My most recent post highlighted BALANCE, a large noninferiority trial showing that 7 days of antibiotics was noninferior to 14 days in many patients with bloodstream infections, even if they’re in the ICU.

While researching that piece, I realized that a manager’s decision to pull a starting pitcher parallels the decision to stop antibiotics. The central question is basically the same: At what point does continued exposure stop adding benefit and start only adding risk?

I was surprised to discover that over the last century, baseball and medicine have tackled this question essentially in tandem.

The stories are fascinating, and as a lover of history, I couldn’t help but write about it.

So this will be different (and quite a bit longer) than my typical posts. The bulk of the column will present the histories side by side as they unfolded.

But when we get to BALANCE, I’ve also included my critical appraisal as a bonus for paid subscribers.

I hope you enjoy this as much as I enjoyed writing it.

From Dead-Ball to Dead Bacteria

Prior to 1920, professional baseball games were low-scoring affairs dominated by pitching. This came to be known as the “dead-ball era.” The expectation then was that starting pitchers, barring injury, would finish what they started. In 1904, Jack Chesbro threw 48 complete games in 51 starts. He pitched 454.2 innings, finishing the season with a 41-12 record.

Relief pitchers were a rarity, but as the game evolved, so did pitcher workload. There were 2,186 complete games in 1904, nearly 88% of games. For context, in 2025, no starting pitcher threw more than two complete games.

But what became apparent over time was that there were advantages to shifting away from starters routinely throwing all nine innings.

A series of rule changes and the arrival of Babe Ruth as a cultural icon ushered in the “live-ball era” in 1920. As the game changed over the next two decades, teams began to roster pitchers who specialized in relief pitching. Before, in the rare case that a reliever was needed, managers typically turned to a starter on his off day. The success of players like Firpo Marberry and Wilcy Moore proved that effective relief pitching could be a valuable asset.

In 1928, Alexander Fleming made one of the most monumental scientific discoveries in history by accident. The Scottish microbiologist noticed that a bacterial culture which had been contaminated by “juice” from a mold called Penicillium notatum didn’t grow properly. His work was largely ignored by the scientific community for the next several years.

Although discovered later, sulfonamides became the first antibiotics to achieve clinical use. In 1932, a German scientist named Gerhard Domagk found that the synthetic dye Prontosil was effective in treating bacterial infections in mice. His own daughter and Franklin D. Roosevelt’s son were among the first test patients for the new miracle drug.

Prontosil was actually a prodrug which was metabolized into sulfanilamide, and many drug companies put formulations of the patent-less sulfanilamide on the market in the 1930s. But it was the S.E. Massengill company’s liquid formulation that would achieve notoriety. Aiming to corner the market, Massengill’s product was marked as a cure for strep throat in an easy-to-swallow elixir.

Over the fall of 1937, 105 patients (34 of whom were children) died of kidney failure soon after taking the product. The culprit was not the sulfanilamide compound, but rather the solvent—diethylene glycol, which we now know can induce proximal tubular necrosis. The Food, Drug, and Cosmetic Act was passed by Congress in 1938 as a response to the tragedy.

Meanwhile, research on the bactericidal compound, penicillin, continued. In 1940, Howard Florey and his colleagues published their findings from treating infections in mice, and this revolutionary drug finally became a clinical reality.

The first known human patient to be treated with penicillin was an Englishman named Albert Alexander in 1941. His condition, as described by his physician, Charles Fletcher, was beyond hope:

“In the septic ward at the Radcliffe Infirmary there was then an unfortunate policeman aged 43 who had had a sore on his lips four months previously from which he had developed a combined staphylococcal and streptococcal septicaemia. He had multiple abscesses on his face and his orbits (for which one eye had been removed): he also had osteomyelitis of his right humerus with discharging sinuses, and abscesses in his lungs. He was in great pain and was desperately and pathetically ill. There was all to gain for him in a trial of penicillin and nothing to lose.”

Fletcher, with Florey’s help, administered penicillin through a slow continuous IV infusion and to their amazement, Alexander showed “striking improvement.” The drug was in short supply, however. Despite extracting and refining excreted penicillin from Alexander’s urine, the penicillin ran out after 5 days. Alexander’s infection relapsed, and he died a month later.

It was a stark early example of the risk of insufficient duration—one of many things medicine would need to figure out in this uncharted age of antibiotic therapy.

The Rise of Specialization

Baseball’s integration in 1947 saw an influx of talent that drastically changed the competitive environment. Managers enjoyed a much deeper pool of pitching talent at their disposal. The calculus shifted such that it became evident that a fresh reliever was often better than a tiring starter. By 1957, the frequency of complete games had dropped to just 28.7%.

But traditions change slowly in baseball, and the iron men of the 1960s and 70s represent the last vestige of the workhorse starter. Fergie Jenkins, Mickey Lolich, Bob Gibson, Steve Carlton, and Nolan Ryan all had seasons with over 25 complete games apiece.

Yet it was during this era that teams began to deploy a “fireman” if their starter ran into trouble late in games. These unshakable men were brought in to “put out the fire” and typically went multiple innings to finish out the game.

The most famous fireman, Mike Marshall, became the first relief pitcher to win the Cy Young award in 1974. He appeared in an MLB-record 106 games—all in relief—with a 2.42 ERA over 208.1 innings.

In the 1980s and 90s, bullpens further evolved as multi-inning firemen transitioned into single-inning closers often paired with an 8^th-inning “set up man.” Middle relievers and long relievers became commonplace. Innovative managers deployed relievers with ultra-specialized roles such as the LOOGY (“lefty one-out guy”) to target specific batters in high-leverage spots.

As baseball entered the 21^st century, pitchers going the distance became a rarity. In 2001, the rate of complete games dropped below 5% for the first time. It has trended down ever since.

The Moneyball era taught baseball teams that there was opportunity in questioning the longstanding dogmas of medicine. Knowing precisely the most optimal time to remove your starter was of great interest to clubs seeking an edge.

In 1999, the question expanded beyond effectiveness on the mound. When the 21-year-old Cubs phenom Kerry Wood tore his ulnar collateral ligament in the spring following his historic rookie campaign, injury prevention became a top priority.

By this time, teams already monitored pitchers for signs of fatigue and pitch count tracking was already in place. In 1998, the flame-throwing Wood routinely threw outings of 120+ pitches, including his 20-strikeout masterpiece against the Astros on May 6^th.

It made sense that pitching while fatigued increased injury risk, but how many pitches were too many? Nobody knew for sure.

The post-penicillin era was a heady time for doctors. New antibiotics were emerging, and armed with powerful new weapons against illnesses for which they were previously helpless, physicians began to feel invincible. In his 1962 book, Natural History of Infectious Disease, the Nobel Laureate Macfarlane Burnet claimed victory:

“One can think of the middle of the 20th century as the end of one of the most important social revolutions in history, the virtual elimination of the infectious diseases as a significant factor in social life.”

Yet, many experts at the time had already been sounding the alarm regarding antibiotic resistance. Fleming himself warned in 1945 about the risk of developing resistant pathogens through insufficient dosing. Worrisome pathogens such as MRSA began to appear in the 1960s along with penicillin-resistant S. pneumoniae.

Maxwell Finland and other leading scientists surmised it was selective pressure from antibiotic misuse driving the emergence of resistant pathogens. Physicians of the era prescribed antibiotics even more indiscriminately than we’re used to today—often recommending them for any ailment they didn’t fully understand. Finland’s team called for physicians to change overzealous prescribing habits but met with little success.

By the 1990s, pathogen resistance began to outpace drug development. Options for treating certain infections were increasingly limited, and the first antibiotic stewardship programs were created to address the growing threat.

Knowing when, and for how long, physicians should use antibiotics became critical.

A Heuristic Emerges

As the new millennium approached, the question of how many pitches can a pitcher safely throw remained unanswered. Famed orthopedist James Andrews—whose name should be familiar to anyone who regularly watched SportsCenter in the 1990s and 2000s—proposed 100 as the limit in 1999. Although arbitrary and lacking supporting evidence, it was a nice round number, and thus made a fine heuristic by which organizations could make decisions about starter limits.

In reality, the point at which a pitcher fatigues varies, but the 100-pitch limit became a hard and fast rule. When baseball Prospectus developed Pitcher Abuse Points (PAP)—which correlated pitcher ineffectiveness and injury risk with higher pitch counts—they chose 100 pitches as the threshold beyond which any additional pitches were too many.

Observational studies using PAP data led to dramatic changes in pitcher usage. In 1998, there were 498 starts where pitchers threw at least 120 pitches. Kerry Wood had 8. The most pitches thrown in a start that season was by Roger Clemens, who threw 149 pitches against the Oakland Athletics on May 13.

A decade later, only 73 of 4856 of MLB starts — a mere 1.5% — exceeded 120 pitches. The 100-pitch heuristic was firmly cemented.

In the same way, prior to BALANCE, conventional durations of antibiotic therapy in bacteremic patients were based on expert opinion rather than solid evidence. These recommendations were mainly informed by historical experience, especially as it relates to staphylococcal bacteremia. Survey evidence cited by the authors of BALANCE suggested that the median duration of antibiotics used in practice was 14 days or longer for critically ill patients.

Lacking better evidence, a conservative approach became the default.

Data Replaces Guesswork

The 100-pitch count was mainly a heuristic for mitigating pitcher injury. Measuring a pitcher’s effectiveness over the course of a start was still based on observational factors and plain old gut instinct.

