Historical Controls in Randomized Trial Augmentation

Why Separate the Historical and Randomized Information?

A randomized trial estimates a treatment contrast by comparing outcomes in participants assigned to treatment and control under the trial’s randomization. When a trial has a relatively small control arm, relevant historical control data may improve precision about the control outcome risk. The key modeling choice is where that historical information enters: it informs the control-arm parameter, but it does not become treatment-arm data and does not replace the randomized comparison.

This distinction matters for interpretation. The randomized treatment-arm outcomes describe participants assigned to treatment in the current trial. Historical controls can inform the current control risk only under an assumption that the historical outcome process is sufficiently comparable to the current control outcome process. Discounting limits the strength of that assumption’s influence; it does not make incompatible studies compatible.

Definition: In a randomized trial with binary outcomes, let \(Y_T\) and \(Y_C\) denote the numbers of participants with the outcome in the current treatment and control arms, with arm sizes \(n_T\) and \(n_C\). Let \(p_T\) and \(p_C\) be the corresponding outcome probabilities. For historical control data, let \(Y_H=y_H\) outcomes be observed among \(n_H\) participants. A fixed power-prior discount \(a\), with \(0\leq a\leq 1\), weights the historical control likelihood by raising it to the power \(a\).

Model and Assumptions

For clarity, consider a binary endpoint and one historical control dataset. The current randomized-arm sampling model is

Y_T\mid p_T\sim\operatorname{Binomial}(n_T,p_T), \qquad Y_C\mid p_C\sim\operatorname{Binomial}(n_C,p_C).

The historical control outcomes are modeled as \(Y_H\mid p_C\sim\operatorname{Binomial}(n_H,p_C)\) for the purpose of borrowing. This representation says that the historical likelihood concerns the same control outcome probability as the current control arm. In practice, that is a substantive transportability assumption, not a consequence of randomization.

Use independent beta baseline priors \(p_C\sim\operatorname{Beta}(\alpha_C,\beta_C)\) and \(p_T\sim\operatorname{Beta}(\alpha_T,\beta_T)\). The fixed-discount power prior for \(p_C\) is proportional to its baseline prior multiplied by the historical likelihood raised to \(a\). The treatment prior is not multiplied by the historical likelihood. Assume independent outcome sampling across arms and studies conditional on the probabilities, and independent baseline priors for \(p_T\) and \(p_C\). The binomial model also presumes that each arm’s participants have a common outcome probability under the model.

The common-\(p_C\) assumption may be questionable if populations, background care, follow-up, endpoint definitions, or calendar periods differ. A discount is a way to temper a borrowed likelihood under the model; it does not itself account for those differences. The earlier material on power priors develops the discount’s role and interpretation.

Deriving the Control-Arm Update

Ignoring factors that do not depend on \(p_C\), the historical binomial likelihood is proportional to \(p_C^{y_H}(1-p_C)^{n_H-y_H}\). Raising it to \(a\) gives a contribution with exponents \(a y_H\) and \(a(n_H-y_H)\). The current control likelihood contributes exponents \(y_C\) and \(n_C-y_C\), and the beta baseline prior contributes \(\alpha_C-1\) and \(\beta_C-1\). Adding the exponents gives the posterior:

\begin{aligned} \pi(p_C\mid y_C,y_H) &\propto p_C^{\alpha_C-1}(1-p_C)^{\beta_C-1} p_C^{y_C}(1-p_C)^{n_C-y_C} \left[p_C^{y_H}(1-p_C)^{n_H-y_H}\right]^a\\ &=p_C^{\alpha_C+y_C+a y_H-1} (1-p_C)^{\beta_C+n_C-y_C+a(n_H-y_H)-1},\\ p_C\mid y_C,y_H &\sim\operatorname{Beta}\left(\alpha_C+y_C+a y_H,\, \beta_C+n_C-y_C+a(n_H-y_H)\right). \end{aligned}

This derivation relies on \(a\) being fixed. For the treatment arm, no historical term is present:

p_T\mid y_T\sim \operatorname{Beta}\left(\alpha_T+y_T,\, \beta_T+n_T-y_T\right).

Under the independent-prior and sampling assumptions above, the joint posterior factorizes into the treatment posterior and the control posterior. Thus the historical data change the posterior for \(p_C\), but do not directly change the posterior for \(p_T\). A treatment contrast, such as \(p_T-p_C\), can change because it includes \(p_C\).

