Application of Bayesian Methods in Multilevel and Longitudinal Mediation Models

Project: Research project

Description

The goal of my proposed research is to develop Bayesian techniques to test mediational effects in cluster randomized designs with multiple measurement waves. Mediational analysis studies the processes through which interventions achieve their effects through intervening variables that are targeted for change. Because the widely used maximum likelihood approach relies on large samples and normal theory to produce valid results, Bayesian techniques are expected to show superior performance in small to moderate sample sizes particularly with the non-normal data common in substance abuse prevention trials. In addition, Bayesian methods can incorporate information from previous studies further adding to their efficiency. Aim 1 will develop Bayesian techniques to test mediation effects in cluster randomized and longitudinal models that can incorporate information from previous experiments and handle non-normal data. The resulting estimates can potentially lead to more reliable estimates and greater statistical power than current approaches. Aim 2 develops Markov Chain Monte Carlo (MCMC) methods to estimate Bayesian mediation models. Also, computer code to implement MCMC methods in publicly free software packages (e.g., R, WinBUGS) will be written, making this work accessible to other researchers. Aim 3 conducts a simulation study to compare the performance of the Bayesian and maximum likelihood estimators using data structures that mimic existing drug prevention cluster randomized trials and longitudinal studies. Of most interest, cluster size, number of clusters, and degree of non-normality will be varied. Aim 4 applies both Bayesian and existing frequentist multilevel methods to three existing data sets on drug prevention to compare the performance of the point and interval estimators of the mediation effects from each model. A Monte Carlo comparison of the performance of maximum likelihood and Bayesian approaches as a function of sample size will be conducted using repeated random samples from a large existing data set. The proposal aims to improve statistical methodology in analyzing data from randomized control trials with multiple waves of measurement in the drug prevention areas. The proposed methodology offers new methods when the number of participants is not large and enhances validity of existing statistical methods and the interpretability of the results in drug prevention and health science.
StatusFinished
Effective start/end date7/27/097/26/11

Funding

  • HHS: National Institutes of Health (NIH): $77,022.00

Fingerprint

Mediation
Bayesian Methods
Drugs
Markov Chain Monte Carlo Methods
Maximum Likelihood
Sample Size
Model
WinBUGS
Estimate
Multilevel Methods
Randomized Trial
Statistical Power
Non-normality
Methodology
Longitudinal Study
Interpretability
Number of Clusters
Bayesian Approach
Software Package
Large Data Sets