: Policy makers rely on accurate assessments of quality of care when making decisions about health care delivery. This task is complicated by the pervasive problem in health services research (HSR) of data that are missing due to client departure prior to completing a full course of treatment and/or loss to follow-up. Analytic methods employed in HSR frequently ignore the mechanism by which data are missing. Biased estimates from the quality of care model could result if the true nature of the missing data is unaccounted for in the analysis, thereby hindering a policy maker's ability to improve quality. The pattern-mixture model (PMM) has been proposed in the statistical literature as an alternative to standard analytic methods that ignore the mechanism by which data are missing.
Our specific aims are to: 1) improve the assessment of the relationships among treatment structure, process, and outcomes when treatment dropout and study attrition occur; and 2) utilize expert opinion and knowledge about the reasons for treatment dropout and attrition in order to inform the PMM-building process.
Aim 1 will be achieved by constructing a pattern-mixture quality of care model.
Aim 2 will build upon Aim 1 by addressing the subjective decisions that are required to build a PMM via Bayesian statistics, properly addressing the uncertainty inherent in these decisions, and measuring their effect on conclusions drawn from the analysis. ? ?
| Paddock, Susan M; Ebener, Patricia (2009) Subjective prior distributions for modeling longitudinal continuous outcomes with non-ignorable dropout. Stat Med 28:659-78 |
