Deriving Treatment Contrasts From Longitudinal Mixed Models

From a Longitudinal Model to a Treatment Contrast

A longitudinal model can estimate treatment differences at several visits while accounting for repeated observations within each participant. But the model’s coefficients are not always the treatment effects a study aims to report. With treatment-by-visit interactions, for example, the treatment difference changes over time and must be formed from more than one coefficient. An average across visits requires another step: combine the visit-specific differences with explicit weights and use their covariances to calculate precision.

The target contrast should be specified before it is calculated. Consistent with the earlier material on aligning estimands with model-based endpoint contrasts, the selected visits, treatment comparison, and any averaging weights need to represent the intended clinical question.

Model and Assumptions

For participant \(i\), collect the repeated outcomes in a vector \(Y_i\). A linear mixed-effects model can be written as

Here, \(X_i\) contains the fixed-effect design variables, \(\beta\) is the vector of fixed-effect coefficients, \(Z_i\) is the random-effect design matrix, and \(b_i\) is the participant-specific random-effect vector. The residual vector is \(\varepsilon_i\). For the usual marginal interpretation, assume the design and covariates are exogenous with respect to the random effects and residuals: in particular, \(E(b_i\mid X_i)=0\) and \(E(\varepsilon_i\mid X_i)=0\). These conditions imply \(E(Y_i\mid X_i)=X_i\beta\). Merely assuming \(E(b_i)=0\) is not sufficient if random effects are associated with the design.

For model-based standard errors, a common specification is \(b_i\sim N(0,G)\), \(\varepsilon_i\sim N(0,R_i)\), with \(b_i\) independent of \(\varepsilon_i\), and observations from different participants independent. Then \(Y_i\mid X_i\) has covariance \(V_i=Z_iGZ_i^{\mathsf T}+R_i\). The covariance structure must be appropriate enough for the fitted model to estimate uncertainty reliably. The mean contrast can still be written as a fixed-effect contrast, but a misspecified covariance structure may make its standard error and interval unreliable.

A convenient fixed-effect specification for \(K\) categorical visits uses visit 1 as the reference. Let \(T_i=1\) for treatment and \(T_i=0\) for control, and let \(I(v_{ij}=k)\) indicate visit \(k\). Omitting other covariates for clarity, the mean model is

At visit 1, the treatment-minus-control difference is \(\beta_T\). At visit \(k>1\), it is \(\beta_T+\beta_{TV_k}\): the visit main effect is shared by the two treatment groups and cancels in the difference. This is a marginal mean contrast over the random effects, conditional on any fixed covariates included in the model. If treatment effects vary with covariates, a population-average contrast also requires averaging the model predictions over a specified target covariate distribution.

Why the Covariance Matrix Determines Precision

Let \(\hat{\beta}\) be the estimated fixed-effect vector and let \(\widehat{\mathrm{Cov}}(\hat{\beta})\) be its estimated covariance matrix. Any contrast that is linear in the fixed effects can be written \(c^{\mathsf T}\beta\), where \(c\) is a vector of coefficients. Its estimate is \(c^{\mathsf T}\hat{\beta}\), and its estimated variance and standard error are

For justification, write the contrast estimate as \(\sum_r c_r\hat{\beta}_r\). Expanding its variance gives

The covariance terms matter: coefficient estimates, and contrasts from different visits, are generally not independent. Under a correctly specified linear model with known \(V_i\), the generalized least-squares estimator is \(\hat{\beta}=(X^{\mathsf T}V^{-1}X)^{-1}X^{\mathsf T}V^{-1}Y\), with covariance \((X^{\mathsf T}V^{-1}X)^{-1}\). In practice, covariance parameters are estimated, so software typically reports an estimated fixed-effect covariance matrix; finite-sample degrees-of-freedom methods may also be used. A normal-based 95% interval is approximately estimate \(\pm1.96SE\) when that approximation is suitable.

Averaging Contrasts Across Visits

An across-visit contrast is not automatically the arithmetic mean. The weights define which visits contribute and in what proportion; they should be specified to match the target summary. If \(\hat{\boldsymbol d}\) is the vector of estimated visit-specific contrasts, \(w\) is the weight vector, and \(\widehat{\Sigma}_d\) is the estimated covariance matrix of those contrasts, then

The diagonal entries of \(\widehat{\Sigma}_d\) are visit-specific contrast variances; its off-diagonal entries are covariances between contrasts at different visits. Ignoring the off-diagonal terms is justified only if those covariances are zero, not simply because the contrasts refer to different visits.

Continuous Time, Diagnostics, and Limitations

With continuous time, the same contrast principle applies, but the visit-specific contrast depends on the chosen time value and the assumed time function. For example, a model with treatment and treatment-by-time terms has a treatment difference \(\beta_T+t\beta_{Tt}\) at time \(t\). That expression is a model-based consequence of the specified linear time trend, not a guarantee that the true treatment difference changes linearly.

Several checks help identify problems before interpreting a contrast:

• Check the coefficient coding. Confirm the reference treatment, reference visit, units, and contrast direction. A reversed subtraction changes the sign.
• Check the target weights. Verify that averaging weights sum to one and correspond to the intended visits and target population.
• Check the covariance calculation. The contrast variance must use the full covariance matrix, including pairwise terms. A negative covariance can reduce the variance; it should not be silently removed.
• Check model fit and assumptions. Residual patterns, implausible fitted trajectories, convergence warnings, or covariance estimates at problematic boundaries may indicate that the mean or covariance specification needs review. Estimated standard errors rely on that specification.
• Check the scope of inference. A model-based marginal mean averages over random effects under the stated exogeneity conditions. It does not automatically average over covariates or missing outcomes in the target population.

An alternative is to fit separate models at each visit or to report visit-specific contrasts without an across-visit summary. Those approaches can make visit-level results direct, but separate analyses do not automatically use the repeated-measure structure jointly and may not answer a prespecified average-contrast question. A longitudinal model offers a coherent set of means and covariances, but its contrast inherits the model’s assumptions. In particular, a poorly chosen time trend, incorrect covariance structure, or unsupported extrapolation can make a precisely calculated contrast misleading. A standard error quantifies uncertainty under the fitted model; it does not measure uncertainty about whether the model itself is correct.