Outcome Regression for Covariate-Adjusted Allocation

Why Model the Outcome?

When allocation uses baseline covariates, a treatment’s observed response rate or mean can reflect both treatment and differences in the patients assigned to it. An outcome regression gives a structured way to estimate responses while accounting for measured baseline characteristics. The central task here is to specify a model for the response and use it to calculate what each treatment’s expected response would be for the same covariate profile.

Let \(A_i\) denote the treatment assigned to participant \(i\), with \(A_i=1\) for treatment and \(A_i=0\) for control, and let \(X_i\) denote the participant’s baseline covariates. Let \(Y_i\) be the observed outcome. A treatment-specific conditional mean is the expected outcome at a given profile \(x\) under treatment \(a\). It is estimated by fitting an outcome model and substituting \(A=a\) and \(X=x\). This substitution is a prediction; it is not an observed outcome for that participant under both treatments.

Models for Continuous and Binary Responses

For a continuous response, a linear model specifies the conditional mean as a linear function of treatment and covariates. For example, with one covariate \(X\), a model without treatment-by-covariate interaction is

Here, the treatment-specific fitted means are \(\hat m_0(x)=\hat\beta_0+\hat\beta_Xx\) and \(\hat m_1(x)=\hat\beta_0+\hat\beta_A+\hat\beta_Xx\). This model assumes a common treatment contrast: at every covariate value, the fitted difference between treatment and control is \(\hat\beta_A\). If the treatment contrast may depend on \(X\), include an interaction term, such as \(\beta_{AX}AX\). Then the contrast is \(\beta_A+\beta_{AX}x\), rather than a constant.

For a binary response coded \(Y=1\) when the specified event occurs and \(Y=0\) otherwise, logistic regression models the probability of that event. The logit link is \(\log\{p/(1-p)\}\), where \(p\) is the event probability:

The inverse-logit function converts a fitted linear predictor \(\eta\) into a probability: \(1/(1+\exp(-\eta))\). Thus, for a binary response, first substitute a treatment and profile into the linear predictor, then apply the inverse logit. A coefficient in a logistic model describes a change in log odds, not a change in probability. The corresponding probability change depends on the starting risk and the covariate profile.

Assumptions and the Meaning of Predictions

A fitted model is useful only in relation to its assumptions. The covariates used for adjustment should be measured before treatment assignment. The model’s functional form should be a reasonable description of how the conditional mean (or, for logistic regression, the log odds) varies with those covariates and treatment. In particular, an interaction should be included when the intended model allows treatment responses to vary across covariate profiles. Predictions for profiles far outside the observed data require extrapolation and may be unreliable.

For a randomized trial, interpreting a treatment-specific model prediction as an expected response under assignment also relies on the trial’s treatment assignment and outcome measurement being well-defined. In an adaptive trial, assignments can depend on previously observed information; the relevant randomization condition is that, given the design history and baseline information used at assignment, the current assignment is randomized with a known, nonzero probability for each treatment. For a causal interpretation, consistency and no interference are also needed: each participant’s observed outcome corresponds to the treatment received, and one participant’s assignment does not alter another’s outcome. These conditions do not guarantee that the regression model is correct. For a causal interpretation of the stated regression \(E(Y\mid A=a,X=x)\), history must be ignorable given \(X\); otherwise, history related to potential outcomes must be included or otherwise appropriately accounted for in estimation and standardization, with positivity.

A model-based estimate for a particular profile answers a conditional question. To describe a population with a mix of profiles, average the treatment-specific predictions over a stated covariate distribution. This is called standardization. The distribution might be the covariate profiles in the trial or a separately specified target population. The choice matters: two populations with different covariate mixes can have different standardized responses even when their profile-specific predictions are the same.

From Profile Predictions to a Population Estimate

For a finite set of \(K\) profiles with specified weights \(w_k\), where \(w_k\geq 0\) and the weights sum to one, the standardized response estimate under treatment \(a\) is the weighted average of the model predictions. The treatment contrast is the difference between the two standardized estimates. The calculation follows directly from averaging the predicted response at each profile:

This is a short but important justification: every profile receives the same weight when predicting under treatment and control, so the comparison refers to one common covariate distribution. If the weights instead differed by treatment, the contrast could mix treatment differences with differences in profile composition. For continuous covariates, standardization uses the same idea, averaging predictions over the chosen distribution rather than over a finite list.

For binary outcomes, average the predicted probabilities on the probability scale. Averaging linear predictors or log odds and then applying the inverse logit generally gives a different result from averaging the profile-specific probabilities. The target is an average event probability, so the profile probabilities must be calculated first.

Checks, Alternatives, and Limitations

Before using fitted responses, check whether the model can support the requested predictions. For a linear model, inspect whether the chosen covariate form and interaction structure are plausible for the outcome and observed range. For logistic regression, check that fitting converged and that the data contain enough outcome events across the treatment and covariate combinations being modeled. Extremely sparse combinations can produce unstable coefficients and probabilities near zero or one. In either model, predictions for profiles with little or no representation in the trial are extrapolations, not well-supported summaries.

Regression is not the only way to obtain treatment-specific estimates. Unadjusted arm means or event proportions are simple and avoid choosing a covariate functional form, but they can be imprecise or reflect different covariate mixes in the arms. Separate estimates within well-populated covariate strata are more transparent about profile differences, but can be noisy when many covariates create many small cells. Regression shares information across profiles through its specified structure; that can improve precision when the structure is reasonable, but can mislead when it is not.

A further limitation is that model-based estimates depend on the response scale and the model specification. In logistic regression, an odds ratio is not a probability difference, and the same odds ratio can correspond to different probability differences at different baseline risks. If covariate effects or treatment effects are misspecified, predictions may be systematically wrong. Outcome regression is therefore a tool for constructing conditional and standardized response estimates, not a substitute for checking data support and the scientific plausibility of the model.

Common Mistakes

• Reading a regression coefficient as a response probability. In a logistic model, coefficients act on log odds. Apply the inverse logit to the full linear predictor to obtain a probability.
• Assuming treatment effects are constant without checking the model. A model without an interaction imposes a common contrast on its modeled scale. An interaction changes that assumption, and in a logistic model probability differences can vary across profiles even without an interaction.
• Averaging on the wrong scale. To estimate an average binary response, calculate each profile’s predicted probability and then average those probabilities using the chosen weights.
• Predicting outside the data and treating the result as established. A numerical prediction can be computed for an unobserved profile, but it relies on extrapolation and may be unstable.