Modeling longitudinal outcomes: A contrast of two methods

Background: Repeated measures analysis of variance (ANOVA) is frequently used to model longitudinal data but does not appropriately account for within-person correlations over time, does not explicitly model time, and cannot flexibly handle missing data. In contrast, mixed-effects regression addresses these limitations. In this commentary, we compare these two methods using openly available tools. Methods: We emulated a real developmental study of elite skiers, tracking national rankings from 2011 to 2018. We constructed unconditional models of time (establishing the “pattern” of change) and conditional models of time (identifying factors that affect change over time), and contrasted these models against comparable repeated measures ANOVAs. Results: Mixed-effects regression allowed for linear and non-linear modeling of the skiers' longitudinal trajectories despite missing data. Missing data is still a concern in mixed-effects regression models, but in the present dataset missingness could be accounted for by skiers' ages, satisfying the missing at random assumption. Discussion: Although ANOVA and mixed-effects regression are both suitable for time-series data, their applications differ. ANOVA will be most parsimonious when the research question focuses on group-level mean differences at arbitrary time points. However, mixed-effects regression is more suitable where time is inherently important to the outcome, and where individual differences are of interest.