Introduction to Bayesian Statistics for Clinical Trials

The treatment effect might be defined as:

A Bayesian analysis can produce the posterior distribution:

From that distribution, we can calculate:

or:

The second quantity asks whether there is a posterior probability that the treatment improves response by at least 10 percentage points.

Relative Risk and Odds Ratios

Bayesian inference is not restricted to risk differences. For example, if:

then the posterior distribution of \(RR\) can be obtained from the joint posterior distributions of the treatment and control response probabilities.

Similarly, the odds ratio is:

and Bayesian inference can provide:

or:

Bayesian Survival Analysis

The Bayesian framework also extends naturally to time-to-event endpoints.

For example, a Bayesian survival model can estimate:

• Survival probabilities
• Hazard ratios
• Median survival
• Restricted mean survival time
• Time-varying hazards

The basic Bayesian workflow remains the same:

The likelihood simply becomes more complex than the binomial likelihood used in the introductory example.

Bayesian Longitudinal Models

For repeated measurements, Bayesian models can incorporate:

• Patient-specific random effects
• Population-level treatment effects
• Correlation between repeated observations
• Missing-data mechanisms
• Nonlinear trajectories

For example:

where \(b_i\) could be assigned a prior distribution representing between-patient heterogeneity.

Bayesian Logistic Regression

For binary clinical outcomes, Bayesian logistic regression can model the probability of response as a function of covariates.

The model can be written:

Bayesian priors are then assigned to the regression coefficients. For example:

The posterior distribution provides uncertainty about each coefficient and about clinically relevant quantities derived from the model.

Bayesian PK/PD Applications

Bayesian methods are particularly useful in pharmacometrics and dose-response modeling.

For example, a Bayesian model can estimate parameters such as:

• Clearance
• Volume of distribution
• Absorption rate
• EC50
• Emax
• Between-subject variability

Prior information from previous studies can be incorporated into the analysis, and posterior distributions can then be used for prediction and decision-making.

Bayesian Evidence Is Not the Same as "Subjective Belief"

A common misconception is that Bayesian statistics means simply allowing investigators to believe whatever they want.

That is not how a properly specified Bayesian analysis works.

A Bayesian analysis requires explicit specification of:

• The probability model
• The prior distribution
• The likelihood
• The posterior distribution
• The decision criteria

These choices can be examined, documented, challenged, and subjected to sensitivity analyses.

What Happens as the Sample Size Becomes Large?

As more data accumulate, the likelihood generally becomes more concentrated.

When the prior is reasonably specified and the likelihood is informative, the current data can increasingly dominate the prior.

Conceptually:

while:

This does not mean that prior specification becomes irrelevant in every large-sample problem. Complex models, weakly identified parameters, and strong priors can behave differently.

Bayesian Evidence Is Conditional on the Model

A posterior probability is always conditional on the assumptions used to construct it.

More formally:

depends on model \(M\).

This includes assumptions about:

• The likelihood
• The prior
• The population
• The endpoint
• The covariates
• The missing-data mechanism

Consequently, Bayesian sensitivity analyses are an important component of credible clinical inference.

R Implementation: The Beta-Binomial Example

The basic Bayesian example can be implemented directly in R.

n <- 40
x <- 14

alpha_prior <- 3
beta_prior  <- 7

alpha_post <- alpha_prior + x
beta_post  <- beta_prior + n - x

alpha_post
beta_post

The resulting posterior parameters are:

alpha_post
# 17

beta_post
# 33

Calculate the Posterior Mean

posterior_mean <-
  alpha_post /
  (alpha_post + beta_post)

posterior_mean

The result is:

# 0.34

Thus, the posterior mean response probability is 34%.

Calculate a 95% Credible Interval

credible_interval <-
  qbeta(
    c(0.025, 0.975),
    alpha_post,
    beta_post
  )

credible_interval

The qbeta() function calculates posterior quantiles from the beta distribution.

Calculate Posterior Probability of \(p>0.30\)

prob_above_30 <-
  1 - pbeta(
    0.30,
    alpha_post,
    beta_post
  )

prob_above_30

This directly calculates:

This is one of the most useful features of Bayesian inference: clinically meaningful probability statements can be calculated directly.

Plot the Prior and Posterior

The prior and posterior distributions can also be visualized.

p <- seq(0, 1, length.out = 1000)

prior_density <-
  dbeta(
    p,
    alpha_prior,
    beta_prior
  )

posterior_density <-
  dbeta(
    p,
    alpha_post,
    beta_post
  )

plot(
  p,
  prior_density,
  type = "l",
  lwd = 2,
  xlab = "Response probability",
  ylab = "Density",
  main = "Prior and Posterior Distributions"
)

lines(
  p,
  posterior_density,
  lwd = 2
)

legend(
  "topright",
  legend = c("Prior", "Posterior"),
  lty = 1,
  lwd = 2
)

The posterior distribution should generally be more concentrated than the prior because the current trial has added information.

