Bayesian Treatment Effect Estimation After Adjustment

Why Bayesian Adjustment Needs an Explicit Model

An external-control comparison can adjust for measured differences between treated trial participants and patients observed in routine care. Bayesian estimation then represents uncertainty about the adjusted contrast with a posterior distribution. This is not a device for making external controls randomized: causal interpretation still depends on whether the measured covariates and the outcome model adequately account for differences between the sources.

Here the target is a treatment contrast conditional on specified covariates, or a common treatment coefficient when the model assumes the contrast is constant across covariate values. The treatment contrast is not automatically an average over the trial population. Defining and estimating such a population-average effect is a separate step. The adjustment methods studied earlier, including outcome regression and doubly robust estimation, motivate the covariates and comparisons; the Bayesian task here is to model uncertainty in the adjusted outcome contrast.

Outcome Model and Identification Assumptions

For a continuous endpoint, a basic model is

where \(Y_i\) is the endpoint for patient \(i\), \(A_i=1\) denotes trial treatment and \(A_i=0\) denotes external control, \(X_i\) is a vector of measured covariates, \(\tau\) is the treatment contrast, and \(\beta_X\) contains covariate coefficients. With this additive model, the conditional mean difference is \(\tau\) at every \(X_i\). The normal likelihood is a working model for the endpoint; the chosen covariate terms and distribution must be suitable for the data and estimand.

A Bayesian analysis assigns a joint prior to the model parameters, for example \(\beta\sim N(b_0,V_0)\), with a prior for \(\sigma^2\) if it is unknown. The posterior is proportional to the likelihood multiplied by the prior. Its treatment-contrast summary must reflect uncertainty in all parameters relevant to the contrast, not merely uncertainty in a point prediction.

The causal interpretation requires more than a well-computed posterior. In particular, it requires consistency of treatment and endpoint definitions across sources, adequate measurement of confounders, conditional exchangeability of treatment and potential outcomes given the modeled covariates, and sufficient covariate support for the comparison. It also requires that the outcome model be adequate for the conditional means being compared. These are substantive assumptions, not consequences of Bayesian computation.

A crucial identification issue arises when the data contain only treated trial participants and external controls. In that design, treatment status and data source are perfectly aligned: every treated patient is from the trial and every control is external. An unrestricted trial-versus-external source effect cannot be estimated separately from the treatment effect. Omitting a source effect therefore amounts to assuming that, conditional on modeled covariates, there is no remaining source difference relevant to the outcome. A prior cannot by itself resolve this lack of information; any posterior separation then depends on constraints or assumptions.

Posterior Calculation and Uncertainty

A useful derivation is available when an adjusted model yields an approximately normal likelihood summary \(d\) for the treatment contrast, with known standard error \(s\). Write that likelihood as \(d\mid\tau\sim N(\tau,s^2)\), and assign \(\tau\sim N(m_0,v_0)\). Multiplying the normal likelihood and prior gives a normal posterior. Completing the square in the exponent gives its variance and mean:

The likelihood summary must come from the adjusted outcome model, not from an unadjusted comparison. This expression also assumes that the likelihood standard error is appropriately estimated and treats it as known. In a full regression with uncertain residual variance, the posterior can instead be sampled jointly over coefficients and variance parameters; the contrast is calculated at each draw and summarized across draws.

For a general linear model with known \(\sigma^2\), design matrix \(Z\), and prior \(\beta\sim N(b_0,V_0)\), the posterior covariance and mean are

If the desired conditional contrast is \(c^\top\beta\), then its posterior mean is \(c^\top b_n\) and its variance is \(c^\top V_n c\). The covariance terms in \(V_n\) matter: coefficients in a contrast may be correlated in the posterior. Discarding those terms can give an incorrect interval.

Contrasts With Interactions and Covariance

If the model includes treatment-by-covariate interactions, the contrast varies with covariates. For example, with a binary severity measure \(X\), a model can include \(\beta_A A+\beta_{AX}AX\). At severity \(x\), the conditional contrast is \(\tau(x)=\beta_A+x\beta_{AX}\). Its posterior uncertainty must use the joint posterior of both coefficients. This is a conditional contrast at a stated covariate value, not a population-standardized effect.

Allowing for Uncertain Source Discrepancy

A primary outcome model can be sensitive to residual differences between the trial and external cohort. A transparent sensitivity model can represent an additive discrepancy \(\delta\) in the adjusted contrast summary: \(d\mid\tau,\delta\sim N(\tau+\delta,s^2)\). The meaning and sign of \(\delta\) must be stated. Assigning it a prior allows uncertainty about this discrepancy to affect the posterior, but does not prove that the prior covers all relevant bias.

Diagnostics, Alternatives, and Limitations

Before interpreting a posterior contrast, examine whether the model is behaving plausibly. Useful checks include posterior predictive comparisons of observed outcomes with replicated outcomes, residual patterns against fitted values and covariates, influential observations, and sensitivity to reasonable changes in priors and model terms. Check whether treated trial patients have comparable external controls across the covariate patterns that inform the contrast. Sparse or absent support means the estimate relies heavily on model extrapolation, even if the posterior interval is numerically narrow. Posterior computation should also be checked for convergence and adequate sampling when simulation is used.

Bayesian outcome regression differs from propensity-score weighting, which constructs a comparison through weights, and from a frequentist regression analysis, which reports sampling-based uncertainty rather than a posterior distribution. Doubly robust estimators combine outcome and treatment-assignment models for particular robustness properties; Bayesian outcome regression alone does not inherit those properties merely by using a prior. These approaches share the need for credible measured-confounding and overlap assumptions. None can correct unmeasured confounding simply by producing an interval.

A central failure mode is source-treatment confounding: if no trial control observations exist, a source difference can mimic a treatment effect. Another is model misspecification, such as an omitted nonlinear covariate relationship that leaves residual confounding. A common mistake is to say that a precise posterior estimate proves the external controls are comparable. The correction is that precision quantifies uncertainty conditional on the model and prior; it does not quantify unmodeled bias unless that bias is explicitly represented.