If teams could reliably recognize when to pull their starter before they ran into trouble, they would have a major strategic advantage. In 2006, The Book: Playing the Percentages in Baseball by Tom Tango, Michael Lichtman, and Andrew Dolphin provided an answer—the times-through-the-order penalty. Their data demonstrated that the opposing team’s wOBA increased with each successive turn through the order. The authors summarized their findings succinctly:

“As the game goes on, the hitter has a progressively greater advantage over the starting pitcher.”

So simple, it’s a wonder that teams hadn’t thought of it previously.

Interestingly, the times-through-the-order penalty became a heuristic of its own. The Book demonstrated that the batting team’s wOBA peaked during the third time through the order. This became a simple cutoff where a manager could turn to the pen. In fact, many now refer to it as the third-time-through-the-order penalty, and it’s such a mainstream concept that even TV commentators cite it regularly.

At the same time, medicine’s paradigm around antibacterial treatment also began to shift. The traditional belief was that stopping antibiotics early would promote resistance. In the early days of my training, patients were commonly counseled to take their full course of antibiotics to mitigate this risk.

Dr. Louis Rice, in a keynote lecture named for Maxwell Finland (later published in the Feb. 2008 issue of Clinical Infectious Diseases) challenged this assumption. In fact, he flipped it on its head—he asserted that the most effective approach to combating antibiotic resistance was to prescribe shorter durations of therapy.

A wave of RCTs began to emerge in the 2000s and 2010s demonstrating noninferiority with shorter courses in treating a myriad of bacterial infections: 5 days for cellulitis, 5 days for community-acquired pneumonia, 4 days after source control for intra-abdominal infections to name a few.

Brad Spellberg, inspired by Dr. Rice, has been a long-time advocate of shorter antibiotic courses. In 2016, he published an editorial in JAMA Internal Medicine that coined the phrase, “Shorter is Better,” and he continues to catalogue the growing evidence base on his website.

Then came the earliest bacteremia studies. Yahav et al (2019) demonstrated that a 7-day course was noninferior to 14 days in uncomplicated patients with gram negative bacteremia. PIRATE (2020) and SHORTEN (2022) followed soon after with similar results. Yet these were smaller studies involving gram negative bacteremia, few ICU patients, and a wide 10-point noninferiority margin.

By now, both medicine and baseball began to have real, quantifiable evidence supporting a massive change in usage. Antibiotics—and starting pitchers—should remain in place only as long as necessary, and not a moment longer.

Breaking the Link

In 2018, the forward-thinking Tampa Bay Rays decided to try something different. On May 19, veteran relief pitcher Sergio Romo took the mound to start the game for the first time in his career. He pitched a 1-2-3 first inning, striking out Zack Cozart, Mike Trout, and Justin Upton, before handing it over to traditional starter Ryan Yarbrough for the next 6.1 innings in relief.

Why this unusual arrangement?

The top of the Angels lineup was heavily right-handed, and Yarbrough a lefty. The idea was to use Romo as an “opener” where he would hopefully pitch an effective inning or two before letting Yarbrough take over.

This has two distinct advantages. First, more runs are scored in the first inning than any other inning. Bringing an effective relief pitcher allows you to potentially exploit favorable matchups against their best hitters in a high-risk inning. Second, it delays the third time through the order for the would-be starter by entering the game after their first plate appearances.

The Rays would use an opener a few dozen times in 2018, and several other teams would follow suit. Tampa Bay also popularized “bullpenning”—a strategy where only relief pitchers were used to get through an entire game. Having 5 to 8 different guys pitching 1-2 innings apiece kept lineups from developing any semblance of familiarity.

It was a clear conceptual shift that broke the cognitive link between a pitcher starting the game and pitching the majority of the game. The new way of thinking was “we need to get 27 outs, so which pitcher should get which outs?”

When it comes to antibiotic duration, medicine is in the midst of making a similar shift. Dr. Spellberg jokes about Constantine Units, suggesting that the 7-day week established by the ancient Roman emperor leads physicians to prescribe antibiotics in intervals of 7 days. Regardless of whether that’s actually the reason, the point is that treatment durations are largely arbitrary, and probably excessively long.

BALANCE was a big step in changing that paradigm. Let’s look at it more closely.

Where we stand today

Openers and bullpen games are somewhat less common than their peak in 2018. Nevertheless, teams have incorporated the principles they established into conventional pitching staffs. Teams are comfortable with shorter starts because they’ve designed their bullpens for maximum efficacy.

Pitch tracking technology (like TrackMan or Rapsodo) and high-speed cameras (like Edgertronic) have helped teams get the most out of their pitchers. The camera captures precisely what the hand, fingers, and seams are doing at release, while the tracking system quantifies it (e.g. velocity, spin rates, movement). Motion capture technology and force plates allow teams to precisely measure how efficiently a pitcher generates force through the kinetic chain.

Pitching labs like these enable pitchers to better understand what their body is doing and the small adjustments they can make to get the most out of each pitch. Now, instead of pacing themselves, pitchers throw at maximum effort to get the best result with each pitch.

This same technology also helps teams track a loss of efficacy in real-time. Coaching staffs can better identify fatigue as metrics like velocity or spin rates start to decline. The decision to change pitchers has never been so data-driven.

Medicine hasn’t quite gotten there yet. We’re still pretty far from knowing precisely the threshold where additional antibiotic treatment stops adding benefit and starts adding harm. Now we know it’s probably sooner than we thought. Perhaps we’ll eventually see something like baseball’s pitch design revolution—RCT data on individualized, biomarker-guided, and potentially shorter durations of therapy.

We’re still in the “bullpenning” stage, but potentially on the cusp of a major shift in thinking—the untethering of antibiotic duration from the calendar and tying it to evidence instead.

If this piece made you think differently about how much of medicine’s conventional wisdom is really just an arbitrary number that stuck, a like ❤️ or restack 🔁 helps this reach others wrestling with the same questions.

Leave a comment

Subscribe now

Share

In this episode, I explore why outcome bias makes a reasonable decision look foolish in hindsight, what the times-through-the-order penalty and win probability actually tell us about Marmol's c…

Read more

July 19^th, 2026. Through five innings, Cardinals starting pitcher Andre Pallante was cruising. He had allowed only two hits, no walks, and the Cardinals had just scored two runs to extend their lead to seven.

Cardinals manager Oliver Marmol did what many managers would do in that situation — he sent Pallante back out for the bottom of the sixth.

This didn’t work out. He allowed three straight hits, and retired only one of the first five batters he faced before getting pulled. The Diamondbacks scored three runs that inning, sparking an Arizona rally that would result in a devastating 10^th-inning loss for St. Louis.

Afterward, some fans criticized Marmol’s decision to stick with Pallante. Outcome bias — the tendency to evaluate a decision based on the result rather than the process — is common and often incorrect.

In this case, it’s easy to look at a historic collapse (this was the largest comeback in Diamondbacks history) and point to this particular decision as a key mistake. Decisions like these are an easy target because the counterfactuals are never seen. Judgment based on a bad outcome tends to presume that the alternative would have been better.

It presumes that a reliever would have held the lead, but there are two plausible alternative realities:

1. Marmol brings in a reliever who gives up five runs — he’s criticized for overmanaging.

2. Pallante pitches a clean inning — nobody bats an eye.

This is decision-making under uncertainty. In these situations, we make the best decision we can based on the information we have at the time. Oliver Marmol had no idea what would happen in the sixth. Call it luck, randomness, whatever — the fact is, the outcome was entirely out of his control.

So instead, let’s evaluate the decision by the quality of the process.

In Marmol’s defense, the Cardinals had a huge lead and teams in that position historically rarely lose. The St. Louis win probability was 98.2%. Additionally, Pallante had been highly effective and had only thrown 67 pitches. The prior probability of this decision resulting in a loss for the Cardinals was very low.

There’s important context to consider. The Cardinals bullpen has been bad this year. It’s not obvious that a more effective pitcher was available. Plus, this was game three of a twenty-one game stretch without a day off. Attempting to milk another inning out of the starter is understandable.

On the other hand, the times-through-the-order penalty is a well-known phenomenon demonstrating that a pitcher’s effectiveness decreases as they face batters a second or third time. Aggregate data from hundreds of thousands of games tell us that subbing them out prior to a third trip through the order is a generally wise policy. The conventional wisdom says to Marmol, “You got five good innings out of him, don’t get greedy.”

I’m typically reluctant to expose starters to a lineup a third time, but sometimes context favors a slower hook. In my opinion, trying to get a little extra length out of Pallante in this instance was a good decision spoiled by luck.

Clinicians face this same trap all the time. We make difficult decisions with real uncertainty, and let outcomes that aren’t under our control determine their quality.

Managing bacterial infection is a common scenario across many specialties. There are two key decisions to be made: which antibiotic to use and for how long. The answer to the latter is typically less clear.

If the duration is too short, it risks treatment failure or relapse. Developing resistance in the target pathogen is also a concern.

A course that is too long can promote resistance in non-target bacteria and potentially increase a patient’s exposure to adverse events, including C. difficile infection.

So we’re looking for the minimum duration of therapy needed to achieve cure. Recent trials have shown that for a variety of infections, a shorter course of antibiotics than previously supposed often does well enough.

Data on the optimal duration of treatment for bacteremia (bacterial bloodstream infection) have been limited, however. Three smaller RCTs suggested that 7 days of antibiotics was non-inferior to 14 days, but these were limited to only gram-negative pathogens and mostly non-ICU patients.

