse, and therefore the relationship of family structure with adult outcomes will seem weaker than it really is. (There is an analogous problem conditioning on the 75^th percentile of parental income—the single parents who have income that high will tend to have unobserved advantages, since they are doing as well as married parents at the 75^th percentile who potentially have two earners.)

This point is relevant for another of the paper’s striking findings. The introduction to the paper reiterated a point made in the 2013 study—that, “Children of married parents also have higher rates of upward mobility in communities with fewer single parents.”²⁴ The correlation between community single motherhood rates and mobility was -0.66 when looking only at children of married parents.²⁵

What to make of this finding? Does it suggest that single parenthood in a community is so harmful that it affects even kids in intact families who are exposed to it? It is not hard to imagine that growing up around single-parent families might, for instance, lead more children with two parents to become single parents as adults. Perhaps the dearth of fathers in a community leads to more crime, affecting the mobility of all kids in a community. Or the disadvantages of growing up with a single parent produce school environments that also hurt the children of married parents. Chetty and his coauthors note this interpretation, saying the correlation for kids of married parents is strong, “perhaps because the stability of the social environment affects children’s outcomes more broadly.”²⁶

Alternatively, perhaps the high correlation between single motherhood rates and mobility among kids of married parents means that being raised by a single mother has, at worst, modest direct negative effects. The correlation between community rates of single motherhood and mobility is nearly as high for children with two parents as for children with single parents.²⁷ The OI researchers take this as evidence against a negative effect of single parenthood at the family level.²⁸ The implicit interpretation is that single motherhood lowers community mobility rates less by being directly harmful to children who grow up without a father than by negatively affecting all children in a community with lots of single mothers.

It seems unlikely that mass fatherlessness hurts all children while individuals growing up without a father don’t especially suffer. A different interpretation of the large correlation for the children of married parents is that perhaps a high rate of single parenthood is mostly the observable product of some other unmeasured community disadvantage that affects children in a CZ regardless of whether both parents are present. That is to say, it’s possible that the strong relationship between community single motherhood rates and upward immobility either reflects a third variable that affects both or reflects some kind of “selection” of certain kinds of parents into certain kinds of family structures and neighborhoods (with the single motherhood being incidental).

For example, perhaps high rates of single motherhood and low rates of mobility both reflect neighborhood cultural values that place a low priority on child economic success. The cultural values might cause low mobility even if every child were in a two-parent family. Perhaps the parents that select into high-single-motherhood neighborhoods would have children with low mobility regardless of the neighborhood in which they lived (because of their values).

This possibility matters because if it’s not really single motherhood per se that is driving the correlation with mobility, then even if we could increase the number of children growing up with two parents, upward mobility would not rise.

Yet another possibility is that in places with lots of single motherhood, growing up fatherless hurts children directly, while the kind of two-parent family that chooses such a community brings its own, different, impediments to upward mobility. After all, we are talking about families with two potential earners who nevertheless are at the 25^th percentile of parental income and choosing neighborhoods with high rates of single motherhood and low rates of upward mobility. So single parenthood might be individually harmful but only look like it harms children with married parents in the community. CZ rates of single motherhood would then be correlated with the mobility of married couples’ children because lower-income married couples are especially disadvantaged, being both at the 25^th percentile of income (despite potentially having two incomes) and living in the sort of area disproportionately chosen by single-parent families.

A final issue is that the Chetty team’s measured correlation of mobility and community single motherhood probably is too strong for children of married parents. The authors assigned kids to family structures based on the marital status of a parent the first time a child was claimed in the IRS data. Some children who had married parents in that single year had already lived without a parent, either because they were born to a single mother before their parent married or experienced a past parental divorce before their parent remarried. Other children experienced parental divorce after the single year in which their family structure was assigned. If it were possible to look at the correlation of mobility with community single motherhood for children continuously in an intact family, it would almost surely be lower than the Chetty team reports of children of parents married in one year.

Moreover, this measurement issue is probably more relevant in CZs with lots of single parenthood, where marriages are probably less stable than elsewhere. That will magnify the correlation between mobility and community single motherhood among kids of “married” parents even more. In sum, the difference in this correlation between kids of single parents and kids of married parents is probably larger than this paper suggests.

“Who Becomes an Inventor in America? The Importance of Exposure to Innovation,” Alex Bell, Chetty, Xavier Jaravel, Neviana Petkova, and John Van Reenen. First released December 2017. Published in the Quarterly Journal of Economics (2019).

We include in our review, for completeness, this paper, which merged IRS data with data on patents to look at predictors of who grows up to become an inventor. It reported incidentally that CZs that produce future inventors “tend to have higher mean incomes (population-weighted correlation ρ = 0.63), fewer single parents (ρ = −0.39), and higher levels of absolute upward intergenerational mobility (ρ = 0.32).”²⁹ (Technically, the family structure measure was single motherhood again.)

“Social Capital I: Measurement and Associations with Economic Mobility,” Chetty, et al. First released July 2022. Published in Nature (2022).

The papers above looked at variation in upward mobility rates across CZs. CZs are relatively large geographic areas—no smaller than a county and sometimes larger than a metropolitan area. In a subsequent paper, OI teamed up with researchers from Meta, Grammarly, Harvard, Stanford, and New York University to analyze a massive amount of data on social capital that looked across counties and ZIP codes.³⁰

The paper linked data on Facebook users in the United States to records assembled by OI for previous studies. The focus of the paper was on “economic connectedness.” Economic connectedness in an area involved the average share of Facebook friendships people in the bottom half of socioeconomic status had that were with people in the upper half of socioeconomic status. “Socioeconomic status” here was based on a prediction of income percentile from a number of Facebook-derived variables, including area median household income, age, sex, college attended, and even phone and carrier type, among others.

Economic connectedness was strongly related to upward mobility, defined as in the earlier papers. Among 22 county-level characteristics the authors examined (9 of them social capital measures), economic connectedness ranked behind only one other predictor—the share of a county’s households headed by a single mother. (The authors again imprecisely use “single parent.”) The correlation of county upward mobility and single motherhood was about -0.65. These correlations were imprecisely measured, and it’s not clear that single motherhood’s relationship to mobility was stronger than that for economic connectedness, or even than the correlations for the share of county residents that were black or for income inequality.

The potential importance of family structure is underplayed in this paper. Of the 13 non-social-capital predictors of mobility, the authors spent three paragraphs discussing the moderate correlations that county median income and poverty rates had with mobility. These correlations became much smaller after controlling for economic connectedness, with the median household income correlation falling to zero.

A paragraph on racial and economic segregation noted that the relatively low correlations these variables had with upward mobility also fell practically to zero after accounting for economic connectedness. In a paragraph on the black population share, which had a strong association with mobility, the authors noted that controlling for economic connectedness eliminated the negative relationship when looking at black adults. Doing so turned the relationship positive when looking at white adults. Controlling for economic connectedness also dropped the strong relationship between income inequality and mobility to practically zero, which the authors discussed in another paragraph.

That left two sentences to discuss the remaining seven county-level predictors of mobility, “ranging from the quality of local schools to job availability to measures of family structure.” The authors minimized the apparent importance of single motherhood by explaining that economic connectedness “is more strongly correlated with upward economic mobility than almost all of those characteristics in univariate specifications.”³¹

They then turned to a multivariate analysis—pitting economic connectedness against single motherhood, median household income, racial segregation, the black population share, income inequality, and third grade math scores. Economic connectedness came out on top, with family structure second. However, the imprecision in these correlations is substantial—enough that it’s reasonably likely that family structure’s correlation with mobility is at least as strong as that of economic connectedness. Other ways of assessing the relative importance of these predictors indicated that single motherhood was either about as important as economic connectedness or third-most important, behind economic connectedness and the black share of the population.³²

Using the data that OI provides publicly, we can further investigate the relative importance of single motherhood. As noted above, controlling for economic connectedness tended to reduce the correlation between county characteristics and mobility to very low levels (sometimes approaching zero). But when we looked at the correlation between single motherhood and mobility, controlling for economic connectedness dropped the association only from -0.65 to -0.44. Moreover, that -0.44 was as strong as the correlation between economic connectedness and mobility after controlling for single motherhood (0.43).

Chetty and his coauthors also reported correlations between the county-level variables and mobility after restricting to counties where over 90 percent of the population was white.³³ Single motherhood was the fourth highest correlation (behind economic connectedness, a social capital index from Penn State University researchers, and income inequality). When they restricted the analyses to the 25 percent of counties with the highest Facebook coverage, single motherhood finished first.³⁴ Finally, at the ZIP code level, single motherhood had the strongest correlation with upward mobility out of 17 predictors.³⁵ When pitted against economic connectedness, the black population share, mean household income, and 3^rd grade test scores, single motherhood was essentially tied with economic connectedness.

“The Opportunity Atlas: Mapping the Childhood Roots of Social Mobility,” Chetty, John N. Friedman, Hendren, Maggie R. Jones, and Sonya R. Porter. First released February 2020. Published in American Economic Review (2026).

Even counties and ZIP codes are large groupings of people. This year has seen the publication of a new OI paper that examined mobility across census tracts.³⁶ Census tracts constitute something like large neighborhoods, averaging a little over 4,000 people. This new paper, authored by Chetty, Hendren, their OI colleague John Friedman, and Census Bureau researchers Maggie R. Jones and Sonya R. Porter, linked IRS data on parental income and the adult income of their children with Census Bureau data. The authors estimated upward mobility rates for each census tract, including separate estimates by parental income, race, and sex.

As in the earlier CZ papers, the authors examined the correlation of upward mobility with a number of community attributes, this time across census tracts within CZs. Not only did they control for the CZ in which a tract was included, they controlled for the racial mix of the CZ by estimating race-specific correlations and then averaging them. Among 13 tract-level predictors of upward mobility (defined again as the expected income percentile for adults raised at the 25^th percentile), single parenthood (this time including single fathers) was essentially tied for first with mean household income as having the strongest correlation. Among other predictors, it beat out four related to employment, two related to education, and tract-level poverty rates. The correlation between tract-level single parenthood and upward mobility remained strong at -0.59.

As in the earlier papers, the authors found that this association between mobility and community single parenthood rates was strong both for the children of single parents and the children of married parents. And they again endorsed the view that, “the correlation is driven not by differences in outcomes between children raised by married versus single parents but rather by ecological (neighborhood-level) factors.”³⁷

In an appendix, the authors also reported that the probability of downward mobility from the 75^th percentile was strongly related to tract-level single parenthood. At nearly -0.55, the strength of the correlation was tied with that for the share of adults who graduated from college and only behind mean household income.³⁸ A footnote indicated that, conditional on having parents at the 25^th percentile, the cross-CZ correlations between single parenthood and other outcomes were “qualitatively similar” to the correlations between single parenthood and upward mobility. Those outcomes included CZs’ average individual income rank, employment rate, incarceration rate, and teenage birth rate.³⁹

“A Practical Method to Reduce Privacy Loss When Disclosing Statistics Based on Small Samples,” Chetty and Friedman. First released March 2019. Published in AEA Papers and Proceedings (2019).

For completeness, we include this short methodological paper, which illustrated its technical point using further results from the tract-level analyses.⁴⁰ Chetty and Friedman found that the higher a census tract’s share of families headed by a single parent, the higher was the teen birth rate among black women who grew up there and had low-income parents. An increase of ten percentage points in the single parent share was associated with a teen birth rate among low-income black girls that was higher by 1.4 points.

The findings in the papers so far reviewed essentially involved correlations, with or without controls for other relevant variables. Within a couple of years of their first paper, the Chetty team would get closer to establishing causal relationships between intergenerational mobility and various community characteristics—including community single parenthood rates.

“The Effects of Exposure to Better Neighborhoods on Children: New Evidence from the Moving to Opportunity Experiment,” Chetty, Hendren, and Lawrence F. Katz. First released May 2015. Published in American Economic Review (2016).

In mid-2015 and late 2016, the Chetty team released three papers addressing the question of whether neighborhood attributes causally affect child outcomes. The fact that some places have higher intergenerational mobility than others does not necessarily mean that the places themselves are responsible for the higher mobility. It may be that the attributes of the families in those places are such that their children would have had high intergenerational mobility no matter where they grew up. This distinction is a crucial one. It addresses the question of whether moving disadvantaged children to a high-mobility place like Salt Lake City will tend to improve their outcomes, or whether those outcomes will be unaffected because Salt Lake City’s disadvantaged children would do well anywhere.

The first paper released by the team—authored by Chetty, Hendren and Harvard’s Lawrence Katz—analyzed a policy experiment that randomly assigned people either to obtain a housing voucher restricted to lower-poverty neighborhoods, to obtain an unrestricted housing voucher, or to receive no voucher.⁴¹ Children in families assigned the restricted voucher had better outcomes on several dimensions than the other children if their family moved before adolescence, implying that living in a higher-income community was relatively advantageous for kids. This so-called “moving to opportunity,” however, was slightly harmful on average if a family used the restricted voucher during a child’s adolescence.

Of note, the authors mention that the lower-poverty neighborhoods that, on average, benefitted younger children had lower rates of single parenthood.⁴² The study design did not allow them to determine the extent to which those lower rates were responsible for the improved outcomes. Nevertheless, the finding that moving to a lower-poverty neighborhood improved outcomes strengthens the interpretation that the correlations of mobility with community-level characteristics found in the other OI papers reflect, at least in part, causal relationships. (Notably, the paper also found that living in a lower-poverty neighborhood made pre-teen girls less likely subsequently to become single mothers and less likely to live in a community with lots of single motherhood.)

“The Impacts of Neighborhoods on Intergenerational Mobility II: County-Level Estimates,” Chetty and Hendren. First released December 2016. Published in the Quarterly Journal of Economics (2018).

The second and third OI papers addressing causality more rigorously were released (and later published) as two parts of a single study and turned back to the IRS data. In Part I, Chetty and Hendren showed that when families moved to a CZ where permanent residents had higher mobility rates, the longer children spent in the new CZ, the more social mobility they enjoyed. This was true even comparing siblings in the same family. This paper only looked at single parenthood incidentally, showing that the effect of moving to a new CZ wasn’t driven by changes in family structure leading to a move.⁴³

Part II, also authored by Chetty and Hendren, included results addressing the importance of family structure.⁴⁴ Building on Part I, the authors focused on families who moved to a new commuting zone. They estimated, for children at the 25^th percentile of parent income, for each CZ, the “fixed effect” of living an additional year there on adulthood income percentile rank.

For a variety of CZ predictors, they then estimated the association across CZs between the predictor and the CZ’s fixed effect on mobility. Chetty and Hendren found that the -0.57 correlation between single motherhood and the causal effect of a CZ on mobility was smaller than the -0.76 correlation they found in their earlier papers (correlating single motherhood with mobility rates), but that’s still a strong relationship. While single motherhood was the strongest predictor of mobility in the earlier papers, it was tied for the 11^th strongest predictor of a CZ’s causal effect on mobility (out of 40 predictors). The CZ’s share of adults married was ranked 16^th, while its share of adults divorced was 37^th.⁴⁵

Chetty and Hendren also estimated the causal fixed effect of counties on mobility. Controlling for the unobserved factors common to counties within a CZ, the correlation across counties of community single motherhood and a county’s fixed effect on mobility was -0.38. That was the fifth highest of the 40 predictors. The share of adults divorced was ranked 10^th, and the share of adults married was ranked 11^th.⁴⁶

Finally, Chetty and Hendren also estimated the “fixed effect” of CZs and counties on the mobility of children raised at the 75^th percentile of income (rather than the 25^th percentile). Interestingly, community single motherhood, divorce, and marriage had relatively small correlations with this causal effect across CZs and across counties within CZs. Only the correlation of the community marriage share with the CZ-level fixed effect was larger than 0.20 (or smaller than -0.20).⁴⁷

Note that these correlations do not necessarily represent the causal effect of, for instance, community single motherhood on community mobility. They are just correlations between community single motherhood and whatever it is about CZs that, causally, make it more or less advantageous for children to spend additional years living in them. However, the results in this paper got closer to estimating the extent to which high levels of community single parenthood reduce upward mobility, compared with the estimates in OI’s other papers.

Three other papers by the Chetty team focused on demographic differences in mobility. The patterns they reported offer additional hints that family structure has causal effects at both the family and community levels, though they are not neatly consistent.

“Childhood Environment and Gender Gaps in Adulthood,” Chetty, Hendren, Frina Lin, Jeremy Majerovitz, and Benjamin Scuderi. First released January 2016. Published in American Economic Review (2016).

Two years after releasing “Where Is the Land of Opportunity?” (and before the two-part study described above) Chetty and Hendren teamed up with three new coauthors from Harvard and Stanford on a study examining how mobility differs for girls and boys.⁴⁸ Using the same IRS data as in their previous papers, they looked at four outcomes: employment, earnings, income rank, and college attendance. Most of the paper focused on employment.

In a first set of findings, the authors reported that in adulthood, men did worse relative to women if they grew up with low income than if their parents had higher income. Men’s employment rates were generally higher than those of women. But if they grew up in poor families, their employment rates tended to be lower than those of women. Men’s earnings and income rank were also generally higher than for women, but the gap was smaller comparing adults who grew up poor than it was comparing adults who grew up rich. Male college attendance rates were generally lower than female rates, but the disadvantage was greater comparing men and women who grew up poor.⁴⁹

This pattern of gender gaps being relatively worse for low-income boys turns out to be linked to family structure, at least in regard to employment. Chetty and his team found that among children of married parents, men were more likely to be employed than women regardless of whether their parents were rich or poor. However, among the children of single parents, if they were raised in the bottom 40 percent of income, women were more likely than men to be employed.⁵⁰ (As of 1998, around the time parental incomes were measured, over two-thirds of single-parent families were in the bottom 40 percent of income.)⁵¹ Single parenthood also looked harmful to upper-income men; rather than the employment gap favoring men, men and women were equally likely to be employed if they grew up with a single parent with income in the top 60 percent. Frustratingly, OI did not include or mention results using earnings, household income percentile, or college attendance as the outcome, so it is not clear whether gender gaps on these dimensions also were confined to or worse for adults with low-income parents.

These patterns once again involve correlations, but they accord with the idea that growing up fatherless (the bulk of single parents are mothers) is more detrimental to boys than to girls. Moreover, the results involve family-level measures of single parenthood and correlations across families rather than community-level measures and correlations across communities.

The same caveats regarding correlations of single parenthood and mobility at the community level apply to family-level correlations. Family structure was measured in a single year of childhood, which means some children who previously or subsequently experienced single parenthood were classified as having grown up with two parents. Conditioning on parental income may dilute the impact of family structure on individual outcomes, since family structure potentially affects parental income percentile. And conditioning on parental income also risks comparing children of single parents with children of especially disadvantaged married parents, such that any harm done by single parenthood gets washed out by the harm done by those unobserved married-parent disadvantages in a correlation.

A second set of findings in this paper looked at variation in employment across CZs among women and men.⁵² Geographic variation in the employment gender gap disproportionately reflected variation in employment rates among men who were raised poor. Further, geographic variation in the gender gap disproportionately reflected variation in employment rates among men who were raised poor and by a single parent.⁵³

Next, the authors focused on adults raised in the bottom quintile and examined what factors were correlated with the gender gap in employment rates for this group across CZs. They found that, “Boys have lower employment rates than girls in areas with three characteristics: (i) a larger fraction of black residents; (ii) greater residential segregation [by race or income]; and (iii) less stable family structures, as measured by the fraction of single mothers and marriage rates.”⁵⁴ (Rounding out the top six correlations, their results also indicated that places with longer commute times had worse gender gaps in employment for men.) These correlations (including the community single motherhood correlation) were sizable even restricting to children with married parents.⁵⁵ CZ divorce rates were not strongly correlated with CZ gender gaps.

The authors then pitted community single motherhood rates as a predictor of gender gaps against the share of a CZ that was black and against community residential segregation by income, using a multivariate model including all three. They found that single motherhood was no longer robustly associated with the employment gender gap, though the other two variables remained correlated with the gap. The same was true for the gender gap in adulthood income ranking. When Chetty and his colleagues looked at the gender gap in a CZ’s causal effect on income ranking, using the approach from the 2018 paper discussed above, only residential segregation remained associated with gender gaps.⁵⁶

These results, interpreted causally, suggest that community single motherhood rates only appear to affect the gender gap among people raised in poor families; the correlation is really just a byproduct of CZs with high income segregation (and perhaps large black populations) having higher single motherhood rates and larger gender gaps. This finding is in tension with the earlier result in the paper that at the national level, the gender gap in favor of women occurs only among adults raised by single parents. That earlier analysis compared adults who, individually, were either raised by a single or married parent, while the analysis at the end of the paper compared community single motherhood rates across CZs, Family structure appears important in the former but incidental in the latter.

“Race and Economic Opportunity in the United States: An Intergenerational Perspective,” Chetty, Hendren, Jones, and Porter. First released March 2018. Published in Quarterly Journal of Economics (2020).

Though focused on racial disparities, the particular mobility problem of boys was highlighted again in this paper coauthored by Chetty and Hendren with Census Bureau researchers.⁵⁷ As in the “Opportunity Atlas” paper, the authors linked IRS data with data from the Census Bureau. The paper first reported that, at every level of parental income, black and American Indian children tended to have lower household income in adulthood than white, Hispanic, and Asian children. Most of the rest of the paper focused on the black-white mobility gap.

