Aligning Estimands With Model-Based Endpoint Contrasts

Start With the Clinical Question

A model can estimate a contrast precisely and still answer the wrong clinical question. For example, a difference between treatment-group means among participants who completed follow-up is not automatically the treatment effect in the randomized population. The endpoint, the handling of events such as rescue medication or treatment discontinuation, and the population over which results are summarized all matter. The estimand makes those choices explicit before the model is selected.

An estimand is the precise quantity a trial seeks to estimate. It is not the fitted model, the estimator, or the numerical result. The model-based contrast is an analysis quantity chosen to estimate the estimand under specified assumptions. The essential task here is to map the clinical question to that contrast and to verify that its interpretation preserves the question.

Definition: An estimand is specified by five attributes: (1) the target population, (2) the treatment conditions being compared, (3) the variable or endpoint, (4) the strategy for each relevant intercurrent event, and (5) the population-level summary measure used to compare treatment conditions. An intercurrent event is an event after treatment initiation that affects either the interpretation or existence of the endpoint, such as starting rescue therapy or discontinuing assigned treatment.

The treatment conditions describe what is compared, including dose, schedule, duration, and comparator. The variable states what is measured and when. The summary measure might be a difference in population means, a ratio of means, or a difference in response probabilities. An intercurrent-event strategy specifies how outcomes after an event contribute to the clinical question; it is not simply a rule for deleting observations.

From Estimand to Model Contrast

For a continuous endpoint, let \(Y^a\) denote the outcome that would be observed under treatment condition \(a\), interpreted according to the chosen intercurrent-event strategy. For two conditions, treatment \(A\) and control \(C\), a mean-difference estimand in target population \(P\) is

\Delta = E_P(Y^A)-E_P(Y^C).

This expression states both the comparison and the population summary. If treatment effects vary across participants, the expectation averages those effects over the target population; it is not necessarily the effect for a participant with average covariates. A model-based estimate must therefore either represent that average directly or use predictions averaged over the intended population.

A useful starting point is an identity-link model with treatment indicator \(T\), where \(T=1\) denotes treatment and \(T=0\) control:

E(Y\mid T,X)=\beta_0+\beta_T T+\beta_X X.

Here \(X\) is a baseline covariate. Under this model, the conditional mean difference at any value of \(X\) is \(\beta_T\), because subtracting the control prediction from the treatment prediction cancels the intercept and covariate term. If the model is correctly specified for the target population, the treatment effect is constant across \(X\), and the strategy defining \(Y\) matches the estimand, then the population mean difference is also \(\beta_T\).

When treatment effects vary with \(X\), a single treatment coefficient need not be the population-average contrast. For a model with a treatment-by-covariate interaction, define the predicted contrast at covariate value \(x\) as the treatment prediction minus the control prediction. The target-population estimand is the average of these contrasts over the target distribution of \(X\):

\Delta = E_P\{m(1,X)-m(0,X)\}.

For a finite set of target-population covariate values \(x_j\) with weights \(w_j\), where \(w_j\geq 0\) and \(\sum_j w_j=1\), this becomes \(\Delta=\sum_j w_j\{m(1,x_j)-m(0,x_j)\}\). This is standardization: predict under each treatment condition for the same target population, then take the specified summary. It prevents the treatment contrast from silently depending on an unintended covariate distribution.

Worked Example: A Baseline-Adjusted Mean Difference

Hypothetical trial. Adults with a chronic condition are randomized to regimen A or placebo for 24 weeks. The endpoint is change from baseline in a symptom score, in points, with lower values indicating improvement. The intercurrent-event strategy is treatment policy for rescue medication: the 24-week endpoint is included regardless of rescue use. The summary measure is the mean difference in change in the randomized target population, A minus placebo.

Suppose a correctly specified identity-link model for change is \(E(Y\mid T,X)=\beta_0+\beta_TT+\beta_X X\), where \(X\) is baseline score centered at its target-population mean. Assume no treatment-by-baseline interaction and that endpoint collection and the model represent the treatment-policy outcome. A hypothetical fitted model gives \(\hat\beta_0=-2.0\), \(\hat\beta_T=-3.2\) points, and \(\hat\beta_X=0.25\) points per baseline-score point.