BALANCE — a large multicenter noninferiority trial — sought to provide stronger and more generalizable data. The results were published in the New England Journal of Medicine in March 2025.

Investigators randomized 3,608 hospitalized patients with positive blood cultures to receive either 7 days or 14 days of antibiotic therapy. Over one-half (55%) were in the ICU. Many patients were excluded, including those with infections requiring prolonged courses (like staph infections, endocarditis, and osteomyelitis) or other complexities such as severe immunocompromise and the presence of prosthetic heart valves or endovascular grafts.

The primary outcome was mortality at 90 days. They estimated a baseline 90-day mortality of 22%, and set the noninferiority margin at 4 percentage points — well below the 10 percentage point margin used by previous RCTs. This means that the upper bound of the confidence interval for mortality in the 7-day arm could be up to 4 absolute percentage points worse and still be considered noninferior.

By 90 days, 261 patients in the 7-day arm (14.5%) had died compared to 286 (16.1%) in the 14-day arm [difference -1.6 percentage points, CI -4.0 to 0.8]. Because the upper bound of the CI fell within the 4 percentage point margin, noninferiority was met. This analysis was by intention-to-treat.

Notably, nonadherence to the protocol was high. In the 7-day group, 23.1% received antibiotics for longer than 9 days. In the 14-day group, 5.8% received antibiotics for less than 12 days and 10.7% for longer than 16 days. Nonadherence can bias results toward noninferiority. Reassuringly, per-protocol analysis (limited to only those who received the assigned duration) also demonstrated noninferiority [difference -2.0 percentage points, CI -4.5 to 0.6].

Rates of relapse, length of stay, and treatment-related adverse outcomes (including C. difficile infections) were similar in both groups.

It’s important to remember that noninferiority does not mean that a 7-day course is equivalent or better than a 14-day course. But, on the whole, BALANCE offers compelling evidence that, for many patients with bloodstream infections, a 7-day course is adequate.

Still, it does not resolve the uncertainty of managing individual patients. Obviously, many bloodstream infections weren’t included, particularly staph infections, which are the second most common type. Regardless, the trial does not tell you the best course of action for any particular patient.

What it does is offer an evidence-based process that can be applied as a default across a population like this one. For many bacteremic patients, 7 days of antibiotics will be enough, but perhaps not for others.

Just as a baseball manager must consider the realities of the in-game situation, we must do the same with managing patients. Within that context, we then decide at what point continued treatment offers little meaningful benefit and only adds risk.

This trial gives us an evidence-informed estimate, but we must make the final call.

Then, after the decision is made, we don’t judge it by the outcome. Remember, mortality was high in both arms of this trial — a lot of good decisions will inevitably result in death.

Instead, we judge it by our process. Did we use the best aggregate evidence available? That is, did we apply both the population-level research and the individual context to make the best decision possible in the moment?

If we can say “yes” to that — to me, that counts as a win, no matter what the final score says.

If this piece made you think differently about how we judge the decisions we make, a like ❤️ or restack 🔁 helps this reach others.

Share

Leave a comment

Read more

I had planned on writing something different for this post. But I decided to shift gears because I haven’t been able to stop thinking about an editorial recently published on Sensible Medicine.

The provocative post was called, “Can you trust your doctor?” — and each of the three editors was given 400 words to answer that question. I highly recommend reading the post if you haven’t already.

The common thread was that doctors themselves are altogether trustworthy people, but the flaws of our healthcare system compound in such a way to make the decisions made by individual clinicians generally unreliable.

I agree with much of what they wrote, but I felt that such an important question warranted a longer answer than was allowed by the word limit.

Few would disagree with the assertion that our healthcare system has many problems. These systemic issues have a profoundly negative impact on individual practice.

First, much of the evidence base doctors work from is compromised. Bias permeates the entire industry — from the original research itself to the decisions policymakers base upon it. Unhelpful, low-quality research is abundant. And as a profession, we place far too much confidence in unreliable observational data and expert opinion.

The end result is care that is “evidence-based” but often suboptimal.

It’s not all bad, though. Clinical research has produced some truly successful therapies which have enabled patients with chronic diseases to live longer than ever before. The tradeoff, though, is that caring for these patients has become much more complex and time-intensive.

This should mean that physicians spend more time with patients, but instead, they’re being pushed to see more patients in less time. Simultaneously, the administrative demands of insurance companies, the government, and the electronic medical record have doctors struggling to keep their heads above water.

It’s no wonder, then, that today’s average clinician doesn’t read important new studies, much less learn how to critically evaluate them. Instead, they depend on filtered sources of information (such as electronic point-of-care references, clinical guidelines, and review articles) to guide their decision-making.

Without critical appraisal from the physicians delivering the care, the bias in the literature propagates unchecked. As the use of AI-infused clinical references grows, this problem will only get worse.

Another complicating factor is that medicine is thoroughly random.

In his book, The Black Swan, Nassim Nicholas Taleb says that “in practice, randomness is fundamentally incomplete information.” Medicine is saturated with it. Many of the variables that make a clinical outcome are hidden, and we can never fully understand a patient’s unique biology.

In the same way, we can’t predict how patients will respond to interventions. As I’ve written previously, even if they get better, it’s impossible to know if the improvement came from the treatment. Natural history, placebo effect, and regression to the mean completely obscure our understanding of what happened.

Add these constraints to our own natural fallibility, and it’s easy to see how often we physicians will get things wrong.

In view of this, can a professional who fails as often as we do still be worthy of your trust? Is that even logical?

I think so, and a baseball analogy helps explain why.

To start, I compiled the batting averages for all MLB qualified hitters (batters who received at least 3.1 plate appearances per team game) from the 2025 season — 145 players in total. I then used Claude Cowork to make this KDE plot to illustrate the distribution.

At one end you have Aaron Judge — an outlier in almost every sense of the word — hitting .331, more than three standard deviations above the mean. At the other, you have Oneil Cruz, a classic “three true outcomes” (home run, walk, strikeout) guy hitting just .200.

It’s not quite a normal distribution but it’s close. And you’ll notice that the plot skews a little to the right. That is, a long tail of elite outliers like Judge pull the average up, while most hitters cluster below it.

In 2025, the majority of qualified major league hitters failed to get a hit roughly 72-80% of their at-bats. You might wonder why teams kept giving these guys at-bats when they failed so often.

It’s not that teams expect them to get hits most of the time, but they trust that their net contribution will be positive over the course of a 162-game season.

The same logic can be applied to doctors practicing in a deeply flawed system. The decisions we make from day-to-day will often be wrong or suboptimal, but we still offer positive expected value in many circumstances.

Much of this is obvious and we generally take it for granted in the US.

For example, if your kid has a fever, you can likely count on your pediatrician or family doctor to diagnose and treat them properly. If you’re having a heart attack, you’re almost always better off going to the emergency department than staying at home. Likewise, I think that you can trust that most general surgeons know when and how to safely remove your appendix.

So maybe your doctor isn’t an expert at critical appraisal, but she follows clinical guidelines. A portion of guideline recommendations are based on high-quality randomized trials. Because of that — in the long run, guideline-directed care still probably helps more patients than it harms.

And just like there’s more to baseball than hitting, a doctor can provide value in other ways that aren’t influenced by the system. They might save patients money by avoiding wasteful tests, be skilled with procedures, or simply be a faithful presence in the midst of illness.

The point is that if you’re seeing a board-certified physician for a common condition within their scope, the expected value over time is probably a net positive.

Perhaps that’s not by a big margin. Maybe as a profession we’re only batting something like .250. That would be tough to swallow, but we don’t have to be satisfied with it.

The healthcare system can and should improve. Some of that is out of our hands. We do what we can though at the individual level. Like any MLB hitter who wants to remain on the roster, we should work hard to get better at our jobs.

After all, hitters have it pretty rough as well — they have baseball randomness and an entire defense working together to prevent them from succeeding.

I think we can learn a lot from that. For them, failure is part of the everyday experience. They’re neither surprised nor beaten by it. It only motivates them to spend even more time in the cage.

If we can emulate that, then I think we can make things a little bit better — and perhaps show ourselves worthy of being trusted.

A .250 hitter stays in the lineup — not because he’s perfect, but because he produces in the long run. The same goes for doctors.

If you know a doctor who works hard and still doubts whether they’re enough — send this their way with a like ❤️ or restack 🔁.

Share

The MLB draft is just a few weeks away, and the speculation over who will be the number one pick always dominates the discussion.

Drafting well is essential to a team’s success. The exclusive opportunity to choose any player from the entire pool of talent in the draft is one that teams rightfully covet.

This year, the consensus top prospect is Roch Cholowsky, and most expect the Chicago White Sox to choose the UCLA shortstop with the first overall pick.

The White Sox have undoubtedly dedicated tremendous resources to determine his worthiness. Their best scouts have spent many hours watching him in person and on film. Their analytics department has meticulously crunched the numbers and analyzed all available metrics. What they ultimately want to know is whether he might become a future franchise player in Chicago.

He’s also ranked as the top overall prospect by the leading media gurus — MLB Pipeline, Baseball America, Keith Law, and Kiley McDaniel.

It’s about as confident an evaluation as the sport produces. But confidence in the evaluation isn’t the same as confidence in the outcome. What’s the probability that their highest-confidence pick actually becomes a franchise player?

To answer this question, I went back and looked at the forty six number one picks from the inaugural draft in 1965 through 2012. I set the definition of a “franchise player” as having accumulated at least 40 career wins above replacement (WAR).