Worked Example: Updating the Control Risk With Discounted Data

Suppose the current randomized trial observes an adverse outcome in 18 of 50 control participants. A hypothetical historical control dataset observes the outcome in 60 of 100 participants. Use independent uniform baseline priors, \(\operatorname{Beta}(1,1)\), and a fixed discount \(a=0.25\). Assume for this calculation that the historical and current control outcomes are sufficiently comparable for the shared-\(p_C\) model.

The current control likelihood contributes 18 outcomes and \(50-18=32\) non-outcomes. The discounted historical contributions are \(0.25(60)=15\) outcomes and \(0.25(100-60)=10\) non-outcomes. Therefore,

p_C\mid y_C=18,y_H=60 \sim \operatorname{Beta}(1+18+0.25(60),\,1+32+0.25(40)) =\operatorname{Beta}(34,43).

The posterior mean control risk is \(34/(34+43)=34/77=0.4416\), or about 44.2%. This is a posterior summary under the stated model, not a claim that 25 historical participants were enrolled in the randomized trial. The weighted historical likelihood contributes fractional information in general; here the weighted exponents happen to be integers.

Keeping the Randomized Treatment Likelihood Distinct

Suppose the same hypothetical trial observes an adverse outcome in 12 of 50 treatment participants. With a uniform baseline prior, its posterior uses only those randomized treatment outcomes:

p_T\mid y_T=12\sim\operatorname{Beta}(1+12,\,1+38) =\operatorname{Beta}(13,39).

Its posterior mean is \(13/52=0.2500\), or 25.0%. Under the independent posteriors in this example, the posterior mean risk difference is \(E(p_T-p_C\mid\text{data})=0.2500-0.4416=-0.1916\). Thus the posterior mean treatment risk is 19.2 percentage points lower than the posterior mean control risk. This summary does not alone describe the uncertainty in the difference; that uncertainty depends on both posterior distributions.

Worked Example: What Changes When Historical Controls Are Added?

Continue with the hypothetical data above. Compare the control posterior with and without historical borrowing, keeping the randomized treatment data and its prior unchanged.

With no borrowing, set \(a=0\). The control posterior is \(\operatorname{Beta}(1+18,1+32)=\operatorname{Beta}(19,33)\), with mean \(19/52=0.3654\). With \(a=0.25\), the posterior is \(\operatorname{Beta}(34,43)\), with mean \(34/77=0.4416\). The historical controls shift the control posterior mean upward because their observed event proportion, 60.0%, is higher than the current control proportion, 36.0%.

In both analyses, the treatment posterior remains \(\operatorname{Beta}(13,39)\), with mean 25.0%. Hence the mean risk difference is \(-0.1154\) without borrowing and \(-0.1916\) with the stated borrowing. Historical controls alter the control-side inference and therefore the contrast; they do not alter the treatment-arm likelihood. These different posterior summaries reflect the different assumptions and information used, not a change in the randomized trial’s observed treatment outcomes.

Assumptions, Diagnostics, and Alternatives

Before borrowing, assess whether the historical data plausibly represent the outcome process for current control participants. Relevant questions include whether the endpoint and follow-up are comparable, whether background care and eligibility differ, and whether the historical outcome rate is compatible with the current control data. These checks do not prove exchangeability, but they can expose serious conflicts with the model.

A practical diagnostic is to compare historical and current control outcome rates, accounting for their sample sizes and uncertainty rather than treating a raw difference as conclusive. In the example, \(60/100=0.600\) historically and \(18/50=0.360\) currently. This visible difference should prompt investigation of design and population comparability and sensitivity to \(a\); discounting does not establish that the difference is harmless. A further diagnostic is to report how the control posterior and treatment contrast change across defensible discount choices. Strong dependence on \(a\) signals that conclusions are sensitive to borrowing assumptions.

Three approaches clarify what augmentation does and does not accomplish:

Approach	Use of historical controls	Main implication
No borrowing	Historical likelihood is omitted from the current control update.	Current randomized control outcomes alone inform \(p_C\).
Discounted augmentation	Historical control likelihood is raised to a chosen power \(a\) and contributes only to \(p_C\).	Historical information can improve precision or shift the control inference, subject to comparability.
Naïve pooling	Historical and current controls are treated as if they were an undifferentiated sample from one process.	It effectively gives historical data untempered influence and can mislead when the studies differ.