Family structure came into play quickly. The authors noted that the black-white mobility gap tended to be quite a bit smaller looking at the individual income of adult children rather than their household income. For instance, the black-white gap using the usual OI upward mobility measure (typical adulthood income percentile conditional on being raised at the 25^th percentile) was 13 percentiles when adults’ household income was examined but 4 percentiles looking at individual income.⁵⁸

This difference reflects the lower marriage rates of blacks, which reduces household incomes. While 55 percent of whites in the study were married as of their mid-thirties, just 16 percent of blacks were.⁵⁹ This marriage gap was only slightly smaller holding constant parental income. Chetty and his coauthors noted that, “White children at the bottom of the income distribution are as likely to be married as black children at the 97th percentile of the parental income distribution.”⁶⁰

These findings related to the family structure of the adult children rather than their family structure when they were growing up. The authors’ remaining analyses examined individual income so as not to introduce complications related to adulthood family structure. However, it’s worth recognizing that most of the OI studies look at household incomes, and in those analyses, family structure can affect mobility not just through parental influences on children but by reducing the number of earners in a household when those children are adults.

Strikingly, the black-white mobility gap using individual income was confined to men. The gap in upward mobility was 10 percentiles for men, while it actually favored black women over white women by one percentile.⁶¹

The authors argued that family-level single parenthood did not seem to be a significant factor in explaining the male black-white mobility gap. The gap among men raised by a single parent was 10 percentiles—the same as for men generally. Among men raised by two parents, the gap was still 8 percentiles.⁶² In multivariate models assessing whether the black-white mobility gap narrowed after controlling for parental family structure, the results for both men and women barely changed after doing so. That was true looking at the racial gaps for both lower-income and upper-income children.⁶³

However, the same methodological caveats already noted in summarizing the other OI papers apply to this one too. In particular, this study offered some evidence about the extent to which using mobility as an outcome—which holds constant parental income—might understate the importance of family structure. Across all boys, the black-white gap in adulthood income was 18 percentiles.⁶⁴ Note that this was the gap without restricting to men who grew up lower-income or upper-income. Controlling for parental income, the black-white gap shrunk to 10 percentiles (for lower-income boys) or 12 percentiles (for upper-income boys).⁶⁵ That’s a narrowing of 6-8 percentiles after conditioning on parental income.

When the authors instead controlled for single parenthood (without controlling for parental income), the gap also narrowed substantially—from 18 percentiles to 13 percentiles, or a reduction of 5 percentiles.⁶⁶ Chetty and his colleagues emphasize that after conditioning on parental income by looking only at lower-income boys or upper-income boys, the racial gap in adult income is roughly the same for kids whether or not they are raised by single parents. But if one of the ways that single parenthood hurts children’s adult income is by lowering parental income, then this test of the importance of family structure will inadequately capture its impact. (To be clear, the reverse argument is also true: if low income leads to family instability and disruption, then controlling for single parenthood would understate the importance of parental income.)

In the rest of the paper, the authors shifted from family-level explanations for the black-white mobility gap to community-level ones. They examined the correlation of upward mobility rates with 35 variables across census tracts, separately for black and white men who grew up poor. (They reported the same correlations for black and white women who grew up poor and for black and white men and women who grew up upper-income, but they focus on explaining the racial gap for boys raised in lower-income families.) Among white men, a tract’s single parenthood rate was tied for third at 0.50, behind a tract’s attitudes toward interracial marriage and mean household income and tied with the share who graduated from high school. A tract’s share of adults divorced had the 6^th highest correlation with upward mobility.⁶⁷

Interestingly, it’s a tract’s overall single parenthood that seems important for upward mobility out of poverty rather than single parenthood among low-income children or even single parenthood among low-income white children. Those correlations are much lower—0.13 or less for tract rates of single motherhood or single fatherhood among low-income white or black children.

The conclusion is that upward mobility out of poverty for white men has more to do with a community’s overall rate of single parenthood than with its rate of single parenthood among poor white families or poor families generally. This curious finding points, again, to the importance of assessing the correlation of single parenthood rates with adult income before conditioning on parental income. If single parenthood hurts upward mobility just by lowering parental income, then tract-level variation in single parenthood among poor families will not be correlated strongly with variation in adult incomes.

Among black men, single parenthood had the highest correlation with upward mobility of the 35 tract-level predictors (0.40). This estimate was lower than for white men, suggesting that community single parenthood was more consequential for white men than for black men. The share of adults married had the 4^th highest correlation. Again, rates of single motherhood or single fatherhood among low-income black or white children were not strongly correlated with upward mobility among black men.

In predicting a tract’s upward mobility gap between white and black boys raised poor, the only variable with a strong correlation was openness to interracial marriage among residents of the state in which a tract was located.

Other findings related to male-female differences in upward mobility. More single motherhood among low-income black or white children in a census tract was associated with less upward mobility for low-income black and white boys but more mobility for low-income black and white girls. Because of this dynamic, lower single motherhood among poor blacks or whites tended to reduce the mobility gap between women and men who grew up poor (both for blacks and whites).⁶⁸ Tract-level single fatherhood among low-income black families was positively correlated with the upward mobility of poor black boys (meaning it was associated with more mobility) but was negatively correlated with the upward mobility of poor black girls. In contrast, single fatherhood among low-income white families seemed to hurt both poor white boys and girls.

Following these analyses, the authors then focused on the 50 percent of census tracts with the lowest poverty rates, looking again at correlations between tract characteristics and male upward mobility. The results related to family structure were similar to those in the full analyses, except that single parenthood and marriage were only weakly correlated with black boys’ upward mobility out of poverty.⁶⁹ The dampening of these correlations may reflect reduced variation in single parenthood once the poorest communities are dropped, once again pointing to the potential importance of family structure in affecting parental income.

Across the low-poverty tracts, the correlation for black boys between a tract’s upward mobility rate and its rate of single motherhood among low-income black children was weak, but it was essentially zero for white boys.⁷⁰ As the share of poor black children headed by single mothers fell from 80 percent to 20 percent, the typical percentile reached by black men who grew up poor rose from the 40^th or 41^st percentile to the 43^rd or 44^th percentile.⁷¹ The typical percentile reached by white men who grew up poor was steady at about the 50^th percentile regardless of the share of single motherhood among poor black children. The black-white upward mobility gap thus was roughly one-third lower in tracts with a 20 percent rate of low-income-black single motherhood than in tracts with an 80 percent rate. These results held after restricting the analyses to native-born children.⁷² Black-white gaps in adult employment and incarceration for children raised poor also were smaller the lower was tract-level single motherhood among poor black children.⁷³ However, to belabor the point, if single motherhood affects parental income (and whether one lives in a non-poor census tract), then the importance of single motherhood in affecting black-white inequality is understated by this exercise.

In terms of male-female mobility gaps, employment rates of black women who grew up poor were unrelated to the share of a tract’s poor black children headed by single mother. As a result, lower tract-level single motherhood among poor blacks also tended to reduce the employment gap between black women and men who grew up poor.⁷⁴

The authors also show that across low-poverty tracts, single motherhood rates among low-income white families have little to no correlation with the mobility of low-income black boys but are associated with less mobility among low-income white boys.⁷⁵

Once again, the evidence suggests that community-level single parenthood is more harmful than family-level single parenthood. The relationship between the upward mobility of low-income black boys and a tract’s single motherhood rate among low-income black children held for boys with married parents as well as for boys with single mothers.⁷⁶

The paper’s appendix tables also provided evidence on mobility for children raised at the 75^th percentile of parental income.⁷⁷ Tract-level divorce and single parenthood rates were moderately correlated with the upward mobility of white boys and among the top five predictors. Family structure variables were less predictive of black boys’ upward mobility, but marriage and single parenthood rates were among the top ten predictors. Among black and white girls, divorce and single parenthood rates were moderately correlated with upward mobility but not among the top ten predictors. Tract-level mobility gaps between black and white boys and girls raised at the 75^th percentile weren’t well predicted by family structure variables.

A final chart in the paper drove home the potential importance of community single motherhood rates in contributing to black-white inequality. Only 3 percent of white children lived in a census tract in which half of white children or more had a single mother. But for black children, 86 percent lived in a tract in which at least half of black children had a single mother.⁷⁸

“Changing Opportunity: Sociological Mechanisms Underlying Growing Class Gaps and Shrinking Race Gaps in Economic Mobility,” Chetty, Will S. Dobbie, Benjamin Goldman, Porter, and Crystal Yang. First released July 2024. Published in Quarterly Journal of Economics (Forthcoming).

This recent paper from Harvard, Cornell, and Census Bureau researchers combined Census Bureau and IRS data.⁷⁹ It examined trends in mobility for lower-income blacks and whites and for upper-income whites. Specifically, the study sought to explain the narrowing of the racial mobility gap for cohorts born between 1978 and 1992 and the widening of the adult income gap comparing whites in lower- and upper-income families as children.

The outcome of interest in both cases was adult household income percentile rank. When comparing lower-income blacks and whites, that amounts to looking at the average income rank conditional on being raised at the 25^th percentile—the same upward mobility measure used in most of the OI research. Upward mobility by this measure rose by 1-2 percentiles among lower-income blacks but fell by over 2 percentiles among lower-income whites. Meanwhile, among upper-income blacks, the expected adulthood income percentile rose by 1.5 percentiles, and it rose by nearly 1 percentile among upper-income whites.

The evidence in this paper suggests that changes in the share of blacks and whites and lower- and upper-income children personally experiencing single parenthood did not much explain changes in these racial and class gaps. Growing up with two parents became rarer among lower-income blacks and whites and upper-income blacks and whites (though the change amounted to only a percentage point for upper-income whites). However, the racial single parenthood gap among poorer children narrowed slightly over the 15 years, from 32 percentage points to 30 points. (The gap in single motherhood was much smaller among black and white children raised at the 75th percentile, and it too fell over time.) The class gap in single parenthood among white children widened by nearly 9 points (while narrowing by 4-5 points among black children).⁸⁰

These changes are directionally consistent with the possibility that changes in family structure narrowed the black-white mobility gap among lower-income children and widened the class gap among white children. Moreover, the racial gap in mobility fell much more among men than among women, largely because it was small to begin with among women. (It favored black women over white women by the end of the 15 years.)⁸¹ That, too, suggests that changes in father absence could be a key factor.

However, two findings in the OI paper cast doubt on this possibility. First, the racial mobility gap narrowed even more among children raised by a single parent and among children raised by married parents than when all lower-income children are pooled.⁸² If changes in family structure had driven changes in the racial mobility gap, we would have expected to see the gap fall much less after conditioning on family structure. Second, in a multivariate model that pooled black and white children raised in lower-income families and controlled for whether someone grew up with two parents, the racial mobility gap was even larger than in the main analyses. The class gap among whites did not shrink much after controlling for single parenthood.⁸³ These analyses are again vulnerable to the critiques we have already raised.

The paper also looked at whether community-level changes in family structure might have affected adulthood income. Here, the evidence suggests that growing up in a community where single parenthood is common may be detrimental.

Across counties, the change in the parental marriage rate was strongly associated with the change in adulthood income. This was true for children pooled together but also separately for lower-income children, whether black or white, and upper-income white children.⁸⁴ The association between the change in parental marriage and the change in adulthood income was a bit smaller than the association between county employment rate change and adult income change. The change in the parental mortality rate was also more strongly correlated with change in adulthood income than was the change in parental marriage.⁸⁵

Looking at the marriage rates of adult children rather than their household income ranks, the change in marriage rates across counties was associated with the change in parental marriage rates. This was true for white and black lower-income children and for upper-income white children. The change in parental marriage rates was a stronger predictor of the change in child marriage rates than was the change in parental employment rates.⁸⁶

When the change in parental marriage rates was pitted against the change in parental employment rates in a single model, parental marriage rates generally had the stronger association across counties with both changes in adulthood income ranks and changes in marriage rates. That was true looking at lower- and upper-income black and white families, except that the correlation with employment rate change was stronger for lower-income black and white children. Controlling for the change in employment rates did not weaken much the strength of the correlation of parental marriage rates with outcomes. Controlling for the change in parental marriage rates did weaken the strength of the correlation of parental employment rates with outcome when looking at upper-income white children.⁸⁷

Conclusion

The OI team has demonstrated ingenuity and creativity in carrying out its analyses over more than a decade [EDITED 2/3/26]. The project of expanding opportunity would benefit greatly from a stronger and intentional OI focus on family structure. While we trust that OI researchers would come up with more clever analyses once they set out to tackle the subject, we offer a few starting ideas for advancing the state of knowledge around family structure and opportunity:

• Measure family structure as ever having lived with a single parent and repeat various analyses conducted in past work. This could also perhaps be done at the community level.

• Measure family structure as having changed family structures versus having always been in an intact family versus always having been in a single-parent family.

• Plot rates of single parenthood by parental income percentile, including separately by race. Also plot rates of community-level single parenthood by parental income percentile.

• Plot adult household income across the parental income distribution, showing the associations for both children of single parents and children of married parents. Do the same for individual income, and also show the associations separately by sex and race. Doing so would clarify that even if these associations are similar regardless of family structure, the family-structure gap in outcomes conditional on parental income is large at all levels of parental income. That would illustrate graphically that family structure can affect not just the correlation between parental income and adult outcomes but the level of parental income.

• Estimate models that predict adult outcomes from single parenthood before conditioning on parental income. Examine these alongside the models predicting adult outcomes from parental income. Then pit the two predictors against each other and see how each of the associations change.

• Examine whether parental income is associated with adult marriage and single parenthood rates. Do the same for parental family structure.

• Examine the extent to which economic connectedness is predicted by family structure.

• Estimate “family fixed effects” models that include children in the same family who either did or didn’t experience single parenthood, controlling for all the family-level influences they share, to assess the causal impact of family structure on adult outcomes.

• Treat single parenthood as the main explanatory variable of interest and explore the pathways by which it might affect adult outcomes (via successively added control variables).

• Look at the relationship between community-level change in mobility and community-level change in single parenthood rates.

It is our hope that this review has clarified the suggestive evidence that already exists across the OI research that family structure may be crucial for upward mobility.

Thanks to Hannah Mayhew for research assistance on this review.

Endnotes

1 David Leonhardt, “Geography Seen as Barrier To Climbing Class Ladder,” New York Times, July 22, 2013, p. A1, https://archive.nytimes.com/www.nytimes.com/2013/07/22/business/in-climbing-income-ladder-location-matters.html.

2 For the print headline, see the image at https://www.nytimes.com/images/2013/07/22/nytfrontpage/scan.pdf.

3 Raj Chetty, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez, “The Economic Impacts of Tax Expenditures: Evidence from Spatial Variation across the US,” unpublished paper, July 2013, https://web.archive.org/web/20130815083624/http://obs.rc.fas.harvard.edu/chetty/tax_expenditure_soi_whitepaper.pdf.

4 See the papers listed at https://opportunityinsights.org/paper/.

5 Scott Winship, “It Takes Two,” Education Next 24(1), 2024, https://www.aei.org/articles/it-takes-two/; Rachel Sheffield and Scott Winship, “The Demise of the Happy Two-Parent Home,” Joint Economic Committee, July 2020, https://www.jec.senate.gov/public/index.cfm/republicans/2020/7/the-demise-of-the-happy-two-parent-home.

6 Raj Chetty, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez, “The Economic Impact of Tax Expenditures: Evidence from Spatial Variation Across the US,” Internal Revenue Service, SOI Working Paper, 2015, https://www.irs.gov/pub/irs-soi/14rptaxexpenditures.pdf. Tables and figures are at https://www.irs.gov/statistics/soi-tax-stats-soi-working-papers#2014.

7 The 2015 version, like the “Where Is the Land of Opportunity?” paper (summarized below) published before it, expanded the number of predictors to 35. See Table XII.

8 See Table 5 of the original paper at https://web.archive.org/web/20130815083624/http://obs.rc.fas.harvard.edu/chetty/tax_expenditure_soi_whitepaper.pdf.

9 See Table 7, Column 2.

10 See Table XIII, Column 3 of the final version at https://www.irs.gov/statistics/soi-tax-stats-soi-working-papers#2014.

11 See Table 7, Column 4.

12 See Table 5 and Table 7, Column 4.

13 See Table 5.

14 Raj Chetty, Nathaniel Hendren, Patrick Kline, and Emmanuel Saez, “Where Is the Land of Opportunity? The Geography of Intergenerational Mobility in the United States,” Quarterly Journal of Economics 129(4): 1553-1623, 2014, https://academic.oup.com/qje/article-abstract/129/4/1553/1853754. The full version, including appendices is also available at https://opportunityinsights.org/wp-content/uploads/2018/03/mobility_geo.pdf.

15 See p. 1616.

16 See Online Appendix Table VIII, Column 1.

17 The county-level mobility and teen birth rate estimates are available at https://opportunityinsights.org/data/?geographic_level=0&topic=0&paper_id=592#resource-listing, “Geography of Mobility: County Intergenerational Mobility Statistics and Selected Covariates.” For the 2013 CZ-level correlation, see Table 5 at https://web.archive.org/web/20130728192706/http:/obs.rc.fas.harvard.edu/chetty/website/IGE/Executive%20Summary.pdf.

18 See Online Appendix Table VIII, Column 5.

19 See Online Appendix Table VIII, Column 6.

20 See Table VI, Column 1.

21 See Table VI, Column 4.

22 See Table VI, Column 2.

23 See Table VI, Column 8.

24 See p. 1558.

25 See p. 1617.

26 See p. 1617.

27 Using the data files provided by OI, we can estimate the correlation across 594 CZs between community single motherhood rates and mobility for three groups: all adults, those raised by a single parent (mother or father), and those raised by married parents. Those correlations are -0.73, -0.75, and -0.61. See https://opportunityinsights.org/data/?geographic_level=0&topic=0&paper_id=592#resource-listing for the data files (Tables 5 and 8).

28 See p. 1616-1617.

29 Alex Bell, Raj Chetty, Xavier Jaravel, Neviana Petkova, John Van Reenen, “Who Becomes an Inventor in America? The Importance of Exposure to Innovation,” Quarterly Journal of Economics 134(2): 647-713, https://academic.oup.com/qje/article-abstract/134/2/647/5218522. See p. 689. Ungated version is available at https://opportunityinsights.org/wp-content/uploads/2019/01/patents_paper.pdf.

30 Raj Chetty et al., “Social Capital I: Measurement and Associations with Economic Mobility,” Nature 608: 108-122, 2022, https://www.nature.com/articles/s41586-022-04996-4#Abs1.

31 See p. 119.

32 See Supplementary Figure 2, Panels A-D at https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-022-04996-4/MediaObjects/41586_2022_4996_MOESM1_ESM.pdf.

33 See Supplementary Figure 5, Panels C and D.

34 See Supplementary Figure 14, Panels A and B.

35 See Supplementary Figure 3, Panel A and Supplementary Figure 4, Panels B and C.

36 Raj Chetty, John N. Friedman, Nathaniel Hendren, Maggie R. Jones, and Sonya R. Porter, “The Opportunity Atlas: Mapping the Childhood Roots of Social Mobility,” American Economic Review 116(1): 1-51, 2026, https://www.aeaweb.org/articles?id=10.1257/aer.20200108. Ungated version available at https://opportunityinsights.org/wp-content/uploads/2018/10/atlas_paper.pdf.

37 See p. 28.

38 See the supplemental appendix at https://www.aeaweb.org/articles/materials/24393, Online Appendix Figure IV.

39 See p. 24-25.

40 Raj Chetty and John N. Friedman, “A Practical Method to Reduce Privacy Loss When Disclosing Statistics Based on Small Samples,” AEA Papers and Proceedings 109, Papers and Proceedings of the One Hundred Thirty-First Annual Meeting of the American Economic Association (May 2019), pp. 414-420, https://www.aeaweb.org/articles?id=10.1257/pandp.20191109.

41 Raj Chetty, Nathaniel Hendren, and Lawrence F. Katz, “The Effects of Exposure to Better Neighborhoods on Children: New Evidence from the Moving to Opportunity Experiment,” American Economic Review 106(4): 855-902, 2016, https://pubs.aeaweb.org/doi/pdfplus/10.1257/aer.20150572. Ungated paper available at https://opportunityinsights.org/wp-content/uploads/2018/03/mto_paper.pdf.

42 See p. 869.

43 Raj Chetty and Nathaniel Hendren, “The Impacts of Neighborhoods on Intergenerational Mobility I: Childhood Exposure Effects,” Quarterly Journal of Economics 133(3): 1107-1162, 2018, https://www.jstor.org/stable/26864972?seq=1. Ungated paper available at https://opportunityinsights.org/wp-content/uploads/2018/03/movers_paper1.pdf.

44 Raj Chetty and Nathaniel Hendren, “The Impacts of Neighborhoods on Intergenerational Mobility II: County-Level Estimates,” Quarterly Journal of Economics 133(3): 1163-1228, 2018, https://www.jstor.org/stable/26864973?seq=1. Ungated paper at https://opportunityinsights.org/wp-content/uploads/2018/03/movers_paper2.pdf.