At any baseline value \(x\), the predicted treatment mean is \(-2.0-3.2+0.25x\), and the predicted placebo mean is \(-2.0+0.25x\). Thus the model contrast is

\widehat{\Delta}(x)=(-2.0-3.2+0.25x)-(-2.0+0.25x)=-3.2\text{ points}.

Because the contrast is constant across baseline values, averaging over the target population still gives \(\widehat{\Delta}=-3.2\) points. In this hypothetical analysis, regimen A is estimated to reduce the mean symptom-score change by 3.2 points relative to placebo. The coefficient has the desired interpretation because the treatment effect is constant on this scale and the model outcome follows the stated treatment-policy strategy.

Interactions Require Population Averaging

The equality between a treatment coefficient and the population mean difference depends on the model and the scale. With a treatment-by-baseline interaction, the treatment coefficient is the contrast at the reference value \(X=0\), not necessarily the average across participants. The required contrast is obtained by predicting both treatment conditions for the target population and averaging their differences.

Worked Example: Standardizing an Interaction Model

Hypothetical setting. Use the same continuous endpoint and treatment-policy strategy as above, but suppose the model allows treatment effects to vary with centered baseline score:

E(Y\mid T,X)=\beta_0+\beta_TT+\beta_XX+\beta_{TX}TX.

Suppose the fitted coefficients are \(\hat\beta_T=-2.0\) points and \(\hat\beta_{TX}=-0.60\) points per baseline-score point. The contrast at \(X=x\) is \(-2.0-0.60x\). Assume the target population has 25% of participants at \(x=-2\), 50% at \(x=0\), and 25% at \(x=2\). The treatment contrasts at these values are:

\begin{aligned} \widehat{\Delta}(-2)&=-2.0-0.60(-2)=-0.80,\\ \widehat{\Delta}(0)&=-2.0-0.60(0)=-2.00,\\ \widehat{\Delta}(2)&=-2.0-0.60(2)=-3.20. \end{aligned}

Weighting by the target-population proportions gives

\widehat{\Delta} =0.25(-0.80)+0.50(-2.00)+0.25(-3.20) =-0.20-1.00-0.80=-2.00\text{ points}.

The treatment coefficient \(-2.0\) happens to equal the standardized contrast because the weighted mean of \(X\) is zero. That equality is not guaranteed. If the target weights were 0.50, 0.30, and 0.20 at \(-2,0,2\), respectively, the weighted mean of \(X\) would be \(-0.60\), and the standardized contrast would be \(-2.0-0.60(-0.60)=-1.64\) points. The estimand depends on the target population distribution, so changing its weights changes the population summary even though the fitted coefficients are unchanged.

Intercurrent-Event Strategy Changes the Outcome Being Compared

Consider rescue medication. Under a treatment-policy strategy, outcomes are used regardless of whether rescue medication is started; the estimand compares the assigned treatment conditions as they operate in the trial, including the consequences of rescue use. Under a hypothetical strategy, the question instead concerns outcomes in a scenario in which rescue medication did not occur. These are different outcomes and generally different estimands. A model contrast cannot make them interchangeable merely by adjusting for rescue use.

Worked Example: A Hypothetical No-Rescue Contrast

Hypothetical setting. Investigators want the mean 24-week symptom-score change that would occur if no participant in either randomized group used rescue medication. The target population is all randomized adults, the treatment conditions are regimen A and placebo, the variable is 24-week change in points, the intercurrent-event strategy is hypothetical no rescue, and the summary is the A-minus-placebo difference in population means.

Suppose a model has been specified for the mean outcome under the no-rescue scenario, conditional on baseline stratum \(X\). It predicts the following means: in the low-baseline stratum, \(-4\) points under A and \(-2\) under placebo; in the high-baseline stratum, \(-6\) under A and \(-4\) under placebo. Suppose the target population is 60% low-baseline and 40% high-baseline.

The stratum-specific model contrasts are \(-4-(-2)=-2\) points and \(-6-(-4)=-2\) points. Standardizing to the target population gives

\widehat{\Delta}_{\text{no rescue}} =0.60(-2)+0.40(-2)=-1.20-0.80=-2.00\text{ points}.