I chose 2012 as a cutoff because most of the number one picks prior to that had completed their careers. Additionally, all the active players included have reached the 40 WAR threshold: Carlos Correa (44.5 bWAR), Gerrit Cole (43.8 bWAR), and Bryce Harper (55.1 bWAR).1

From 1965-2012, only ten of the forty six (21.7%) number one draft picks reached the 40 WAR threshold.

Most number one draft picks do not become franchise players. It’s not a failure of the evaluation — all of these guys had the raw skills to make it — it’s that a franchise-level player is rare to begin with.

According to Baseball Reference, there have been a total of 23,670 players who have ever played in Major League Baseball. Only 561 have accumulated at least 40 career WAR. Only 2.37% of the players who have ever reached the big leagues were good enough to be called a franchise player.

Even the most confident evaluation can’t overcome that. When the outcome is rare enough, most positive results — even the most justified ones — will be wrong.

Read more

Now I am going to tell you why that was a bad idea.

It’s not that the writing was bad — it was mediocre, which is a completely different adjective — but an objective decision analysis completed 366 days too late shows that I was clearly out of my mind.

First consider the base rate:

A 2024 study of 75,000 Substack newsletters found that 45% of them were just plain dead. Having a little better than a coin flip’s chance of surviving isn’t exactly a promising start. The experts agree: a highly unscientific Statcast & Stethoscopes poll found that 99 out of 100 doctors believe a 55% survival rate is, in fact, less than ideal.

And what was the prior probability that there would even be an audience for a Substack like this? I’ll be the first to admit that mixing baseball analytics and medicine is at least as intuitive as putting mustard on an Oreo. And yet, according to some deeply concerning corners of the internet, that apparently works too.

So if you’re reading this right now, that means you’re a unique individual — and there’s a pretty good chance that you’d enjoy a thin smear of Grey Poupon on your Oreos.

Another symptom of my insanity was that I thought writing would be easy. I told myself, “Bobby, you’ve written multiple articles for academic journals and at least a handful of well-worded emails. Writing for Substack is not that hard. Tell him, Wash.”

*Sighs* You were right, Ron Washington. You’re always right.

So what if it was crazy to start this newsletter? I’m a year into it now, and there’s no turning back at this point.

(The astute reader might flag this as an example of the sunk cost fallacy — and to you, wise reader, I award fifty Kahneman points!)^⁠1

But if I follow my own advice and take a Bayesian approach, I need to update the priors based on new evidence that came in — most of which I didn’t see coming.

While I’ve always understood that the odds of becoming the next Atul Gawande were infinitesimally small, I didn’t anticipate how writing on Substack would change the way I thought about myself. Despite my jokes a few paragraphs earlier about my professional writing experience, I’ve never felt like a real writer.

But after a year of consistently putting words out there, I think I can call myself an honest-to-goodness writer. Sure, I may have established myself as the oddball doctor writing in baseball analogies, but hey — I’m a writer!

And then there are the unexpected connections. Writing on Substack has allowed me to interact with and learn from some truly talented writers in both medicine and sports. Some have even recommended me to their audiences.

Plus, I think that writing about uncertainty, evidence interpretation, and clinical reasoning has made me a better doctor. One thing about writing is that it forces you to think deeply about your ideas. Doing this regularly has brought clarity to my decision-making, teaching, and communication that I wouldn’t have developed otherwise.

These things are bigger than any objective analysis could have captured a year ago, something I think Bayes himself would have understood.

A less commonly known fact about Thomas Bayes is that he was a Presbyterian minister who believed that the probabilistic manner in which our world seems to operate is a reflection of a divinely ordered world — one that we gradually understand more deeply as new evidence arises^⁠2.

For Bayes, probability wasn’t a sign of chaos. It was a framework for better understanding the big and important things in life.

And so many of those things are improbable.

This past weekend, my family and I were invited to watch a Charlotte Knights game. Charlotte is the AAA affiliate of the Chicago White Sox. It was a hot afternoon for baseball, and eventually my 9-month-old son decided he’d had enough of the heat.

We were there for a corporate event, and fortunately our tickets included access to the suite-level concessions and bathrooms. I thought a walk through the air-conditioned hallway might help him fall asleep so I could go back and watch the game.

There I was, casually strolling the hallway, and looking at old team photos and jerseys on the walls. My son was quiet now (nailed it) and suddenly, standing in front of me, there was Hall of Fame manager Tony La Russa.

I caught his attention. I asked an idiotic question: “Are you Tony La Russa?” (Of course, he was.)

I stammered through an explanation of my being a lifelong Cardinals fan (La Russa managed the Cardinals to World Series wins in 2006 and 2011), and how honored I was to meet him.

Despite having been waylaid by a bumbling stranger with a baby, he was uncommonly gracious. He told me which suite he was watching the game from and invited me to come back later if I wanted him to autograph something.

I couldn’t believe my luck. I rushed back to share my experience with my wife and kids.

I did go back to his suite a bit later — this time with my two oldest kids (ages 12 and 10) in tow.

Reluctantly, I opened the glass door, poked my head through with my hand raised as if to say, “Hey, do you remember me — the big doofus with the baby?”

He remembered me. He welcomed us in and introduced me to his guest, whom I didn’t immediately recognize, but who just so happened to be former Cardinals manager Mike Shildt.

Incredible. Here I am, diehard Cardinals fan, hanging out with two former Cardinals managers in a luxury suite at a minor league baseball game.

We chat for a few minutes. Tony is especially engaging with my kids — particularly in answering my son’s question of how he felt after David Freese’s legendary walk-off home run in Game 6 of the 2011 World Series.

Then we line up for a photo, and all of a sudden, Tony says, “Wait a second, if we’re gonna do this, we’ve gotta do it right.”

He turns to my 12-year-old son, pulls off his 2006 World Series ring, and tells him to put it on.

I’m absolutely floored. My son’s grinning with excitement, but he cannot fully grasp the rarity of the moment.

We take our photos, chat a bit more, and not wanting to overstay our welcome, we politely excuse ourselves.

An event like this could never be captured by a model. It was far too improbable — the big moments are often like that.

It’s being down to your last strike (twice) and coming back to win one of the greatest World Series games in history while a hometown kid transforms himself into a franchise legend.

It’s the unexpected connections made from simple choices, like starting a newsletter or taking a walk to soothe a crying baby.

It’s taking your family to a baseball game — and walking away with an experience that will forever deepen the bond between father and son.

Perhaps I should have done a bit more analysis before clicking publish on that first article. But I’m glad I didn’t — now that I know what I would have missed.

Sometimes, it’s better just to see what happens. We live in a probabilistic world, where the seemingly improbable happens all the time.

We can’t predict when those moments will happen. And they can only be experienced if we choose to show up.

Which is, of course, a very Bayesian way of thinking.

¹ Kahneman points can be redeemed for merchandise in the official Statcast & Stethoscopes store just as soon as it exists.

² See The Signal and the Noise by Nate Silver, Chapter 8.

“We just felt at this point something needed to be done.”

This was how Phillies President of Baseball Operations Dave Dombrowski justified his decision to fire manager Rob Thomson just twenty eight games into the season.

Thomson got the job in 2022 after his predecessor Joe Girardi was also fired midseason by Dombrowski. Since then, Thomson had led the Phillies to the postseason every year, including a World Series appearance in his first season.

They’ve won the highly competitive National League East division in each of the last two seasons — and they were the favorite to do it again in 2026.

But then they lost ten games in a row. And it was apparently enough to override everything that Thomson had done previously.

“I think it was just that we need a different voice in there,” Dombrowski said. “A few years ago… he was the right voice for us… I think we needed a different voice with this group where we are right now.”

There’s little evidence to support the idea that firing Thomson would truly change the Phillies’ trajectory this year. An analysis by FanGraphs’ Dan Szymborski suggests that midseason managerial firings have historically had little effect.

A front-office executive in today’s analytic era should know not to over-interpret small sample sizes. Twenty-eight games is a mere 17% of a full season — and a 9–19 stretch can be highly influenced by factors outside the manager’s control.

Managers have limited impact on outcomes — they obviously can’t hit, field, or pitch for the players. The randomness of batted ball luck, hit sequencing, and injuries can easily transform an otherwise slow start into what looks like a disaster to a demanding fanbase.

For a preseason division favorite, a 9–19 start is almost certainly an outlier prone to regression to the mean — the statistical phenomenon where extreme values will tend to drift back toward average regardless of intervention.

At least from the outside, there was very little evidence that a “different voice” was the missing piece. In fact, the Phillies’ clubhouse reacted with disbelief and guilt, believing Thomson to be a scapegoat taking the fall for their collective failure.

Maybe Dombrowski knew something outsiders could not see. Maybe the clubhouse really did need a different voice. But the public explanation points to a familiar decision-making trap: when the situation is deteriorating and the causes are hard to isolate, leaders feel pressure to make their concern visible.

He felt he needed to make a visible move, and changing the manager was one that everyone could see.

That impulse to act in a visible way under uncertainty — even when waiting is at least as well-supported by the evidence — is called action bias. Like other heuristics, it exists for a reason. Often, doing something really is better, so our brains naturally reach for action when we’re unsure. Under pressure, our brains may substitute an easier question.

Instead of asking, “What action is most likely to improve the outcome?” we ask, “What action will prove that I am taking this seriously?”

As of May 11, the Phillies have gone 10-3 since Thomson was replaced by bench coach Don Mattingly. Some in the media are already crediting the managerial change for providing the “spark” the Phillies needed.