45 See Table A.11 at https://www.jstor.org/stable/get_asset/10.2307/26864973?supp_index=0.

46 See Table A.12.

47 See Tables A.13 and A.14.

48 Raj Chetty, Nathaniel Hendren, Frina Lin, Jeremy Majerovitz, and Benjamin Scuderi, “Childhood Environment and Gender Gaps in Adulthood,” American Economic Review 106(5): 282-288, 2016, https://www.jstor.org/stable/43861030. Ungated paper at https://opportunityinsights.org/wp-content/uploads/2018/03/gender_paper.pdf.

49 Many of these results are presented in appendix tables and figures, available at https://opportunityinsights.org/wp-content/uploads/2018/03/gender_paper.pdf.

50 See Appendix Figure 3.

51 U.S. Census Bureau, Current Population Reports, P60-206, Money Income in the United States: 1998, U.S. Government Printing Office, Washington, DC, 1999. https://www2.census.gov/library/publications/1999/demographics/p60-206.pdf. See Table 5.

52 These analyses were restricted to people whose parents didn’t change CZs. See Appendix Figures 6 and 7.

53 Among children of married parents, the cross-CZ variation in employment rates was lower among men than among women, regardless of parental income. But among children of single parents, the cross-CZ variation in employment rates was larger among men than among women for people raised in the bottom 60 percent of parental income, and the variation was greater among men the lower was parental income. While variation in male employment across CZs was lower the higher parental income for both men raised by married parents and men raised by single parents, the gradient was larger for the latter. See Appendix Figure 7.

54 See p. 286-287.

55 See Appendix Table 2.

56 See Appendix Table 3.

57 Raj Chetty, Nathaniel Hendren, Maggie R Jones, and Sonya R Porter, “Race and Economic Opportunity in the United States: an Intergenerational Perspective,” Quarterly Journal of Economics 135(2): 711-783, 2020, https://academic.oup.com/qje/article/135/2/711/5687353. Ungated version at https://opportunityinsights.org/wp-content/uploads/2018/04/race_paper.pdf.

58 See Figures II and IV.

59 See Online Appendix Table VI, available at https://www.jstor.org/stable/26935273?seq=1.

60 See p. 738.

61 See Figure V.

62 See Online Appendix Figure VI.

63 See Figure VIII.

64 See Online Appendix Figure VII.

65 See Figure V.

66 See Online Appendix Figure VII.

67 See Online Appendix Table XI.

68 See Online Appendix Tables XI and XIII.

69 See Online Appendix Table XII.

70 See Online Appendix Table XII.

71 See Figure XII.

72 See footnote 31, p. 763.

73 See Figure XII.

74 See Figure XII.

75 See Figure XII and Table II.

76 See Table II.

77 See Online Appendix Tables XI and XIII.

78 See Figure XIV.

79 Raj Chetty, Will S. Dobbie, Benjamin Goldman, Sonya Porter, and Crystal Yang, “Changing Opportunity: Sociological Mechanisms Underlying Growing Class Gaps and Shrinking Race Gaps in Economic Mobility,” Quarterly Journal of Economics (Forthcoming), https://academic.oup.com/qje/advance-article-abstract/doi/10.1093/qje/qjaf057/8417176. For working paper draft, see https://www.nber.org/system/files/working_papers/w32697/w32697.pdf.

80 See Tables A.3 and A.4.

81 See Figure A.8.

82 See Figure A.9, Panels I and J.

83 See Figure IV.

84 See Figures V and A.28.

85 See Table A.23, Figures V, A.24 (Panel F), A.27, and A.28.

86 See Table A.23.

87 See Table A.23.

In his speech at Turning Point USA’s Student Action Summit earlier this year, Tucker Carlson expressed the widely-held view that homeownership has become out of reach, to the detriment of family formation:

If you want a measure of how your economy is doing….my measure is really simple. I’ve got a bunch of kids. Can they afford houses with full-time jobs at like 27, 28? And the answer is no way. And the answer is that 35 year olds with really good jobs can’t afford a house unless they stretch and go deep into debt. And I just think that’s a total disaster….Nobody wants to raise their kids in an apartment. People do it because they have to. Nobody wants to. People want a little house, not some McMansion, just a little normal house. That is the actual American dream. And that is what is totally unattainable for young people.

However, there’s just one problem with blaming unaffordable homeownership for declining family formation: it gets the causality backwards.

It’s clear that homeownership has declined among young adults. I turned to the Annual Social and Economic Supplement (ASEC) to the Current Population Survey to look at the share of Americans ages 25-34 that were both homeowners and either a household head or the head’s spouse or partner.1 The data indicate that homeownership within this age group peaked in 1980 at 54 percent. By 2025, the share had fallen about 20 points to 35-36 percent.2

As an aside, this decline in homeownership was hardly confined to the young. Among adults ages 35-54, the homeownership rate that I calculated from the ASEC fell from 77 percent in 1980 (also the peak) to 60 percent in 2025. That 16.8-point drop is quite close to the 18.2-point drop among adults ages 25-34, though it is from a higher starting point.

Also clear is the fact that the share of Americans ages 25-34 that are married fell dramatically over this period—from 67 percent in 1980 to 37 percent in 2025. The question is whether the decline in the affordability of homeownership is responsible for the fall in family formation among young adults.

Most Young Newlyweds Have Always Been Renters

Let’s start by wiping the rose tint off the lens through which many view this question. Nostalgia about the halcyon past when homeownership and family formation went hand-in-hand is, as nostalgia tends to be, overly sunny. We can consider young first-time newlyweds using decennial census data and the American Community Survey (ACS). In 1960, among Americans under 35 years old who were one year into their first marriage (and still married), 83 percent were renters. As late as 1980, 70 percent were renters. As of 2023, 58 percent were renting—still a majority.3

If we look at first-time parents under 35 who are married and have a child under two years old, the share who rent tends to be lower than among young newlyweds. But renting was still the norm for new parents during the 1950s and 1960s. In 1960, 69 percent of young, married new parents rented their home. By 1970, it was still 65 percent. Only by 1980 had it fallen to a minority (47 percent). Today it’s 35 percent. If we include single parents, 49 percent of young first-time parents are renters.

That newlyweds and new parents are more likely to be homeowners today than in the past doesn’t necessarily mean that it’s become easier for young adults to own a home. If it really has become more difficult to own a home, we might see less marriage and fewer parents but more homeownership among the advantaged few who do marry or have kids. But the numbers do reveal that it’s simply a myth that most people in the mid-twentieth century delayed marriage and parenthood until they could buy a home. When the oldest boomers became first-time parents in the 1960s and 1970s, they were more likely to be renters than homeowners.

Nor do these figures suggest that it’s become harder for young married couples to afford a home after getting married but before having children. Subtracting the newlywed estimates from the new-parent ones, in 1980 there was a 23 percentage-point difference in homeownership. In 2023, the difference was also 23 points.

Let’s also say it again: a majority of young newlyweds and a third of new married parents today are renters; many young adults continue to marry and have children despite not owning a home.

Homeownership among Young Married Couples and among Young Single Adults Has Declined Modestly or Not at All

Homeownership trends for young adults look decidedly less dire if we take account of the decline in marriage. A simple way to do that is to separate out single and married Americans ages 25-34. I show these trends in the figure below, going all the way back to 1900 using decennial census and ASEC data.

Source: Decennial census public use microdata (1900-1980) and Annual Social and Economic Supplement to the Current Population Survey (1976-2025). University of Minnesota, www.ipums.org. Cohabiting partners of household heads are excluded from the analyses; other cohabiting partners are counted as single renters. Adults who are married but with an absent spouse or who are separated are also excluded. The “all” trend includes everyone regardless of marital status or cohabitation arrangements.

Note, first, the trend for single young adults (the bottom lines). Simply put, it has never been the case that a large share of single young people has been able to afford homeownership. Not when marriage was common, not when it was rare. To be young and a homeowner is generally to be married, and that has always been true.

That said, homeownership among single young adults rose steadily through 1980 and has remained elevated since. Homeownership has not declined relative to past generations. That’s the opposite of what we’d expect to see if home affordability were declining.

Turning to young married couples (the top lines), we see that a majority have been homeowners since the 1950s. The homeownership rate did decline after 1980, but by much less than the overall homeownership rate (the middle lines). The overall homeownership rate fell from 54 percent in 1980 to 35-36 percent in 2025, or by about a third. (The calculation is 1 - 35.5/53.7.) Among married couples, homeownership only fell from 71 percent to 59 percent. That 17 percent decline was about half as large as the overall homeownership trend suggests.

As recently as 2023, 63 percent of young married couples were homeowners. That was the same as in 1983 and only 3 percentage points lower than at the height of the 2000s housing bubble. The 2023 rate was also higher than in any year through 1970 and any year from 1985 to 1999.

Most of the decline among young married adults occurred between 1980 and 1983, due in large part to high interest rates that Federal Reserve Board Chair Paul Volcker imposed to arrest runaway inflation. That rate hike made mortgages costlier and also induced a deep recession. Homeownership among young married couples never recovered, though for the most part, it remained above 1970 levels.

The relatively high homeownership rate around 1980, like the subsequent peak during the pre-financial-crisis housing bubble, may be an unrealistic benchmark. In part, it reflected the interaction of the sustained high inflation of the 1970s with tax policy, which reduced the cost of owning relative to renting. That said, in the current inflationary environment, this relative cost of owning is even lower than it was in the 1970s.4

One might object to my separating out married young adults on the grounds that perhaps they became a more elite group over time, which could prop up the trend. However, if we look at the 56 percent of married young adults in 1980 with less than one year of college and compare them with the 51 percent of married young adults in 2025 with less than a bachelor’s degree, the drop in homeownership was from 68 percent to 54 percent instead of from 71 percent to 59 percent (a 21 percent drop instead of a 17 percent drop).

Simply put, married couples are in a better position than single young adults to afford a home and always have been. A big reason for the 18-percentage-point drop in young adults’ home ownership rate is the massive 30-point fall in marriage. For purposes of affording a down payment or a mortgage, sharing expenses and pooling incomes have always been important. Indeed, if we look just at those young married couples with two earners, the drop in homeownership has been even smaller than the decline for young married couples generally. It fell from 72 percent in 1980 to 62 percent in 2025 (and was 67 percent in 2023).

It’s Not Clear the Age of First-Time Homebuyers Has Increased

According to the National Association of Realtors, the median age of homebuyers has risen over time, from 31 in 1981 to 59 by 2025. Among first-time homebuyers that increase has been smaller, but still from 29 to 40 years old. However, according to a second data source, from the Federal Reserve Bank of New York (FRBNY), the median age of first-time homebuyers declined from 2000 to 2024, from 38 to 36 years old, even though the NAR estimates rise from 32 to 38 years old over this period. The FRBNY estimates come from a large sample of credit reports, while the NAR results are from a relatively small survey with low response rates.

Unfortunately, the FRBNY data do not go further back in time, so we don’t know what they would show for 1981. We do know that from 1981 to 2000, the NAR estimates for first-time buyers increase from 29 to 32 years old. If the FRBNY age estimates also would have risen 3 years over this period, the 1981 to 2000 change would be from a median age of 35 to 38 years old, and the change over the entire 43 years would be from 35 to 36 years old.

Even using the NAR estimates, the increase from 1981 to 2021 for first-time buyers was only from 29 to 33. Indeed, the NAR data indicate that in the four years from 2021 to 2025, the median age rose by 7 years—nearly twice as much as over the preceding 40 years. To the extent this data is accurate (despite clear reasons to prefer the FRBNY data) the jump after 2021 probably was initially due to rising home prices. That increase stemmed from the Federal Reserve Board’s large purchase of mortgage-backed securities during the COVID-19 pandemic and the broader post-pandemic inflation. When the Fed increased interest rates to rein in that inflation, it caused a jump in mortgage rates, making homebuying even more unaffordable.

If the NAR data are right, the rise in median age of homebuyers was modest until very recently, suggesting that if the cost of homeownership has been a barrier to marriage, it has been so mainly for the past 4 years rather than the past 45 years. If the NYFRB data are right, any increase in median age of homebuyers had stopped by 2000, and it’s likely to have changed little over 45 years.

Home Sale Prices Have Risen Relative to Incomes, But Not as Much as It Might Appear

As evidence of an affordability crisis facing young adults, Patrick Brown of the Ethics and Public Policy Center and others have pointed to the rise in the median sale price of homes sold relative to median household income. Brown cites an increase in that ratio from something like 3.6 in the mid-1980s to around 5.3 by 2023. I combined my own ASEC figures using family income (including cohabiting couples as families) with the same median home price data used by Brown. I get a similar result, finding a rise in median home price to median income from 3.8 in 1980 to 5.4 in 2024.

Certainly, one reason for this increase is that the supply of homes that are affordable to young adults has diminished with the spread of land use and zoning regulations. This factor is especially important in particular metropolitan areas and regions of the country. But it doesn’t seem to be the most important one for understanding national affordability trends among young adults. If we divide median home sales price by the median family income of married adults ages 25-34, that ratio rises between 1980 and 2025—but only from 2.9 to 3.5. The typical home seems only a little more out of reach for young married couples than it used to be.5

The Cost of Buying a Home Has Risen Less Than Home Prices (And Pay)

It’s useful to think of the cost of homeownership as comprising the cost of buying a home and the cost of living in a home. The initial cost involves the various amounts paid at closing, as well as the down payment. (Note that a down payment really is just a transfer of assets from, say, a checking account to real estate, though of course one must initially have enough assets to transfer.) When we combine data on down payment amounts with sales price data, we find that this part of the cost of buying a home has risen by less than sales prices. I have located data only back to 1990, but from then to 2019, while the median sales price of homes sold rose by 162 percent, down payments rose only 73 percent.6 For comparison, the nominal (non-inflation-adjusted) hourly wages of private production and nonsupervisory workers—a common measure of the typical worker’s hourly pay—rose 130 percent over the same period. Other research (see Figure 5) indicates that the average down payment as a share of home price increased slightly among first-time homebuyers between 1980 and 1990, suggesting that if I could extend my analysis back to 1980 it would tell the same story.

The bottom line is that home prices outpaced wages, but down payments did not. The result supports the importance of declining marriage in explaining falling homeownership: it’s not that down payments have become more expensive, or even that they’ve become more expensive relative to wages; it’s that fewer people can afford a down payment because they are not combining their wages with a spouse’s income.

The cost of living in a home is more complicated, but mortgage interest is the largest expense for most homeowners with a mortgage. Those rates fell steadily until the past 4 years and were 71 percent lower in 2019 than in 1980.7 Even after the recent spike, the average 30-year fixed rate mortgage interest rate was 51 percent lower in 2024 than in 1980.

Moreover, rising home values increase the value of home equity for existing homeowners, and this increase constitutes income (or a negative cost) in the form of accrued capital gains. Rising home prices also increase the value of the “service flow” from not having to pay rent to a landlord for the roof over one’s head, another form of income. While they will also tend to increase the cost of property taxes and homeowner’s insurance, the relative cost of living in an owned home versus renting is lower today than in 1980 (Figure 6).8

Young Men Are As Marriageable As In the Past

Adjacent to claims that the inability to afford a home is behind declining marriage rates is the assertion that men have become less “marriageable” than in the past—that they are economically worse off. If men are less marriageable today, that would make them less able to afford a home, which would lead to less marriage.

In past research, I tested this assertion by defining “marriageability” as a fixed percentile (the median or the 25^th percentile) of the earnings distribution of married fathers ages 25-29 in 1979 who were sole earners. Young men whose inflation-adjusted income clears this threshold are marriageable by this definition, since half to three-quarters of sole-earner married fathers of the same age were making at least that much 45 years ago.

I found that in 2020 men were either more marriageable than in 1979 (using a higher marriageability threshold) or modestly less marriageable (using a lower threshold). In ongoing analyses that treat non-workers differently and use an improved price index to adjust earnings for inflation, I find that young men are unambiguously more marriageable than in 1979. In another recent paper that uses these improved methods, I report that median earnings of men ages 25-29 rose 21 percent from 1973 to 2023 after adjusting for inflation and 25 percent from 1989 to 2023. (Not incidentally, median earnings among young women rose 60 percent from 1973 to 2023. No marriageability problem there.)

Claims that Many Young Adults Delay Marriage or Family Formation Until They Can Buy a Home, Thus Explaining the Decline in Marriage, Are Invariably Anecdotal

I searched the massive Roper Center iPoll public opinion archive (subscription required) for polling questions that directly ask people whether they are delaying marriage or childbearing until they can afford to buy a home. There is plenty of evidence that many people would like to own a home but can’t afford to buy one. There is plenty of evidence that many married couples would like to start a family or expand their family but cannot afford to do so and that many people would like to get married but can’t afford it. These questions on family formation and marriage tend to find that sizable minorities of people who have not achieved their family goals cite financial concerns.

For instance, an especially relevant survey conducted this past fall indicated that 27 percent of adults ages 18-28 who did not own their own home but wished to be a homeowner said they were delaying getting married or formalizing a long-term partnership until they could afford to buy a home.9 Of course, this is one side of the question—we don’t know how many homeowners in this age group did buy a home before getting married or how many are renting but didn’t delay marriage.

More importantly, there is essentially no trend evidence that such perceptions or preferences have changed over the long run. Is that 27 percent of young renters delaying marriage higher or lower than in the past? In theory, it could be at an all-time low, which would hardly be consistent with the claim that fewer people are marrying because of falling home affordability.

Nor did I find any contemporary evidence about the share of young adults who will only marry once they can afford a home. Invariably, when people claim that so many young adults won’t marry until they can afford to buy a home that it has produced a large decline in marriage, it’s simply an assertion or justified with anecdotes.

Conclusion

It matters a lot for important questions of policy which is closer to the truth: economic factors, such as the cost of homeownership, are depressing marriage; or falling marriage is making it harder to afford major purchases, such as buying a home. Many people agree with venture capitalist Peter Thiel, who believes that the cultural resentments of young adults are rooted in housing affordability. Theil recently offered that

I think you can reduce 80 percent of culture wars to questions of economics—like a libertarian or a Marxist would—and then you can reduce maybe 80 percent of economic questions to questions of real estate. It’s extremely difficult these days for young people to become homeowners.

If the decline in marriage and other aspects of social breakdown—from political polarization to widespread loneliness to falling social trust—have economic causes, it may be possible to identify villains at fault (neoliberal politicians or corporate elites), taxes to cut, subsidies to increase, and regulations to tweak. Problem solved.

However, if the decline in marriage is more to do with cultural change—with shifting preferences or other nebulous societal trends—economic fixes won’t do (and we may need to collectively look in the mirror to assign blame).

I’ll close by noting one group of young adults for whom homeownership has plummeted—married couples with one earner. Their homeownership rate fell from 70 percent in 1980 (about the same as for two-earner married couples) to just 49 percent in 2025. That was lower than in 1970 (65 percent) and even in 1960 (60 percent).

Arguably, these families have been hurt by the greater purchasing power that dual-earner families command, which leaves them at a financial disadvantage when looking for family-friendly homes. The evidence reinforces a conclusion I have made elsewhere: it is difficult for one earner to afford what two earners can buy, and the increase in dual-earner families has left sole-earner families feeling worse off. Indeed, due to housing supply restrictions, it’s possible that sole-earner families are actually worse off than they’d otherwise have been if not for the popularity of the dual-earner model, at least in regards to homeownership.

This nostalgia for the sole breadwinner model seems to be behind much of the new right’s economic agenda. But the share of young adults pursuing the one-earner model was already small by 1980. According to decennial census data, in 1960, married people ages 25-34 in one-earner couples were 52 percent of all young adults, and they were 41 percent in 1970. By 1980, according to the ASEC, that was down to 22 percent, falling further to just 10 percent in 2025. That drop partly reflected the decline in marriage, but even among married young adults, the share in one-earner couples fell from 62 percent in 1960 to 50 percent in 1970 to 31 percent in 1980 and 25 percent in 2025.

It’s unlikely that this decline in the one-earner model reflects increased economic difficulty. As I’ve argued elsewhere, the patterns of increased employment among married women are similarly impressive among those with more and less educational attainment, those with husbands who have more and less educational attainment, and those with higher- and lower-earning husbands. The rise in labor force participation among women baby boomers was preceded by an increase in their educational attainment. These trends long predate any slowdown in wage growth among men, and they occur across the rich world (see Figure 5.1). As the opportunity cost of keeping one spouse out of the workforce has risen, fewer families have done so.

The culture has changed as the nation has become richer. We would do well to recognize that, to accept (or not accept) the trade-offs that our collective choices have produced, and to stop squinting for signs of long-run economic deterioration to wash the reality of these complicated cultural changes away. That doesn’t mean we shouldn’t try to increase incomes or lower housing costs. But it does mean we should perhaps worry less about those things and more about social breakdown.

I’ll spend the bulk of this post critiquing the LWC numbers, but it’s first worth emphasizing how misleadingly Green characterized his numbers.

Subscribe now

First, he claims to have “built a Basic Needs budget for a family of four” himself. No citation of the LWC at all. (Actually, as I was writing this, the Free Press version of his post was updated to include a link to the LWC site. His original post was unchanged when I published this.) It’s a data-analytic equivalent of plagiarism not to cite someone whose numbers you use and to pass them off as your own.

