Bayesian Borrowing of Historical Control Data

Introduction

Clinical trials often contain less information than investigators would ideally like. This is particularly important when the control treatment has a well-established response rate from previous studies, registries, natural-history cohorts, or earlier trials.

Suppose a new Phase II study is evaluating an experimental treatment against a control treatment. A conventional analysis may use only the concurrently randomized control patients.

If the current trial contains relatively few control patients, however, a substantial amount of relevant information from previous studies may be left unused.

Bayesian borrowing provides a formal framework for incorporating historical control information into the current analysis.

The important qualification is that historical patients cannot automatically be assumed to be exchangeable with current patients.

Differences in patient population, disease severity, eligibility criteria, endpoint definitions, follow-up, standard of care, geography, calendar time, and trial conduct can make historical information less relevant to the current study.

Bayesian borrowing is therefore not simply a mechanism for increasing the sample size. It is a framework for quantifying how much historical information should influence the current inference.

The Basic Problem

Suppose a randomized clinical trial contains an experimental treatment arm and a concurrent control arm.

Let:

• \(p_C\) = current control response probability
• \(p_T\) = current treatment response probability

The treatment effect on the risk-difference scale is:

or on the risk-ratio scale:

Historical studies may contain information about:

where \(p_H\) represents the historical control response probability.

The fundamental question is:

If the historical and current populations are highly comparable, borrowing can be substantial.

If they differ materially, borrowing should be reduced or eliminated.

Why Borrow Historical Controls?

There are several reasons historical borrowing can be attractive.

• Control patients may be difficult or expensive to recruit.
• The disease may be rare.
• Randomization may require a large number of concurrent controls.
• Historical control information may be highly informative.
• The control response rate may be relatively stable across studies.
• Additional control enrollment may expose patients to an established treatment without directly increasing information about the experimental treatment.

For example, if 100 historical control patients have been studied under conditions very similar to the current trial, it may be inefficient to pretend that those observations contain no information whatsoever.

Current Data Versus Historical Data

Suppose the current control arm contains:

where \(Y_C\) is the number of responses among \(n_C\) concurrent control patients.

Suppose the historical dataset contains:

The key Bayesian modeling question is whether:

or whether they should instead be treated as related but potentially different parameters.

Complete Pooling: The Strongest Form of Borrowing

The simplest approach is to assume that historical and current controls have exactly the same response probability:

Under this assumption, the historical and current observations can be pooled.

Suppose the historical study has:

and the current trial has:

Complete pooling would treat all 150 observations as coming from the same Bernoulli probability.

The combined observed response rate would be:

This can be statistically efficient if the exchangeability assumption is credible.

However, complete pooling can be dangerous if the historical study is not representative of the current trial.

Why Complete Pooling Can Fail

Imagine that the historical control response rate was 60%, while the current control response rate is actually closer to 30%.

If historical data are treated as fully exchangeable, the historical observations will pull the current estimate toward 60%.

That can create a seriously misleading analysis.

Bayesian Prior Distributions

In Bayesian inference, uncertainty about an unknown parameter is represented with a probability distribution.

For a binary control endpoint, a natural prior is the Beta distribution:

Its density is proportional to:

The prior mean is:

and the prior effective sample size is commonly interpreted as:

This makes the Beta distribution especially convenient for explaining historical borrowing.

Historical Data as a Prior

Suppose historical controls produce:

Under a Beta-Binomial model, a simple historical-data prior is:

when the baseline prior is:

This is equivalent to treating the historical observations as prior evidence about the current control response probability.

However, using all historical observations with full weight is often too strong.

Bayesian borrowing methods therefore introduce mechanisms for discounting historical information.

The Power Prior

One of the most intuitive approaches is the power prior.

Let the historical likelihood be:

The power prior raises the historical likelihood to a power \(\alpha\):

where:

• \(D_H\) = historical data
• \(\alpha\) = borrowing weight
• \(\pi_0(p_C)\) = baseline prior

The borrowing parameter satisfies:

Interpretation is straightforward:

Power Prior for a Beta-Binomial Endpoint

Suppose:

and:

The historical likelihood is proportional to:

After applying the power parameter:

Therefore the resulting prior is:

This makes the interpretation particularly transparent.

A Worked Example

Consider a randomized Phase II study evaluating a new therapy for a rare disease.

The primary endpoint is binary response at a prespecified assessment time.