That’s possible. But it is exactly the kind of story our brains are wired to believe. When a good team plays poorly for a month and then starts winning, the new manager benefits from catching regression to the mean on the upswing.

We’ll never know what would have happened if Thomson hadn’t been fired. That’s what makes action bias so subconsciously attractive — often the decision cannot be honestly graded later.

Action bias appears wherever uncertain decisions must be made under pressure, meaning medicine is a natural home for it.

It’s the inverse of a bias I wrote about previously — omission bias, the tendency to view inaction as safer than action even when it isn’t.

Both biases describe the same pattern but in opposite directions.

Consider the following case:

Mary is a 68-year-old woman with type 2 diabetes. She’s done well on metformin for 12 years. Six months ago, her A1c was 7.1%, but today it came back at 8.2%.

For reasons that don’t really matter to the decision in front of you, neither a GLP-1 receptor agonist nor an SGLT2 inhibitor is on the table. The allotted appointment time is almost over, and you have more patients waiting. What will you do?

Will you add pioglitazone, or maybe sitagliptin?

Basal insulin?

Glipizide?

The impulse is to prescribe. The A1c is out of range, and we feel compelled to correct it. Adding a medication feels satisfying.

And there’s real pressure at play.

Your clinic is slammed. You’re running behind.

If you prescribe, and Mary does well, you’ve helped. If you don’t, and she gets worse, you feel you’ve missed something.

But there may be legitimate reasons to stay the course.

Perhaps her orthopedist gave her a steroid injection 6 weeks ago. Perhaps she is under a lot of stress caring for a declining parent. Perhaps she ran out of her metformin for a month.

Or maybe there’s no concrete reason — and she just had a few months where she let her diet get away from her.

And what’s the potential cost? The medications available to Mary are not interchangeable substitutes for the therapies we would most prefer. They lack the robust outcome benefits that elevate GLP-1 RAs and SGLT2 inhibitors in many patients.

They may also introduce more harm. Sulfonylureas and insulin may cause life-threatening hypoglycemia. Insulin often causes weight gain. Pioglitazone can worsen fluid retention and heart failure risk. DPP-4 inhibitors have also had mixed safety signals around heart failure, though the concern is less clear.

Prescribing them isn’t clearly wrong. There are scenarios where adding them might make a lot of sense. But once started, these meds rarely come off the patient’s list, so the decision should always be made after thoughtful deliberation.

The alternative to prescribing is not passivity. It is diagnosing the reason for the change, deciding how much uncertainty is acceptable, and making a plan that can be revisited.

So instead of allowing our subconscious brain to take the easy route, here are three practical ways to mitigate action bias:

• Reframe the choice. Choosing to defer a medication is also a real decision. Treating the visible action like it’s the only “active” choice is how our brains deceive us in the moment.

• Bring the patient in. “Mary, your A1c went up to 8.2% this time — what do you think happened?” Patients know things the lab test doesn’t.

• Make follow-up a real plan. It’s not: “Mary, we’ll just see how things go in 3 months.” It’s: “Mary, this may have just been an outlier. Let’s get back to work with your diet and exercise and we’ll recheck your A1c in 3 months. If we’re still out of range then, we might talk about another medication.”

Yielding to action bias probably happens more than we’d like to admit, and it should humble us because we share the same vulnerabilities as our patients.

The patients who demand antibiotics after two days of respiratory symptoms may be feeling the same pressure: work piling up, children to care for, and no way to know when — or if — they will feel better.

Consider, also, family members pushing for aggressive treatment at the end of life. They may be facing the unbearable possibility of life without someone they love and are desperate for one more visible thing to try.

Whether we’re a patient, a physician, or an MLB executive — ultimately, we’re more alike than we are different.

We’re all trying to answer uncertain questions under pressure. The skill is noticing which question we’re actually answering.

In this episode, I explore why the Phillies fired their winning manager twenty-eight games into the season, what a 68-year-old with a rising A1c reveals about the medications we add too quic…

Read more

In this episode, I walk through the full trial record on left atrial appendage closure devices: what PROTECT AF, PREVAIL, PRAGUE-17, and CLOSURE-AF actually showed heading into CHAMPION-AF, why CH…

Read more

Being a doctor is harder than it looks.

Not technically hard—though it is that too. But hard in a subtler way.

Every day I make decisions with incomplete information, knowing that the outcome isn’t entirely in my hands.

Two careful physicians can look at the same patient and reach different conclusions, and neither of them is wrong. The right call today can look like the wrong one in hindsight, not because anyone made a mistake, but because outcomes are partly based on luck.

The culture doesn’t make this easy to admit. Medicine trains you to project confidence. Doubt is something to overcome, not navigate. So most of us carry the uncertainty quietly—which feels lonely, and a little dishonest.

It’s not a great model to put in front of the residents I train every day.

I’ve loved baseball my whole life.

I really mean that—there’s a photo of me as a toddler wearing a Cardinals hat and proudly holding a sign that said, “Cubbies stinky.”

But what pulled me deeper as an adult was the analytics revolution. Sabermetrics showed that conventional wisdom about baseball was often just wrong, and that open-minded thinkers willing to look more carefully could find value everyone else had missed.

Then came the player development revolution, which went further: you didn’t just find hidden value, you could create it. Mechanics got rebuilt. Careers got rescued. Players became something they hadn’t been before.

The idea that careful thinking could overturn assumptions and unlock hidden potential has never gotten old for me.

The connection came to me on a ruck through my neighborhood.

I knew I wanted to write, but I couldn’t figure out what about.

I’d been listening to a podcast where Kevin Kelly offered this advice: don’t be the best, be the only. Figure out what combination of skills and interests is uniquely yours, and build something nobody else could.

I’m a family physician. I teach residents and students. I’ve spent years immersed in baseball analytics. I put those three things together and thought, this is it. Nobody else was writing about this—using the ideas and concepts of baseball analytics to think more carefully about uncertainty in medicine.

That’s how Statcast & Stethoscopes came to be.

What this newsletter is

I’m Bobby Scott—an academic family physician training the next generation of physicians, which means I spend a lot of my time watching smart people struggle with something medicine doesn't teach well: how to reason under uncertainty.

Each post takes a concept from baseball analytics and applies it to a real problem in clinical reasoning—the kind of thinking medical training never quite got around to teaching.

How do you make good decisions when outcomes are uncertain or evidence is incomplete? How do you separate signal from noise in a small sample?

Baseball spent decades learning answers to these questions. I’m bringing those answers to medicine.

I write it because I needed those answers. I think a lot of physicians do.

What to expect

A new post lands in your inbox roughly every ten days — sometimes more, sometimes less, depending on what’s happening in my clinical life. Most of what I write is free.

Members get the deeper work: close reads of landmark studies, applied decision science, step-by-step breakdowns of real clinical decisions, and audio versions of every post.

About Membership

If you’re a clinician who has felt the quiet discomfort of uncertainty and wanted a better way to think through it—this is for you. If you just love baseball and good reasoning, you’re also welcome here. Below are a few of my favorite posts to get you started.

Either way, I’m glad you found this.

— Bobby

Four years later, they released him, preferring to eat the $20 million still owed to him rather than keep him on the roster for one more year.

The Phillies signed him hoping that he would be a key piece in a championship team. They knew that they were getting a well-rounded hitter with power to all fields. His batting metrics were among the league’s best in 2021—a career-best season where he produced a 140 wRC+

They also knew that he couldn’t field his position. His Outs Above Average (OAA) has been negative throughout his career, the lowest being -23 in 2018—the worst in Major League Baseball. Since 2016, Statcast estimates that Castellanos’ defense has cost his teams a cumulative 39 runs.

For a player like that, the math only works one way. His bat has to be good enough that he’s still better than the alternative. The standard for a player this bad in the field isn’t “above average.” It’s “so good offensively that nobody cares what he does in the field” (See: Soto, Juan).

But he didn’t maintain his offensive production. In his first season in Philly, Castellanos hit just below league average, and as a result, generated -0.2 wins above replacement (WAR). His best season in Philadelphia came in 2023 when he hit slightly above average yet still only produced 1.3 WAR.

In baseball, WAR measures the value a player adds above that of a readily available replacement player. A roughly league-average bat was not enough to offset his defensive liabilities. Thus, Castellanos quickly became a replacement-level player1.

If you’re new to these posts, Meds Above Replacement is my adaptation of baseball’s Wins Above Replacement (WAR) statistic to compare the marginal value of medical interventions.

Here’s how it works:

This is the first installment that evaluates a medical device—specifically the left atrial appendage closure (LAAC) device. The most commonly implanted LAAC device is the WATCHMAN, produced by Boston Scientific.

Atrial fibrillation (AF) affects nearly 12 million people in the US. Anticoagulation for stroke prevention is the standard of care for AF patients with elevated stroke risk.

The original niche for LAAC was AF patients who had an indication for anticoagulation but couldn’t tolerate it long-term. A new RCT, CHAMPION-AF seeks to evaluate LAAC as a first-line alternative to direct oral anticoagulants (DOACs). It’s an ambitious claim that Boston Scientific is eager to prove.

I have no ties to Boston Scientific. Likewise, I’m not a cardiologist, so I won’t be putting this device in anyone.

I am a family physician, and my patients will see the WATCHMAN commercials on TV. They’ll come to me asking whether they should get it so they can stop taking their apixaban. I’ll have to decide whether to refer my patients to the interventional cardiologist.

That decision isn’t neutral. Just a few clicks in the EMR will send my patient into a system with strong incentive pressures toward implantation.

It turns out to be the highest leverage moment in the whole process.