This calculation targets the hypothetical estimand only if the predictions identify the outcomes that would have occurred without rescue. That interpretation requires, among other things, a defensible model for outcomes under no rescue, adequate information to predict those outcomes across the target population, and assumptions that address why rescue was started. The arithmetic alone does not establish those assumptions. By contrast, a treatment-policy analysis would use outcomes after rescue as observed and would answer a different question.

Justification and Practical Checks

The standardization formula follows directly from averaging conditional means. For treatment condition \(a\), partition the target population into strata \(j\) with proportions \(w_j\). The population mean under \(a\) is the weighted sum of the stratum-specific means; subtracting the corresponding control mean yields the weighted average of stratum contrasts:

\begin{aligned} E_P(Y^A)-E_P(Y^C) &=\sum_j w_j E_P(Y^A\mid X=x_j)-\sum_j w_j E_P(Y^C\mid X=x_j)\\ &=\sum_j w_j\left\{E_P(Y^A\mid X=x_j)-E_P(Y^C\mid X=x_j)\right\}. \end{aligned}

This derivation explains why both predictions must use the same target-population weights. Using treatment-group-specific covariate distributions instead can mix treatment effects with differences in the populations being averaged. In a randomized trial such distributions may differ by chance, and in adjusted or nonrandomized analyses they may differ systematically.

Before interpreting a model contrast, check that the analysis matches the estimand:

Population: Are predictions averaged over the intended target population, rather than just the observed completers or an implicit reference profile?
Endpoint: Does the modeled variable match the specified measurement, time point, and direction of benefit?
Intercurrent events: Does the analysis use outcomes consistent with the chosen strategy, or does it implicitly switch to a different one?
Model and scale: Does the contrast on the model scale equal the requested summary, or must predictions be transformed and averaged on the outcome scale?
Support and fit: Are the needed predictions supported by observed data, and do residual patterns, influential observations, or poor calibration suggest model inadequacy?

A common mistake is to report a treatment coefficient as the estimand simply because it is labeled “treatment effect.” In an interaction model, it may describe only the effect at the covariate reference value. In a nonlinear model, such as a logistic model, a coefficient on the log-odds scale is not automatically a population-average risk difference. The correction is to write the estimand first, obtain predicted outcomes under each treatment condition, and calculate the specified summary over the target population.

The main alternative to a direct coefficient contrast is model-based standardization of predicted outcomes. A direct coefficient is compact and efficient when its scale and assumptions align with the estimand. Standardization is more transparent when effects vary or when the estimand is marginal, but depends on a suitable prediction model and target-population weights. Neither approach repairs a poorly defined endpoint or an unsupported intercurrent-event assumption. Model-based contrasts also remain limited by misspecification, limited data in parts of the target population, and uncertainty about hypothetical scenarios.

Key takeaway: A model contrast estimates an estimand only when its population, treatment conditions, endpoint, intercurrent-event strategy, and summary measure agree with the clinical question. When effects vary, predict under each treatment condition and average the contrast over the target population.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A trial compares two regimens on a 12-week mean endpoint, using all observed outcomes after treatment discontinuation. Which estimand attribute does this event-handling decision specify?
A model has treatment contrast \(-1.5-0.4X\), and the target-population mean of \(X\) is 1. What is the standardized mean contrast?
Why might a treatment coefficient in an interaction model fail to equal a population-average treatment effect?
Explain why deleting everyone who uses rescue medication does not, by itself, estimate a hypothetical no-rescue effect.

Solutions. (1) It specifies the intercurrent-event strategy; using outcomes regardless of discontinuation is treatment-policy handling. (2) Averaging the contrast gives \(-1.5-0.4E_P(X)=-1.5-0.4(1)=-1.9\) outcome units. (3) The coefficient gives the contrast at the model’s reference value \(X=0\); the population average also depends on the target distribution of \(X\). (4) Rescue use may be related to prognosis and treatment response. Removing those participants changes the analyzed population and can induce selection; it does not generally predict their outcomes in the scenario where rescue was prevented.