The current trial contains 50 concurrent control patients.

Historical studies contain 100 control patients who received the same nominal standard of care under broadly similar eligibility criteria.

The historical data contain:

so the historical response rate is:

The current trial observes:

giving a current observed response rate of:

The historical and current estimates are noticeably different.

This makes the example particularly useful because it illustrates why discounting historical information can matter.

Step 1: Specify a Weak Baseline Prior

Start with a weak Beta prior:

This is the uniform distribution on the interval from 0 to 1.

Its effective sample size is:

The baseline prior therefore contributes relatively little information compared with the current and historical datasets.

Step 2: Apply a 50% Historical Borrowing Weight

Suppose the investigators decide that a power-prior weight of:

The historical contribution becomes:

and:

Therefore the historical-data prior is:

or:

The prior mean is:

Thus the historical information centers the prior around approximately 59.6%.

Step 3: Add the Current Control Data

The current likelihood contributes:

responses and:

nonresponses.

Because the Beta prior is conjugate to the Binomial likelihood, the posterior is also Beta:

Therefore:

Step 4: Calculate the Posterior Mean

The posterior mean is:

The current observed response rate was 30%, while the historical response rate was 60%.

The Bayesian posterior lies between these values because both sources of information contribute to the inference.

Step 5: Understand the Effective Sample Size

The historical effective sample size under the 50% power prior is:

so:

The current trial contributes 50 actual patients.

The approximate total information is therefore comparable to:

current-equivalent observations, apart from the small baseline-prior contribution.

This illustrates why historical borrowing can materially increase precision.

Compare Different Borrowing Weights

One of the most useful ways to understand Bayesian borrowing is to repeat the analysis under different values of \(\alpha\).

The exact posterior distribution should be recalculated for each borrowing weight rather than interpreting the weight itself as a posterior probability.

No Borrowing: The Reference Analysis

A useful sensitivity analysis begins with:

This means that the historical dataset contributes no information.

With the baseline prior:

and current data of 15 responses among 50 patients:

The posterior mean is:

which is close to the observed current response rate of 30%.

Full Borrowing

At the other extreme:

the historical observations are given full weight.

The prior becomes:

After adding the current data:

with posterior mean:

Notice what happened.

The current data suggested a control response rate of 30%, while the historical data suggested 60%.

Full borrowing produces a posterior centered at 50%.

Dynamic Borrowing

A major development beyond fixed power priors is dynamic borrowing.

Instead of deciding in advance that:

the model allows the amount of borrowing to depend on how compatible the historical and current data appear to be.

Conceptually:

while:

This is particularly attractive because the investigator does not need to pretend that historical information is either perfectly exchangeable or completely irrelevant.

Commensurate Priors

The commensurate prior approach explicitly models the similarity between the historical and current parameters.

Suppose:

represents the historical control rate and:

represents the current control rate.

A commensurate model might specify:

where \(\tau\) controls how closely the two parameters are expected to agree.

A high degree of commensurability means:

while weak commensurability allows:

to differ substantially from:

The model can therefore learn how compatible the datasets are.

The Concept of a Commensurability Parameter

A useful conceptual model is:

where:

• \(\theta_H\) = historical parameter
• \(\theta_C\) = current parameter
• \(\tau\) = precision describing how closely they are expected to agree

Large \(\tau\) implies strong similarity.

Small \(\tau\) implies greater uncertainty about whether the historical and current parameters are comparable.

Robust Bayesian Borrowing

Another important strategy is the robust mixture prior.

Instead of assuming that the historical prior is always appropriate, construct a mixture of an informative historical prior and a weakly informative or vague component.

For example:

where:

• \(\pi_H\) = historical-data prior
• \(\pi_V\) = weakly informative prior
• \(w\) = prior probability assigned to the historical component

This creates protection against complete dependence on historical information.

Why Robust Priors Are Useful

Suppose historical data strongly suggest a 60% response rate.

But there is a possibility that the historical population is not comparable to the current population.

A robust prior can represent both possibilities:

The posterior probability assigned to the components can then shift as current data accumulate.

Historical Borrowing Does Not Mean Replacing Randomization

Historical borrowing is often discussed in the context of randomized trials, but it should not be confused with abandoning concurrent controls.

A randomized trial may still contain a concurrent control arm while borrowing additional information about the expected control response.