The Trials So Far

Two RCTs comparing LAAC to warfarin served as the basis for LAAC’s approval by the FDA in 2015.

PROTECT AF (2014) met noninferiority to warfarin for a composite of stroke, systemic embolism, and cardiovascular death. However, it raised serious procedural safety concerns—namely pericardial effusions requiring drainage and procedure-related stroke.

PREVAIL (2014) did a bit better with safety, cutting the 7-day complication rate to 4.5% compared to 8.7% in PROTECT AF. However, in this smaller trial (n=407), LAAC failed to meet noninferiority to warfarin in its primary efficacy endpoint (composite of stroke, systemic embolism and CV/unexplained death). It did meet noninferiority for its secondary endpoint of ischemic events after 7 days.

PRAGUE-17 (2020) was the first to compare LAAC to DOACs, and studied it in a population at high risk for both stroke and bleeding (mean CHA₂DS₂-VASc 4.7, mean HAS-BLED 3.0).

LAAC met noninferiority for a broad primary composite outcome that now included clinically significant bleeding. Event rates were low, and as a result, confidence intervals were wide. The inclusion of bleeding in the composite allowed LAAC to demonstrate a potential net benefit over time, but it obscured the original question of its efficacy in preventing strokes.

CLOSURE-AF (2026) enrolled a similar population as PRAGUE-17 (mean CHA₂DS₂-VASc 5.2, mean HAS-BLED 3.0), and this time, LAAC failed to meet noninferiority to DOACs.

CLOSURE-AF observed over three times as many primary outcome events (similar composite as PRAGUE) which was reflected in a much tighter confidence interval. It offers stronger evidence that LAAC may not be a viable alternative in higher risk patients—those who theoretically should have the strongest rationale for getting the device.

In sum, the prior evidence heading into CHAMPION-AF ranges from inconclusive to actively discouraging.

CHAMPION-AF

Funded by Boston Scientific, CHAMPION-AF randomized 3000 patients to either the WATCHMAN FLX or DOAC. The population had a mean CHA₂DS₂-VASc 3.5 and a mean HAS-BLED of only 1.3—a lower risk population that were more likely to do well on an oral anticoagulant.

They prespecified the following outcomes:

• Primary efficacy outcome: composite of CV death, stroke, or systemic embolism at 3 years.

  • A separate composite of ischemic stroke or systemic embolism will be evaluated at 5 years follow-up.

• Primary safety outcome: Non-procedural bleeding

• Secondary outcomes:

  • Composite of procedure-related bleeding and non-procedure-related major bleeding

  • Net clinical benefit: Composite of CV death, stroke, systemic embolism, or non-procedure-related bleeding)

To understand the results of this trial, one must first understand what a noninferiority trial tells us.

It does not tell us if a treatment is better, or even as good, as the comparator. Noninferior means that the potential treatment is not unacceptably worse than the standard.

The idea is that a new treatment might have some other advantage that makes it a reasonable trade-off despite lower efficacy. In the case of LAAC, a patient might accept a little less efficacy in stroke prevention for less bleeding or the chance to stop a lifelong medication.

The question becomes how much worse can LAAC be and still be acceptable in light of those potential advantages?

This is defined by the margin. This is the maximum allowable risk difference for the treatment to be considered noninferior.

The margin can be set in terms of absolute risk difference, relative risk difference, or both. As long as the upper band of the confidence interval falls within that margin, noninferiority is met.

In CHAMPION-AF, the margin was set at 4.8 absolute percentage points after estimating that a primary outcome event would occur in 12% of patients. This corresponds with a 40% relative increase of the events at 3 years.

So what were the results?

There were 65 primary outcome events in the DOAC group versus 81 in the LAAC group. The point estimate of the absolute risk difference was a 0.9% increase in events. The 95% CI ranged from -0.8% to 2.6%.

Since 2.6% (the upper bound of the CI) is less than the set margin of 4.8%, LAAC met statistical noninferiority.

The primary safety outcome (non-procedural bleeding) was assessed on superiority rather than noninferiority. LAAC was found to be superior by 8.1% (95% CI -10.8 to -5.5, p<0.001).

So the headline becomes that WATCHMAN FLX is noninferior to DOAC in terms of efficacy and causes less bleeding. Boston Scientific has already announced that they will pursue an expanded indication for the WATCHMAN based on the result.

Read more

For the first time in history, there’s a higher authority than the home plate umpire in deciding balls or strikes. Players may now appeal to a computerized “robo-ump” that employs advanced motion-tracking technology for a final judgment.

Teams are given only two challenges per game, but they retain the challenge if the call is overturned. Only pitchers, catchers, and hitters are allowed to initiate a challenge, and baseball teams are trying to figure out when challenges should occur and by whom^⁠.

And because we are nerds, there’s already a Statcast leaderboard to measure which players are the best at it.

Atlanta Braves pitcher, Chris Sale, made up his mind in Spring Training to never challenge a pitch^⁠⁠1. Here, he explains why:

He knows that his bias—all borderline pitches look like strikes to him—makes him an unreliable judge.

It’s a bias rooted in egocentric anchoring—the human tendency to start from our own perspective as the default reference point and adjust too little from there.

In the heat of competition, Sale knows that he can’t be objective. From his place on the mound, a pitch that looks and feels like a strike will never be anything but a strike to him. When coupled with a self-serving bias (what he calls being “greedy”), his read of the pitch is less reliable than that of the umpire or the catcher.

Sale recognizes his egocentric anchoring and doesn’t trust it. Physicians deal with the same bias, but we trust it far too much.

A basic component of a clinical encounter is the information transfer between the physician and patient. Sick patients come to us because they need a diagnosis, prognosis, and a treatment plan.

It follows that effective communication of this information is essential. Unfortunately, we often struggle to do this well.

Estimates suggest that patients forget or misunderstand between 40 to 80 percent of the information received during a clinical encounter.

Part of that results from patient factors such as low health literacy or language barriers. And on the physician side, we tend to use medical jargon far too freely.

But even if we attempt to mitigate these, a gap between what we think the patient understands and what they actually understand, often remains. You might think your patient’s getting it, but this confidence is often misplaced. Research suggests that those who rate themselves highly as an educator are likely to have more patients with comprehension gaps.

We miss the mark because we prematurely conclude that adequate information transfer has occurred. Egocentric anchoring leads us to feel confident that the patient understands even when they don’t. This is Sale on the mound—from our viewpoint, the explanation felt like a strike.

A 2004 paper by Nicholas Epley, et al.^⁠ introduced egocentric anchoring as an expansion of Kahneman and Tversky’s anchoring heuristic from the 1970s. The paper proposed that when we try to understand another person’s perspective, we tend to start from our own perspective as a reference point (the anchor).

From there, we try to mentally shift toward their perspective, but this adjustment is almost always insufficient. We typically settle in a place that’s closer to our own perspective than that of the other person.

The most vivid illustration comes from the work of Elizabeth Newton, who at the time was a PhD student in the psychology department at Stanford. In her dissertation, she asked participants to finger-tap popular songs for uninformed listeners. Beforehand, they were asked to estimate how often listeners would recognize the song they tapped.

The tappers were drastically overconfident. They predicted a correct guess rate of 50%, but it was actually only 3%.

The tune was so vivid in their minds that they believed that everyone else would easily hear it too.

We make the same mistake with patients. What we say is so clear to us from our own vantage point that we presume they are following along just as easily. Then, as soon as we reach a point where it seems that the patient should understand, we stop. Typically, that’s too soon.

Making matters worse, Epley’s paper found that this problem compounds under time pressure. It’s yet another way that patients suffer from the rushed time constraints of modern medicine.

One might be tempted to think that if we try harder to understand our patients—“putting ourselves in their shoes”—we might adjust more accurately. Although this seems intuitive, it doesn’t really work.

In a follow-up paper in 2018, Epley demonstrated across 25 different experiments that efforts at perspective-taking don’t move the needle. In many cases, it made things worse.

What does seem to work is perspective-getting: asking for the other person’s perspective instead of guessing.

In the 25^th experiment, Epley found that people spending five minutes asking their romantic partner about their opinions were far more accurate in predicting their responses on an opinion survey compared to trying to imagine the world through their partner’s eyes.

This seems like a no-brainer. The best way to know what another person thinks is to ask them.

The “teach-back” method is a perspective-getting intervention that has been shown to improve patient comprehension of health-related information^⁠2,3. It’s asking patients to repeat back what they understood in their own words. It’s quick, it’s easy, and it works.

I’ll confess that I have known about teach-back for a long time, but I’ve used it far less than I should. I’m still growing as a doctor.

I’m not alone, though. Data suggests that teach-back is highly underutilized. Common barriers include lack of time, insufficient training, and the fear that patients might find it condescending—although evidence suggests that patients greatly prefer it.

I suggest that the lack of uptake is because we have a major blind spot as physicians.

We’re so confident in our ability to assess our patients’ understanding that we think we’re doing well enough without it. In Epley’s experiment, the participants who imagined their partner’s perspective were just as confident as those who actually asked.

Confidence is high, but comprehension is often low.

I didn’t expect a pitcher’s throwaway remark in spring training to reveal such a glaring blind spot in my practice. As a result, I’m making a conscious effort to use teach-back more regularly. Here’s an example of how I’m phrasing it:

“Just to be sure I explained things clearly, can you tell me in your own words what the plan is?”

It takes hardly any time at all, and what I get back is real, objective data. Teach-back is the ABS challenge system, but I get to use it as many times as I like.

