|
Abstract:
|
One of the common complications in randomized trials is noncompliance of study participants. Individuals with different compliance types often benefit from treatments differently. Therefore, analyses without considering heterogeneity in compliance tend to understate treatment efficacy and power to detect treatment effects. To better estimate treatment efficacy, the possibility of estimating treatment effects only for compliers has been explored under the label Complier Average Causal Effect (CACE) estimation (Angrist, Imbens, & Rubin, 1996; Bloom, 1984; Frangakis, Rubin, & Zhou,1998; Hirano, Imbens, Rubin, & Zhou, 2000; Imbens & Angrist, 1994; Imbens & Rubin, 1997; Little & Yau, 1998). Methodological difficulties lie in estimating CACE, because compliance status is not usually completely observed. In estimating CACE, this presentation will demonstrate how to apply the Mplus program (Muthén & Muthén, 1998-2001), which accommodates broad ranges of latent class and finite mixture models. Various CACE modeling possibilities will be demonstrated within the ML-EM framework. Topics include: CACE estimation for continuous and binary outcomes, CACE estimation with covariates, missing at random, nonignorable nonresponse, clustering (multilevel data), and repeated measures. Other modeling possibilities under investigation will also be discussed (e.g., estimating dosage effects, allowing for nonlinear relationship between compliance and covariates, outcome and covariates, and combining exploratory and confirmatory analyses).
|
