This dissertation compares different methods for the analysis of longitudinal data obtained from randomized controlled trials (RCTs) in a series of three papers that share the characteristic of combining applied analyses based on empirical examples of health behavior change interventions with a simulation study.
In study 1, traditionally used methods for the analysis of RCT data are compared to modern computer intensive methods in respect to a standard 2x3 (group*time) design. Empirical analyses of two independent samples of three health behaviors (smoking cessation, dietary intake, and sun protection) showed that the commonly assumed compound symmetrical (CS) pattern provided inadequate fit. The results of the simulation study showed that Type I error rates were inflated for analyses assuming CS. Modern approaches (i.e., fixed effects model and conditional latent growth model (LGM)) exhibited accurate Type I error rates and superior statistical power (power=0.80 at approximately N=470).
In study 2, the same design was used to compare methods in respect to three approaches to missing data. The last observation carried forward (LOCF) approach produced conservatively biased results. Power approached adequate levels for combined sample sizes of N=720 using all available data, N=730 for complete case analysis and N=750 for LOCF.
In study 3, the four methods were compared in regards to their statistical power of detecting two different long-term intervention effect patterns (i.e., maintained or diminishing) for two groups measured over five occasions. Both patterns had an effect of d =0.20 at t =3. Statistical power was found to be highly dependent on the intervention effect pattern. For the maintained pattern, N =400 or larger resulted in adequate statistical power. For the diminishing pattern, only N =600 sufficed, and only for select models for the mean (i.e., approximately N=670 for quadratic and N=610 for piecewise linear models).
In conclusion, applied researchers are discouraged from using repeated measures ANOVA or similar analyses using the compound symmetry assumption, and encouraged to use either LGMs or fixed effect models with appropriate covariance structures. Sample size guidelines for cases involving missing data or nonlinear intervention effect patterns are offered. SAS syntax for all analyses is provided.
