Journal of Data Science

Abstract: Good inference for the random effects in a linear mixed-effects model is important because of their role in decision making. For example, estimates of the random effects may be used to make decisions about the quality of medical providers such as hospitals, surgeons, etc. Standard methods assume that the random effects are normally distributed, but this may be problematic because inferences are sensitive to this assumption and to the composition of the study sample. We investigate whether using a Dirichlet process prior instead of a normal prior for the random effects is effective in reducing the dependence of inferences on the study sample. Specifically, we compare the two models, normal and Dirichlet process, emphasizing inferences for extrema. Our main finding is that using the Dirichlet process prior provides inferences that are substantially more robust to the composition of the study sample.

Abstract: We investigate whether the posterior predictive p-value can detect unknown hierarchical structure. We select several common discrepancy measures (i.e., mean, median, standard deviation, and χ2 goodness-of-fit) whose choice is not motivated by knowledge of the hierarchical structure. We show that if we use the entire data set these discrepancy measures do not detect hierarchical structure. However, if we make use of the subpopulation structure many of these discrepancy measures are effective. The use of this technique is illustrated by studying the case where the data come from a two-stage hierarchical regression model while the fitted model does not include this feature.