Discounted augmentation is not equivalent to increasing the randomized sample size: the historical participants were not randomized in the current trial, and the treatment likelihood still contains only current treatment outcomes. Nor does borrowing turn an observational historical comparison into a randomized one. The model’s treatment contrast combines randomized treatment information with an external assumption about current control risk.

Worked Example: A Sensitivity Check Across Discounts

Using the same hypothetical control data and uniform prior, consider \(a=0\), \(a=0.25\), and \(a=0.50\). The control posterior means are calculated from the beta parameters:

\begin{aligned} a=0:\quad&\operatorname{Beta}(19,33),& E(p_C\mid\text{data})&=19/52=0.3654,\\ a=0.25:\quad&\operatorname{Beta}(34,43),& E(p_C\mid\text{data})&=34/77=0.4416,\\ a=0.50:\quad&\operatorname{Beta}(49,53),& E(p_C\mid\text{data})&=49/102=0.4804. \end{aligned}

For \(a=0.50\), the discounted historical contributions are 30 events and 20 non-events, which combine with the current 18 events and 32 non-events and the uniform prior to give \(\operatorname{Beta}(49,53)\). The posterior mean increases as more weight is assigned to historical data with a higher event rate. Since the treatment posterior mean remains 0.2500 in all three calculations, the corresponding posterior mean risk differences are \(-0.1154\), \(-0.1916\), and \(-0.2304\). These calculations show sensitivity; they do not identify which discount is scientifically appropriate.

Common Pitfalls and Limits

Counting historical participants as randomized participants. The power-prior contribution is weighted likelihood information about \(p_C\), not a change to \(n_C\), \(n_T\), or the trial’s randomization.
Letting historical controls inform the treatment likelihood. Under this setup, historical outcomes enter the control update only. They can affect a treatment-control contrast through \(p_C\), but not the posterior update for \(p_T\).
Assuming discounting fixes incompatibility. A smaller \(a\) reduces historical influence under the specified model; it cannot repair a mismatched endpoint, population, or care process.
Choosing a discount only because it gives a preferred conclusion. Specify and justify the discount using design knowledge and sensitivity analysis, rather than selecting it to produce a desired posterior contrast.

The beta-binomial calculation is transparent, but the same conceptual separation applies in richer models: the randomized treatment likelihood remains associated with current treatment participants, while borrowing is targeted to the control outcome model. The formulas here do not account for covariates, multiple strata, varying historical study effects, or uncertain discount parameters. Such extensions require their own models and assumptions; the simple shared-\(p_C\) update should not be treated as automatically adequate for them.

Key takeaway: Historical controls can augment inference about the randomized trial’s control risk through a discounted historical likelihood. Keep the randomized treatment-arm likelihood distinct, state the transportability assumption, and show how conclusions respond to the borrowing choice.

Check Your Understanding

Answer each question using the model and hypothetical data in this tutorial. Brief solutions follow the questions.

With \(y_C=10\), \(n_C=40\), \(y_H=30\), \(n_H=60\), a uniform baseline prior, and \(a=0.20\), what is the posterior distribution for \(p_C\)?
In the same trial, does adding historical control data change the posterior distribution for \(p_T\) under the independent-prior model? Why?
For the data in Worked Example 1, explain why the historical event rate shifts the control posterior mean upward when \(a=0.25\).
What is one substantive assumption required to interpret historical controls as information about current \(p_C\)?
Why is it incorrect to describe discounted historical participants as additional randomized control participants?

Solutions. (1) The weighted historical contributions are \(0.20(30)=6\) events and \(0.20(30)=6\) non-events. Therefore \(p_C\mid\text{data}\sim\operatorname{Beta}(1+10+6,1+30+6)=\operatorname{Beta}(17,37)\). (2) No. The treatment posterior uses the current treatment likelihood and its prior; the historical likelihood is applied only to \(p_C\). (3) The historical event rate is \(60/100=0.600\), above the current control rate \(18/50=0.360\), so its weighted likelihood pulls the control update toward a higher event risk. (4) Historical and current control outcomes must be sufficiently comparable to be modeled as sharing a control outcome probability, possibly conditional on modeled factors. (5) Historical participants were not assigned by the current trial’s randomization, and discounting weights their likelihood contribution rather than changing the randomized arm sizes.