The current randomized comparison remains the primary source of protection against many forms of bias.

Borrowing and Treatment Effect Estimation

Suppose the treatment arm contains:

and the current control arm contains:

The treatment effect is:

Bayesian inference produces a posterior distribution for both parameters.

The clinically important quantity may therefore be:

or, more usefully, the probability that the treatment exceeds a clinically meaningful effect:

where \(\delta\) is the minimum clinically important treatment effect.

Example: Probability of a Clinically Meaningful Benefit

Suppose the clinically meaningful improvement is:

The Bayesian decision criterion could be:

This is fundamentally different from asking whether a conventional p-value is less than 0.05.

The Bayesian analysis directly describes the posterior probability that the treatment effect exceeds a clinically meaningful threshold.

Historical Borrowing and Effective Sample Size

Effective sample size is one of the most intuitive ways to communicate the strength of historical borrowing.

For a power prior:

For example, if:

and:

then:

The historical dataset contains 200 patients, but under the selected borrowing scheme it contributes information approximately equivalent to 50 fully exchangeable observations.

Historical Data Quality Matters

Before borrowing historical information, investigators should examine the historical dataset itself.

Relevant characteristics include:

• Eligibility criteria
• Baseline disease severity
• Demographic characteristics
• Prior treatment history
• Concomitant medications
• Endpoint definition
• Assessment schedule
• Follow-up duration
• Protocol requirements
• Standard of care
• Geographic region
• Calendar period
• Study conduct

A historical response rate should not be treated as a universal biological constant.

Calendar-Time Drift

One particularly important source of historical-data incompatibility is calendar time.

Standard-of-care therapy can change substantially over several years.

Diagnostic criteria can change.

Supportive care can improve.

Patient referral patterns can change.

The apparent response rate can therefore change even if the underlying treatment has not changed.

Endpoint Compatibility

Historical borrowing is especially sensitive to differences in endpoint definitions.

For example, "response at 12 weeks" is not necessarily interchangeable with "best overall response during treatment."

Similarly:

• Independent central review may differ from investigator assessment.
• Different response criteria may produce different response rates.
• Different assessment windows can change the observed event probability.
• Different missing-data rules can alter the estimand.

Historical data should therefore be harmonized to the extent possible before they are incorporated into the model.

Exchangeability

The statistical concept underlying historical borrowing is exchangeability.

Informally, exchangeability means that after accounting for relevant characteristics, the historical and current observations can reasonably be treated as arising from related populations.

It does not necessarily mean that every patient has exactly the same risk.

Rather, it represents a modeling assumption about the relationship between datasets.

A useful conceptual hierarchy is:

Historical Data Can Be Highly Informative—and Highly Misleading

This is one of the most important concepts in Bayesian historical borrowing.

Suppose the historical sample size is enormous.

One might assume that this automatically makes the historical information valuable.

But if the historical study is systematically different from the current study, the large sample size can actually make the resulting inference more strongly misleading.

Prior-Data Conflict

A useful Bayesian diagnostic is prior-data conflict.

Suppose the historical data strongly favor:

while current data favor:

This discrepancy should not simply be hidden by the model.

Instead, investigators should investigate whether it reflects:

• Random variation
• Population differences
• Changes in standard of care
• Different endpoint definitions
• Different follow-up
• Changes in trial conduct
• True temporal changes in the disease or treatment effect

Sequential Updating

One advantage of the Bayesian framework is that historical information can be combined with current information sequentially.

Conceptually:

As additional current patients are enrolled, the posterior can be updated.

The historical information does not have to be treated as if it were newly collected current data.

Power Prior Versus Commensurate Prior

Robust Mixture Priors

Bayesian Borrowing Is a Modeling Decision

It is tempting to think of historical borrowing as a computational trick for reducing sample size.

That is not the correct perspective.

The borrowing model encodes a scientific belief:

That belief must be justified using clinical and statistical knowledge.

Type I Error Under Bayesian Decision Rules

Bayesian analyses do not automatically have a frequentist type I error of 5%.

For example, suppose a decision rule is:

The posterior probability threshold of 0.95 is not automatically equivalent to a frequentist one-sided type I error of 5%.

The frequentist operating characteristics must be evaluated through simulation or exact calculation under a specified data-generating process.