Green also indicates that the numbers he provides involve, “conservative, national-average data.” But he revealed to @MTSInsights that the figures were for Essex County, New Jersey. (In fact, they are here.) Now, you might think “conservative, national-average data” reflects the national average or, being conservative, somewhere that is poorer than the national average. You would be wrong. Out of 3,114 counties with per capita income estimates available from the Bureau of Economic Analysis for 2023, Essex County, New Jersey is the 224^th richest. (It includes Newark and, more importantly, its western suburbs.) Per capita income in Essex County was $80,460 in 2023, 15 percent higher than the US figure ($69,810). Jeremy Horpedahl noted on X.com that median family income in Essex County for a married couple with children was $174,644 in 2024—nearly one-third higher than the national median of $132,959. We should expect that expenses in Essex County will be 15-30 percent higher than national figures too. And, indeed, I found in my last post that his averages were too high by 20 percent. (We’ll see below that they’re even more off than these estimates.)

Finally, while I noted that Green said nothing about his “Other essentials” category in his post (a $21,857 expense), I can tell from the LWC documentation what it represents: “Entertainment: fees and admissions; Audio and visual equipment and services; Pets; Toys, hobbies, and playground equipment; Entertainment: other supplies, equip., & services; Reading; and Education,” “Apparel and services; Housekeeping supplies; Personal care products and services; Household furnishings and equipment; and Miscellaneous household equipment,” and “Cellular phone service.” (These are the components of the LWC categories, “Civic Engagement,” “Internet & Mobile,” and “Other Necessities.”)

It takes a bit of digging to uncover this, but these categories reflect the entirety of spending tracked in the federal Consumer Expenditure Survey (CEX) on categories outside the other ones Green lists (excepting alcohol and tobacco, and a couple other small categories). So whether we call them essentials or necessities, what they really are is “all other spending,” more or less, as I anticipated in my last post. In particular, while Green says his budget involves “no vacations, no Netflix,” that’s simply not true; they are implicitly included in his “other essentials.”

OK, let’s move on from Green’s obfuscation and look at the Living Wage Calculator.

First of all, lest one make too much of the authority of the “MIT” on Glasmeier’s CV, it’s worth emphasizing that DUSP is not part of MIT’s renowned economics program. It sits within the university’s School of Architecture and Planning. Just so it’s clear where we’re starting from, here’s the mission of DUSP:

We prepare leaders to plan, design, and create communities and places that are socially, economically, and environmentally just.

We pursue this mission through four evolving strategic priorities: achieving racial justice, enhancing multi-racial democratic governance, tackling the climate crisis, and closing the wealth gap.

We aim to fulfill our mission through all aspects of our teaching, research, department culture, and external collaborations.

Now, the clear point of view doesn’t necessarily mean that the department has sacrificed rigor for ideological goals and doesn’t necessarily mean that the Living Wage Calculator is bunk. We need to dive into the calculator’s methods to ascertain that. But when you do that, you find that the Living Wage Calculator is bunk.

Let’s look at each of the LWC categories in turn. There are two different potential problems here: the fact that Essex County isn’t representative, and possible issues with the LWC methodology. The LWC doesn’t seem to provide national estimates, so I’ll compare the Essex County figures to those for Montgomery County, Alabama, which is roughly the median county for per capita personal income. I’m going off of the documentation LWC provides.

I’ll report the Essex County amounts (Green’s), the Montgomery County amounts, amounts from national CEX data for married parents, and a preferred measure to use. At the end, we’ll sum the category amounts and compare them to incomes. As I emphasized in my last post, this is a bad thing to do! But it is the bad thing Green does, so let’s kick the tires a bit.

Food

The LWC gives the typical food expense in Essex County for a family with two workers and two children as $14,717. (Everything I cite from the LWC below will be for two worker families with two children.) This estimate is superior to several of the other categories we’ll get to in that it is not simply an average expenditure amount but based on some standard of need (from the US Department of Agriculture). The USDA amounts vary by family size and type and reflect the cost of a given standard if food is purchased and then prepared at home. It is adjusted by the LWC folks to reflect county-level amounts using other data. It’s not an unreasonable approach.

The amount for Montgomery County is not much lower than the amount for Essex County—$12,811. That said, nationally, married couples with children spent $15,059 on food in 2023. Forty percent of that was on food away from home; just over $9,000 was spent on food at home. These are rough bounds on how much the average married couple family with children would have to spend if they bought all their food at home. Let’s just go with the LWC amount for Montgomery County ($12,811) as the preferred measure.

Childcare

According to the LWC, typical childcare expenses in Essex County for a family with two workers and two children amount to $32,773. The LWC childcare amounts are based on the cost of center-based care. Nationally, just 37 percent of children under age 5 and not yet in kindergarten are in center-based care. So this component of the “cost of existing” is relevant for barely a third of children. Since center-based care is more expensive than other forms of child care, it’s no wonder that even among families that pay for childcare, the average amount is less than half what the LWC indicates for Essex County.

If we look at Montgomery County, the LWC reports the typical expense as $15,006. Again, this is the amount for people who pay for center-based care; it would be lower for families who don’t pay for care or who pay for less formal arrangements.

In the national data, I obtain a value of under $16,000 for 2022 for people paying for care, an amount that falls to $5,300 in 2023 for married-couple families with no child over age five (including those not paying for any care and those with fewer or more than two children). It’s just $1,525 for married-couple families with a child older than five, and if you just look at all married-couple families with children, it’s $1,704. That’s the most appropriate estimate to use, because families with children do not spend an average of $5,300 every year a child lives at home. Remember, though, that this is an average and not a standard of need (unless you think everyone should be able to afford the average).

Medical

The amount for Essex County is $10,567. It is based on (1) the employee share of premiums for health insurance sponsored by private employers and (2) out-of-pocket spending on prescription drugs, medical services, and medical supplies. These are state and national averages that are adjusted to the county level with other data sources. The two county amounts are then presumably added together.

Again, these are averages, not standards of need. The insurance premium component is the amount that people with employer coverage spend, which reflects the comprehensiveness of the coverage offered. If some people reject employer coverage or are not offered it, they might obtain public coverage, including subsidized premiums for many middle-class families. About three in four married non-elderly adults who have health insurance get it from their employer.

Two other potential problems to flag. First, adding these two averages (premium expense and out-of-pocket costs) overstates average spending on health care. People who pay more for their premium will tend to have better coverage and lower out-of-pocket costs. Second, the use of averages is also likely to be particularly problematic for out-of-pocket costs given that a small number of people who are sick or disabled may have very large spending that pulls the average up so that it doesn’t reflect the typical person. (In 2022, five percent of people accounted for half of health spending.) Even for employee premiums for family coverage, the mean is 18 percent higher than the median (see Table X.D.1). Third, the out-of-pocket averages are not confined to those with health insurance. Basing the averages only on people with employer coverage would probably increase them.

In Montgomery County, the LWC gives the typical expense as $9,570. Nationally, average healthcare spending for married couples with children was just $7,600. Since the LWC is using the same source as me for the out-of-pocket amounts, the difference is due to the insurance premium estimate (or the county-level adjustment). Let’s use $9,570 for the preferred amount, just to be conservative.

Housing

The LWC estimate for Essex County is $23,267. This estimate is based on Department of Housing and Urban Development data on the 40^th percentile of rents in metropolitan areas. It is actually an estimate for the entire Newark metro area, which includes three counties other than Essex. There are 4,764 entries in the latest HUD spreadsheet with rents (multiple ones per county), and when I ranked them by the rent for a 3-bedroom apartment, Essex County was at the 94^th percentile (meaning one of the most expensive).

While the 40^th percentile of rents is something like a needs standard (rather than simply an average), it’s also definitionally true that 40 percent of people will be below it. If you think that 40 percent of people can’t meet their housing needs, that might make sense to you, but if you think only 20 percent of people can’t, you’d want to use the 20^th percentile. You see the problem here.

Moreover, homeowners face very different housing costs than renters. Fully 80 percent of married couples with children are homeowners. So at best, this measure gets it “right” for the one-fifth of families that rent. The costs homeowners face include mortgage interest payments, homeowners insurance, maintenance and depreciation, and property taxes.

In Montgomery County, the LWC reports a typical housing amount of $12,707—much lower than in Essex County. The national average for 2023 for married parents was $19,946 for shelter. This is higher than the Montgomery County amount because it is an average, not the 40^th percentile of shelter spending. The 40^th percentile would be a much more useful needs standard. Let’s use $12,707 for the preferred estimate.

Transportation

It’s not at all clear to me how the LWC estimates its transportation costs, and I view this component as the least reliable of the major ones. It uses an elaborately (not necessarily well-) modeled set of county-level estimates from another source and then adjusts them further. The estimate for Essex County is $14,828, compared with $16,392 for Montgomery County. (Of note is that the latter is higher, perhaps because it is more rural and requires more driving.) Nationally, the average spending on transportation for married parents was $19,343. I think that’s largely due to differences in how the cost of buying a new car are treated. Absent net outlays on vehicle purchases, the national figure is $10,800. That’s too low, but it’s also not the case that buying a new car entails the full cost of the outlay—one is partly transferring assets from a bank account to a vehicle, though the vehicle depreciates rapidly and one will incur interest expenses if a car loan is used to finance the purchase. All of these estimates are based on averages, not any needs threshold. Sure, let’s go with $16,392 for the preferred estimate.

Civic Engagement

This is entirely based off of the CEX (the same source I’m using) and reflects average spending amounts (adjusted to the county level). Specifically, it includes spending on “audio-visual equipment; education; fees and admission; other entertainment; pets; reading; and toys, hobbies, and playground equipment.” It’s an odd set of goods and services to lump under “civic engagement,” and it’s part of Green’s “other essentials” category. In the CEX, these spending amounts are subsumed in just three categories: entertainment, reading, and education. The reading amount is negligible.

If Green and the LWC were being consistent with their approach to childcare, they would (wrongly) determine how much families with non-zero spending on education pay and assign that to “education” (as if that amount were paid every year rather than primarily occurring when children are in college).

You can think some level of spending on these things represents a necessity, but it’s probably not the average amount people spend. The Essex County amount is $8,810, while in Montgomery County it is $6,450, again, much lower. The national average for married parents for these categories is $8,872. If we’re talking about “essentials,” let’s stick with the education amount for our preferred estimate, which is $3,301.

Internet & Mobile

Mobile service again comes straight from the CEX and represents average spending amounts (adjusted to the county level). Internet spending is based on county-level data on low-cost internet service plans. The “low-cost” distinction makes it a useful needs standard.

In Essex County, the typical spending was $2,001; it was $2,142 for Montgomery County. Nationally, for married parents, I estimate it would be about $2,500 in 2023, based on the amounts for internet service and cell service for all families and the spending on mobile service for married parents. Let’s use mine to be conservative.

Other Necessities

This category basically includes the remaining categories of spending, and uses averages from the CEX. Specifically, it includes spending on “apparel; household furnishings and equipment; housekeeping supplies; personal care products; and miscellaneous household equipment.” At this point, the LWC is declaring that all categories of spending, in fact, are necessities, and at least for these ones, everyone should be able to afford what the average family affords.

I’m pretty sure that the LWC is double-counting “miscellaneous household equipment,” since it is part of “household furnishings and equipment” in the data we’re both using.

The typical amount in Essex County was $11,046, while it was $9,120 in Montgomery County. Nationally, spending on these categories among married parents amounted to $8,881 ($10,675 if I double-count miscellaneous household equipment). We’ll use the correct national total for the preferred amount.

Taxes

The LWC assumes that everyone spends their entire after-tax income, then it runs the amounts for different family types through a widely-used tax simulator called TAXSIM to get income and payroll taxes. However, I’m pretty sure the LWC folks are understating these taxes, because they assume that after-tax income is pre-tax income and then compute taxes on the after-tax income. That’s the only way I can interpret the methods that are offered (and being familiar with TAXSIM, I think it would be very difficult to go from after-tax income and back into pre-tax income). On the other hand, by summing the amounts in each of the spending categories, they are overstating how much people really spend, or need to spend, as I discussed in my last post. That will tend to push the tax estimates too high.

The LWC reports $18,488 for Essex County and $11,538 for Montgomery County. Nationally, the average for married parents is $22,862—much higher than the Montgomery County amount, which suggests I’m right about the error in the LWC. We’ll use my estimate below.

Summing Up (Remember: Don’t Do This!)

OK, we can now sum these categories and see what we get. As a reminder: this is a bad idea, for the reasons I went through in my last post.

We already know that the total for Essex County—Green’s total—is $136,498. That’s his “cost of existence.” If we use the estimates for Montgomery County instead, we get $95,735. That’s lower by 30 percent. Put another way, Green’s estimate is higher than the median county by 43 percent. If we use the preferred estimates I flag for each of these categories for married parents, we get $90,728.

These amounts are too high as indicators of what people “need” to spend because some or all components reflect average spending rather than minimum needs and because averages are skewed by outliers and don’t reflect what’s even typical. They’re even too high as indicators of average spending: average spending does not equal the sum of component averages (since people trade off different kinds of expenses, especially at different points in their lives).

Nevertheless, we can look at how these amounts compare to incomes and see whether Green is right that much of the population can’t participate in American society. Green mentions in his post that median family income is “roughly $80,000.” Median household income is $81,604. That’s below his needs threshold by 40 percent. Apocalypse Now!

But of course, it’s a no-no to compare the national income amount to the higher-than-average Essex County spending amount. In Essex County, median family income among married parents is $174,644, which is 28 percent higher than Green’s (lousy) “needs” threshold. In Montgomery County, median income is $116,929, or 22 percent higher than the (still lousy) median-county threshold. And nationally, among married parents, median income is $132,959. That’s 47 percent higher than the (still lousy!) preferred spending threshold. [Added 11/30/25: I should also have emphasized that the median incomes here are for all married parents, not married parents with two workers, nor married parents with two children. Incomes for married parents with two workers and two children would be that much higher than these thresholds!]

I can’t emphasize enough that you should not use any of the thresholds I’ve reported here: they are all too high as indicators of what middle class families “need” because they are all much too close to simply being measures of what middle class families spend (and they are overstated measures of what middle class families spend). If everyone’s income doubled tomorrow, it’s likely that everyone’s spending would also roughly double. Would our “needs” also be required to double in that event? Would we be no better off as a society? No more secure? Of course not. But that’s the sort of logic behind Michael Green’s (and the Living Wage Calculator’s) analyses.

Finally, I have to say that it is remarkable that of the 3,000+ counties Green could have chosen, he found one that reinforced his $140,000 poverty line claim from his separate (also flawed) analyses.

I’ve spent more time on Green’s post than I have wanted to, and it’s possible I’m not done. This stuff needs to be discredited because it conveys a completely inaccurate picture of how the American economy is doing. It’s nice to get hundreds of thousands of followers—I wouldn’t know!—but followership is not the same as accuracy. Nor are vibes.

Here I want to address some of the pushback I received on the critique. A number of people on X said that my technical nerd-talk about poverty measurement and my rejection of Green’s claim that the poverty line should be $140,000 missed the point. Green wasn’t really saying that $140,000 should be the new poverty line. He was just making a point about the cost of living being so high today that it has hurt even Americans making six figures.

Pardon me, but I think I do pretty well with reading comprehension. There’s really no ambiguity in these passages from Green’s essay:

Which means if you measured income inadequacy today the way Orshansky measured it in 1963, the threshold for a family of four wouldn’t be $31,200. It would be somewhere between $130,000 and $150,000. And remember: Orshansky was only trying to define “too little.” She was identifying crisis, not sufficiency. If the crisis threshold—the floor below which families cannot function—is honestly updated to current spending patterns, it lands at $140,000.

The official poverty line for a family of four in 2024 is $31,200. The median household income is roughly $80,000. We have been told, implicitly, that a family earning $80,000 is doing fine—safely above poverty, solidly middle class, perhaps comfortable. But if Orshansky’s crisis threshold were calculated today using her own methodology, that $80,000 family would be living in deep poverty.

“So when I say the real poverty line is $140,000, I’m being conservative.”

But look at that chart through the lens of the real poverty line. If the cost of basic self-sufficiency for a family of four—housing, childcare, healthcare, transportation—is $140,000, then that top light-blue tier isn’t “Upper Class.” It’s the Survival Line. This chart doesn’t show that 34% of Americans are rich. It shows that only 34% of Americans have managed to escape deprivation. It shows that the “Middle Class” (the dark blue section between $50,000 and $150,000)—roughly 45% of the country—is actually the Working Poor.

The Lie

So that’s the trap. The real poverty line—the threshold where a family can afford housing, healthcare, childcare, and transportation without relying on means-tested benefits—isn’t $31,200. It’s ~$140,000.

And that’s without even noting the subtitle of the condensed piece he wrote for The Free Press: “Why $100,000 Is the New Poverty.”

Others complained that I didn’t have anything to say about Green’s analyses using expenses on different categories of spending. Well, actually, I did have something to say: “Much of the essay involves other arguments about middle-class expenses, and I’ll try to address those claims separately. Here I just want to convince you that Green’s claim about poverty is, I’m sorry, hot garbage.”

In the meantime, Jeremy Horpedahl published an excellent critique of those analyses, which I encourage you to read. But there’s more to be said!

At two different points in his essay, Green reports average spending for a number of product categories and then explicitly or implicitly adds them to show how unaffordable life is today. There are several problems with this methodology. First, let’s remember that averages, as opposed to medians, are pulled upward by a relatively small number of high values (in this case, big spenders). Median household income in the US is under $84,000, but the mean is $121,000. The average in this case is 45 percent higher than what the typical household gets. Using averages overstates what typical people spend and, therefore, what typical people need.

Second, as Horpedahl notes, averages should not be floors. Indeed, medians should not be floors. By definition, half of families will have incomes below median family income. If the median is the floor, half of families will be poor by definition.

More to the point, if the average family is spending some amount, we should not assume that this is the amount they “have to” spend, as Green does. If everyone’s incomes doubled, would they now “have to” spend twice as much on, say, housing? One suspects that in this world of abundance, Green would decry how the poverty line has risen to $280,000. This is the same problem I flagged in my last post—the issue of spending going up because we get richer and buy more and better things, versus spending increasing because we have to pay more for the same things. It’s only the latter that is bad.

Third—and this holds for the median too—someone spending the average of total expenditures typically does not spend the average amount for each product category. In the real world, some families spend relatively more on some things and relatively less on others, depending on their life stage and unique needs and preferences. Few people spend the average or typical amount on everything.

Take childcare. As Horpedahl notes, childcare costs are incurred over a relatively small number of years in an adult’s lifetime. In many years, they are $0—even for parents (and future parents). Many adults don’t become parents and incur $0 in childcare costs. Many others have a stay-at-home parent during the years that they’d otherwise need childcare; their childcare costs are also $0. The same arguments can be made for higher education expenses.

In general, people make tradeoffs in life—if they want or have to spend more on, say, higher education, they spend less on retirement savings. Then they can spend more on retirement savings later once they are done paying for higher education. If you want to spend more than average on food at restaurants, you may need to spend less than average on vacations. If you are spending the average amount on every category of product purchased by citizens of one of the richest nations the world has ever known, you are probably not poor.

There’s also a fourth problem, which is that Green doesn’t cite a source for any of his averages. I truly hope he rectifies this issue, as it would help shut down the online speculation that he just used AI to write the piece.

To see how these issues affect Green’s analyses, consider the following passage:

The composition of household spending transformed completely. In 2024, food-at-home is no longer 33% of household spending. For most families, it’s 5 to 7 percent. Housing now consumes 35 to 45 percent. Healthcare takes 15 to 25 percent. Childcare, for families with young children, can eat 20 to 40 percent.

Again, where these numbers come from is unclear. But if we take the midpoints of each of these ranges, then food at home, housing, healthcare, and childcare consume 96 percent of household spending. That leaves practically no room for education, clothes, food away from home, transportation, entertainment, vacations, or retirement savings. Yet, as discussed in my last post, the share of spending devoted to non-necessities has risen quite a bit over time. How can that be?

One answer is that these estimates appear inflated. The US averages for 2023 are 8 percent for food at home, 33 percent for housing (including utilities, cell phones, housekeeping supplies, appliances, furnishings, and other stuff), and 8 percent for healthcare. (Childcare is included in the “housing” category.) That sums to 49 percent, or about half of what Green estimates.

If we want to separate out childcare from that housing category, then something less than one percent of expenditures was on childcare. Yes, you read that correctly. The reason for it being so low is that lots of households don’t incur childcare expenses. Green is reporting an amount (again, unsourced) for families with young children.

However, even for married-couple parents in which the oldest child is under 6 years old, under 5 percent of spending goes to childcare. (“Personal services,” which includes childcare, accounts for 5.2 percent.) Add in the rest of “housing,” food at home, and healthcare and you have accounted for 48 percent of total spending—half what you’d get from Green’s numbers. If you pull the other stuff out of housing that the reader is unlikely to think is included, the remainder (“shelter”) is just 19 percent of spending (rather than Green’s 35-45 percent). The total accounted for by Green’s categories then falls to 33-38 percent of spending (with the uncertainty due to the unknown childcare share). [Update 11/27/25: I realized one can use another spreadsheet to get the breakdown for 2022. Nearly all of “personal services” involves child care, so 38 percent is the better estimate. That spreadsheet also lets one calculate how much people with positive spending on childcare spend. It was under $16,000 in 2022—less than half of Green’s estimate of $32,773.]

