How To Speed Up the Search for Cures Through a Change in Probability Theory

Marty Makary was appointed commissioner of the Food and Drug Administration (FDA) in 2025. The prominent surgeon, medical researcher, bestselling author, and critic of the medical establishment—including the FDA and the Centers for Disease Control and Prevention—has made the incendiary claim that medical errors are the third-leading cause of death in the United States, roughly equal to the next four causes combined: smoking-related lung disease, suicide, firearms, and automobile accidents.

In January 2026, the FDA issued a draft for public comment advocating more use of Bayesian statistics in drug approvals and other FDA actions. “Bayesian methodologies help address two of the biggest problems of drug development: high costs and long timelines,” Makary wrote. “Providing clarity around modern statistical methods will help sponsors bring more cures and meaningful treatments to patients faster and more affordably.” The proposal also claims a Bayesian approach would result in more accurate and nuanced advice for patient subgroups. 

This move would be roughly analogous to regulators moving away from rigid rules specifically defining mandated and forbidden behavior and toward trying to achieve the same regulatory goals with sometimes market-based tools to help nudge people’s behavior, such as auctions, subsidies, and taxes. 

Rather than central planners making one-size-fits-all choices for people they never met, don’t understand, and wouldn’t like if they did meet, that approach allows individuals to choose for themselves in individualized reaction to the carrots and sticks wielded by planners. It’s a kinder, gentler central planning.

The public health/medical complex is a battleship, and it takes more than one commissioner waving a draft proposal to change its direction. Makary’s proposal is incremental, not revolutionary. While the FDA’s DNA is what statisticians call “frequentist,” Bayesian methods have been sneaking in through the back door for decades. The Makary proposal is to open the front door for them, but to watch them carefully so they don’t steal the frequentist silver. Only Bayesian tools are permitted entry, not Bayesian theory. 

A key theoretical divide runs between Bayesians and frequentists. Frequentists define probability as the long-run frequency of repeated independent trials. There are some severe philosophical issues with that, but it does seem to make some sense for events such as coin flips, dice rolls, and roulette spins. It’s harder to apply that method to things that happen once, like the probability of rain tomorrow or of the Seattle Mariners winning the 2026 World Series. 

Bayesians define probability as subjective belief—loosely speaking, the value you would place on a bet. For example, according to recent odds posted at FanDuel, a claim that pays $100 if the Mariners win the World Series is worth $5.78 today, and FanDuel will sell you one for $7.14 (the difference is FanDuel’s expected profit). Now that’s a betting line offered by a commercial venture, not any individual’s subjective belief, but it suggests a reasonable Bayesian estimate of the probability of the Mariners winning is 5.78 percent if you have no strong knowledge on the subject to outguess the bookies.

To see the difference between the two methods played out in medical science, consider how the Moderna COVID-19 vaccine was approved. The FDA negotiated with Moderna to enroll 30,420 healthy adults in the Phase 3 clinical trials. The volunteers had no history of COVID-19 but were at high risk of exposure. Earlier phases involved animal and smaller-scale human testing to establish safety and some evidence of efficacy. Half the volunteers were given the vaccine, half a placebo. 

Neither the volunteers nor the medical staff evaluating the results knew which ones got which. The trial was to continue until at least 151 participants had contracted symptomatic COVID-19. If at that point, fewer than 63 of the treatment group had had COVID-19, the vaccine would be approved. 

By the end of the trial, 196 cases had arisen among the study participants, 185 in the placebo group and only 11 in the treatment group. The trial was deemed a spectacular success for the vaccine.

I have omitted reams of details and technicalities here to focus on the basic idea. This trial was conducted under the new rules of Operation Warp Speed, which cut the vaccine development and testing time from an average of 10 to 20 years to under one year.

The numbers involved—30,240 subjects, 151 events, 63 treatment events—came from frequentist calculations. But the important point is that every detail of the test was set in advance. All patients in the treatment group got the same two vaccine doses at the same two-week intervals. The primary decision depended on a binary criterion—either fewer than 63 subjects got symptomatic COVID-19 or more than 63 did. (There were important secondary considerations, including the severity of cases and side effects, but I’m focused on the basics.)

A pure Bayesian test would be run differently. You would begin by using all information from prior phases, theory, similar vaccines, and other sources to identify the best patients to begin with—those with the most chance of being helped relative to current alternatives, the least likely to be harmed, and the ones whose outcomes would provide the most useful information. 

A much smaller number of initial patients would be tested with a much larger number of investigators, because each patient would be carefully selected and watched by multiple people—not just a head researcher or specialist staff members but multiple caregivers, the patient herself, and perhaps friends and family. Doses, timing, and supportive therapies would be adjusted by trial and error. All aspects of patients’ reactions would be recorded and considered.

The more you learned, the broader the range of patients would be tested by additional researchers. Conclusions could be nuanced for different patient groups and different individual preferences, such as between longevity and quality of life. There is no fixed goal as in the frequentist study, which was estimating the percentage reduction in COVID-19 infections from two fixed doses of the vaccine two weeks apart. 

