Covariate-Adjusted Historical Control Borrowing

Why Adjust for Covariates Before Borrowing?

Historical and current control groups can differ in their mix of prognostic characteristics. If those characteristics predict outcome, a difference between the groups’ crude averages may reflect that mix rather than a change in outcomes conditional on covariates. Borrowing historical outcomes without accounting for this distinction can pull an estimate toward an inappropriate benchmark. Covariate adjustment makes the intended comparison explicit: it models outcomes at the same covariate values, then discounts the historical likelihood.

The model does not make historical data exchangeable with current data by itself. It relies on a defensible conditional relationship between covariates and outcomes, and on adequate representation of the relevant covariate values in the historical data. A power-prior discount can reduce the influence of historical information; it cannot repair a poorly chosen covariate set or justify extrapolation.

Model and Assumptions

Let \(y_{Hi}\) and \(y_{Cj}\) denote outcomes for historical and current control participants, and let \(x_{Hi}\) and \(x_{Cj}\) be their covariate vectors. A standard starting point for a continuous outcome is

Here, \(x\) can include an intercept and prespecified prognostic covariates, and \(\beta\) is the common regression-coefficient vector. For the conjugate calculation below, \(\sigma^2\) is treated as known. In practice, covariates should be chosen using clinical knowledge and specified before examining outcome differences that might motivate borrowing. The model assumes independent observations conditional on the covariates and coefficients, a correctly specified conditional mean, and a common conditional outcome relationship across historical and current controls.

A common relationship is a substantive transportability assumption, not a consequence of fitting a regression. It may be implausible if treatment standards, eligibility criteria, measurement procedures, or unmeasured prognostic factors changed. The covariates also need sufficient overlap: the historical data should include participants whose covariate profiles cover those relevant to the current trial. If current participants occupy a region absent from the historical data, the model’s prediction there is extrapolation.

How the Discount Enters the Posterior

Write \(X_H\) and \(X_C\) for the design matrices and \(y_H\) and \(y_C\) for the outcome vectors. For fixed \(a\), the posterior density for \(\beta\) is proportional to its prior density times the current likelihood and the discounted historical likelihood:

This fixed-discount expression is the relevant power-prior update; the earlier material on power priors gives the broader framework. With normal outcomes and known \(\sigma^2\), the historical likelihood is proportional to \(\exp\{-\tfrac{1}{2\sigma^2}(y_H-X_H\beta)^{\mathsf T}(y_H-X_H\beta)\}\). Raising it to \(a\) multiplies its quadratic and linear contributions by \(a\). Thus, historical information is discounted in the regression likelihood, not by first treating a crude historical mean as if it were a current observation.

Suppose the prior is \(\beta\sim N(m_0,V_0)\), with proper covariance matrix \(V_0\). Completing the square gives a multivariate normal posterior with precision matrix \(\Lambda\) and mean \(m\):

To see why, expand each squared residual. The terms involving \(\beta\) in the log posterior are

Collecting quadratic and linear terms yields \(-\tfrac12\beta^{\mathsf T}\Lambda\beta+\beta^{\mathsf T}\Lambda m\), which is the kernel of \(N(m,\Lambda^{-1})\). The current data contribute at full weight; each historical contribution is multiplied by \(a\). At \(a=0\), historical observations contribute no likelihood information. At \(a=1\), they contribute at full likelihood weight under the shared model. These endpoints do not remove the need to assess model fit or compatibility.

Worked Posterior Update

The next hypothetical calculation uses a simplified normal regression with two covariate-basis coefficients, a known reference mean of 50 points, and known residual variance \(\sigma^2=4\) score-points squared. The coefficients describe deviations from that reference mean. The two design columns are orthogonal, so the information matrices are diagonal. A proper prior has mean \((0,0)^{\mathsf T}\) and precision matrix \(\operatorname{diag}(0.01,0.04)\). For this example, the historical cross-product matrix is \(\operatorname{diag}(3,3)\), the current cross-product matrix is \(\operatorname{diag}(2,2)\), and \(a=0.5\). Suppose the corresponding score vectors are \(X_H^{\mathsf T}(y_H-50)/\sigma^2=(1.2,-0.6)^{\mathsf T}\) and \(X_C^{\mathsf T}(y_C-50)/\sigma^2=(1.2,0.8)^{\mathsf T}\). These are hypothetical sufficient-statistic summaries for the regression update.

Diagnostics, Alternatives, and Limitations

Before interpreting a borrowed estimate, inspect whether the shared regression is credible in both sources. Useful checks include plotting outcomes against continuous covariates by source, examining residual patterns, checking influential observations, and comparing covariate ranges and combinations across sources. For binary outcomes, the analogous conditional model may be logistic; its powered historical likelihood generally does not yield the normal-model conjugate update shown above. Whatever the outcome model, a poor fit in either source makes the shared conditional relationship questionable.

One direct diagnostic is to compare the estimated conditional relationships across sources before imposing a common coefficient. Large differences may indicate changes in care, outcome ascertainment, or effect modification. Such differences do not automatically identify their cause, but they are a warning against uncritical borrowing. Sensitivity analysis over plausible values of \(a\) is also informative: if conclusions change substantially over a modest range of discounts, the result depends materially on historical information. A discount cannot compensate for severe lack of covariate overlap, since the model may still rely on unsupported extrapolation.

The main alternative is unadjusted borrowing, which combines historical and current outcomes without conditioning on prognostic covariates. It is simpler but can confound source differences in covariate mix with outcome differences. Regression adjustment instead compares outcomes conditional on included covariates, but depends on the chosen functional form and common relationship assumptions. A stratified approach can address categorical covariates by separate groups, but may become unstable when many covariates or sparse groups are involved. Neither regression nor stratification guarantees comparability; each requires adequate data and careful specification.

A further limitation is that a model with a shared \(\beta\) rules out source-specific covariate effects unless those are explicitly added. Adding source interactions may reveal departures from the shared model, but increases model complexity and may be weakly identified when historical and current covariate ranges differ. Covariate adjustment also cannot control differences in unmeasured prognostic factors. The model is most defensible when covariates are measured consistently, are prognostic, are prespecified, and have substantial overlap between sources.