And again, not everyone spends the average amounts on each of these items. The average amount spent on childcare is surely greater than 5 percent among the families using paid childcare, but it’s precisely 0 percent for a lot of families. For those families, they may spend quite a bit more on categories Green has left out. That’s also true of some people who spend less than the average on healthcare or their home.

Green later shows “conservative averages” (again, without citing sources) for eight categories of expenses, totals them, and finds that $136,500 is “required” to participate in society. It is the “cost of existence,” the “floor.” Conveniently, it is also close to the $140,000 he gets from his…innovative new poverty measure.

But again, these are category averages, and few people spend the average amount on every category. If they did, it’s hard to see how they would be poor.

Moreover, Green is exaggerating costs again. Let’s try and recreate his total, using the 2023 data for married-couple parents with no child age 6 or older. That will maximize the share of spending going to childcare. I’ll leave out his mysterious “other essentials” category for now, since who knows what’s in there. That leaves his total at $114,652. My total is $92,222—20 percent lower. If I compute total spending plus taxes, I only get to $122,707. That suggests that Green’s “other essentials” category may just be “all other spending.” (By the way, if I use the figures for all households, single or married, with or without kids, the spending total is just over $91,000, so the issue isn’t my focusing on married parents with young children.)

Finally, I’d be remiss if I didn’t marvel at how close Green’s mysterious $136,500 is to the $140,000 poverty line he estimates using the “multiplier” approach I criticized in my last post. And to think that without the undocumented “other essentials” category, it would fall short by $22,000.

I don’t think this will be the last of my posts on Green’s long essay. There’s too much bad information in there to stop now.

The essay is by Michael W. Green, who is chief strategist and portfolio manager for Simplify Asset Management. The Free Press version is subtitled, “Why $100,000 Is the New Poverty,” but Green actually argues—I am not making this up—that the new poverty is $140,000. Since the typical household only makes a little over $80,000 a year in income, that means roughly two-thirds of Americans are poor.

It is…The Worst Poverty Analysis I Have Ever Seen. (And I’ve read Matthew Desmond!)

Much of the essay involves other arguments about middle-class expenses, and I’ll try to address those claims separately. Here I just want to convince you that Green’s claim about poverty is, I’m sorry, hot garbage. Rather than the Free Press piece, I’ll be critiquing his longer Substack post.

According to the official poverty measure (OPM) used by the federal government, 19.5 percent of Americans were poor in 1963, falling to 10.6 percent in 2024. The official poverty line is supposed to reflect a constant standard of living, after adjusting for inflation over time. There are actually 48 official poverty lines, depending on how many people are in a family, how many children there are, and whether the household head is under age 65. For a family of four with two children, the threshold was $31,812 in 2024. If a family’s income was below that line (or, rather, the line specific to their family type), they were officially poor.

These thresholds are not entirely arbitrary, but nor are they especially meaningful. The first attempt at a kind of official federal poverty rate came in the 1964 Economic Report of the President, which adopted a $3,000 poverty threshold for families and a $1,500 threshold for individuals living without family. (My discussion draws on a definitive history by Gordon Fisher.) The $3,000 was consistent with research by Mollie Orshansky, a government economist who had estimated a $3,165 threshold for a nonfarm family of four. Orshansky had gotten to this figure from the starting point that in 1955, families with at least three people spent an average of one-third of their after-tax income on food. Using food budgets for different kinds of families that reflected an adequate diet for “temporary or emergency use when funds are low,” she created poverty thresholds by multiplying the budgets by three. (This is something of an oversimplification, but you can read Fisher if you want to go deep.)

A number of other contemporary studies, including by staff of the Council of Economic Advisers, which produced the Economic Report of the President, had also gotten to around $3,000 using different approaches. In no small measure, the choice of $3,000 and $1,500 thresholds seems to have reflected a desire on the part of President Lyndon Johnson to deem one-fifth of Americans poor when he declared war on poverty.

Over the next five years, Orshansky’s approach, with its numerous poverty lines, was informally adopted until an interagency committee in 1969 agreed on an official measure. That measure took Orshansky’s estimates for 1963 and extended them backward and forward based on the change in the cost of living. That has been the basis for updating the poverty thresholds annually ever since.

At no point have the poverty thresholds ever been adjusted by revisiting either the cost of a minimally adequate diet or the food share of spending (the inverse of which is the “multiplier” that turns the cost of the diet into a poverty line). The thresholds are best conceived in the way that Orshansky herself described them: as “arbitrary, but not unreasonable.” One can set poverty thresholds higher or lower, but whatever the levels, they should be updated annually to account for the change in the cost of living. (Don’t talk to me about relative poverty measures.) Then one can see whether we have more or less poverty over time, whether it’s higher in some places than others, or whether there’s more hardship among some groups than others. And if you measure poverty this way, with an “absolute” measure, it doesn’t really matter what levels you pick—far fewer Americans are poor today than in the past. Let’s come back to that.

Green claims that Orshansky’s “principle” is “that poverty could be defined by the inverse of food’s budget share.” However, she never would have agreed with this as a principle for poverty measurement. We know because Orshansky was well aware—as Green apparently is not—of something called “Engel’s Law.” As she articulated (in two different 1965 publications), “the proportion of income allocated to the ‘necessaries,’ and in particular to food, is an indicator of economic well-being” and “a low percentage of income going for food can be equated with prosperity and a high percentage with privation.”

Engel’s Law says that as societies grow richer, they spend a smaller share of their income on food. For instance, according to one study, in the US the share fell from about 17 percent in 1960 to under 10 percent by 2019.

Engel’s Law renders the way Green is proposing to revise the poverty line absurd. He’s saying that rather than multiplying a minimally adequate food budget by three, as Orshansky did, we should multiply it by a larger number because we spend a smaller share of what we make on food today. He seems to think, wrongly, that that smaller share of spending on food is because everything else has gotten expensive without our being able to buy more and better non-food goods and services. In that case, the more expensive non-food purchases would crowd out spending on food and lower its share of spending. Multiplying a food budget by three would produce a low poverty threshold relative to the increased cost of living. It would fail to capture the increased hardship from simply paying more for the same non-food goods and services while being unable to afford to buy the same food. Updating the poverty measure by using a new multiplier would then make sense.

But in reality, we devote a smaller share of our spending to food because we can afford what we could in the past plus a lot more non-food spending. In 2023 dollars, the average household spent about $10,000 on food in 1960 and about $10,000 on food in 2023. (See here for 2023 and here for 1960, then adjust the latter for inflation using this index.) But average family income rose by over $85,000 in inflation-adjusted dollars and median family income by nearly $55,000.

The multiplier Green advocates is larger than three because we spend a larger share of income on things that aren’t necessities. For example, one (dated) federal study found that American spending on non-necessities was just over 20 percent of all expenditures in 1901, about 36 percent as of 1960, and 50 percent by 2002. It is undoubtedly larger today. The multiplier Green wants to use is so large because we’re much better off than in 1963.

Let’s look at Green’s calculation. He says, again, without citation, that “most families” today spend “5 to 7 percent” of total expenditures on food prepared at home. Taking the inverse of 5 and 7 percent, he says that today we have to multiply the minimum adequate food budget not by the three that Orshansky used, but by 16. (The inverse of 7 percent is 14.3, while the inverse of 5 percent is 20.0.) Green indicates that the true poverty line is between $130,000 and $150,000, settling for $140,000.

I was able to replicate his $130,000-$150,000 estimates, but not using a multiplier of 16. Green takes the federal poverty guideline for a family of four, $31,200 in 2024. He divides it by three to get back to Orshansky’s minimum adequate food budget. Then he has to multiply it by the inverse of the food share of the budget to get the new poverty line. In 2023, food at home accounted for 7.8 percent of spending. If I assume the food share of the budget is 8 or 7 percent, then the resulting poverty thresholds are $130,000 and $148,570. (Those are multipliers of 12.5 and 14.3, not 16. Using a multiplier of 16 produces a poverty threshold of $171,360. Using 20 yields $214,200!)

Note that Orshansky was using not just food at home as the basis for her multiplier, but all food. The total food share of spending in 2023 was 12.9 percent. Plug that into Green’s equation and the poverty line is “just” $80,620. But this is only somewhat crazier than saying the poverty line is $140,000.

The whole exercise is crazy. In 1901, the food share of the budget was 42.5 percent. Plug that into Green’s equation and the poverty line is $24,470 in 2024 dollars. That would be much easier to surpass than $80,620.

But does anyone think that the reason we spend such a smaller percentage of our income on food today than in 1901 is because everything else has become more expensive without our lives improving? In that case, we’d be better off if non-food prices declined and we could afford to get our food share of spending back up to 42.5 percent. But if the food share has gone down because we can afford more and better things than the days when fewer than one in ten homes had electricity, then why on earth would we better off spending less on non-food goods and services in order to spend more on food?

Before moving on, I also want to direct your attention to the fact that while the share of expenditures going toward food prepared at home fell from 14 percent to 5 percent between 1960 and 2019, the share spent on food enjoyed outside the home rose. Increased visits to restaurants is…not what we’d expect to see if hardship were rising.

Orshansky’s poverty lines are arbitrary but reasonable, but Green’s are arbitrary and unreasonable. No one in their right mind should think that a meaningful poverty line can be set at $140,000. But say you’re not in your right mind. We could still set the poverty line at that level in 2025. But then to figure out how bad we’re doing at alleviating hardship, we would want to adjust it for inflation, create the equivalent poverty lines going back to, say, 1963, and determine how much “poverty” we have today compared with the past. (The answer is we’d have less.)

As it happens, returning to statistics for people in their right mind, though the logic of the OPM’s thresholds in 1963 was reasonable, in practice the measure severely understates the extent to which poverty has declined. This is so for several reasons.

First, the price index it uses to adjust the thresholds each year overstates the rise in the cost of living. Green cites (without any real numbers) increases in the cost of housing, health care, child care, and higher education to suggest that the OPM has become obsolete. But guess what? While some costs have gone up, others have gone down (or risen more slowly than incomes have). Price indexes account for all of these products rather than cherry-picking the specific items that Green mentions. Price indexes also distinguish between people spending more for better stuff and people paying more for the same stuff. It’s the latter that represents an increase in the cost of living.

Adjusting the OPM for inflation is exactly what one should do to account for any increase in the cost of living. If costs go up, the poverty thresholds will go up, putting more people into poverty. But it’s important that the price index accurately reflects the true change in the cost of living. Most experts agree that our most-used indexes actually overstate inflation.

Green uses tired and facile arguments to cast doubt on the relevance of inflation adjustment for the questions he’s posing. But his “price of participation” has all the fatal flaws of Oren Cass’s “cost of thriving,” and I won’t spend thousands more words shooting down his claims—go here for that. Nor will I take up more of your time by batting down the idea that the economy used to support a single-earner model but no longer does, another argument Green trots out without evidence to say the price of participation has risen. Here you go if you’re interested.

A second reason why the OPM understates the decline in poverty is that many resources are missing from the “money income” measure that it compares against the poverty thresholds. The missing resources include employer-provided health insurance, noncash government transfers (such as food stamps, health coverage, and housing subsidies), and refundable tax credits.

A third factor that causes the OPM to understate poverty declines is that money income is a pre-tax measure, but spending depends on after-tax income. The latter has increased more than pre-tax income.

The most comprehensive effort to address these issues is a recent paper by Richard Burkhauser, Kevin Corinth, James Elwell, and Jeff Larrimore (ungated version here). These authors use a complete after-tax income measure, set the poverty line so that 19.5 percent of the population is poor (as the Orshansky-based OPM does), and update that threshold for inflation with a better price index than the OPM uses. They find that poverty fell to 1.6 percent by 2019—much lower than the decline to 10.5 percent found by the OPM.

At this point, there’s a good chance you’re saying, “Who would take seriously any poverty measure that says only 1.6 percent of Americans are poor?” But remember: poverty lines are arbitrary (but should be reasonable). The authors of the paper also show results setting the 2019 poverty rate to 10.5 percent and then looking back at what the 1963 rate was using this more generous threshold. The answer is that the poverty rate was 70 percent. Either way, poverty has fallen dramatically.

The idea that populist rage is rampant because of widespread economic dissatisfaction is almost always an assumption not backed up by evidence. But evidence does exist in this case too, and it tells us that people are not actually any more dissatisfied with their own economic situation than in the past. Rather, they think—contrary to what people say about themselves—that everyone else is doing poorly. That impression doesn’t develop in a vacuum. It’s fed by inaccurate—sometimes wildly inaccurate—claims that are taken up by the very online and spread like wildfire among the disaffected on social media. I won’t follow Green in accusations of bad faith (“don’t let them gaslight you”). But his argument about poverty is simply bad.

What should we make of earnings trends in the United States? People can differ in their opinions about whether some reported trend is impressive or cause for concern, of course. However, you might think that it’s at least straightforward to determine what the relevant trend is and how to measure it. But that turns out not to be the case, and claims about what has happened to earnings range widely from significant declines to sizable gains.1

It’s tempting to either say that these disparate results all have merit or that they render it impossible to say anything about long-term earnings trends. However, some estimates are decidedly less relevant and more misleading than others. We need to be able to characterize the status of workers’ economic well-being in meaningful ways. Fortunately, the measurement and data issues are clear enough that careful consideration leads to fairly unambiguous conclusions about what has happened (if less clarity about why).

Consider the change in median earnings over the 50-year period from 1973 to 2023. The median refers to the earnings of the “typical” person—the one with earnings higher than half the population and lower than half the population. In a yearly supplement to the federal Current Population Survey, which I’ll analyze here, everyone above the age of 14 is asked about their earnings.2 Earnings include the wage and salary income of employees as well as self-employment income (after business expenses).

The most naïve approach to the question is to simply compare the earnings of all people in 1973 and 2023. It turns out that the median person over the age of 14 in 2023 had earnings 11 times higher than the 1973 median. To be clear, we are not anything like 11 times as rich as we were fifty years ago. But how much better off are we? We’ll need a series of adjustments to arrive at the most meaningful answer, but feel free to skip to the conclusion if you’re not interested in the journey. The gist is that the median man is 40 to 50 percent better off than he was in 1973, and the median woman is doing more than twice as well as she did fifty years ago.

Step One: Account for Inflation (First Pass)

So, how do we get a more reasonable answer to the question of how much median earnings have risen? To start with, the cost of living has, of course, risen over the past fifty years; a dollar today buys less than it did in 1973. When examining earnings trends, it is essential to adjust “nominal” values for inflation to account for the declining purchasing power of the dollar. I’ll start off by using the “personal consumption expenditures price index,” (PCEPI), though I’ll revise this choice below.3 According to the PCEPI, the cost of living rose by more than a factor of five. Still, if we look at median inflation-adjusted earnings for all Americans age 15 and older, it more than doubled between 1973 and 2023, rising 113 percent over 50 years.

Step Two: Separate Women and Men

That figure combines men and women, and their earnings trajectories look quite different. Median earnings among women rose 44-fold, while the figure for men fell by 18 percent. To be clear, neither of these estimates reasonably depicts well how women or men have fared, but trends between the two groups will continue to differ as we get closer to the best depiction. (Use Tables 1 and 2 to follow along as we go. Figures 1 and 2 show the trends for the inflation-adjusted pre-tax estimates.)

Step Three: Address Non-Workers, Part-Year Workers, and Part-Time Workers

The figures so far have included everyone at least 15 years old, whether they have earnings or not. That includes retirees, as well as teenagers and young adults in school. It includes the disabled and sick. It includes homemakers. Moreover, it includes part-year and part-time workers. If the relative sizes of these groups change over time, that can affect earnings trends.

For instance, a basic reason why women older than 14 saw a 4,316 percent increase in median earnings is that nearly half of them had no earnings at all in 1973. Among those who were employed, many worked only part-time, part of the year, or both. Today, there are many fewer homemakers than fifty years ago—a change that reflects increased economic opportunities for women and cultural shifts in gender roles.

More generally, if we are interested in how the economy is doing for workers, it may make sense to exclude from the analysis people who are too sick or disabled to work or people who have chosen to be in school, retirement, or to handle domestic responsibilities. At the same time, some people work only part-time, part of the year, or not at all due to labor market conditions, the availability of safety net benefits, or the interaction between the two. If labor supply isn’t entirely voluntary or is affected by policy incentives, that complicates the question of how to convey what has happened to earnings.

As a first attempt to address this issue of changing labor supply, we can confine the analyses to Americans of prime working age. By excluding the youngest and oldest age groups, focusing on people between the ages of 25 and 54 partly addresses several labor supply complications. These issues include rising school enrollment, changing norms related to teen employment, declining fertility and delayed childbearing, and earlier retirement.

After confining to this narrower group, rather than women’s median earnings rising by a factor of 44, it increases by a factor of eight to nine— from $4,400 in 1973 (expressed in terms of 2023 purchasing power) to $38,000 fifty years later. Among prime-age men, median earnings fell not by 18 percent, but 3 percent (from $55,300 to $53,600).

Table 1. Median Earnings and Compensation Trends, Women, 1973-2023

Source: Author’s analyses of the Annual Social and Economic Supplement to the Current Population Survey. The 1989-2023 trend is adjusted for an administrative break in the CPS between 2013 and 2015, as described in end note 2.

Table 2. Median Earnings and Compensation Trends, Men, 1973-2023

Source: Author’s analyses of the Annual Social and Economic Supplement to the Current Population Survey. The 1989-2023 trend is adjusted for an administrative break in the CPS between 2013 and 2015, as described in end note 2.

Figure 1. Median Inflation-Adjusted Pre-Tax Earnings and Compensation, Women, 1973-2023

Source: Author’s analyses of the Annual Social and Economic Supplement to the Current Population Survey.

Figure 2. Median Inflation-Adjusted Pre-Tax Earnings and Compensation, Men, 1973-2023

Source: Author’s analyses of the Annual Social and Economic Supplement to the Current Population Survey.

These estimates still include nonworkers, as well as part-year and part-time workers. All we have done is exclude age groups within which such adults are overrepresented. Rather than restricting to prime-age workers, we can instead go back to looking at all workers older than 14, except this time leave out people who don’t have positive earnings.4 Among women, the increase in the median is smaller than before, but the rise from $16,700 to $43,500 was still 160 percent, while men’s median earnings rose 20 percent (from $45,900 to $55,000). Restricting to prime-age adults while dropping nonworkers produces yet another set of estimates: a 126 percent rise among women, and a 5 percent increase among men.

In the appendix, I describe experiments I conducted to determine the most effective way to address the issues created by the changing number and nature of nonworkers, part-year workers, and part-time workers. The best compromise is to examine trends for year-round workers (including part-time workers) aged at least 14 years old.5 Doing so, the increase in median earnings for women was 75 percent from 1973 to 2023 ($22,000), and it was 13 percent for men ($7,300). These are not yet the best estimates, but we are getting closer.

Interlude: Understand the Trajectories of Earnings

Before we turn to the next set of adjustments, it’s worth making a point about the basic shapes of these 50-year trends. Imagine that median earnings were the same in 2023 as in 1973 among some group, so that the 50-year increase was 0 percent. The implications of this long-term trend would be significantly different if median earnings remained constant every year for 50 years than if the median rose substantially for 25 years but then fell by an equivalent amount over the most recent 25 years. In the latter case, we might reasonably draw much more negative conclusions about the state of the economy. We could, at the very least, say that things are getting worse, and we would want to consider whether various economic and public policy changes are factors.

However, a third possibility is that the median might have fallen for 25 years but then experienced an equally large recovery more recently. In that case, we might conclude that while the economy was underperforming for a stretch, that period has receded into the distant past, and things are on the upswing. We would not want to blame economic or policy factors from the past 25 years for “stagnation.”

These distinctions are particularly relevant to trends in male earnings. If we compare business cycle peaks for the median earnings of men working year-round, we find that the median rose 17 percent from 1973 to 2021.6 But if we look at changes between each business cycle peak, median earnings were flat from 1973 to 1979 and flat between 1979 and 1988. They rose by 0.3 percent per year from 1988 to 2000, then by 0.3 percent per year from 2000 to 2007, and subsequently by 0.7 percent annually from 2007 to 2021.

All told, as shown in Table 2, during the 16 years from 1973 to 1989, men’s earnings were flat, but during the 34 years from 1989 to 2023, they rose by 0.4 percent per year. The biggest problem for men over the past 50 years was a period of stagnation or decline that ended 40 years ago, with 1983 marking its low point. Millennial and Gen Z social-media-fueled complaints about the economy aside, it was baby boomer men who experienced historically low earnings growth.

Step Four: Account for Inflation (More Accurately)

The picture is better than these figures would suggest, because the PCEPI overstates the increase in the cost of living. In a recent paper, I summarized the evidence supporting this conclusion, reviewed studies on the magnitude of bias in popular price indexes, and documented the statements of federal statistical agencies that prefer certain measurement choices over others in assessing inflation.7 I concluded that the PCEPI overstates the rise in the cost of living by about 0.4 percentage points every year. Over extended periods of time, this has a sizable impact on measured inflation, and it makes earnings and income trends look worse than they truly have been. The paper developed a “more accurate consumer price index” (MACPI) based on evidence of bias in conventional measures.