Bayesian Decision Rule in R

Suppose the trial declares success if:

The decision rule can be implemented as:

decision_threshold <- 0.90

success <-
  prob_above_30 >= decision_threshold

success

The result is a logical value indicating whether the prespecified Bayesian criterion has been met.

Bayesian Predictive Probability in R

For the beta-binomial model, future responses have a beta-binomial predictive distribution. Suppose:

• 14 responses have been observed among 40 patients.
• 20 additional patients will be enrolled.
• The final success criterion is at least 22 total responses.

At least:

additional responses are therefore required.

The posterior predictive probability can be calculated using the beta-binomial distribution.

library(stats)

future_n <- 20
current_x <- 14
final_required <- 22

future_required <-
  final_required - current_x

predictive_probability <-
  1 -
  pbetabinom(
    future_required - 1,
    size = future_n,
    shape1 = alpha_post,
    shape2 = beta_post
  )

predictive_probability

This calculates the probability that the final trial will meet the response criterion, conditional on the current posterior distribution.

Posterior Predictive Thinking

The distinction between posterior and predictive inference is worth emphasizing.

Suppose the posterior probability that:

is very high. That does not automatically mean that the trial is guaranteed to succeed.

Future patients still have uncertain outcomes.

Predictive probability accounts for both:

• Uncertainty about the underlying treatment effect
• Randomness in future patient outcomes

This makes predictive probability particularly useful for interim decision-making.

Bayesian Clinical Trial Workflow

Bayesian Operating Characteristics

A Bayesian clinical trial should still be evaluated using simulation before the trial begins.

Important operating characteristics can include:

• Probability of declaring success
• Probability of declaring futility
• Probability of an incorrect decision
• Expected sample size
• Maximum sample size
• Probability of early stopping
• Predictive probability behavior
• Posterior probability calibration

For example, a Bayesian design might be simulated under several true treatment effect values:

The resulting decision probabilities provide a detailed picture of how the trial behaves under different possible realities.

Bayesian Design Is Not "No Type I Error"

A common misconception is that Bayesian trials eliminate false-positive concerns.

That is not correct.

Bayesian designs can be evaluated for the probability of incorrectly declaring success under clinically relevant null scenarios.

For example:

can be estimated through simulation.

This quantity is not automatically identical to a frequentist type I error, but it can be an important operating characteristic for clinical trial design.

Bayesian Calibration

A Bayesian clinical trial design should be calibrated against realistic clinical scenarios.

For example, investigators may simulate trials under:

• Clearly ineffective treatment effects
• Marginally ineffective treatment effects
• Clinically meaningful treatment effects
• Highly effective treatment effects

For each scenario, the proportion of simulations resulting in each decision can be recorded.

Common Bayesian Trial Design Mistakes

1. Choosing the prior after seeing the data. The prior should be specified before the current trial data are used for the primary analysis.
2. Using an overly informative prior without justification. A strong prior can materially influence the posterior.
3. Reporting only the posterior mean. The full posterior distribution contains information about uncertainty.
4. Confusing credible intervals with confidence intervals. They have different interpretations.
5. Confusing posterior probability with predictive probability. One describes the current parameter uncertainty; the other incorporates uncertainty about future outcomes.
6. Ignoring prior sensitivity. A conclusion that changes substantially under plausible priors requires careful interpretation.
7. Assuming Bayesian means subjective or arbitrary. Bayesian analyses require explicit statistical assumptions that can be evaluated and challenged.
8. Failing to simulate the design. Complex Bayesian trial designs often require extensive simulation to evaluate their operating characteristics.
9. Changing decision thresholds after observing the data. This can invalidate the planned operating characteristics and undermine the interpretability of the trial.
10. Ignoring model misspecification. A posterior distribution is only as credible as the model assumptions used to construct it.

Bayesian vs. Frequentist: A Practical Comparison

A More Intuitive Way to Think About Bayesian Statistics

Imagine that before a clinical trial begins, the investigator has a probability distribution describing plausible treatment effects.

The trial then produces evidence.

The evidence reshapes that distribution.

The result is not:

Instead:

That updated uncertainty can then be used to make decisions.

The Bayesian Clinical Trial as a Learning System

This perspective is particularly useful for clinical development.

A development program can be viewed as accumulating evidence:

Bayesian methods provide a mathematical framework for representing this accumulation of information.

What Bayesian Statistics Does Not Mean

Several misconceptions are worth eliminating.