It’s fast, it’s easy, and it shows me whether I’m throwing balls or strikes.

Teach-back takes hardly any time and returns objective data. It’s an ABS challenge you get to use every time.

If this piece gave you something you’ll actually use, a like ❤️ or restack 🔁 helps it reach other physicians who are ready for it too.

Leave a comment

Share

Subscribe now

¹ Through two starts in 2026, Sale has yet to initiate an ABS challenge.

² Slater BA, Huang Y, Dalawari P. The Impact of Teach-Back Method on Retention of Key Domains of Emergency Department Discharge Instructions. J Emerg Med. 2017;53(5):e59-e65. doi:10.1016/j.jemermed.2017.06.032

³ Parikh P, Jaisinghani P, Kaloth S, Konakanchi A, Yanamala N, Kim S. Enhancing Patient Understanding of Hospitalization and Post-Discharge Needs: The Impact of Physician-Led Verbal Communication and Teach-Back Method. J Gen Intern Med. 2025;40(9):2103-2110. doi:10.1007/s11606-025-09510-w

In this post, I explore egocentric anchoring as a cognitive bias in clinical communication, what Chris Sale's spring training decision reveals about physician self-awareness, and why perspective-taking — try…

Read more

In this episode, I explore why the scouts in Moneyball made the same epistemic error that physicians make every day, what vertebroplast…

Read more

Periodically though, something unexpected happens that forces the game in a new direction.

Perhaps that’s why Moneyball is one of those books I just can’t leave behind—not just because change happened, but because of what had to be overcome to make it happen.

In the early 2000s, A’s general manager Billy Beane challenged a century’s worth of baseball convention. The opposition came chiefly from his scouts—lifelong baseball men who believed that what they saw with their eyes was all they needed to know about a player.

They thought that a baseball player should do things a certain way and look a certain way doing it. This ideology was so deeply engrained in baseball that no one was willing to consider anything different—except for Billy Beane, who used it to his advantage.

As author Michael Lewis describes it:

“The old scouts are like a Greek chorus; it is their job to underscore the eternal themes of baseball. The eternal themes are precisely what Billy Beane wants to exploit for profit—by ignoring them.”^⁠1

It’s not that the scouts were dumb, they were reasoning based on decades of personal experience and wisdom held by generations of knowledgeable baseball insiders. Their problem wasn’t their experience. It was what they thought they learned from it.

It’s the same risk we face as physicians.

Clinical Experience and Its Limits

We’ve already seen how we cannot know ahead of time who will benefit from treatment, nor can we confirm, afterward, whether the treatment caused the improvement we observed.

What we haven’t fully reckoned with is how observing a response to treatment affects what we believe about it. If we aren’t careful, these beliefs may find us confident in a therapy even when the evidence doesn’t support it. Epistemic humility—a recognition of the limitations of our knowledge—is an essential safeguard, because we are so easily led astray by our clinical experience.

We like to think that we’re always Billy Beane, but often we’re the scouts.

It’s not that having more experience is a bad thing—but it’s important to understand its strengths and limitations. Experience genuinely teaches us pattern recognition, diagnostic intuition, rapport-building, recognition of nonverbal cues, and how to navigate clinical uncertainty.

But the critical flaw of learning by experience is that there’s no control arm. No matter how many times you’ve seen a treatment work, you cannot isolate treatment effect from statistical randomness.

A randomized controlled trial (RCT) can. A well-designed RCT largely eliminates statistical variance and provides real, objective evidence that a treatment does actually cause the outcome.

But clinical experience is a persuasive teacher, and it’s tempting to equate it with actual evidence. If we build convictions about the effects of therapy on a foundation of experience alone, we make the same categorical error the scouts made in Moneyball—believing that our eyes tell us all that we need to know.

A History of Wrong Conclusions

Bloodletting was standard medical practice for thousands of years. In the 20^th century, surgeons intentionally collapsed lungs to treat tuberculosis, obstetricians performed routine episiotomies, and cardiologists used to recommend antiarrhythmics after heart attacks.

We know now that all of these practices were wrong. But in their day, they were the conventional standards of care—good medicine endorsed by some of the best clinicians of their generation. Their error wasn’t in character or skill, but was in what they believed their experience taught them.

Experience isn’t reliable evidence, but the error of building practice standards upon it continues today. One physician reasoning this way is a problem. But an entire profession reasoning like this has the potential for serious harm.

The leap to declare standards before they are proven is not a result of ignorance. And it’s not as simple as physicians failing to learn from the past.

It’s that seeing is believing.

When we see a treatment work, it feels just like knowledge.

The Pattern Continues

The conditions that produced the errors of our past haven’t fully left us.

Clinical experience generates false confidence. Larger observational studies amplify the signal. At this critical juncture, we may or may not get high quality RCTs, but often the practice has already become widespread.

Consider vertebroplasty—a well-known example highlighted in Ending Medical Reversal by and . Hundreds of thousands of them had been performed on patients before a pair of sham-controlled RCTs in 2009 showed no benefit.

Colonoscopy was (and still is) the standard for colon cancer screening for nearly three decades before the first randomized controlled trial comparing it to no screening was finally published in 2022. We don’t have a definitive answer yet—additional RCTs are pending—but these results were underwhelming, to say the least.

Not every version of this error is traceable to a missing RCT. A subtler, but more common form is confirmation bias: when false confidence in clinical experience takes root, we elevate weak evidence so long as it confirms our deeply held beliefs.

Case in point: nearly half of the almost seventy Class 1 recommendations—the things that should be performed broadly— in the new ACC/AHA dyslipidemia guidelines are based on weak (level B-NR) evidence.

The implications for practice have been detailed far more thoroughly than I can do here— and have written excellent commentaries on this problem, and they’re worth reading carefully.

Sometimes the conclusions we draw based on experience are later confirmed high-quality RCTs. Yet a correct conclusion reached by faulty reasoning is still faulty reasoning.

But where it gets messy is when RCTs present us with disconfirming evidence. The powerful influence of our observed experience—what we’ve seen work in real, live patients—makes us resistant to change.

Abandonment of reversed practices is often delayed for years, and patients are harmed at scale. Vertebroplasties are still being performed today, and a similar treatment (kyphoplasty) has yet to be tested by a sham-controlled RCT.

What Epistemic Humility Looks Like

The conditions for large-scale medical reversal are structural, and they won’t be fixed by policy alone.

Our part requires a change in how we reason.

What we need is epistemic humility—that we hold our experiential convictions loosely enough that we’re willing to let go of them when contradictory evidence develops.

None of this is an argument for withholding treatment until every intervention has survived a randomized trial. Although that would happen in a perfect world, some problems are too urgent, too rare, or too logistically complex (consider the HIV epidemic in the 1980s) to make that standard realistic. I’m also not advocating for clinical paralysis or suspending clinical judgment altogether.

What I am suggesting is this: remind ourselves constantly of our biases and maintain a healthy skepticism of interventions that are based purely on observational data or expert opinion.

A test question to ask yourself is, “If a treatment I believe in is later shown to be no better than placebo, would I receive that as helpful information or as a threat?”

The Stakes

Epistemic humility is not an easy posture to take. It requires restraint in a profession eager to help suffering patients as soon as possible with the latest and greatest.

It means that all of us practicing clinicians must develop skills in critical appraisal.

Guideline-makers are just as biased as we are. They need front-line physicians to help keep them in check. We can’t do that if we don’t understand the literature.

By playing our part, we hold our profession to a higher standard which protects patients by avoiding unnecessary large-scale reversals of practice.

Medicine doesn’t need more “Greek chorus” members underscoring the eternal themes of clinical experience.

We need more Billy Beanes.

The scouts weren’t dumb. They just confused a century of watching for a century of knowing.

If this piece made you think differently about what clinical experience can and cannot teach us, a like ❤️ or restack 🔁 helps this reach others wrestling with the same questions.

Leave a comment

Share

Subscribe now

¹ The Greek chorus is an apt analogy. The first baseball game played under modern rules was played on June 19, 1846 at a public park in Hoboken, NJ named the Elysian Fields^⁠—after the eternal paradise reserved as a final resting place for the heroes of Greek Mythology.

In the US, the MMR vaccine is routinely given in two doses at 12 months (MMR-1) and between 4 to 6 years of age (MMR-2). The state’s guidance now recommends that physicians consider giving an additional early dose (MMR-0) between 6-11 months to infants who live in—or are traveling to—counties with community transmission.

My county made the list.

I routinely care for infants in my family medicine practice. I also have a 6-month old at home, so understanding the data became a priority to me as both a physician and a parent.

And as a parent who planned to vaccinate anyway, following the guidance seems reasonable enough. If there have been cases in the area, why not go ahead and get the vaccine a few months early?

But like most decisions in medicine, it’s not as straightforward as it seems.

This post is a snapshot of how I am actively reasoning through this decision for both my son and my patients—an “inside baseball” look, if you will.

Here’s the decision in a nutshell

Measles cases in North Carolina have been rising—23 confirmed new cases in the last three months, with a larger outbreak in neighboring South Carolina counties. More cases are expected.

It’s a dangerous disease: one in five children are hospitalized, and one to three per thousand cases are fatal. Severe neurologic complications, such as encephalitis or acute disseminating encephalomyelitis, occur in 0.1% of cases. A lesser-known complication is subacute sclerosing panencephalitis—a rare (approximately 4-11 cases per 100,000), but debilitating condition that presents 8 to 11 years after infection and causes progressive neurocognitive deterioration.