Simulation-Based Operating Characteristics

Suppose the trial declares success when:

To evaluate frequentist type I error, simulate trials under a null scenario in which:

For each simulated trial:

1. Generate historical data according to the assumed historical scenario.
2. Generate current control data.
3. Generate current treatment data.
4. Fit the Bayesian model.
5. Calculate the posterior probability.
6. Record whether the decision criterion is crossed.

After many simulated trials:

provides an empirical estimate of the type I error.

Power Simulation

Similarly, power can be estimated by simulating under a clinically meaningful alternative.

For example:

Then:

This is often essential for Bayesian clinical trial designs involving nonconjugate models, dynamic borrowing, multiple endpoints, or hierarchical structures.

R: Basic Beta-Binomial Historical Borrowing

The simple conjugate example can be implemented directly in R.

y_hist <- 60
n_hist <- 100

y_curr <- 15
n_curr <- 50

alpha_borrow <- 0.50

a0 <- 1
b0 <- 1

a_prior <- a0 + alpha_borrow * y_hist
b_prior <- b0 + alpha_borrow * (n_hist - y_hist)

a_post <- a_prior + y_curr
b_post <- b_prior + (n_curr - y_curr)

a_prior
b_prior
a_post
b_post

For the example:

a_prior
# 31

b_prior
# 21

a_post
# 46

b_post
# 56

Posterior Mean

posterior_mean <- a_post /
  (a_post + b_post)

posterior_mean
# approximately 0.451

The posterior mean is approximately 45.1%.

Posterior Probability That the Control Rate Exceeds 40%

A Bayesian analysis can answer direct probability questions.

For example:

In R:

prob_above_40 <- 1 - pbeta(
  0.40,
  a_post,
  b_post
)

prob_above_40

This is a posterior probability, not a p-value.

Posterior Credible Interval

A 95% equal-tail credible interval can be calculated with:

qbeta(
  c(0.025, 0.975),
  a_post,
  b_post
)

This interval has a Bayesian interpretation: conditional on the model and prior assumptions, it contains 95% of the posterior probability for the parameter.

That interpretation differs from the frequentist interpretation of a 95% confidence interval.

Compare Borrowing Weights in R

posterior_summary <- function(
  borrow,
  y_hist,
  n_hist,
  y_curr,
  n_curr,
  a0 = 1,
  b0 = 1
) {

  a_prior <- a0 + borrow * y_hist
  b_prior <- b0 + borrow * (n_hist - y_hist)

  a_post <- a_prior + y_curr
  b_post <- b_prior + n_curr - y_curr

  c(
    borrowing = borrow,
    prior_ess = a_prior + b_prior,
    posterior_mean =
      a_post / (a_post + b_post),
    lower =
      qbeta(0.025, a_post, b_post),
    upper =
      qbeta(0.975, a_post, b_post)
  )
}

sapply(
  c(0, 0.25, 0.50, 0.75, 1),
  posterior_summary,
  y_hist = y_hist,
  n_hist = n_hist,
  y_curr = y_curr,
  n_curr = n_curr
)

This provides a simple sensitivity analysis showing how the posterior changes as historical borrowing increases.

Plotting the Prior and Posterior

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

prior_density <- dbeta(
  p,
  a_prior,
  b_prior
)

posterior_density <- dbeta(
  p,
  a_post,
  b_post
)

plot(
  p,
  prior_density,
  type = "l",
  lwd = 2,
  xlab = "Control response probability",
  ylab = "Density"
)

lines(
  p,
  posterior_density,
  lwd = 2,
  lty = 2
)

legend(
  "topright",
  legend = c("Historical-data prior", "Posterior"),
  lty = c(1, 2),
  lwd = 2
)

The prior represents what the historical information contributes before the current data are observed.

The posterior represents the updated inference after incorporating the current control data.

Dynamic Borrowing With a Grid Over the Power Parameter

A simple way to explore dynamic borrowing is to evaluate several possible borrowing weights.

weights <- seq(0, 1, by = 0.05)

posterior_means <- sapply(
  weights,
  function(w) {

    a_prior <- 1 + w * y_hist
    b_prior <- 1 + w * (n_hist - y_hist)

    a_post <- a_prior + y_curr
    b_post <- b_prior + n_curr - y_curr

    a_post / (a_post + b_post)
  }
)

data.frame(
  borrowing = weights,
  posterior_mean = posterior_means
)

This creates a borrowing sensitivity curve.