I subsequently validated the MACPI against long-term trends in the subjective belief that one’s family is better off financially than their parents were at the same age.8 It is possible to measure the share of adults who objectively have higher inflation-adjusted income than their parents did. However, the level and trend differ depending on the inflation measure used. I showed that using the MACPI to adjust for inflation produced trends and levels of “absolute mobility” that more closely matched trends and levels of subjective feelings of having experienced absolute mobility.

When we use the MACPI to adjust for the rise in the cost of living, median earnings among women who worked year-round rose between 1973 and 2023 not by 75 percent, but by 116 percent. Instead of rising by 13 percent, male year-round workers’ earnings rose by 40 percent. These are increases of $27,700 for women and $17,900 for men. Men’s earnings rose 0.7 percent per year from 1973 to 1989 and 0.7 percent per year from 1989 to 2023.

Step Five: Account for Non-Wage Compensation

Earnings measures don’t include the full compensation that workers receive from their employers. Those employers are largely indifferent between paying employees $1,000 in wage and salary income and paying them $1,000 in health care or retirement benefits. But only the wage and salary income is included in earnings in the CPS and most other surveys.

We can adjust median earnings estimates to account for non-wage compensation by multiplying all employees’ wage and salary income by the ratio of aggregate compensation to aggregate employee earnings, as reported in the national accounts data.9 (The earnings of the self-employed are unadjusted.) This approach assumes that the aggregate ratio is appropriate for adjusting all workers’ incomes, and that it is the same for less-paid workers as for higher-paid workers. In previous work, I tested this assumption and found considerable evidence to support it.10

The median compensation of year-round female workers rose by 122 percent from 1973 to 2023—a bit more than the 116 percent increase in earnings. Among men, the increase bumped up from 40 percent for earnings to 47 percent for compensation. In dollar terms, these were increases of $31,500 for women and $22,300 for men over fifty years.

Step Six: Account for Earnings Taxed Away by Corporate and Payroll Taxes

Conventional measures such as those presented so far are “pre-tax” in the sense that they reflect earnings before deducting the income and payroll taxes that are statutorily imposed on employees and the self-employed. Looking at pre-tax earnings is entirely appropriate if we are mainly concerned about how labor markets and employers deliver to workers, apart from the question of how tax policy then affects what they have left to spend or save.

However, at least some fraction of the corporate income and payroll taxes assessed on employers comes at the expense of worker pay. If those taxes were to disappear tomorrow, some of the savings to firms would be allocated to profits, but much of it would be passed on to workers. In this sense, conventional earnings measures are not entirely “pre-tax”, because we see the earnings paid to workers only after corporate income and payroll taxes have been deducted. These measures are effectively “pre-individual-tax, post-employer-tax” estimates.

The next set of adjustments, therefore, adds back not only nonwage benefits to wage and salary income of employees, but employer-paid payroll taxes and 25 percent of corporate income taxes.11 These adjustments are again based on aggregate data and assume that the ratio adjustment is appropriate for all workers.12 Among women year-round workers, this measure of full compensation once again shows an increase of 121 percent from 1973 to 2023 ($33,600), while the increase among men is 46 percent ($23,200).

Optional: Deduct Taxes

While we may be interested in the question of how markets are serving workers, we might also want to know how much workers have to spend and save after the state claims some of what they earn. Answering this question requires examining post-tax compensation, which is possible using the CPS estimates from 1979 onward. I estimate the federal and state income taxes applied to earnings, as well as the employee’s share of payroll taxes (and self-employment taxes).13 I subtract these from compensation, along with employer-paid payroll taxes and 25 percent of corporate income taxes. Before taxes are deducted, median compensation rose 96 percent for year-round female workers between 1979 and 2023, and that was unchanged after taking taxes into account. Among men, the median rose 33 percent before taxes are deducted and 38 percent after.

Step Seven: Separate by Age Group

We saw earlier that restricting the age group to prime-working-age adults and excluding non-workers both have independent effects on earnings trends. Among men, for instance, excluding non-workers made the earnings trend stronger. Restricting to prime-age men also strengthened the trend, except that when we initially excluded non-workers, restricting to prime-age men then weakened the earnings trend.

One way the choice of age range can affect trends in the median is that when birth cohorts differ in size, a larger or smaller one can influence trends among the entire group as it ages, since earnings tend to vary predictably with age. To account for this issue, we can look at trends separately for narrower age groups. Since it’s not clear that the adjustments for non-wage compensation and earnings taxed away by employer taxes should be the same for each age group, and since these adjustments did not have a significant impact on the trend estimates, I concentrate on MACPI-adjusted earnings without these additional adjustments.

As a reminder, among women who worked year-round, median earnings rose 116 percent from 1973 to 2023, according to this measure. But as shown in Table 3, the increase was smaller for women under age 35 (ranging from 46 percent to 89 percent) than for older women, among whom the increase ranged from 111 percent to 133 percent). For men, the overall increase was 40 percent, but earnings growth generally improved steadily with age. For men 20 to 24 years old, the increase was just 17 percent, while it was 62 percent for men 60 to 64 years old.

Table 3. Median Inflation-Adjusted Pre-Tax Earnings Trends by Age Group, 1973-2023

Source: Author’s analyses of the Annual Social and Economic Supplement to the Current Population Survey.

Interesting patterns emerge when we examine the 1973-1989 and 1989-2023 periods separately. Since the two periods are of different lengths, it’s more informative to look at the annual percentage changes, which are also shown in Table 3. Among women, earnings rose faster after 1989 than before for the youngest and oldest workers. Women between the ages of 30 and 49 saw larger gains from 1973 to 1989, probably reflecting an increase in work and career opportunities for mothers with older children.

The youngest men did worse than other groups from 1973 to 1989 but better than older groups from 1989 to 2023. Among men under 35, earnings fell during the earlier period but increased by about one-third from 1989 to 2023. Grouping all under-40 year-round male workers, median earnings rose by 28 to 29 percent from 1989 to 2023, or by $12,200 (not shown in Table 3). In contrast, the annual rates of change in Table 3 indicate that earnings growth for men aged 40 and above was significantly stronger before 1990 than after.

Aside: Other Adjustments are Possible

Of course, there are other adjustments one could make, although some are fraught with the potential to make things worse. Some analysts show median earnings trends by educational attainment. However, that tends to produce results that are too gloomy because Americans’ educational attainment has been rising. Comparing adults with less than a high school education in 1973 to the same group in 2023 captures a relatively large group in the initial year but a significantly smaller group today. Among adults ages 25 to 29, for instance, it’s essentially comparing the least-skilled 5 percent of today’s young adults to the least-skilled 25 percent in 1970.14 A better approach is to look at trends for specific percentiles of the earnings distribution—not just the median (or 50^th percentile), but the 20^th and 80^th percentiles, for instance. The 20^th percentile of earnings refers to the earnings of the person who earns more than 20 percent of workers but less than 80 percent of workers.

Among women who worked year-round, the 20^th percentile of earnings rose 105 percent from 1973 to 2023, compared with 147 percent for the 80^th percentile (and 115 percent for the median). For men, the figures were 35 percent and 70 percent (and 40 percent for the median). These amounts are MACPI-adjusted and do not add nonwage compensation or employer-paid taxes to earnings.

One might argue that these long-term trends should be adjusted to reflect changing immigration patterns, too—rising immigration over the period added a steady supply of low-wage workers to the labor force. From 1993 to 2023, the data allow us to restrict the analyses to native-born Americans. Surprisingly, immigration appears to have little impact on the median trend.

Among women working year-round, median earnings increased by 59 percent overall, 59 percent among native-born women, and 60 percent among foreign-born women. The situation is somewhat more complicated for men, but still suggests that adjusting for rising immigration is unnecessary. Median earnings rose 29 percent among male year-round workers generally, 27 percent among native-born men, and 65 percent among foreign-born men. Within that native-born group, earnings rose 31 percent for non-Hispanic native-born men and 53 percent among Hispanic native-born men.

Note that these results do not answer the question of how immigration affects the wages of native-born Americans—if it lowers those wages, then in the absence of rising immigration, the median increases among the native-born would have been even larger than these figures suggest.

Finale: Do It Yourself!

I have argued here that the trend in median pay for year-round workers is a superior way to understand changes in the economic fortunes of both men and women. Tabulations of this measure are not readily available from public sources, and most people lack the background required to produce their own estimates from the publicly available data. Fortunately, the Census Bureau annually publishes a closely related measure that appears to track my preferred measure reasonably well. Specifically, “Historical Income Table P-38” presents the median pre-tax earnings of year-round and fulltime workers at least 15 years old.15 The main difference from my preferred series is that mine includes part-time workers who worked year-round. An added bonus of the Census Bureau series is that it goes back to 1960.

Unfortunately, the published series employs an inflation adjustment that overstates the long-term rise in the cost of living.16 However, the Census Bureau table includes nominal amounts too (unadjusted for inflation). That makes it (relatively) easy for researchers to adjust the amounts using the PCEPI or MACPI.17 You can use an artificial intelligence, such as ChatGPT, to do it.

When I adjusted the Census Bureau amounts using the MACPI, I found median earnings rose 38 percent for men from 1973 to 2023, increasing 7 percent from 1973 to 1989 and rising 29 percent from 1989 to 2023. My own estimates for year-round workers (without adding nonwage compensation or employer-paid taxes) are 40 percent, 11 percent, and 26 percent. For women, the Census Bureau estimates rose 100 percent from 1973 to 2023, 29 percent from 1973 to 1989, and 55 percent from 1989 to 2023. My estimates rose 115 percent, 27 percent, and 70 percent from 1989 to 2023.

The published Census Bureau series also includes just-released 2024 estimates, which show the median earnings of full-time, year-round men at the same level as 2023 and the earnings of women up two percent.

Conclusion

Without the context behind a trend, it’s easy to draw the wrong conclusion about whether the labor market is improving or deteriorating in its ability to provide well-paying jobs and whether workers are better or worse off than in the past. While we have seen that it requires a fair amount of space to convey a meaningful picture of trends, the results to which we’ve arrived are straightforward to describe.

After considering complications such as inflation, non-work, part-year and part-time work, disability, nonwage compensation, and employer-paid taxes, the typical female worker earns more than twice what her counterpart did fifty years ago. The typical male worker is better off by around 45 percent compared to the 1973 standard. In terms of today’s dollars, these gains translate to about $33,600 for women and $23,200 for men who work year-round. Most of this increase came after 1979, totaling almost $30,000 for women and over $18,000 for men since that year. After accounting for taxes, the median rose $20,200 for women between 1979 and 2023 and $14,200 for men.

These are substantial improvements. Moreover, the findings are reinforced by trends in hourly wages.18

Critics who wish to ignore the arguments in favor of using the MACPI to adjust for the rise in the cost of living might prefer the earnings trends presented earlier using the PCEPI, which showed men’s earnings rising by just 13 percent. Even in that case, though, it’s not that men’s pay has been flat for 50 years—it was flat and then rose. Had men’s earnings grown at the same annual rate from 1973 to 1989 that they did from 1989 to 2023, the median would have risen 20 percent over the whole 50 years rather than 13 percent—52 percent more.

In other words, MACPI skeptics who wish men’s earnings growth over the past half century had been faster have a quarrel with the 1970s and 1980s. That means that explanations for why growth wasn’t stronger have less to do with the China Shock, immigration, the global financial crisis of the aughts, the Great Recession, or rising income concentration than they do with the stagflation of the 1970s and the deep recessions of the early 1980s.

While women’s earnings growth was unambiguously strong across age groups, growth varied for men of different ages. Among men under age 30, the increase in MACPI-adjusted median earnings from 1973 to 2023 was only 24 percent rather than the 40 percent for all men using the same measure of earnings. But again, the problem here dates to the 1970s and 1980s. Young men’s earnings fell 2 percent from 1973 to 1989 but rose 26 percent from 1989 to 2023. Since 1989, men’s earnings growth has been stronger for younger men than for older men.

A plausible and straightforward explanation for the male trends emerges from the interaction of trends in men’s educational attainment and women’s labor force participation. If we look at older men, between 1973 and 1989 the share who graduated from college rose relatively rapidly. For instance, 21 percent of 45- to 49-year-olds who worked year-round in 1973 had a bachelor’s degree, while that was true of 32 percent of such men in 1989. That was an 11-point increase over 16 years. This increase in human capital was layered onto the relatively high levels of experience possessed by older workers, and because of historical patterns of female labor force participation, these men faced relatively low levels of competition for the best jobs. The result was solid earnings growth.

In contrast, from 1989 to 2023, educational attainment rose more modestly for older male workers. By 2023, the college graduation rate had risen by another 10 points to 42 percent, but spread over 34 years rather than the 16 years from 1973 to 1989. Meanwhile, there were many more experienced and well-educated older women in the workforce, competing for the best jobs. Earnings growth for men slowed.

Switching to younger men, the story looks quite different. The college graduation rate for men 25 to 29 years old who worked year-round was 24 percent in 1973, but just slightly higher at 26 percent in 1989. Stagnant educational attainment, low levels of work experience, and rising competition from young women with greater professional opportunities were a recipe for male earnings declines.

However, from 1989 to 2023, the college graduation rate among men aged 25 to 29 increased by 11 percentage points. Combined with stabilized professional competition from young women (and the temporary 1990s productivity boom), that was enough to produce strong earnings growth.

Otherwise, two other factors seem likely to be important in explaining the pre-1990 decline in the earnings of young men.19 First, young boomer men during this period of societal transition may have been the first generation not to receive what I have called “breadwinner rents”—pay over and above their value to an employer, a bonus rooted in the paternalistic norm that men should ideally be sole breadwinners of their families. Second, the decline in marriage and fertility left an increasing share of young men without family responsibilities, which could have reduced their incentives to maximize earnings and invest in human capital.

Solutions to the economic problems of contemporary men that focus on international trade, lack of union jobs, immigration, opioid addiction, safety net policies, the absence of a national industrial policy, “financialization” of the economy, rising income concentration, explosive housing costs, higher education debt, or other changes over the past 35 years may have their merits. However, these issues are irrelevant as explanations for stagnant or declining male earnings, for the simple reason that no such trend has existed for over 35 years, rendering them unnecessary.

One way to address the question of how to deal with nonworkers, part-year workers, and part-time workers in estimating meaningful median earnings trends is to exploit a little-used feature of the CPS—the fact that it collects income information two different times for most households. The way the CPS works is that everyone is surveyed in four consecutive months, before taking a year off and then being interviewed again in the same four months the following year. Therefore, if someone is observed as a nonworker, part-year worker, or part-time worker in one year, we may be able to leverage information from their earnings one year earlier or one year later.

For example, in my sample of people at least 15 years old, I am able to match 71 percent of them to another earnings record in 1973 and 64 percent in 2022.1 (I stop at 2022 instead of 2023 because 2024 data was unavailable at the time of writing, so I could only check if men without earnings in 2023 had them in 2022, not 2024.) Among people who have no earnings, I can match 75 percent of them to another earnings record in 1973 and 67 percent in 2022. Most of those with no earnings who can be matched also have no earnings in the other year they are observed; just 12 percent in 1973 and 9 percent in 2022 can be matched to a positive earnings amount.2 (Note that many of these people with no earnings in consecutive years are retired.)

Let’s focus, first, on men, since their earnings trends look much worse than those of women and assessing their trends is more sensitive to measurement issues. For those who we can match across surveys, we can use their linked earnings information to get a better sense of what earnings trends would look like if we could observe everyone’s earnings. If someone does not have earnings in one year but does in an adjacent year, we can swap in the nonzero amount.

Of course, some men have no earnings in a given year because of subpar economic conditions. It’s not informative to ask, “What if the economy were great?” to the extent that most nonworking men only wish they had a job. But we know from the CPS why someone went an entire year without working, why they only worked part of the year instead of the whole year, and why they worked part-time instead of full-time. Since we can identify men who say the economy was to blame for their working less than year-round or less than full-time, we can keep them in the data as below-median earners and let the strength of the economy affect the results. That’s what I do in these experiments; values are replaced with earnings from a linked survey only if someone had a low initial value for some reason other than economic problems.3

Importantly, the sample that can be linked has similar earnings growth to the full sample of men if all non-positive earnings are dropped. Among men with positive earnings, the median increased 48 percent from 1973 to 2022 for the full sample and 47 percent for the linked sample.4 (Note that the analyses in this appendix use the MACPI, described in the paper, to adjust for inflation. Using the PCEPI instead would affect the magnitudes but not the relative comparisons of different measures.)

However, when including men with no earnings or negative earnings, growth is worse in the linked sample than the full sample. Among all men, the median fell by 1 percent in the full sample but by 15 percent in the linked sample.5

So, staying with the linked sample, which is closer to The Truth: a 47 percent increase, or a 15 percent decline? For starters, if we replace earnings of $0 with positive values from the year before or after—but only for men who were nonworkers for reasons other than difficulty finding a job—median earnings falls by 1 percent rather than 15 percent.6 That suggests that if we could link more nonworkers to past or future earnings, the earnings trend would improve substantially relative to the trend when all nonworkers are included.

Some men who don’t work clearly could but are doing other things instead. If we drop the remaining men with non-positive earnings (except for those who had no earnings due to economic conditions), the median increases 40 percent from 1973 to 2022.7 That’s a bit less than the 47 percent we got before we added people whose earnings we could pull from a second interview (or even the 44 percent we’d have obtained from that sample if we’d left in those with no earnings because of economic conditions, not shown). That suggests that the rising share of men with no earnings causes men’s trends to be too strong relative to what it would be if everyone’s earnings could be observed in a given year.

Addressing people with negative or no earnings isn’t the only complication to address. Some people work only part of the year for voluntary reasons having nothing to do with the state of the economy. For instance, graduating seniors may only work for half the calendar year. Parents may take time off after the birth of a child. Some part-year workers are older and retired midway through the year. These part-year jobless spells are often just continuations of the full-year jobless spell that we are trying to address.

Replacing a value of $0 with earnings from the year before or after someone is observed may only replace it with a value that itself still reflects (voluntary) time out of the labor force, understating what the person would make as a full-year worker. If we only replace values of $0 using full-year earnings from the year before or after, dropping all other non-positive earnings (unless due to economic problems), the increase in median earnings bumps up to 45 percent from the 40 percent in the last step.8

Further, we can also check what happens when we replace earnings of men working only part of the year with their observed full-year earnings from the year before or after (where possible). If we make that replacement too—again, only for men whose part-year employment had nothing to do with economic conditions—on top of replacing non-positive earnings with positive earnings from the year before or after, the increase in median earnings is 43 percent.9

If we also drop the remaining workers who worked only part of the year for reasons other than economic conditions (those who weren’t full-year workers the other year they were observed), median earnings rise by 42 percent.10 Note that this estimate still includes non-working men and part-year men if economic conditions were the reason they didn’t work or only worked part of the year.11

These estimates have dropped all non-workers and part-year earners for whom we can’t pull full-year earnings from an adjacent year, unless their status is due to economic conditions. That seems reasonable in the case of men who consistently aren’t working because they are in school, taking care of home or family, or retired. It also seems reasonable for unambiguously disabled adults.

However, over the long run, while disability rates of men have been fairly constant, the likelihood of working conditional on reporting a disability has fallen.12 At the same time, federal disability benefit rolls have increased substantially.13 Disability benefit receipt increases predictably during economic downturns, suggesting that some men experiencing economic problems respond by turning to disability benefits. If the increased rate of disability benefit receipt were suppressing a bigger rise in the number of men saying they don’t work because of the economy, then dropping disabled non-workers would overstate the gains that able-bodied men have made over time.

One reasonable way of addressing this issue is to add back to the analysis “excess” non-workers and all part-year workers who cite illness or disability as the reason for not working year-round. I determine the share of adults in 1973, by age group, who were nonworking and disabled, then I add back in the excess share of non-working disabled adults (the number above the 1973 share) in every year.14 Then I add back all part-year disabled workers. (I ignore any changes in part-time work, since that information is unavailable for 1973.) This final sample is 27 percent smaller in 1973 and 35 percent smaller in 2022 than the full sample of men ages 15 and older and linkable to a second earnings record. It is 9 percent smaller than the sample of such men with positive earnings in 1973 and 1 percent smaller in 2022.15

After making this final adjustment, median male earnings rises by 32 percent from 1973 to 2022—significantly lower than the 42 percent increase before adding back these sick and disabled men.16

To recap where we have arrived: some men either don’t work or work only part of the year for reasons having nothing to do with economic conditions. In the data, for a subset of men, we can observe earnings from one year earlier or one year later. Some of the men who didn’t work year-round in one year did so in an adjacent year, and some who worked part-year worked year-round in an adjacent year. We can (1) swap in their full-year earnings and (2) exclude the remaining non-workers and part-year earners whose status had nothing to do with economic conditions nor with their health, and (3) exclude a number of disabled non-workers who would have been non-workers even in 1973.

The resulting 32 percent increase in median male earnings is closer to the 47 percent rise in the median when we looked at all men with positive earnings than to the 15 percent fall when we looked at all men with or without earnings. If we just look at full-year workers (without swapping in any full-year values from adjacent years for non-workers and part-year workers), median earnings increase by 36 percent.17 That suggests that the simple trend for full-year earners is a solid basis for assessing male earnings trends, at least when comparing similar points in the business cycle. (There will be more part-year workers and non-workers when times are bad than when times are good, but the checks here suggest that taking that into account wouldn’t change the 50-year trend comparing two strong economic years.)

