Abstract:
|
In the clustered randomized trial setting, researchers are often interested in the effects of both trial-level (e.g., experimental setting) and cluster-level (e.g., subject age) predictors. Classification and regression trees are widely-used nonparametric predictive modeling approaches, but most algorithms do not allow the user to properly account for clustered data. We adapt rpms, a tree algorithm designed for analysis of survey data, to analyze data from an experimental study. Participants were asked to select the appropriate child restraint system (CRS) for a child of a given height and weight, install the CRS, and secure a specially designed doll. Each participant performed four trials, under varying experimental conditions; installations were checked for errors after each trial. Previous analyses of the data have shown clear trial-level differences in error rates, but it is also important to determine which participant-level characteristics (e.g., age, sex, experience) are predictive of higher error rates, or of certain types of errors. We compare the performance of rpms to traditional classification trees and to mixed-effects modeling, and make recommendations for practice.
|