Historical Borrowing in a Randomized Trial

Now suppose the experimental treatment arm contains:

The observed treatment response rate is therefore:

The current control arm contains:

with an observed response rate of 30%.

The naive observed risk difference is:

Bayesian borrowing changes the uncertainty around the control parameter and therefore changes the posterior distribution of the treatment effect.

Bayesian Treatment Effect Model

One simple model is:

and:

where \(\pi_C\) is constructed using historical borrowing.

After observing current data:

while the control posterior incorporates both historical and current control information.

Posterior draws can then be used to calculate:

for posterior draw \(s\).

Monte Carlo Calculation of Treatment Benefit

set.seed(123)

S <- 100000

p_treatment <- rbeta(
  S,
  1 + y_curr_treatment,
  1 + n_curr_treatment - y_curr_treatment
)

p_control <- rbeta(
  S,
  a_post,
  b_post
)

delta <- p_treatment - p_control

mean(delta > 0.10)

quantile(
  delta,
  c(0.025, 0.50, 0.975)
)

The first result estimates:

This is often an intuitive quantity for clinical decision-making.

Borrowing Can Increase Precision

Suppose the historical and current control populations are sufficiently comparable.

Borrowing then effectively adds information about the control response rate.

The posterior distribution of \(p_C\) can become narrower than the posterior based only on the current control patients.

This can increase precision for the treatment effect:

may decrease because:

is reduced.

But Precision Is Not the Same as Accuracy

This distinction is fundamental.

A narrower posterior distribution is not necessarily better if it is centered around the wrong value.

If historical controls are systematically different from current controls, aggressive borrowing can create a highly precise but biased inference.

Conflict Scenario

Consider two hypothetical situations.

The appropriate borrowing strategy may be very different between these situations.

Scenario A provides little evidence that the historical and current settings differ.

Scenario B requires investigation and sensitivity analysis.

Historical Borrowing Sensitivity Analysis

A robust analysis should generally examine how conclusions change under different assumptions about historical relevance.

For example:

• No borrowing
• Weak borrowing
• Moderate borrowing
• Strong borrowing
• Robust dynamic borrowing

The investigator should then examine whether the clinical conclusion is stable.

A Useful Sensitivity Table

If the conclusion changes dramatically between these analyses, the dependence on the historical-data assumption should be explicitly discussed.

Historical Borrowing and Sample Size

Historical borrowing can also affect prospective trial design.

Suppose the current trial requires a target precision for the control response rate.

If sufficiently relevant historical information is available, the required number of concurrent control patients may be reduced.

However, sample-size calculations should be based on the actual Bayesian model and decision criterion.

It is not generally sufficient to say:

because the historical observations may be discounted or dynamically weighted.

Bayesian Sample Size by Simulation

A typical simulation-based sample-size workflow is:

What Scenarios Should Be Simulated?

A comprehensive simulation should not assume that the historical data are always perfectly compatible.

For example, consider:

These scenarios reveal whether the proposed borrowing method behaves appropriately when exchangeability is violated.

Bias Under Historical Mismatch

Suppose the true current control response rate is 30%, but the historical response rate is 60%.

If the historical sample is large and receives substantial weight, the posterior for the current control rate can be pulled upward.

This can make the experimental treatment appear less effective if the treatment response is compared against the borrowed control.

The opposite can also occur.

If the historical control rate is lower than the current control rate, excessive borrowing can make the treatment appear more effective than it actually is relative to the current control population.

Borrowing Multiple Historical Studies

Often there is not one historical study but several.

Suppose there are:

• Historical Study 1
• Historical Study 2
• Historical Study 3
• Historical Study 4

A naive approach would pool all observations.

A hierarchical Bayesian model can instead allow study-specific control rates:

where \(j\) indexes historical studies.

The studies then share information through the population-level distribution rather than being assumed to have exactly identical response rates.

Hierarchical Historical Borrowing

A conceptual normal-scale model is:

where:

• \(\theta_j\) = study-specific treatment or control parameter
• \(\mu\) = overall population mean
• \(\tau\) = between-study precision

The amount of shrinkage depends on the between-study heterogeneity.

If studies are highly consistent, they borrow strongly from one another.

If studies are heterogeneous, the model reduces the amount of pooling.

Partial Pooling

This is known as partial pooling.

Compare three approaches:

Bayesian hierarchical models are particularly useful when several historical datasets exist.