What Should Be Prespecified?

For a clinical trial using Bayesian methods, the protocol and statistical analysis plan should clearly document the inferential framework.

Depending on the design, this may include:

• Estimand
• Statistical model
• Likelihood
• Prior distributions
• Prior rationale
• Primary posterior quantity
• Credible interval method
• Decision thresholds
• Predictive probability definitions
• Interim analysis rules
• Stopping rules
• Missing-data assumptions
• Sensitivity analyses
• Simulation methodology
• Operating characteristics

Prior Sensitivity Should Be Part of the Analysis Plan

A useful Bayesian analysis does not merely report one posterior distribution.

It should also demonstrate how robust the conclusions are to reasonable alternative assumptions.

For example, the analysis might compare:

The goal is not necessarily to make every posterior identical. The goal is to understand whether the clinical conclusion is robust to plausible prior assumptions.

Bayesian Model Checking

Bayesian inference requires more than calculating a posterior distribution. The model itself should be examined.

Questions include:

• Does the likelihood adequately represent the endpoint?
• Are the prior assumptions reasonable?
• Does the model reproduce important characteristics of the observed data?
• Are there important patient subgroups?
• Are the residual assumptions reasonable?
• Is the model sensitive to influential observations?

Posterior predictive checks can be useful for evaluating whether the fitted model can reproduce features of the observed data.

Posterior Predictive Checking

Suppose:

represents replicated data generated from the posterior predictive distribution.

A posterior predictive check compares characteristics of:

If the model systematically fails to reproduce important features of the observed data, the model may require reconsideration.

From Simple Conjugate Models to Modern Bayesian Computation

The beta-binomial example is intentionally simple. In real clinical trials, models may not have closed-form posterior distributions.

Modern Bayesian computation can use methods such as:

• Markov chain Monte Carlo (MCMC)
• Hamiltonian Monte Carlo
• Sequential Monte Carlo
• Variational inference
• Numerical integration

These methods allow Bayesian inference for complex hierarchical, longitudinal, survival, PK/PD, and dose-response models.

MCMC Conceptually

Suppose the posterior distribution cannot be written in a convenient closed form. MCMC generates a sequence of draws:

that approximately represents the posterior distribution.

Posterior quantities can then be estimated from these draws. For example:

where \(I(\cdot)\) is an indicator function.

Why Bayesian Computation Matters in Clinical Biostatistics

Modern clinical development increasingly involves models that are too complex for simple textbook formulas.

Examples include:

• Bayesian hierarchical models
• Longitudinal mixed-effects models
• Population PK models
• PK/PD models
• Joint longitudinal-survival models
• Bayesian dose-escalation models
• Adaptive platform trials

The fundamental Bayesian equation remains the same:

The computational machinery simply becomes more sophisticated.

A Practical Bayesian Clinical Trial Workflow

Worked Example Summary

The beta-binomial example can be summarized as follows.

The Most Important Concept

The most important concept in Bayesian statistics is not the beta distribution, MCMC, or even Bayes' theorem by itself.

It is the idea of updating uncertainty as evidence accumulates.

Before the trial:

After observing data:

The posterior distribution then becomes the basis for probability statements, prediction, and decision-making.

For clinical trials, this can provide an intuitive way to answer questions such as:

• How probable is a clinically meaningful treatment effect?
• How uncertain is the treatment effect?
• What is the probability that the trial will ultimately succeed?
• How much evidence supports stopping for futility?
• How much evidence supports stopping for efficacy?
• How should prior information influence the analysis?

References

Berry, D.A. (1993). A case for Bayesianism in clinical trials. Statistics in Medicine, 12, 1377–1393.
Spiegelhalter, D.J., Abrams, K.R. & Myles, J.P. (2004). Bayesian Approaches to Clinical Trials and Health-Care Evaluation. Wiley.
Berry, S.M., Carlin, B.P., Lee, J.J. & Muller, P. (2010). Bayesian Adaptive Methods for Clinical Trials. CRC Press.
Gelman, A., Carlin, J.B., Stern, H.S., Dunson, D.B., Vehtari, A. & Rubin, D.B. (2013). Bayesian Data Analysis. 3rd ed. Chapman & Hall/CRC.
Hobbs, B.P., Carlin, B.P. & Mandrekar, S.J. (2011). Bayesian design of clinical trials. Clinical Trials, 8, 141–151.
Berry, S.M., Viele, K. & Arguello, M. (2013). Bayesian clinical trials. Nature Reviews Drug Discovery, 12, 19–25.
Lee, J.J. & Chu, C. (2012). Bayesian clinical trials. Clinical Trials, 9, 13–19.