Continuous Efficacy Outcomes with Repeated Measurements

Why Repeated Continuous Outcomes Need Their Own Model

Blood pressure, symptom scores measured on a continuous scale, and biomarker concentrations are often recorded at baseline and at several visits after treatment begins. These outcomes provide a trajectory, not just a single endpoint. Measurements from one participant are generally correlated: a participant with a relatively high blood pressure at one visit may also tend to have a relatively high value at the next. Treating all measurements as independent can therefore misstate uncertainty and give misleading inferences about efficacy.

A longitudinal model separates two questions. The mean model describes how the expected outcome changes with time, baseline status, and exposure. The covariance model describes how measurements from the same participant vary together around that mean. The model estimates an exposure-response relationship conditional on its covariates; a causal interpretation additionally depends on the causal assumptions and design considered in the earlier material on causal assumptions in individual-level exposure-response.

A common starting point is a linear mean model with time represented by visit indicators, and exposure allowed to have a different association at each visit:

In this expression, \(\bar{B}\) is the mean baseline outcome in the analysis population, \(I(\cdot)\) is an indicator, and the \(K\) postbaseline visits are compared with the reference at baseline. The coefficient \(\beta_B\) adjusts for baseline outcome. At visit \(t_k\), \(\theta_k\) is the modeled change in expected outcome per unit of exposure, holding the other model terms fixed. If exposure is constant for each participant, \(E_{ij}\) can be replaced by \(E_i\). Exposure units must be stated, and centering or rescaling exposure can make coefficients easier to interpret without changing fitted contrasts.

The equation gives a mean structure, not a complete probability model. A standard model assumes that, conditional on modeled covariates, the vector of repeated outcomes has mean given by the equation and a specified covariance matrix. A multivariate normal distribution is often used for likelihood-based estimation, though a correctly specified mean and covariance can also be analyzed with methods that provide robust standard errors. The choice of covariance structure is substantive: it should represent plausible dependence without asking the data to estimate more parameters than the design can support.

Baseline Adjustment and Within-Participant Correlation

Baseline adjustment includes the baseline outcome as a predictor of postbaseline outcomes. It can improve precision because baseline is often strongly associated with later outcome, and it accounts for chance baseline imbalance between treatment or exposure groups. This is not the same as subtracting baseline from every follow-up value. A change score \(Y_{ij}-B_i\) fixes the coefficient on baseline at one in that subtraction; an adjusted model estimates the relationship using the data.

For repeated measurements, a covariance model can be stated directly for the vector \(\boldsymbol{\epsilon}_i=(\epsilon_{i1},\ldots,\epsilon_{iK})^\mathsf{T}\). For example, a random-intercept model writes \(\epsilon_{ij}=u_i+e_{ij}\), where \(u_i\) is a participant-specific deviation and \(e_{ij}\) is visit-specific error. If \(u_i\) and the \(e_{ij}\) are independent, with variances \(\tau^2\) and \(\sigma^2\), then

This implies the same correlation for every pair of visits. That may be reasonable when persistent participant differences dominate, but it may be inadequate if nearby visits are more strongly correlated than distant visits. Alternatives include an autoregressive structure, where correlation declines with time separation, or an unstructured covariance, where each visit variance and pairwise covariance is estimated separately. An unstructured model is flexible but requires many parameters and may be unstable with few participants or many visits.

Deriving the Precision Benefit of Baseline Adjustment

A simple derivation shows why a relevant baseline measurement can improve precision. Suppose a postbaseline outcome \(Y\) and baseline \(B\) have a linear relationship with slope \(\beta_B\), and the unexplained part of \(Y\) has variance \(\sigma^2\). In a bivariate normal setting, the population linear regression slope is \(\beta_B=\operatorname{Cov}(Y,B)/\operatorname{Var}(B)\). The variance remaining after linear adjustment is

where \(\rho_{YB}\) is the correlation between baseline and follow-up. The last equality follows from \(\operatorname{Cov}(Y,B)=\rho_{YB}\sqrt{\operatorname{Var}(Y)\operatorname{Var}(B)}\). Thus, when the linear relationship is useful, baseline adjustment removes predictable variation. In a fitted clinical model, the slope is estimated and the analysis must still account for the repeated measurements and study design.

Estimation, Diagnostics, and Alternatives

In practice, the mean and covariance parameters are commonly estimated jointly by maximum likelihood or restricted maximum likelihood for a specified repeated-measures model. A treatment or exposure contrast at a visit is a linear combination of estimated coefficients. Its standard error must use their estimated covariance; using only the standard error of one coefficient is insufficient when the contrast combines several terms. Confidence intervals describe uncertainty in the estimated model contrast under the model and sampling assumptions, not a probability that a fixed parameter lies in a particular calculated interval.

Diagnostics should be tied to the model’s two components. Plot observed and predicted outcomes by visit and exposure range to look for systematic mean-model departures, such as curvature, an omitted time pattern, or an exposure relationship that differs across visits. Examine residuals for changing spread and extreme values. Compare residual correlations by visit lag with those implied by the covariance model, and check for convergence warnings, boundary estimates, or implausibly large standard errors. A covariance model that cannot be estimated reliably may be too complex for the available data. These checks identify possible problems; they do not prove that the model is correct.

Repeated-measures ANCOVA is often a useful primary approach because it adjusts postbaseline outcomes for baseline and uses all modeled visits while representing their correlation. A change-score analysis is simpler and can be interpretable, but constrains baseline’s contribution through subtraction and may discard information if only one visit is analyzed. A random-intercept model captures persistent participant differences parsimoniously, but its constant-correlation implication can be unrealistic. An unstructured covariance is more flexible, but costs parameters and may fail in sparse designs. The best choice depends on the visit schedule, sample size, scientific question, and whether the resulting fit is stable and credible.

The earlier material on placebo and background-therapy effects in longitudinal models addresses how those components enter the expected trajectory. Here, the additional focus is continuous outcomes and the joint handling of baseline, exposure, and within-participant dependence. A central limitation remains: a well-fitting repeated-measures model does not repair biased exposure comparisons, measurement error in exposure, or an inappropriate exposure metric.