Multivariate Longitudinal Research — Tracking Multiple Variables Over Time

A minimum of three waves is generally recommended for modeling change trajectories with latent growth curve models, as two waves can only estimate linear change and provide no information about trajectory shape. For cross-lagged panel models, two waves are technically sufficient but three or more are strongly preferred because they allow testing whether cross-lagged effects are stable across intervals and reduce the risk of spurious estimates.

Repeated-measures ANOVA tests whether mean levels of one or more outcomes differ across time points in a single sample; it is primarily a significance test for group-level mean change. Multivariate longitudinal research encompasses a broader set of questions and models — including individual growth trajectories, reciprocal predictive relationships between variables, mediation of change, and change as a function of time-varying covariates — using techniques such as latent growth curves, cross-lagged panel models, and multilevel models. ANOVA is a narrow special case within this broader design framework.

Modern best practice is full information maximum likelihood (FIML) estimation or multiple imputation (MI), both of which use all available data under the missing-at-random assumption and outperform listwise deletion. If dropout is suspected to be non-random (missing not at random), sensitivity analyses using pattern-mixture models or selection models should be reported. Always test whether completers and dropouts differ on baseline characteristics and report the results transparently.

Yes. Before comparing latent means or correlations across waves, configural, metric, and scalar invariance should be tested using confirmatory factor analysis. Metric invariance (equal factor loadings) is the minimum required for comparing relationships among variables; scalar invariance (equal item intercepts) is required for comparing latent means. If full scalar invariance fails, partial invariance models can sometimes salvage meaningful comparison with appropriate caveats.

There is no single answer, but multivariate longitudinal models are parameter-heavy. A commonly cited heuristic for latent variable models is 10–20 observations per estimated parameter, though Monte Carlo power analyses provide more accurate guidance. After accounting for expected attrition per wave, a starting sample that might seem adequate can become underpowered by the final wave. Plan conservatively and over-recruit.