Reporting Fixed and Random Effects Clearly

Why Clear Reporting Matters

A mixed-model results section should let a reader answer two different questions: how does the expected biomarker level vary with measured predictors, and how much variation is associated with subjects and within-subject measurement error? A list of coefficients without scales, units, or covariance details rarely answers either question. For example, a time coefficient depends on whether time was measured in days or years, while a random-intercept variance has squared biomarker units and is not itself a typical deviation.

This tutorial uses a hypothetical biomarker study to develop a concise reporting approach. It focuses on what to report and how to interpret it, rather than on fitting the model. The earlier material on random-effects covariance parameterizations, within-subject residual correlation, heteroscedastic residuals, and residual diagnostics supplies the modeling background needed here.

A Model and Its Reporting Targets

Suppose \(Y_{ij}\) is a biomarker measurement for subject \(i\) at visit \(j\), measured in milligrams per liter (mg/L). Let \(t_{ij}\) denote time in years since baseline, and let \(G_i=1\) for the treatment group and \(G_i=0\) for the comparison group. Consider the following linear mixed model:

The fixed effects \(\beta_0,\ldots,\beta_3\) describe the group-level mean trajectory. With this coding, \(\beta_0\) is the comparison-group mean at baseline, \(\beta_1\) is its mean change per year, \(\beta_2\) is the baseline treatment-versus-comparison difference, and \(\beta_3\) is the difference between the groups’ mean yearly changes. The treatment-group mean change per year is \(\beta_1+\beta_3\), and its mean at time \(t\) differs from the comparison-group mean by \(\beta_2+t\beta_3\).

The subject-specific random intercept and slope are \(b_{0i}\) and \(b_{1i}\). Assume their vector has mean zero and covariance matrix \(D\), and that the random effects are independent across subjects. In this example, residuals have an AR(1) correlation across equally spaced visits: their variance is \(\sigma^2\), and the correlation between residuals \(k\) visits apart is \(\rho^{|k|}\). Assume residuals are independent across subjects and independent of the random effects. These assumptions describe the particular model to be reported; other covariance structures require different descriptions.

A useful report therefore identifies the fixed-effect estimates and uncertainty intervals, the random-effect variance-covariance parameters, and the residual variance and correlation structure. State the time origin, time units, biomarker units, group coding, and interval method. Random-effect variances have squared units: \(\tau_{00}\) has units (mg/L)\(^2\), \(\tau_{11}\) has units (mg/L per year)\(^2\), and \(\tau_{01}\) has units (mg/L)\(^2\) per year. The correlation \(\rho\) is dimensionless.

Uncertainty for Effects and Contrasts

A coefficient estimate should ordinarily be accompanied by an uncertainty interval and, when relevant to the question, a test result. For a large-sample Wald interval, an estimate \(\hat\beta\) with standard error \(SE(\hat\beta)\) has an approximate 95% interval \(\hat\beta\pm1.96SE(\hat\beta)\). If inference uses a \(t\) reference distribution, report the chosen degrees of freedom and critical value instead. State the method rather than implying that every software package uses the same interval calculation.

A derived contrast is often more useful than a coefficient alone. For a linear contrast \(c^\mathsf{T}\beta\), its estimated variance is \(c^\mathsf{T}\widehat{\operatorname{Var}}(\hat\beta)c\). In particular, the uncertainty for a treatment difference at a specified time must account for the covariance between \(\hat\beta_2\) and \(\hat\beta_3\), not just their separate standard errors.

Reporting Random Effects and Residual Structure

Report random effects as a covariance structure, not as an unqualified statement that “subjects varied.” For an intercept-and-slope model, give the intercept variance, slope variance, and their covariance or correlation. The correlation is calculated as \(\tau_{01}/\sqrt{\tau_{00}\tau_{11}}\). Variances describe dispersion on squared scales; their square roots, the corresponding standard deviations, are often easier to interpret as typical scales of between-subject variation. Do not confuse a standard deviation with a standard error of an estimate.

Residual reporting is separate. Give the residual variance or residual standard deviation and specify any modeled serial correlation, such as an AR(1) correlation with estimated \(\rho\). Random intercepts and slopes represent between-subject variation in trajectories; correlated residuals represent remaining within-subject dependence after accounting for the model. Reporting one does not substitute for reporting the other.

Intervals for Variance Parameters and a Complete Summary

Uncertainty intervals for variance components need special care. Variances cannot be negative, and their sampling distributions may be skewed, especially near zero. A symmetric estimate-plus-or-minus-standard-error interval can therefore extend below zero or poorly represent uncertainty. Prefer a method appropriate to the fitted model, such as a profile-likelihood interval or a suitable bootstrap interval, and name the method. If a reported interval is for a variance, retain squared units; if it is transformed to a standard deviation, take square roots of both nonnegative endpoints and label the interval as an interval for the standard deviation. Boundary estimates and singular fits also deserve explicit disclosure rather than a falsely precise summary.

The following hypothetical results combine fixed effects, covariance estimates, and residual structure. The fixed-effect intervals use the normal-based Wald calculation; the variance-component intervals are assumed to come from profile likelihood. These methods are stated separately because they are not interchangeable.

Diagnostics, Alternatives, and Common Pitfalls

A polished summary should reflect a model whose assumptions and fit have been examined. Check residual plots and the diagnostics appropriate to the fitted residual structure; look for systematic patterns, unequal spread, influential subjects, or remaining serial dependence. Check whether random-effect estimates are near a boundary, whether the fit is singular, and whether the data support estimating a random slope. A reported AR(1) structure should match the visit indexing and spacing used by the model. If the observed visits are irregularly spaced, an ordinary discrete-visit AR(1) structure may not represent correlation as a function of elapsed time; describe and justify the actual structure used.

There is no single reporting format for every scientific question. A coefficient table is efficient for presenting the model’s parameter estimates, standard errors, intervals, and units. A contrast-focused summary is more direct when the question concerns a group difference at a particular time or a change over a specified interval. A table of standard deviations and correlations can make random-effect scales more intuitive, while a variance-covariance table preserves the parameters used by the model. Whichever format is chosen, do not omit the residual structure or mix parameter scales without labels.

• Do not report only a p-value. Include an estimate, its units, and an uncertainty interval so the magnitude and precision are visible.
• Do not interpret an interaction as a universal group difference. A time-by-group coefficient is a difference in slopes; the group difference at time \(t\) is \(\beta_2+t\beta_3\).
• Do not confuse fitted population effects with subject-specific predictions. Fixed effects describe the modeled group-level mean; conditional predictions incorporating estimated subject effects answer a different question.
• Do not omit the reference point for time. The intercept and group main effect depend on the time origin; identify baseline or the chosen centering value.
• Do not claim the chosen covariance structure is established merely because it was fitted. State the structure, check its adequacy, and acknowledge weak identification or singularity when present.

These reporting choices improve interpretability but do not remove limitations. Estimates depend on the model specification, observed follow-up, and assumptions about the mean and covariance structure. A confidence interval describes uncertainty under the stated procedure and model; it does not account automatically for model misspecification, selection bias, or unmeasured confounding. When the results are used in a broader joint analysis, report the longitudinal submodel clearly without implying that its summary alone describes event-time outcomes.