Baseline Covariates in Longitudinal Models

Why Include a Baseline Covariate?

Subjects can differ at the start of follow-up in characteristics related to the longitudinal outcome. For example, baseline disease severity may be associated with later biomarker values. Including such a measurement as a fixed-effect covariate lets the model compare expected outcomes at different baseline values while holding the other predictors fixed. This is covariate adjustment: it changes the model for the mean, not the definition of the subject-specific random effects.

A baseline covariate is measured once per subject and is therefore shared by that subject's repeated observations. Repeating the same value on every visit does not turn it into a time-varying predictor. Its coefficient describes a between-subject association: subjects with different baseline values are compared, conditional on the other included predictors.

An Additive Mixed Model With a Baseline Covariate

Let \(Y_{ij}\) be the outcome for subject \(i\) at visit time \(t_{ij}\), let \(A_i\) indicate treatment group, and let \(x_i\) be a baseline covariate. For a simple linear-time example, consider

Here \(\beta_0,\beta_A,\beta_t,\beta_{At},\beta_x\) are fixed-effect coefficients. The subject-specific random intercept and slope are \(b_{0i}\) and \(b_{1i}\); \(\epsilon_{ij}\) is the residual. The model includes a common baseline-covariate effect: there is no \(x_i\)-by-time interaction, so the model assumes that the covariate's mean association is the same at every time.

For the population mean to have the stated fixed-effects form, assumptions about conditional means matter. In particular, suppose the random effects have conditional means zero given the subject's covariates and visit times, and the residual has conditional mean zero given those predictors and the random effects:

These are conditional mean assumptions; an unconditional statement that the random effects or residuals have mean zero would not, by itself, imply the conditional expectations used below. The random-effects covariance and residual covariance must also be valid covariance structures for the observations. Residuals need not be independent across visits for the mean interpretation, but their dependence must be handled appropriately for estimation and uncertainty calculations.

Taking the expectation conditional on \(A_i,x_i\), and the visit times, and averaging over the random effects and residuals gives

If instead we condition on the random effects as well, the subject-specific expected outcome includes \(b_{0i}+b_{1i}t_{ij}\). Thus the fixed mean and the subject-specific deviation have distinct roles: the covariate shifts the modeled mean according to \(x_i\), while the random effects represent subject-level departures from that mean.

Why the Coefficient Has This Interpretation

The interpretation follows directly by comparing the model means for two baseline values while holding treatment and time fixed. Let \(x_i=x\) for one subject profile and \(x_i=x+\Delta\) for another. Subtracting the two population means cancels every term that is unchanged between the profiles:

For a one-unit increase, \(\Delta=1\), the difference is \(\beta_x\). This is a comparison of conditional means under the specified model. It does not imply that changing a subject's baseline covariate would cause the outcome to change by \(\beta_x\); such a causal interpretation requires additional design and identification assumptions.

Adjustment Is Not Random-Effect Variation

A fixed baseline-covariate coefficient and a random intercept can both contribute to differences among subjects, but they represent different things. The fixed term \(\beta_x x_i\) describes a systematic mean difference associated with an observed predictor. A random intercept \(b_{0i}\) represents a subject-specific deviation not explained by the fixed predictors. Similarly, a random slope \(b_{1i}\) represents subject-specific deviation in the time trend; it is not a substitute for including a measured baseline covariate in the mean.

A baseline predictor is constant within subject, so its main effect is identified by differences between subjects, not by changes within a subject over time. If the scientific question is whether the baseline association itself changes during follow-up, the additive model is not sufficient: a covariate-by-time interaction would be needed. That extension changes the interpretation and should be justified by the question and data rather than added automatically.

Inference, Coding, and Practical Checks

A coefficient estimate should be read with its uncertainty and units. For example, a confidence interval for \(\beta_x\) describes plausible values for the adjusted mean difference per one-unit increase in the covariate, under the model and sampling assumptions. It is not a probability statement about the fixed parameter after the interval has been computed.

Centering a covariate, for example defining \(x_i^c=x_i-\bar{x}\), changes the reference point for the intercept but not the fitted values or the slope for that covariate. With an uncentered covariate, the intercept refers to \(x_i=0\), which may be clinically impossible or far from the observed data. Centering at a meaningful value makes the intercept easier to interpret. Scaling also matters: a coefficient per one unit may be numerically small, so a ten-unit contrast can be more informative when it lies within the observed range.

Before interpreting the adjustment, check that the covariate is measured consistently, that its coding and units are clear, and that subjects with different covariate values overlap sufficiently in the other predictors for the intended comparison. Strong collinearity with treatment or another subject-level predictor can make estimates imprecise. Repeated outcomes do not create new independent baseline-covariate values: the information about a purely subject-level covariate is fundamentally between subjects. A few subjects or a narrow range of covariate values can therefore limit the precision and scope of the estimated association.

Common Pitfalls and Model Limits

• Assuming a zero unconditional mean is enough. To obtain the conditional mean formula, require \(E(\epsilon_{ij}\mid A_i,x_i,\{t_{ij}\},b_{0i},b_{1i})=0\), and require conditional mean-zero random effects given the predictors and visit times. Unconditional mean zero alone does not justify those conditional expectations.
• Calling adjustment a causal effect. A fixed coefficient is an adjusted association under the model. Confounding, selection, measurement error, or an inappropriate model can prevent a causal interpretation.
• Treating the covariate as if it varied within subject. Because \(x_i\) is constant across visits, its main effect does not describe within-subject change. A time interaction is required if the modeled association is to vary over time.
• Using a random intercept as a replacement for a measured predictor. Random effects represent residual subject-to-subject variation; they do not estimate the systematic mean association with a specific observed covariate.
• Extrapolating beyond observed covariate values. A linear coefficient may summarize the observed range well but need not describe subjects far outside it. Check the data range and the adequacy of the assumed mean relationship.

The main alternatives answer different questions. Omitting the covariate gives a less adjusted mean structure and may leave relevant systematic differences unmodeled. Stratifying by categories of a continuous covariate can avoid a linearity assumption but discards information and depends on cutpoints. Adding a covariate-by-time interaction allows the association to vary with time but uses additional parameters and requires evidence and interpretation at specific times. A random intercept or slope addresses unexplained subject-level variation, not adjustment for the measured baseline characteristic. These approaches are not interchangeable.