Bayesian Thinking in Clinical Research

Part 1: What Do We Actually Mean by a “Prior”?

Image is made by GPT by author’s prompts.

In clinical research, results are often treated as if they emerge cleanly from data, self-contained, objective, and independent of what came before. A study is conducted, an analysis is performed, and conclusions follow.

But in practice, this is rarely how decisions are made.

Long before results are available, there is already a sense, sometimes subtle and sometimes strong, of what might be expected. A therapy may appear promising based on its mechanism. Earlier trials may have suggested a signal. Clinical experience may point in one direction or another. Even skepticism has a structure, shaped by prior failures or inconsistencies.

These expectations are not formally part of most statistical analyses. They exist in the background, informing interpretation and shaping judgment, while remaining largely unspoken.

Bayesian statistics begins by bringing that background into the foreground. It offers a structured way to incorporate prior knowledge and update it as new data becomes available. Over the next few months, we will build on this idea and explore how Bayesian thinking is applied in clinical research, using concrete examples from practice.

A shift in what we ask of data

Traditional statistical approaches are built around a specific question:

“If the treatment had no effect, how unusual is the data we observed?” (This is what the p-value attempts to quantify.)

This question is mathematically precise, but it is not always aligned with how clinical reasoning unfolds.

Bayesian thinking reframes the problem:

“Given the data we observed, what is the probability that the treatment is beneficial?”

This shift may seem subtle, but it changes the role of data, from something we test against a fixed assumption to something we use to update what we believe.

In Bayesian analysis, results are expressed in direct terms such as:

Pr(treatment effect > 0 | data), a probability statement about the treatment effect itself, conditional on the observed data.

Making prior knowledge visible

At the center of this framework is the idea of a prior.

A prior is a structured way to represent what is already known, or assumed, before analyzing the current study.

In oncology, this prior knowledge is often substantial:

• signals from early-phase trials
• understanding of biological pathways
• performance of related therapies
• accumulated clinical experience

In most analyses, this information remains implicit.

Bayesian methods make it explicit. Rather than treating each study as isolated, they allow prior knowledge to be formally incorporated into the analysis, creating a starting point that reflects existing evidence.

Learning as an ongoing process

The process that follows is straightforward in principle. Prior knowledge is combined with new data to produce an updated understanding.

What begins as an initial belief is revised as evidence accumulates, resulting in what is known as the posterior, a probability distribution that reflects both past information and current observations.

This updating process is formalized through Bayes’ theorem, where prior information is mathematically combined with observed data.

Importantly, this is not a rigid system. When new data strongly disagrees with prior expectations, the influence of the prior can diminish, allowing the observed evidence to take precedence.

A concrete oncology example

Consider a trial evaluating a new immunotherapy agent in breast cancer. Before the trial begins, there is already context: similar therapies have shown response rates in a certain range, the biological mechanism suggests potential effectiveness, and early studies provide preliminary signals. This context forms the prior.

As patient outcomes are observed, the analysis updates this prior understanding. Strong responses reinforce confidence in the treatment, while weaker or inconsistent results reduce it.

The result is not a binary conclusion, but a continuously updated assessment of how likely the treatment is to be effective.

Why this matters now

The relevance of this approach is increasingly visible in modern clinical trial design.

Adaptive oncology trials, such as the I-SPY2 platform in breast cancer, use Bayesian methods to evaluate treatments as data accumulates, allowing decisions to be made during the trial about whether a therapy should continue, expand, or stop.

At the same time, regulatory guidance has begun to formally recognize Bayesian methods as valid for primary analyses, provided that assumptions, including priors, are clearly specified and rigorously evaluated.

This is only the starting point. In the next bulletin, we will move from concept to practice and examine how these ideas are implemented in real studies.

Statistical Wisdom of the Month

Statistical analysis does not begin when data arrives. It begins with what we believe before we see it, and how willing we are to update that belief.