Recall that the median earnings trend for the CPS sample linkable across years was somewhat different than the trend for the full CPS sample. If we go back to the full CPS sample and look at median earnings among men who worked all year, the increase from 1973 to 2022 was 41 percent, and the increase from 1973 to 2023 was 40 percent, or $17,900. From 1973 to 1989, the median rose by 11 percent, or $5,000. It rose by 26 percent from 1989 to 2023, or $12,900.18 The 1973-to-2023 increase using the PCEPI instead of the MACPI was 13 percent.

When we conduct the same exercise for women, we get similar results. The median rises by 117 percent from 1973 to 2022 among full-year workers in the linked sample and by 124 percent after we make all the adjustments, including “excess” disabled adults.19 If anything, using the year-round-worker trend understates gains for women. In the full sample (linkable or not), median earnings among year-round workers increases 115 percent from 1973 to 2022 and 115 percent from 1973 to 2023.20 Using the PCEPI, the 1973-to-2023 increase is 75 percent.

Thus, the trend for year-round workers is an appropriate approximation of the increase in median earnings among both men and women using more careful adjustments to account for changes in non-work, part-year work, and part-time work.

By the end of the day, President Donald Trump had fired BLS Commissioner Erika McEntarfer. The White House quickly released a memo criticizing her “lengthy history of inaccuracies, incompetence.” It cited multiple instances of revisions to jobs estimates going back to April 2024 and referenced several breaches of data security and data-sharing protocol that occurred in 2024.

But any ambiguity about the specific sins that led to McEntarfer’s fall was cleared up later that afternoon, when Trump took to his preferred social media platform, Truth Social:

In my opinion, today’s Jobs Numbers were RIGGED in order to make the Republicans, and ME, look bad — Just like when they had three great days around the 2024 Presidential Election, and then, those numbers were “taken away” on November 15, 2024, right after the Election, when the Jobs Numbers were massively revised DOWNWARD, making a correction of over 818,000 Jobs — A TOTAL SCAM.

Let’s stipulate that the three operational breaches in question were serious matters. Moreover, the acting secretary of the Department of Labor appears not to have been sufficiently cooperative with House oversight of BLS, and it’s not clear that McEntarfer herself responded to a similar inquiry from Senators Bill Cassidy and Susan Collins.

Even so, the incidents involved mistakes and judgmental lapses on the part of lower-ranking staff rather than intentional wrong-doing by McEntarfer or other senior BLS officials. BLS acknowledged the breaches promptly and looped in the Department of Labor’s inspector general. The incidents in question took place between February and August of last year. The last one drew a Truth Social post, but Trump had not criticized BLS in the ensuing months, before or after becoming president.

In fact, if his social media posts are to be believed, Trump only became aware of McEntarfer after the latest jobs report:

I was just informed that our Country’s “Jobs Numbers” are being produced by a Biden Appointee, Dr. Erika McEntarfer, the Commissioner of Labor Statistics, who faked the Jobs Numbers before the Election to try and boost Kamala’s chances of Victory. This is the same Bureau of Labor Statistics that overstated the Jobs Growth in March 2024 by approximately 818,000 and, then again, right before the 2024 Presidential Election, in August and September, by 112,000. These were Records — No one can be that wrong? We need accurate Jobs Numbers. I have directed my Team to fire this Biden Political Appointee, IMMEDIATELY. She will be replaced with someone much more competent and qualified. Important numbers like this must be fair and accurate, they can’t be manipulated for political purposes. McEntarfer said there were only 73,000 Jobs added (a shock!) but, more importantly, that a major mistake was made by them, 258,000 Jobs downward, in the prior two months. Similar things happened in the first part of the year, always to the negative.

Tellingly, the President’s posts make no mention of the BLS operational missteps—only the revisions to the jobs estimates and the implied political bias behind them. However, far from McEntarfer being guilty of running a political shop, her firing was itself primarily political. The President’s accusations don’t stand up, and his actions should send a chill down the spine of Americans of all ideological stripes.

Understanding the BLS Monthly Employment Revisions

The monthly jobs estimates come from the Current Employment Statistics (CES) program—a massive survey of about 630,000 worksites each month. The first Friday of every month, estimates for the previous month are released on what is often colloquially called Jobs Day. These are hot off the presses—the survey obtains employment information about the pay period the previous month that includes the 12^th of the month, so the turnaround time is a tough challenge.

For example, on June 6, BLS reported the initial estimates for May. To do so, it looked at how much employment changed among the subset of establishments that beat the reporting deadline for the May numbers and that also reported April employment. It used this percentage increase to update the April estimate, and it made some model-driven assumptions about how many establishments were created in May and how many closed. The result was the “first preliminary estimate” of nonfarm employment for May.

Employers continue to report to BLS after this first estimate is released, so on Jobs Day the following month, BLS updates them as “second preliminary estimates.” It does the same one month later, releasing “final sample-based estimates.” I’ll just call these numbers the first, second, and third estimates.

The July 3 Jobs Day provided the second employment estimates for May and the first estimates for June. The data indicated that the US economy had added 144,000 nonfarm jobs in May and another 147,000 in June. That May figure was revised up from the 128,000 reported on Jobs Day in June. In the August release, however, the numbers were revised down to 19,000 jobs added in May and 14,000 in June.

These revisions are standard, routine updates. BLS has released these twice-updated estimates since the mid-twentieth century. They are not unusual “mistakes”; they follow initial estimates deemed “preliminary.” When the White House memo criticizing McEntarfer singled out her BLS for releasing initially optimistic estimates “only for those numbers to be quietly revised later,” it unjustifiably cast standard operating procedure in a sinister light.

An important issue for these monthly revisions is that small changes to the employment estimates can produce relatively large changes in the estimates of jobs added. Take the May-to-June jobs added estimate. In last week’s jobs report, the estimate for nonfarm employment in May was revised downward by one-tenth of one percent (actually 0.08 percent). The estimate for June was revised downward by two-tenths of one percent (0.16 percent). But the resulting increase in nonfarm jobs from May to June after these revisions was smaller by over 90 percent than had been reported the previous month.

If the June revision had come in lower by 0.10 percent instead of 0.16 percent, the jobs-added estimate would have shrunk not by 90 percent but by 24 percent. Rather than it being revised downward by 133,000 jobs, it would only have been revised downward by 35,000.

How Have the Monthly Employment Revisions Changed Over Time?

We can assess how the month-to-month revisions have affected the nonfarm employment estimates over time. In Figure 1, I’ve plotted for each month the difference between the first employment estimate and the twice-revised estimate reported two Jobs Days later (the third estimate). This difference is expressed as a percentage of the third estimate, so the figures tell how much the first estimate was too high or low relative to the eventual one. The chart covers over ten years—enough to include McEntarfer’s recent predecessors as BLS commissioner. I have shaded and labeled the chart to indicate the terms served by the various commissioners.

Figure 1. Percentage by Which the First Nonfarm Employment Estimate was Off, Relative to the Twice-Revised Third Estimate, by BLS Commissioner Term

(Positive = First Estimate too High, All Estimates Seasonally Adjusted)

Note: Estimates from the Federal Reserve Board of St. Louis (https://alfred.stlouisfed.org/series/downloaddata?seid=PAYEMS). All of the estimates are seasonally adjusted and are those reported before the benchmark revision. Note that there are no third estimates for November and December that are published separately from the benchmark revision; the third estimate for November and the second and third estimates for December incorporate the benchmark revision. Therefore, for the November points in the chart, I compare the first estimate to the second estimate, and I omit entirely the December points. For June 2025, since we will not know the third estimate until next month, I compare the first and second estimates. The April 2020 and September 2021 points are outliers (0.49 percent and -0.20 percent) and outside the range of the y axis.

An initial point to emphasize is just how small the revisions to the monthly estimates are. Over the past twelve and a half years, the first employment estimate nearly always has been within -0.1 to 0.1 percent of the third estimate. Note that this is not -10 percent to 10 percent; this is off by plus or minus one-tenth of one percent. Only nine estimates fall outside this range.

A second point is that despite the typically small revisions, there has been a trend toward the first estimates overstating employment. Importantly, that trend predates Trump’s current presidency or McEntarfer’s time at BLS. Since March 2022, with just three exceptions, the first estimates have tended to be too high.

We can compare how accurate the first estimates have been across BLS commissioners. Erica Groshen was appointed by President Obama and unanimously confirmed by the Senate. She served her full four-year term, from January 2013 to January 2017. During the 49 months in this span, the first employment estimate was, on average, too low by 0.01 percent relative to the third estimate. That’s pretty much spot-on.

Bill Wiatrowski, a long-time BLS employee, then served as acting commissioner for over half of Trump’s first term, until March 2019. The average accuracy of the first employment estimate was again impressive during these years—undershooting the third estimate by an average of just 0.01 percent over 26 months.

Bill Beach was nominated by President Trump in October 2017, but the Senate did not confirm him until March 2019 (on a divided 55-45 vote). He, too, served his full four-year term, finishing a couple of years into the Biden Administration. Beach had previously served in the conservative Heritage Foundation, on Republican Senate Budget committee staff, and at the libertarian Mercatus Center.

Beach’s tenure was marked by substantial data collection challenges during the COVID-19 pandemic. The accuracy of the first employment estimates were markedly volatile. Still, as with Groshen and Wiatrowski before him, Beach’s tenure saw the average first estimate too low by 0.01 percent. (Throwing out the March and April 2020 COVID-affected estimates, the average was low by 0.02 percent.)

However, from March 2022 to March 2023, the last year of Beach’s term, the first estimates were consistently too high. This was also true of Wiatrowski’s second stint as acting commissioner, from March 2023 to January 2024. During the last 13 months of Beach’s term, the first employment estimate was too high by 0.04 percent, on average, and it was high by an average of 0.05 percent for the 10 months of Wiatrowski’s term.

That brings us to McEntarfer, who was nominated by President Biden in July 2023 and confirmed in a vote of 86-8 by the Senate in late January 2024. McEntarfer previously had worked in the Census Bureau and served in the Obama Treasury Department and on Biden’s Council of Economic Advisers. Over the 17 months from February 2024 to June 2025, the first employment estimates were too high by an average of 0.04 percent—just as in the latter months of the Beach term and slightly lower than during the interim Wiatrowski term.

If the June estimate stays where it is in the next revision, it would be the biggest non-holiday, non-COVID downward revision going back at least to the start of 2013. (There were comparable or larger upward revisions three times in 2021, though they were likely related to data collection problems during the pandemic too.) It’s worth remembering that the June employment estimate could be revised upward next month, which would make the accuracy of the first estimate less of an outlier.

Moreover, the first May estimate was not an outlier in terms of how much it ultimately overstated employment. The revision of the March estimate was even larger, as were five revisions covering May 2023 to May 2024, with only the last two occurring on McEntarfer’s watch. The other three took place under Wiatrowski during the Biden presidency. (With McEntarfer gone, by the way, Wiatrowski is once again acting commissioner.)

Trump’s Truth Social post accused BLS of overstating jobs growth “right before the 2024 Presidential Election, in August and September.” He’s referring to the jobs report released on November 1, four days before Americans went to the polls, which revised job growth downward in those two months by a combined 112,000. Though Trump claimed this was a record, the downward revision in March was 167,000, and the July downward revision was nearly as high at 111,000. There were four downward revisions exceeding 100,000 in 2023.

The New York Times headlined its feature on the new August and September numbers, “U.S. Job Growth Stalls in Days Before Vote”—evidence, if it were needed, for the obvious conclusion that the revision was hardly a boon for the Kamala Harris campaign. And the same BLS report indicated that October had seen a pathetic 12,000 jobs added. Meanwhile, the first post-election jobs report found a gain of 227,000 jobs in November and revised upwards the September and October gains. If someone should have been fired for incompetence, it should have been whoever was supposedly in charge of getting Harris elected.

It’s not even clear what about the most recent jobs numbers Trump thinks is wrong. The White House memo criticizing McEntarfer implied that BLS had reported out overly optimistic job numbers back in July so that the Federal Reserve Board’s Federal Open Market Committee could justify higher interest rates. But the White House trumpeted these overly optimistic numbers when they came out. Are we to believe that if the May and June reports had accurately indicated job gains of only 19,000 and 14,000 (as the revised estimates show), that the July report with the meager 73,000 increase would have been OK?

Would McEntarfer even have lasted to the July report’s release? I mean, the May and June gains would have been the smallest since the pandemic—except for the tiny October gain reported right before the 2024 election…in order to…elect Kamala Harris. Am I getting this right?

Or perhaps Trump wishes BLS hadn’t revised the May and June numbers at all. But even in that case, the mediocre July number would have remained. Rather than reporting out an increase of 73,000 jobs in July, McEntarfer would have announced that jobs declined by 185,000. You have to go back to the earliest days of the pandemic to find anything remotely as bad as that. Here, too, it’s hard to see how McEntarfer would have kept her job.

What’s Gone Wrong with the Monthly Revisions?

Here’s the simple answer to that question in one chart. Figure 2 shows the rates at which establishments have cooperated with BLS as it administers the CES. The upper line displays establishments who responded in time for the release of the first employment estimate as a share of establishments who responded in time for the release of the third estimate. Response rates fell notably during 2020, 2021, and 2022 and have remained low since. Rather than around 80 percent of establishments that eventually report employment numbers to BLS doing so in time for the first estimate, today it’s more like 65 percent. Apparently, establishments with relatively weak job growth became increasingly likely to report their estimates late. First estimates deriving from on-time reporters therefore tend to paint too rosy a picture of the economy. Revisions that incorporate laggard employers give a more accurate but gloomier impression.

Figure 2. Response Rates to the Establishment Survey, by BLS Commissioner Term

Note: Estimates from the Bureau of Labor Statistics Current Employment Statistics program (https://data.bls.gov/series-report?redirect=true, Series CEU00000000C1, CEU00000000C3, and CEU00000000RR). The upper line is estimated as the first preliminary release collection rate divided by the final sample-based release collection rate. The lower line is estimated as the first preliminary release responders divided by establishments selected for the CES sample. Computationally, it is the first preliminary release collection rate multiplied by (final response rate + (final response rate * (1 – final sample-based release collection rate) / final sample-based release collection rate)).

In an important regard, these estimates understate the response rate problem, because a substantial share of employers never gets back to BLS. (In most states, participation is voluntary.) The lower line in Figure 2 shows establishments reporting in time to be included in the first estimates as a share of all employers contacted for the CES. Rather than about half of establishments participating, as was the case a decade ago, today just 25 to 30 percent do. That raises the concern that since only a minority of employers are cooperating, they may not reflect broader employment trends as well as in the past.

The Benchmark Revision

That brings us to the third main revision to the CES employment estimates—the annual “benchmark revision.” Each year, as better data come in from other sources, BLS updates its March employment estimate. It then propagates the revision backward and forward in time, publishing the new results on Jobs Day the following February. Again, this is a regular, planned revision, not any indication an error has been made.

Trump’s allusions to an 818,000-job correction in 2024 relate to the benchmark revision that was preliminarily estimated in August of that year. BLS publicly released their “preliminary benchmark announcement” as in previous years, which they do once the data behind the new March estimate first become available. Those new figures indicated that nonfarm employment in March 2024 was lower by 818,000 than the CES had estimated. When compared with the previously-benchmark-corrected March 2023 employment estimates, the implication was that growth in employment over those 12 months was lower by that same amount.

Trump and other McEntarfer critics have suggested that this revision, too, was a political one on behalf of the Democratic Party. However, the BLS announcement came the day before Harris formally secured the nomination to be the Democrats’ candidate for president—literally the last day before the general election campaign. Why this would be good for Democrats is…not obviously apparent.

Indeed, the clear political liability it constituted was best expressed (if unscrupulously and irresponsibly) at the time by one Donald Trump, who posted on Truth Social,

MASSIVE SCANDAL! The Harris-Biden Administration has been caught fraudulently manipulating Job Statistics to hide the true extent of the Economic Ruin they have inflicted upon America. New Data from the Bureau of Labor Statistics shows that the Administration PADDED THE NUMBERS with an extra 818,000 Jobs that DO NOT EXIST, AND NEVER DID.

Perhaps the nonsensical logic of claiming BLS’s timing helped Harris is why Trump has…revised…the history of when the release occurred. In one of those Truth Social posts from the day he fired McEntarfer, he wrote:

Just like when [Democrats] had three great days around the 2024 Presidential Election, and then, those numbers were “taken away” on November 15, 2024, right after the Election, when the Jobs Numbers were massively revised DOWNWARD, making a correction of over 818,000 Jobs — A TOTAL SCAM.

The November 15 date suggesting a quiet post-election revision appears to be completely made of out of whole cloth. What really happened is that on the eve of the general election, a full two-and-a-half months from Voting Day, BLS made a routine announcement that happened to show that the economy had been even worse under the Democratic Party than voters thought.

It gets even better, though, because as it happens, Trump’s talking points are not just temporally challenged but out-of-date. In February 2025, when the benchmark analyses were complete, BLS reported that the actual March 2024 revision was 589,000. If there was a conspiracy to elect Kamala Harris, the downgrade of the miss from 818,000 to 589,000 sure could have been timed better.

In Figure 3, I compare the third estimates (before any benchmark revision) to what they are today. The jumps apparent in these figures require some additional explanation for context. The estimates for most months (all except January, February, and March) are subject to two benchmark revisions. Each published revision in February begins with March of the previous year. Those revisions then are propagated back 11 months from March and forward through the just-past December.

So an employment estimate for, say, October 2023 was subject to a benchmark revision published in February 2024 (which updated the March 2023 employment estimate, previous ones through April 2022, and subsequent ones through the rest of 2023). Then it was also modified by the benchmark revision published in February 2025 (which updated the March 2024 estimate, previous ones through April 2023, and subsequent ones through the rest of 2024). Seasonally adjusted estimates may also change for up to five years, because the model for this adjustment is updated with each benchmark revision and applied to past years.

Figure 3. Percentage by Which the Third Nonfarm Employment Estimate was Off, Relative to Benchmark Revisions, by BLS Commissioner Term

(Positive = First Estimate too High, All Estimates Seasonally Adjusted)

Note: Estimates from the Federal Reserve Board of St. Louis (https://alfred.stlouisfed.org/series/downloaddata?seid=PAYEMS). All of the estimates are seasonally adjusted. The estimates compare the final sample-based estimates to the estimates as of Jobs Day, August 1, 2025. Chart runs to October 2024, since the November and December 2024 third estimates incorporate the initial benchmark revision applied to them and 2025 estimates have yet to be subject to a benchmark revision. The 2025 benchmark revision, which will be published in February 2026, will modify the April-October 2024 estimates shown in this chart (as well as the November and December 2024 estimates omitted).

Each year, the revision lifts or lowers 21 months’ of data together, causing jumps between years. Jumps show up systematically in Figure 3 between October and November. That’s because the third estimates for November and December already incorporate the benchmark revisions published in February. The third employment estimate for any November is the initial post-benchmark estimate. The only reason Figure 3 shows anything other than zeroes each November is that the November estimate may be revised by the subsequent benchmark revision (and by changes in seasonal adjustment). Meanwhile, the second estimate for any December is identical to the initial post-benchmark estimate, while the third estimate will differ because of new reporting from laggard establishments. It will then be revised again by the subsequent benchmarking.

It's clear that the benchmark revisions have a bigger impact on the employment estimates than the one- and two-month-out revisions do. Still, the third estimates are generally off by less than one half of one percent. As noted, the third estimates for 2024 were too high relative to the benchmark estimates. But that was also true in 2015, 2018, 2019, 2020, and 2023, as well as during much of 2016. The third estimates have been too high since November 2022, and getting worse. That’s a trend predating McEntarfer’s term though.

Last year’s benchmark revision was unusually large. The downward revision to the March estimate (from which the adjustments to other months are propagated) was larger in 2024 than in any year since 2009, 15 years earlier, and you have to go back another 18 years before that benchmark revision, in 1991, until the 2024 revision is exceeded again.

But Figure 3 shows that the cumulative effect of revisions was not off the charts relative to recent experience. One reason the absolute revision is so much larger than in the past is that the labor market is larger than in the past, so a given percentage miss is larger today. Moreover, the effects of any one benchmark revision can be moderated (or exacerbated) by the subsequent, second benchmark revision. The third estimates from January to October 2019 were too high relative to the final ones by an average of 0.32 percentage points, while those from January to October 2024 were too high by 0.40 points. Nor was the year-to-year jump an unusual departure from recent history. The swing from the 2022 revision, which raised employment estimates that had been too low, to the 2023 revision, which lowered estimates that had been too high, is much more striking.

It's worth remembering that the April-December 2024 estimates will be modified by the 2025 benchmark revision, which will be published next February. That revision will also modify the contentious May, June, and July figures for this year. In the meantime, we will get the preliminary estimate of the 2025 benchmark revision on September 9.