Historical Borrowing for Continuous Endpoints

The same principles apply to continuous endpoints.

Suppose the current control mean is:

and the historical control mean is:

A normal-normal model might specify:

with historical information incorporated through the prior distribution.

The same concerns remain:

• Exchangeability
• Temporal drift
• Population differences
• Endpoint compatibility
• Borrowing strength
• Prior-data conflict

Historical Borrowing for Time-to-Event Endpoints

Historical borrowing can also be used for survival endpoints.

For example, historical information may be incorporated into a model for a baseline hazard or regression parameter.

A simplified proportional-hazards representation is:

Historical data might inform:

• Baseline hazard parameters
• Regression coefficients
• Control-arm survival parameters
• Frailty or random-effect distributions

The modeling is considerably more complex than the conjugate Beta-Binomial example, but the conceptual principle remains the same.

Historical Borrowing and Estimands

An important issue is that the historical and current studies must address compatible estimands.

Suppose the current trial targets:

Historical data should provide information relevant to that same quantity.

If the historical study instead measures a different endpoint, population, or treatment strategy, borrowing may not be scientifically justified even if the numbers appear superficially similar.

Historical Borrowing and Missing Data

Historical datasets may have different patterns of missingness than the current trial.

For example, the historical response rate may have been calculated among evaluable patients while the current trial uses an intention-to-treat estimand.

Simply borrowing the historical response count without accounting for the different estimand can create incompatibility.

Historical Borrowing and Prior Predictive Checks

Before analyzing current data, investigators can examine what the historical prior predicts.

For a Beta prior:

the prior predictive distribution for a future sample can be calculated.

This asks:

If the prior predicts implausible current outcomes, the prior may be too strong or poorly constructed.

Prior Predictive Simulation in R

set.seed(123)

S <- 100000

p_prior <- rbeta(
  S,
  a_prior,
  b_prior
)

y_prior_predictive <- rbinom(
  S,
  size = n_curr,
  prob = p_prior
)

quantile(
  y_prior_predictive,
  c(0.025, 0.50, 0.975)
)

This provides a simple diagnostic for whether the historical-data prior implies reasonable numbers of responses in a future sample of the current trial's size.

Posterior Predictive Distribution

After observing the current data, the posterior can be used to predict future control outcomes.

For a future sample of size \(m\):

This allows questions such as:

which can be useful for predictive trial planning and decision-making.

Borrowing and Predictive Probability

Historical information can also be incorporated into predictive probability calculations.

For example, suppose a trial is currently underpowered based only on concurrent controls.

A Bayesian model can calculate:

This can support adaptive decisions such as:

• Continue enrollment
• Stop for futility
• Modify the design
• Declare success

provided those rules were appropriately prespecified and their operating characteristics evaluated.

Common Mistakes

1. Treating historical data as automatically exchangeable. Similarity must be justified rather than assumed.
2. Using all historical observations at full weight without sensitivity analysis. Historical mismatch can produce substantial bias.
3. Choosing a borrowing weight after seeing the current outcome. This can introduce undesirable operating characteristics unless the adaptation is explicitly incorporated into the design.
4. Calling a posterior probability a p-value. They have different interpretations.
5. Assuming a 95% posterior probability implies 5% type I error. That equivalence does not automatically hold.
6. Ignoring historical study differences. Endpoint definitions, eligibility, standard of care, and follow-up can all matter.
7. Reporting the posterior without describing the prior. The posterior depends on both the prior and the likelihood.
8. Reporting an effective sample size as if it were literal enrollment. ESS is a model-dependent measure of information.
9. Failing to simulate prior-data conflict. Historical mismatch should be part of design evaluation.
10. Optimizing only under the assumption of perfect historical compatibility. A borrowing method should also be examined under plausible departures from exchangeability.

A Practical Historical Borrowing Workflow

What Should Be Prespecified?

A Bayesian historical-borrowing analysis should be fully described before the current outcome data are available.

The statistical analysis plan should specify:

• Historical datasets included in the analysis
• Eligibility criteria for historical observations
• Endpoint harmonization rules
• Baseline prior
• Borrowing model
• Hyperpriors
• Borrowing parameters
• Decision criterion
• Posterior probability threshold
• Effective sample-size diagnostics
• Prior predictive checks
• Sensitivity analyses
• Simulation scenarios
• Handling of missing historical data
• Handling of missing current data
• Computational methods
• Convergence diagnostics for MCMC models, when applicable

When MCMC Is Required

The simple Beta-Binomial power-prior example is conjugate and can be solved analytically.

