Treatment integrity concerns in comparative education studies

摘要

Treatment integrity is the degree to which treatments or interventions are implemented as intended in comparative education studies (Meyer & Fienberg 1992, Boruch 1997). In this dissertation I examine two specific treatment integrity concerns. The first involves differential subject attendance, whereby subjects experience varying intensity levels of interventions. This can lead to confounding of the learning outcome of interest. I extend results from the setting of clinical trials with subject non-compliance by developing bounds related to those of Robins (1989) and Manski (1990) for the causal effect of attendance level a in intervention t versus attendance level b in intervention s. I examine a particular statistical model developed in the non-compliance setting that seems attractive for modeling causal effects in an educational setting. Parameter estimates from this model are sensitive to untestable assumptions, however, and I develop an alternative model to account for confounding due to observed covariates.;The second treatment integrity concern I consider is the between-subject dependence inherent in education studies where interventions are often assigned and administered to groups of subjects rather than to individual, non-interacting subjects. I extend the causal modeling framework developed by Rubin (1974, 1977, 1978, 1980) to include assumptions to address the nested designs of many education studies. Under these assumptions, causal effects of treatment can be specified, and I relate these assumptions to the construction of hierarchical Bayes models for the data (Lindley & Smith 1972). In addition, I argue that such assumptions are necessary for making causal inferences from these models. Throughout the dissertation, these treatment integrity concerns, and developments to address them, are illustrated using a dataset from a health education experiment (Helgeson, Cohen, Schulz & Yasko 1999).