Treatment-Arm Regression and Reference-Arm Borrowing

Why the Covariate Regression Matters

Reference-based imputation specifies how the distribution of an outcome after dropout relates to outcomes in a reference arm. When a baseline covariate predicts that outcome, the imputed value depends not only on which arm supplies the reference distribution but also on the regression used to relate the outcome to the covariate. A separate slope for each arm and a common slope across arms can therefore produce different predictions for the same participant.

Consider a single post-baseline outcome \(Y\), a baseline covariate \(X\), and randomized arm \(A\), where \(A=R\) denotes the reference arm and \(A=T\) the treatment arm. For a treatment-arm participant who drops out, a reference-based rule uses a reference-arm conditional distribution for the post-dropout outcome. The choice between separate and shared slopes concerns how that conditional distribution is estimated. It does not, by itself, change the rule that the post-dropout distribution borrows from arm \(R\).

For several covariates, replace \(x\) by a vector \(\mathbf{x}\) and the scalar slope by a vector of regression coefficients. The same distinction remains: arm-specific models allow each coefficient to vary by arm, whereas shared-slope models constrain the covariate coefficients to be equal across arms. This tutorial uses one covariate to make the consequences visible.

Two Regression Specifications

The arm-specific model is

For a post-dropout treatment-arm outcome whose distribution borrows from the reference arm, the conditional mean is \(\alpha_R+\beta_R x\). The treatment-arm slope \(\beta_T\) describes the observed treatment-arm regression; it does not replace \(\beta_R\) in that reference-based prediction.

The shared-slope model is

Here the reference-based mean is \(\alpha_R+\beta x\). The common slope may be estimated using information from both arms, while the intercept for the prediction remains the reference-arm intercept. Thus, “shared slope” does not mean “use the treatment-arm mean after dropout,” and “reference borrowing” does not necessarily mean that every regression coefficient was estimated from reference-arm observations alone.

These equations describe conditional means, not complete predictive distributions. An imputation procedure also needs a suitable conditional residual distribution and its uncertainty. The present comparison focuses on the mean regression; it does not imply that residual variances or other distributional features must be shared.

Assumptions and the Common-Slope Derivation

Both specifications assume that a linear conditional mean is adequate over the covariate values used for prediction. Estimating the regression from observed outcomes additionally requires an appropriate observed-data analysis: for example, dropout must be ignorable for that regression after conditioning on the covariates and any observed history included in the model, or the estimation must otherwise account for the dropout process. A regression fitted to observed cases does not, by itself, identify the unobserved post-dropout outcomes.

The shared-slope model makes a further substantive restriction: at a given \(x\), the expected change in \(Y\) for a one-unit increase in \(X\) is the same in both arms. Arm-specific intercepts permit different arm means at a given covariate value, but cannot represent different linear covariate effects. Both models also require enough covariate variation to estimate their slopes. Reference-based prediction at covariate values far outside the reference arm’s observed range is extrapolation, regardless of which model is chosen.

A simple least-squares derivation clarifies what pooling slopes means. Let \(\bar{x}_a\) and \(\bar{y}_a\) be the observed means in arm \(a\), and define \(S_{xx,a}=\sum_{i:A_i=a}(x_i-\bar{x}_a)^2\) and \(S_{xy,a}=\sum_{i:A_i=a}(x_i-\bar{x}_a)(y_i-\bar{y}_a)\). With an intercept for each arm, the fitted intercept for any proposed common slope \(\beta\) is \(\bar{y}_a-\beta\bar{x}_a\). Substituting these intercepts into the least-squares criterion leaves the within-arm residual sum of squares as a function of \(\beta\). Differentiating and setting the derivative to zero gives the common-slope estimate below.

The estimate combines within-arm covariate-outcome information; it does not use differences between arm means as if those differences were covariate effects. In contrast, the separate-slope estimates are \(\hat{\beta}_a=S_{xy,a}/S_{xx,a}\), provided \(S_{xx,a}>0\) in each arm. This derivation is for ordinary least squares with the stated linear mean model. Other estimation procedures may use different weights or likelihoods.

Comparing the Choices and Checking Their Fit

Arm-specific slopes are more flexible: they can represent different covariate-outcome relationships in the two arms. Their cost is that each slope is estimated from one arm’s information, which can make estimates imprecise when an arm is small or has little covariate variation. A shared slope uses information across arms and may improve precision if the equality restriction is reasonable. If the true slopes differ, however, pooling can distort the reference-arm relationship that supplies the post-dropout prediction.

A useful diagnostic is to examine arm-specific residual plots and fitted relationships, particularly whether the covariate trend appears systematically different by arm. An interaction model, with an arm-by-covariate term, expresses the separate-slope specification and can help assess whether the data show evidence against a common slope. A failure to detect a difference is not proof that the slopes are equal: limited sample size or restricted covariate variation may conceal real heterogeneity.

Also inspect the covariate ranges and distributions in both arms. If a treatment-arm participant’s \(X\) is rarely or never observed in the reference arm, the reference-based mean depends on extrapolating the reference regression. A shared slope does not solve this lack of support; it substitutes a pooled trend and still requires that trend to be credible at the target value. High leverage, strong curvature, or residual patterns can likewise make either linear specification unreliable.

Common Mistake and Practical Interpretation

A common mistake is to reason, “The post-dropout distribution borrows from the reference arm, so the regression slope must be estimated only in that arm.” That conclusion does not follow. The reference-based restriction determines which arm’s conditional distribution supplies the post-dropout prediction; a shared-slope specification can estimate the covariate slope using both arms while retaining a reference-arm intercept. Conversely, a shared slope does not make the post-dropout distribution a treatment-arm distribution.

The choice is a modeling assumption, not a mechanical consequence of randomization. State it explicitly, report which arm information contributes to each coefficient, and check whether the fitted conditional relationships and covariate support make the resulting predictions plausible. Neither greater flexibility nor greater precision automatically makes a model preferable: flexibility can increase uncertainty, while pooling can introduce bias if the common-slope restriction is wrong.