|
Abstract:
|
When study sample is not representative of the target population, selection bias can occur. In this paper, we aim to extend inverse probability weighting by introducing a latent construct, “motivation” which summarizes all omitted critical covariates. We considered both multinomial and continuous motivation and evaluated model performance in Monte Carlo simulated random controlled trials with continuous endpoints. EM based algorithm and marginal MLEs were developed to estimate latent motivation level, using each subject’s best linear unbiased predictor. The estimated motivation was then used to calculate modeling weights, and between arm differences were estimated using a survey weight regression. Treatment effect estimations were almost unbiased, and MSE reduced by over 80% compared to a naïve analysis. Power and false positives were well controlled. Conventional mixed effect model showed robust performance under less skewed distributions with lower computational cost. Therefore, a normality test is recommended to BLUPs for model selection. When motivation was truly continuous, multinomial motivation models yielded unsatisfactory results.
|