Infants younger than 12 months are at higher risk for these complications. Waiting to vaccinate carries real risk.

Simultaneously, there are risks to early vaccination.

First, it may not be highly protective. Circulating maternal antibodies may prevent the vaccine from stimulating an immune response.

Second, there is evidence, albeit limited, that infants who receive early vaccination may demonstrate a precipitous drop in antibodies over time—potentially resulting in inadequate protection.

There are legitimate risks on both sides—risks across different time horizons and involving different types of uncertainty.

It’s actually the kind of decision that baseball organizations make frequently.

A question of timing and development

Deciding when to promote a young prospect to the big leagues is exceedingly difficult.

The crux of the issue is developmental risk. Bring them up too early and you risk them struggling at a formative stage. Some young players handle this just fine, while others never recover.

The problem is you don’t know which type of player you have in advance.

Jackson Holliday, the top overall draft pick in 2022 and the son of a seven-time All-Star Matt Holliday, seemed to be a “can’t miss” prospect.

Having successfully fast-tracked top prospects Adley Rutschman and Gunnar Henderson in recent years, the Baltimore Orioles attempted to replicate that success with Holliday. The 20-year-old was promoted on April 10, 2024 with only 18 Triple-A games under his belt.

In his first ten games, Holliday struggled mightily. He went 2-for-34 and struck out 18 times.

His first Major League stint lasted only sixteen days.

He got a second chance in July, but he ended his rookie season with a dismal slash line of .189/.255/.311. Later, Baltimore general manager Mike Elias expressed regret over his decision to fast-track Holliday:

“I think some of our call-ups that (sic) didn’t go smoothly, Jackson being one of them, we didn’t give them enough time in Triple-A.”

Holliday showed improvement in 2025, but his below-average production suggests the early promotion may have left a mark—whether he reaches his ceiling remains genuinely uncertain.

The decision Elias faced had the same structure as the one facing physicians and parents right now. But as we’ll see, the feedback is harder to come by.

Short-term risks of waiting to promote:

• The Orioles, having won 101 games in 2023, appeared to be in the midst of a competitive window.

• Fast-tracking Rutschman and Henderson worked out well. Repeating that success with Holliday would solidify the Orioles as a World Series contender.

• They could miss this rare opportunity by being too cautious with their top prospect.

Long-term risks of early promotion:

• Exposing an immature player to failure at a formative moment may negatively impact his future trajectory.

• They had little data to go on—just a small sample of success at lower levels of competition.

• A repeat of the team’s success from 2023 was not a foregone conclusion. Injuries, poor performance, and bad luck could keep them out of playoff contention regardless.

Short-term risks of waiting to vaccinate:

• Measles in a 6-month-old is genuinely dangerous, with higher rates of severe complications than in older children.

• Cases are already confirmed in the area—exposure risk is real and growing. However, exposure before 12 months is still unlikely at current rates.

• We could miss the opportunity to prevent severe illness by being too cautious.

Long-term risks of early vaccination:

• Early vaccination may blunt the immune response to later doses, potentially leaving children with diminished protection years down the road.

• The evidence supporting this concern comes from a small cohort on a different vaccination schedule than the one used in the US.

So what does the evidence actually say about these risks?

Read more

In this episode, I walk through what thre…

Read more

In this episode, I explore why natural history, placebo effect, and regression to the mean are always competing with yo…

Read more

Now let’s say we prescribe a treatment, and the patient gets better. Surely that proves that the treatment worked, right? There are problems with that inference, and understanding why starts with a patient I saw recently.

Those who fail to learn from natural history are doomed to repeat it

One of my older patients presented to my clinic with two weeks of dizziness. His presentation was classic for benign paroxysmal positional vertigo (BPPV)—a common condition caused by the dislodgment of inner-ear crystals known as canaliths, which disrupts the brain’s mechanism for sensing head position.

The most effective methods for treating BPPV are canalith repositioning maneuvers, the most well known being the Epley maneuver. Most patients can perform this at home and successfully treat their vertigo, but this patient’s advanced neck arthritis kept him from doing so safely on his own. I referred him for vestibular therapy where these maneuvers could be performed under the guidance of a trained physical therapist.

As it turned out, his appointment was delayed for a month. By that point, his symptoms had fully resolved without any additional treatment. This is typical for BPPV, as most episodes self-resolve within days to weeks. It’s the natural history as it occurs in pure form.

It’s a straightforward case in and of itself, but the counterfactual is an illuminating thought experiment.

What if I had prescribed meclizine at the first visit? Though it is commonly prescribed to treat BPPV, there is little evidence to support its use. Again he improves before his PT visit, but this time he credits the medication. I see him in follow-up, and now he’s pleased with me because the medication I prescribed “fixed” him. In fact, he’s told all his friends about me and my magic meclizine.

He was going to get better anyway—in this scenario, neither of us knows that—yet we come away with the false conclusion that meclizine cured his vertigo.

Certain conditions follow a predictable course. Mechanical low back pain will usually resolve within six weeks. Viral respiratory infections typically peak at days five to seven and resolve by day ten. A physician who intervenes in these situations—as we are prone to do—will almost always appear to succeed no matter what they do.

It may be placebo effect, but it’s an effect nonetheless

Placebo effect is a phenomenon in which patients demonstrate a real improvement despite receiving an inactive treatment or a sham procedure. Many randomized trials compare an intervention to a placebo, and remarkably, a portion of both groups will have the same outcome.

Placebo effect is truly fascinating when you dig into it. We don’t know why it happens, but something about the experience of treatment triggers measurable biological changes. We used to think that some level of deception was essential, that participants must believe they are receiving the actual treatment to get the effect. That isn’t the case.

One of my favorite studies is this RCT from 2019. It found that patients with chronic back pain who took placebo for three weeks had reduced pain and disability compared to usual care. But there’s a catch—participants were told ahead of time that they would be receiving placebo. It didn’t matter that they knew it was placebo. Incredibly, the effect still held.

You might be asking, “What does this have to do with me? I’m not going to be prescibing placebo.” Well, maybe not intentionally—but placebo effect is embedded in every treatment we prescribe. The ritual of treatment has an impact whether we realize it or not. We rarely pause to acknowledge that in blinded, placebo-controlled RCTs, the results in the intervention group reflect the treatment plus placebo.

Placebo effect and natural history both make it impossible to prove that our treatment caused any one patient to get better. The third reason is much less intuitive, and in some ways more insidious. As I’ve found to be true, a baseball analogy usually makes it clearer.

The slump was always going to end

Rafael Devers got off to a historically bad start in 2025. He went hitless in his first 23 plate appearances and struck out fifteen times—the most strikeouts by any player in his first five games. After 27 games, the Boston slugger had only two home runs and a .194/.331/.327 slash line.

It was a baffling start for a player considered the cornerstone of the Red Sox offense. In his previous five seasons, Devers had posted a wRC+ of 131—meaning he was 31% better than the average MLB hitter. That he was now producing at a rate that was 18% below average was inexplicable.

But that didn’t keep people from trying to explain it. Some speculated it was because he had been forced to move to full-time DH after Boston signed Alex Bregman to play third base. Others thought he still hadn’t fully recovered from the shoulder injuries that ended his 2024 season.

And with great confidence, Sports Illustrated writer Tom Verducci attributed his disastrous first five games to a swing fault:

“The problem is mechanical… Devers has developed a loop with his hands in the load phase of his swing. Instead of sliding his hands slightly back and near to his head, Devers is moving them in a larger loop, triggered by the back elbow raising higher than normal.”

Actually, it was a statistical near-certainty that Devers would break out of his slump. Regression to the mean—a statistical tendency where extreme results in a sample are typically followed by results that are closer to average—practically guaranteed it. His start was an extreme outlier from career averages. Improvement was to be expected. It came: Devers finished the season with a wRC+ of 135.

This happens all the time in medicine, but we rarely recognize it. For many conditions, patients come in when their symptoms peak—their blood pressure has spiked, their chronic back pain has flared, or their anxiety has dramatically increased. These conditions are highly influenced by external factors (such as stress), meaning that they are noisy metrics prone to fluctuation. Regression to the mean tells us that extreme values will soon drift back toward average regardless of what we do. And the larger the outlier, the more dramatic the regression we can expect.

This is distinct from natural history because it is a statistical, rather than biological, phenomenon. It arises from the selection process by which patients tend to present. Because they often show up at the height of their symptoms, we cannot know how much of a patient’s subsequent improvement was from our intervention or from the expected regression to the mean.

The convergence

All three of these forces operate in clinical medicine such that seeing a patient improve after treatment is not proof that the treatment caused it. This is why randomized controlled trials are so important. By controlling for all three of these effects simultaneously, RCTs isolate treatment effects where clinical experience cannot.

And that’s important to understand because it affects how we think about everything we do. Physicians are shaped by clinical experience, and seeing a patient improve (or fail to improve) is powerful feedback. Yet many of those lessons were built partly on perception rather than reality. That’s an unsettling thought in a science-based profession.

But it’s not just a philosophical problem. What happens inside each of us has a ripple effect for the field of medicine as a whole—in its culture, the way we teach, and its evolution. In the final part of this series, we’ll explore what it looks like when misplaced confidence perpetuates, and why epistemic humility is an essential clinical skill.

Even a 0-for-21 stretch eventually ends. But knowing why it ended — that's the harder question.

If this piece made you think differently about what clinical experience can and cannot teach us, a like ❤️ or restack 🔁 helps this reach others wrestling with the same questions.

Subscribe now

Share