TL45: Logitudinal ANCOVA vs. Constrained Longitudinal Data Analysis Model
*Xiaoming Li, Merck Research Laboratories  *Jiajun Liu, Merck Research Laboratories 


In a constrained longitudinal data analysis (cLDA) model (Liang and Zeger), both the baseline and post-baseline values are modeled as dependent variables, as opposed to a longitudinal ANCOVA model in which the baseline value is included as a covariate. Although the baseline measure is included in the response vector in cLDA, the true baseline means are constrained to be the same for different treatment groups due to randomization, and this analysis provides an adjustment for the observed baseline difference in estimating the treatment effects. When there are no missing data, it is shown that both the cLDA and longitudinal ANCOVA models produce identical point estimates for the treatment difference, while the cLDA can have a slightly smaller estimated variance for the treatment difference and hence can be more efficient and powerful for testing the treatment difference. When there are missing data, both cLDA and ANCOVA model should produce unbiased estimates and inference if the missing data mechanism is missing completely at random (MCAR); however, if missing data are missing at random (MAR), the cLDA model is valid but the ANCOVA model can produce biased results.