Outcome Regression With External Controls

Why Model the Outcome?

An external control group can provide information about outcomes under control when a concurrent randomized control group is unavailable or limited. Outcome regression uses measured baseline covariates to model the outcome in relation to treatment and covariates. It then compares the model’s treatment and control predictions at the same covariate values. This adjusts the comparison for the covariates included in the model, but it does not make unmeasured confounding disappear.

The central modeling choice is whether to assume a common treatment contrast across covariate values. If treatment effects or risks may vary, a model with only treatment and covariate main effects can be too restrictive. Treatment-by-covariate terms allow the fitted contrast to change with baseline characteristics. Their interpretation depends on the outcome model’s link function.

Model Specification and Assumptions

For a continuous outcome, a linear mean model is often written as

Here, \(\beta_X^\mathsf{T}X\) represents covariate main effects, and \(A\gamma^\mathsf{T}X\) represents treatment-by-covariate interactions. With an identity link, the conditional mean difference at \(X=x\) is \(\beta_A+\gamma^\mathsf{T}x\). If \(\gamma=0\), the model assumes the conditional mean difference is constant over the modeled covariates. The same idea applies to generalized linear models:

where \(g\) is a chosen link function. For a binary outcome, the logit link models log odds; for a count outcome, a log link can model a rate when an exposure-time offset is included. Interactions then describe variation on the link scale. For example, a treatment-by-covariate interaction in a logistic model means that the conditional log odds ratio varies with that covariate; it does not directly say that the risk difference varies by the same amount.

An external-control analysis also needs assumptions beyond a well-fitted regression. The covariates should be defined and measured comparably across sources, refer to an appropriate baseline, and capture the important measured differences affecting both treatment selection and outcome. Conditional on those covariates, the potential outcomes under the treatment and control strategies must be comparable across the groups being contrasted. There must also be adequate covariate support: predictions for one group should not rely heavily on covariate patterns absent from the other. Finally, the outcome model must represent the conditional outcome relationship adequately for the intended comparison.

In a Bayesian implementation, a likelihood for the selected outcome model is combined with priors on its coefficients to obtain a posterior distribution. Priors should be specified on interpretable scales and should not be used to disguise weak covariate support or a poorly justified model. The details of posterior treatment-effect estimation are a separate step from choosing and interpreting the outcome regression.

Worked Models

Why the Interaction Changes the Contrast

The treatment contrast follows directly by evaluating the model twice at the same covariate value and subtracting on the link scale. In the general model, the linear predictor at treatment \(A=1\) is \(\beta_0+\beta_A+\beta_X^\mathsf{T}x+\gamma^\mathsf{T}x\), while at \(A=0\) it is \(\beta_0+\beta_X^\mathsf{T}x\). Subtracting gives

For an identity link, this difference is a mean difference. For a logit link, it is a log odds ratio, and exponentiating gives the conditional odds ratio. For a log link, exponentiating gives a conditional rate ratio when the modeled mean is a rate. This derivation explains why an interaction must be interpreted on the model’s link scale first. To report an absolute risk difference from a logistic model, calculate the two predicted risks and subtract them, as in the binary example.

Fitting, Comparisons, and Diagnostics

A practical model-building sequence is to define the outcome and baseline covariates first, choose a likelihood and link appropriate to the outcome, and then specify main effects and clinically justified interactions before examining fitted contrasts. Avoid selecting interactions solely because they produce a favorable result. Continuous covariates may need nonlinear terms; a single linear term assumes a particular shape on the model scale. If a treatment effect is expected to vary across a covariate, include the corresponding treatment-by-covariate term and ensure that the relevant covariate patterns are represented in both groups.

Outcome regression differs from matching and weighting in what it models. Regression specifies the outcome conditional on treatment and covariates, then uses that model for adjusted comparisons. Matching and weighting instead construct a comparison by pairing or reweighting participants based on treatment assignment and measured covariates. Each approach can fail if its assumptions or implementation are poor. A regression model may extrapolate when groups have little covariate overlap; weighting can be unstable when estimated weights are extreme; matching can discard participants or leave residual imbalance. Combining approaches is possible, but it does not remove the need to state the target population and assumptions.

Check whether the fitted model is plausible for the observed outcome and covariate patterns. For continuous outcomes, inspect residual patterns and the adequacy of the mean structure. For binary and count outcomes, compare fitted predictions with observed outcome frequencies or rates across meaningful covariate groups, and check for sparse data or separation. Examine whether interaction estimates are supported by enough observations in each treatment-by-covariate pattern. A large coefficient or a precise-looking model output cannot compensate for a comparison based on unsupported extrapolation.

A common mistake is to fit only treatment and covariate main effects, then interpret the treatment coefficient as an effect that must be constant for every patient. That interpretation is true only on the model’s link scale under the no-interaction specification. If the outcome relationship plausibly varies by baseline risk, omitting an interaction imposes a common contrast and may distort subgroup predictions. Conversely, adding many interactions to sparse data can make estimates unstable. Use a limited, clinically motivated specification and report the scale on which each contrast is interpreted.

The main limitation remains substantive: adjustment only addresses differences represented adequately by measured covariates and the model. It cannot establish that external controls are exchangeable with trial participants, repair incompatible outcome measurement, or resolve a treatment-source aliasing problem. Regression is a transparent adjustment strategy, not a substitute for design scrutiny.