More sophisticated borrowing models are often not.

MCMC may be required when the model contains:

• Commensurate priors
• Hierarchical study effects
• Robust mixture priors
• Multiple historical datasets
• Covariate adjustment
• Time-to-event models
• Nonlinear treatment effects
• Complex missing-data models

In those settings, the posterior is represented by a large collection of simulated draws rather than a closed-form distribution.

MCMC Diagnostics

When MCMC is used, investigators should examine:

• Trace plots
• Effective sample size
• Potential scale reduction statistics
• Autocorrelation
• Monte Carlo error
• Posterior predictive checks

A model that produces apparently plausible posterior estimates but has not converged adequately should not be considered reliable.

Borrowing and Regulatory Considerations

Historical borrowing can be particularly attractive in rare diseases, pediatric populations, and other settings where randomized control enrollment is difficult.

However, the use of historical information requires careful justification.

The statistical documentation should make it possible to understand:

• Where the historical data came from
• Why they are relevant
• How they were incorporated
• How much they were allowed to influence the current analysis
• What happened when historical and current data disagreed
• How operating characteristics were evaluated
• How sensitive conclusions were to borrowing assumptions

Bayesian Borrowing Versus Ignoring Historical Data

The Bias-Variance Tradeoff

Historical borrowing can be understood through the familiar bias-variance tradeoff.

Without borrowing:

With strong appropriate borrowing:

With strong inappropriate borrowing:

The statistical objective is therefore not simply to minimize posterior variance.

A Conceptual Comparison of Borrowing Strategies

What Makes a Good Historical Dataset?

Historical information is particularly attractive when:

• The disease definition is the same.
• The patient population is highly comparable.
• The treatment regimen is the same.
• The control treatment is the same.
• The endpoint is defined identically.
• The assessment schedule is comparable.
• The standard of care has not materially changed.
• The historical study was conducted recently enough to remain relevant.
• The historical analysis population is compatible with the current estimand.

What Makes Historical Data Less Suitable?

Borrowing should be treated cautiously when:

• Eligibility criteria differ substantially.
• Endpoint definitions differ.
• Control treatment changed.
• Supportive care changed.
• The disease population changed.
• Historical follow-up differs substantially.
• There is evidence of strong temporal drift.
• The historical study had substantially different ascertainment procedures.
• The historical population is much more or less severe than the current population.

The Most Important Diagnostic: Does the Conclusion Depend on Borrowing?

Suppose the primary Bayesian analysis concludes:

but the analysis without historical borrowing produces:

That difference is important.

It means the conclusion depends materially on the historical information.

That does not automatically invalidate the Bayesian analysis.

It does mean that the relevance of the historical data becomes especially important to the scientific interpretation.

A Robust Reporting Framework

A transparent report could present:

This makes the contribution of the historical information transparent.

Worked Example Summary

The Most Important Concept

Bayesian borrowing of historical control data is fundamentally about information relevance.

Historical observations can contain valuable information about what the control response rate is likely to be in the current trial.

But their value depends on how exchangeable they are with the current population.

A useful conceptual sequence is:

Bottom Line

References

Ibrahim, J.G. & Chen, M.H. (2000). Power prior distributions for regression models. Statistical Science, 15(1), 46–60.

Hobbs, B.P., Carlin, B.P., Mandrekar, S.J. & Sargent, D.J. (2011). Hierarchical commensurate and power prior models for adaptive incorporation of historical information in clinical trials. Biometrics, 67(3), 1047–1056.

Neuenschwander, B., Capkun-Niggli, G., Branson, M. & Spiegelhalter, D.J. (2010). Summarizing historical information on controls in clinical trials. Clinical Trials, 7(1), 5–18.

Viele, K., Berry, S., Neuenschwander, B., et al. (2014). Use of historical control data for assessing treatment effects in clinical trials. Pharmaceutical Statistics, 13(1), 41–54.

Schmidli, H., Gsteiger, S., Roychoudhury, S., O'Hagan, A., Spiegelhalter, D. & Neuenschwander, B. (2014). Robust meta-analytic-predictive priors in clinical trials with historical control information. Biometrics, 70(4), 1023–1032.

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.