You might end up discovering the treatment was beneficial for something other than the initial purpose or for different types of patients in different ways. There’s no sharp distinction between study and practice. You use the treatment for more things as you learn more about it, with continual learning and improvement.

Perhaps most importantly, the end goal is not an approved/not-approved decision but a compendium of information that allows individual patients and doctors to apply what is known to their specific circumstances and make their own choices. 

This is, in idealized form, what happens after a treatment is approved by the FDA. It’s also how traditional medicine evolves. In real life, not every patient and caregiver is observant, not all data are recorded or shared, lots of bad information creeps in, and lots of good information is suppressed for various reasons. The process has many failures—worthless or harmful traditional remedies, and medical disasters of the sort Makary documented in his book Blind Spots. But it’s also the main way medical practice advances. It’s far more influential than official decisions by the FDA or conclusions of academic medical research.

The Bayesian approach has many advantages over the frequentist approach. It’s faster and cheaper. Patient care improves continuously, not just in a jump at the end of the study. Every patient is given the best available treatment given the state of knowledge at the time. Bayesians use all available information, not just the narrow data that are the primary focus of frequentist testing. 

Bayesians can test any kind of holistic or complex treatment, whereas frequentist methods work best for single-ingredient drugs with known dosage patterns that work independently of other treatments. The frequentist approach can be modified for treatments that cannot be double-blinded—such as massage therapy—or ones that have many free parameters to calibrate, but these complications can rapidly increase its expense and reduce its reliability. Often, treatments not amenable to straightforward frequentist testing are ignored in favor of simpler treatments for regulatory convenience rather than optimal medicine.

The Bayesian approach has its flaws as well. It opens the door to human prejudices, biases, wishful thinking, and delusion. These are precisely the things frequentism was invented to combat, with the goal of making science objective. 

The double-blind, controlled, frequentist trial is still the gold standard for getting objective answers to narrow questions, such as if 1,000 people are chosen at random to get a specific vaccine dose regime, how many of them will get severe COVID-19 in the subsequent six months compared to a matched random sample of people without the vaccine? That question may be relevant to a greatest-good-of-the-greatest-number public-health collectivist. It might seem less so to an individualist libertarian. 

The question is not whether the Bayesian approach is better or worse overall than the frequentist, but at which points in the medical research process should we apply each technique, each with its own dangers and benefits? In the current system, researchers use Bayesian approaches to come up with new conjectures and candidate therapies to be tested with simulations, laboratory work, and animals. 

But once humans are being tested, medical science has traditionally required strict frequentism until a regulator says the medicine is permitted. After that, doctors and patients are free to experiment with approved treatments, and other researchers do frequentist studies to continue monitoring safety and effectiveness.

The Makary FDA report suggesting a more Bayesian approach does not advocate switching to pure Bayesian clinical trials. Rather, these new proposals would allow more Bayesian-flavored methodologies within what remains a basically frequentist process. The idea is to incorporate more human judgment into the process—making it less objective—to get cheaper, faster, and better results for a wider variety of treatment types. 

But the proposal describes a system with sober negotiations between researchers and regulators on what judgments to credit, resulting in committee decisions, not crediting pure individual subjective belief as Bayesian theory demands.

The difference in the two approaches in medicine can be described like this: Frequentists want to estimate the long-term frequency of patients who will benefit from a treatment, while Bayesians want to price bets on outcomes for individual patients. This leads to profoundly different regulatory goals. Frequentists imagine an infinite series of faceless patients without volition or preferences. The researcher is not one of the patients, but an expert above them, reading arcane entrails to make a choice for all of them, maximizing the regulator’s opinion of aggregate benefit.

Frequentists reason in this inhuman abstraction because it allows rigorous, objective decision making. It bears no relation to reality. Whether or not a treatment will benefit you is not a random event. It depends on your precise condition, your genetics, and other factors. It can be hard to predict, but it’s not random.

Bayesians accept that a patient’s response to a treatment is not random, just unknown. Bayesians are like bookies. They want to hand patients a list of bets they can make with different payouts—cure, death, unpleasant side effects, no change, etc.—and prices attached. Patients can consult with doctors and other experts to choose good bets based on their own additional information, if any, and their personal preferences. Importantly, the Bayesian statistician considers herself one of the people for whom she gives advice—she bets for herself and her loved ones using the same odds she quotes for the public.

Most people who work with data are agnostics, however. They use whatever tools seem to work for different applications without worrying much about the underlying philosophy. Most of the time, the data speak for themselves, and frequentists and Bayesian methods give similar answers; and two frequentists or two Bayesians are as likely to disagree with each other as a frequentist is to disagree with a Bayesian.

If the FDA follows through with the proposed guidelines, and they are not fatally twisted by pressure from the medical establishment and health care industry, it should bring fresh air and sunlight into the approval process. It should save money and speed innovation, with better health outcomes.

Most ideal would be an FDA with Bayesian DNA that left room for some double-blind, controlled frequentist studies to provide a skeleton to the Bayesian flesh of medical practice. This kind of regulation would focus on empowering patients with information and choice, rather than using information about outcomes in a narrowly conceived experiment to make choices for others.