Concluding Outrage

The idea that Erika McEntarfer was fired for cause does not stand up. It is clear that she was dismissed not because of any operational breaches of protocol or because she is objectively doing a worse job at the challenging task of collecting timely and accurate employment data. That challenge has grown and has become more of a problem. But replacing the commissioner won’t fix it. It’s a problem that started well before McEntarfer’s term began. Fixing it will require giving BLS more resources to develop new solutions to the difficulties presented by low survey nonresponse.

An American president has removed the person running a statistical agency because he did not like the numbers it was producing. He has accused the agency of “rigging” official statistics, without any credible evidence. He clearly hopes to find someone who will deliver not the accurate numbers necessary for businesses, policymakers, savers, investors, students, and parents to make informed decisions, but the numbers that will flatter him and get him his preferred policies. Everyone else be damned. If that doesn’t upset you, you may not be as much of a patriot as you imagine yourself to be.

Trade with China began accelerating in the 1990s due to liberal reforms to the Chinese economy. However, anti-trade critics emphasize decisions at the turn of the century made or led by US policymakers: the granting to China of permanent normal trade relations status in 2000 and the concurrent acceptance of China into the World Trade Organization (WTO), which it joined at the end of 2001. Populists have blamed the subsequent wave of imports—the “China Shock”— for destroying manufacturing jobs, reducing employment generally, hurting marriage, and increasing receipt of safety net benefits and deaths of despair.

In this regard, no research has been more influential in providing fuel for the populist bonfire than the work of economists David Autor, David Dorn, and Gordon Hanson (ADH).1 However, their research has been widely misunderstood, in part due to questionable claims made in their papers. Those papers are primarily about how the China Shock affected some geographic areas relative to others in the US, not the overall impact of increased trade with China. Their findings have rarely been contextualized in a way to inform policymaking. Moreover, other researchers have contested them.

Many observers and experts believe the diffuse benefits from trade with China, such as lower prices, outweigh more-concentrated costs in the form of reduced manufacturing employment. What is perhaps not appreciated enough is that the China Shock may have hurt manufacturing employment much less than conventional wisdom suggests—or even boosted it.

The ADH Estimates of the China Shock’s Effect on Manufacturing Employment Are, Arguably, Small

ADH first assessed what they called the “China Syndrome” in a paper published in 2013.2 They leveraged the fact that different geographic areas (“commuting zones,” or CZs) had more or less susceptibility to import competition from China depending on their pre-Shock mix of industries. ADH assigned Chinese import growth in different industries to CZs based on the areas’ initial share of national employment in each industry. In other words, they assumed a CZ that initially had 4 percent of US employment in some manufacturing industry was hit twice as hard by increased Chinese imports within that industry as a CZ that initially had 2 percent of employment in the industry.

ADH summed these amounts across all manufacturing industries to get a measure of each CZ’s overall exposure to Chinese import growth. Finally, they scaled the growth in imports for each CZ by the area’s initial employment level. (Absorbing $1 million in imports is a bigger deal in a CZ with 50,000 workers than in one with 500,000 workers.) They found that stronger growth in Chinese imports in some CZs than in others caused those CZs to have worse manufacturing employment trajectories relative to the others than would have been the case absent the China Shock. If this seems like a very particular way to word the conclusion, the reason will become clearer below.

Assume for now that ADH’s estimates are unassailable. Do they suggest effects large enough to cause the economic, social, and political outcomes that are often blamed on the China Shock? To put them into context, imagine two large commuting zones each with 200,000 working-age people and 20,000 manufacturing workers in 2000.3 Imagine one of them was at the 10^th percentile of exposure to the China Shock—meaning that it was relatively unexposed to rising Chinese imports—and the other was on the other end, at the 90^th percentile. For simplicity, imagine further that they experience no population growth, and that in the absence of the China Shock, they would both have continued to have 20,000 manufacturing workers.

The 2013 paper implies that the CZ with the greater import exposure would have had about 2,700 fewer manufacturing workers than the other in 2007.4 On the one hand, that means that one out of every seven manufacturing workers would have lost their job in the one place but wouldn’t have in the other. On the other hand, it would mean a relative decline in the manufacturing employment rate of 1.4 percentage points—14 would-have-been manufacturing workers for every 1,000 working-age people.

This hypothetical example compares a very heavily hit CZ to a little-affected area. Put the more [ed.: originally said “less”] exposed CZ at the median level for import competition exposure instead of the 90^th percentile and the job loss in the more exposed place would have been just 900 workers who would have remained manufacturing employees in the other CZ. That’s one in 22 manufacturing workers, or 4-5 employees per 1,000 working-age people.

To be clear, we absolutely should care about manufacturing workers who lose their job because of import competition. That’s true whether they number 2,700, 2.7 million, or 27 workers. And we can almost certainly do a better job crafting policies to help them than our current safety net offers. But effects of this size hardly seem large enough to inspire a populist backlash against trade on their own.

Other studies by ADH have found smaller or larger effects. In a paper published two years after their initial China Shock study, ADH reported an effect on manufacturing employment that was 15 percent lower than their earlier estimate.5 It accounted for the distinct impact of automation on manufacturing employment differently than did the 2013 paper.

In another follow-up paper (Acemoglu et al., 2016), ADH and two coauthors used an import competition measure defined somewhat differently than in the earlier papers.6 The new measure produced much larger cross-CZ effects—more negative by nearly a factor of three.7 It’s not obvious, however, that the more recent measure is an improvement, and ADH have offered little justification for the change.8

More to the point, a number of subsequent studies have identified issues with the ADH analyses. When addressed, the magnitude of the China Shock effect tends to decline by over 50 percent.

Studies Reassessing the ADH Analyses Have Found the China Shock Had a Much Smaller Negative Impact on Exposed CZs’ Manufacturing Employment—Or a Positive One

One relatively minor issue relates to combining pre- and post-China Shock evidence. Rothwell (2017) separated out the 2000-2007 data used by Autor, Dorn, and Hanson (2013) from the 1990-2000 (pre-China Shock) data that they combined it with in their analyses. He found an effect of import exposure for the China Shock era that was smaller by 21 percent than the effect they reported.9 The effect for the pre-Shock era was 63 percent smaller. That is to say, simply separating the two periods out produced smaller effects for each period than reported by ADH, whose statistical model assumed the effect was the same for both periods.

Jakubik and Stolzenburg independently affirmed the Rothwell result for 2000-2007 in a 2018 working paper. The published paper (2020), using somewhat different data than ADH, found an effect one-third larger than the estimate reported in Autor, Dorn, and Hanson (2013). However, the variation in import exposure across CZs was also smaller than in ADH’s paper, so the hypothetical exercises above produce smaller job loss totals using their estimates.

As an example, we can compare two CZs with import exposure differing by a standard deviation, and otherwise plug in the same numbers from the earlier exercise. The implied difference in manufacturing employment in 2007 is 1,677 using the Autor, Dorn, and Hanson results but 1,477 using the Jakubik and Stolzenburg results.10

The Jakubik and Stolzenburg paper’s main contribution was in identifying the importance of an issue largely neglected in the research literature. They noted that the China Shock papers have generally measured imports in such a way as to assign imports entirely to the industry of the final good (for instance, smart phones) rather than recognizing that other goods and services went into the final good (glass, product design). Moreover, some “Chinese” imports include components made in America; counting the entire value of such an import as coming from the China Shock is perverse. In addition, imports get double-counted if, for instance, a US firm imports some good from China, uses it as in input to production, and sends the resulting intermediate good to a factory in China, which then produces a final good that becomes a US import. The value of the original good imported by the US also gets counted in the value of the final good imported.

When Jakubik and Stolzenburg focused on the “value added” of imports from China (after subtracting out the US value added), the negative impact of Chinese import competition on manufacturing employment across geographic areas increased by 50 percent compared with their estimate using “gross trade”. However, the variation across CZs in exposure to the value-added part of imports was lower by two-thirds than the variation in exposure to imports using gross trade.11 Therefore, the hypothetical examples above comparing two CZs would show much smaller differences in manufacturing employment trajectories using the value-added of imports. Instead of the difference in 2007 manufacturing employment being 1,477, it would be 762 jobs.12 That’s 55 percent lower than the corresponding estimate from Autor, Dorn, and Hansen.

In another paper, Borusyak, Hull, and Jaravel (2022) made two improvements to the analyses in Autor, Dorn, and Hanson (2013) and found that the effect of the China Shock on manufacturing employment was 55 percent smaller than the latter had estimated.13

A paper by Feenstra, Ma, and Xu (2019) included cross-CZ methods that seemingly replicated those in Acemoglu et al. but obtained very different results. They estimated a China Shock effect on manufacturing employment that, for reasons that aren’t clear, were about 35 percent smaller than the cross-CZ estimate from Acemoglu et al.14 Consistent with the Borusyak, Hull, and Jaravel results, when they implemented that paper’s improvements, their estimated effect was 50 percent smaller than in Acemoglu et al.15

Bloom et al. (2024) improved ADH’s industry coding and found the effect of the China Shock on manufacturing employment fell by 23 percent.16 Another change produced an even larger impact on the estimate. As noted, ADH looked at changes from 1990 to 2000 and from 2000 to 2007. Bloom and his colleagues pointed out that the underlying establishment data used by ADH is better during years when the economic census is conducted (every five years). When they looked at changes from 1992-1997, 1997-2002, and 2002-2007 (all census years), the China Shock effect on manufacturing employment fell by 38 percent. When they looked only at changes from 1997 to 2007, which they argued is the cleanest comparison, the effect was lower by 58 percent than when looking at 1990-2000 and 2000-2007.

A paper by Chaisemartin and Lei (2023) addressed other assumptions embedded in the ADH methods. Their estimates of the China Shock’s effect on American manufacturing employment were imprecise, but two out of three were positive rather than negative.17 That is, the evidence was consistent with more exposure to Chinese imports being better for a CZ’s manufacturing employment.

Magyari (2017) used the same methodology of Autor, Dorn, and Hanson (2013), except that she looked across manufacturing firms rather than across geographic areas. That is, she leveraged the fact that different firms (potentially with multiple establishments) had more or less susceptibility to import competition from China depending on their establishments’ pre-Shock mix of industries.

Magyari assigned Chinese imports in different industries to firms based on firms’ initial share of national employment in each industry. She found that between 1997 and 2007, manufacturing firms with greater exposure to the China Shock saw bigger increases in manufacturing employment (or smaller declines). Cheaper imports reduced firms’ costs, and firms reallocated employment to establishments in industries less impacted by import competition. Magyari’s results don’t invalidate ADH’s commuting zone results—firms may have reallocated manufacturing jobs across commuting zones—but they cast doubt on whether the China Shock, on net, was bad for American manufacturing employment.

Note that the critiques in these papers have generally not been taken up by the other papers. Presumably, addressing multiple issues treated separately by the studies would have a bigger impact than any of the individual modifications.

Other Studies Have Found the China Shock Had a Negative Impact on Exposed CZs’ Manufacturing Employment but a Positive Impact on Their Total Employment

The ADH papers have found not only that greater exposure to the China Shock was worse for CZs’ manufacturing employment, but for their employment generally. Any employment gains outside of manufacturing were not enough to prevent overall employment growth from worsening. But other studies have contradicted this conclusion too.

We’ve already seen that Chaisemartin and Lei (2023) and Magyari (2017) found evidence that the China Shock wasn’t even necessarily bad for manufacturing employment. Other research using methods similar to ADH has found negative effects for manufacturing employment but more-than-offsetting positive effects on employment outside manufacturing.

Wang et al. (2018), using a cross-CZ approach and import concentration measure based on Acemoglu et al., found that while the China Shock had a negative effect on manufacturing employment, it increased total employment in the exposed CZs. It did so primarily by making intermediate goods in non-manufacturing firms cheaper (computers and office equipment, for example). The Autor research missed this result by failing to distinguish between imported inputs used by American firms and imported final goods.

Bloom et al. (2024) found that after accounting for the varying quality of the establishment data in different years (depending on if the data came from an economic census year), the negative effects of the China Shock on manufacturing employment were more than offset within CZs by positive effects on non-manufacturing employment.

ADH’s Estimates of the China Shock’s Impact on National Employment Dubiously Apply Relative Effects Based on Comparing Geographic Areas

The ADH findings that have garnered the most attention are not the somewhat subtle estimates of relative job declines in more-trade-exposed areas compared to less-exposed ones. Rather, the headline results have been ADH’s estimates of national job loss due to the China Shock. However, these estimates come from dubious calculations using the cross-CZ estimates (or similar cross-industry estimates).

What the original ADH paper found, precisely, was that experiencing stronger Chinese import competition caused greater declines in manufacturing employment in some places or smaller increases relative to places less exposed to import competition. The research design of the paper can’t tell us whether, in the aggregate, Chinese import competition reduced total American manufacturing employment or by how much; it only provides information about how the resulting change differed across geographic areas.

We know that American manufacturing employment has declined for decades (starting well before the China Shock). But imagine that increased imports from China slowed this decline in a small number of very large CZs not strongly exposed to competition from the specific imports China sent our way. Maybe cheap imported inputs reduced business costs and allowed many manufacturers to expand. In that case, we would say that the China Shock increased manufacturing employment in these CZs (relative to a counterfactual in which no Shock occurred). The hypothetical increase caused by the China Shock wouldn’t necessarily have been enough to actually raise manufacturing employment in these CZs, but it would have partly countered the other factors leading those jobs to become rarer.

At the same time, in this hypothetical, it might have been that many smaller CZs were more exposed to import competition and saw relatively large declines in manufacturing employment. Consistent with ADH, we would find that greater exposure to Chinese import competition was associated with worse manufacturing employment trends. But it could very well have been that the “increased” employment in the less-strongly exposed CZs exceeded the reduced employment in the more-strongly exposed ones. That could have been true even as the long-term manufacturing employment trend continued to decline.

Alternatively, perhaps the China Shock raised manufacturing employment in all CZs (relative to there not being a China Shock)—just not as much in the ones with more exposure to imports. That might have occurred if imports universally increased some kinds of manufacturing employment through cheaper inputs, leading to lower business costs. At the same time, import growth might have put workers in other kinds of manufacturing out of a job in more-exposed CZs. Such a universal effect would be netted out in estimating a cross-CZ effect and leave the negative effect of greater import exposure in some CZs relative to others.

Or perhaps Chinese imports lowered consumer prices enough to increase demand for domestic manufacturing (and thereby boost manufacturing employment) by increasing purchasing power. If we just assume that effect, if it occurred, was constant across CZs, it gets netted out in estimating the cross-CZ effect. But if we want to know the national effect of the China Shock, we need to incorporate that universal-to-CZs demand-boosting effect.

Of course, it’s possible the China Shock hurt all CZs, but the magnitude is nonetheless overstated because the relative effect comparing more- and less-exposed CZs nets out universal-to-CZs positive effects.

We can put the issue in terms of the first ADH paper’s findings. It is one thing to conclude that a CZ experiencing $1,000 per worker more in Chinese imports than another area saw its manufacturing employment rate fall 0.6 percentage points more than the other (or grow 0.6 points less). It is quite another to say that because the US saw an increase in Chinese imports of $1,839 per worker from 2000 to 2007, that -0.6 effect implies the China Shock reduced the national manufacturing employment rate by 1.1 percentage points.18 The second conclusion does not necessarily follow from the first.

Nevertheless, ADH applied that estimate to argue that the China Shock reduced national manufacturing employment by 2 million workers. Their headline estimate of 982,000 is a downward adjustment to account for the fact that not all the increase in Chinese imports was due to changes in the supply of, rather demand for, imports. (The cross-CZ effects are all estimates that focus on the China Shock as a supply-side phenomenon.)

In their follow-up papers, ADH similarly estimated an absolute national China Shock effect from their relative cross-CZ estimates. Acemoglu et al. found that the China Shock reduced employment in manufacturing by 2.35 million workers between 1999 and 2011 (or 36 percent). Autor, Dorn, and Hanson (2021) found that 59 percent of the drop in the manufacturing employment rate from 2001 to 2019 was due to the China Shock.19

In fact, Acemoglu et al. argue that the China Shock’s true impact on national manufacturing employment was larger than the 2.35 million suggested by applying their cross-CZ estimates. In a second set of analyses, ADH rely on a cross-industry estimate of the China Shock’s impact on manufacturing employment and (again, questionably) use that to obtain a national estimate. This total comes to 985,000 workers. Because neither of the two approaches in the paper accounts for all the ways that the China Shock might have reduced manufacturing employment, Acemoglu et al. assert that the true effect is larger than either implies.

Even if we were to accept the validity of using cross-CZ or cross-industry China Shock effects to estimate a national effect on manufacturing employment, there are reasons to think the ADH guesses are too negative. First, the smaller cross-CZ effects other studies have found imply that putting improved estimates through the ADH formulas would yield national totals much lower theirs. That’s to say nothing of the studies finding positive employment effects.

Second, expanded trade with China may have increased American manufacturing jobs through greater exports.20 All the research discussed above has focused on the impact of increased Chinese imports. Feenstra, Ma, and Xu (2019) used cross-industry methods intentionally comparable to Acemoglu et al. to look not just at the employment effects of increased imports from China but of increased exports to China. Their results are imprecise but suggest that to a great extent, the two effects balanced out—perhaps fully so. (And they estimate an effect of imports that is very similar to that of Acemoglu et al.)21

Finally, all the research estimating “effects” of the China Shock suffers from an important real-world weakness. They all envision a counterfactual world in which not only do Chinese imports stay frozen at, say, 1990 or 1999 levels, but imports from other developing countries fail to expand to meet the American demand that the China Shock filled. Imagine if China had never liberalized its economy or if it had been kept out of the World Trade Organization. Heck, imagine that we had levied insane tariffs against China in 2000. What would have happened?

If you think that the China Shock reduced employment, do you think all those jobs would have been saved if it could have been prevented? Or do you think that trade with Vietnam and other countries would have grown at a faster rate in the absence of greater trade with China? Implicitly, all this research has in mind that trade with other countries would not have ramped up in the absence of Chinese import competition. But that seems highly dubious. If at least some ramp-up would have happened, then all the research discussed above overstates any negative impact of the China Shock.

A similar argument could be made about automation. Suppress the China Shock and perhaps manufacturer reliance on automation would have accelerated to keep costs and prices down. If that would have been the counterfactual history, and automation would have reduced manufacturing employment more than it did in the presence of the China Shock, then the studies above overstate the effect of the Shock.

Indeed, when we remember that the cross-CZ studies are estimating relative effects (how some places were affected by the China Shock relative to others), we must also take seriously a different sort of counterfactual if the China Shock had been suppressed. If places vulnerable to import competition had avoided those rising imports, they still might have lost manufacturing employment to places elsewhere in the US to which manufacturing had already been migrating for decades. Some places would have lost out relative to others even in the absence of the China Shock, and if the reallocation had accelerated in a world without the Shock, the research discussed above gives the China Shock too much blame. If the cross-CZ estimates of the China Shock are too large in this sense, then applying them to get a national estimate will carry forward the problem.

In the 30 years from 1969 to 1999, the share of the working-age population employed in manufacturing fell at a rate of 1.9 percent per year. Over the 20 years from 1999 to 2019, the drop was 2.0 percent per year; the decline in manufacturing employment has proceeded steadily for decades, with no obvious break created by the China Shock.22

Conclusion

The Autor, Dorn, and Hanson research has been influential, and for a reason. It uses sophisticated analyses to answer one of the most-discussed policy questions of the day. I have only discussed their findings on employment, but their papers also look at a range of other outcomes potentially related to the China Shock. The critique offered here largely applies to those results too.

Even regarding employment, I have stuck to papers that are in conversation with the ADH research, ignoring studies that have relied on, for instance, macroeconomic modeling.23 I do not claim to have summarized the entire relevant body of research on the China Shock. Moreover, there are important distributional questions discussed in the papers cited here that I have ignored for sake of length. These include questions about how costs and benefits are distributed geographically and demographically. Similarly, I have avoided discussion of other effects of trade, such as on prices or welfare.

I do not mean to suggest that this critique nor any of the research I have cited refutes the ADH research. Their findings should absolutely inform policymakers’ and citizens’ thinking about the costs and benefits of increased trade with China. But we should not oversimplify, misinterpret, or fail to contextualize those results. Nor should we necessarily elevate them over equally sophisticated studies that point to other conclusions.

The message from this review is that the debate around the effect on workers of increased trade with China since it joined the WTO and was granted permanent normal trade relations status has been unduly alarmist given the ambiguity of the evidence we have at hand. Economic theory offers a variety of reasons to think that voluntary exchange benefits both parties in a trade and that freer trade tends to benefit a country even in the face of trade barriers erected by other countries. If the evidence clearly contradicted the theory, we would want to strongly question the theory. But even confining ourselves narrowly to the evidence on manufacturing employment and Chinese imports—ignoring all the other channels whereby trade affects Americans—we see no such clarity.

A clear-eyed look at the existing evidence should inform both our policy debates around trade and tariffs and our political debates about how increased trade has affected voter preferences. Too often, we grant anti-trade sentiment a greater role in fueling populism than may be warranted. Or we are too quick to assume that sentiment corresponds with a widespread lived experience instead of simply reflecting commitment to ideological priors.

We neglect the importance of great cultural divisions and of social breakdown in assessing what ails our nation to the extent that we focus unduly on trade and economics during a time in which Americans are better off in material terms than humanity has ever known.