Modelling effectiveness of internally heterogeneous organizations in the presence of survey non-response: an application to the ULTRA study.

Levels of turbulence and uncertainty in family medicine practices participating in the ULTRA study are measured via a survey from heterogeneous viewpoints within each practice, including clinicians, nursing staff and office staff. In order to examine the effect of the average practice 'chaos' score on screening provided to patients, the analysis needed to account for large amounts of missing survey responses. Patterns of missing data included portions of missing responses within a job group (e.g. one out of three clinicians) to missing responses from whole subgroups within a practice (e.g. all three clinicians) to a whole practice of missing survey responses (e.g. all clinicians, nursing staff and office staff). The expectation-maximization (EM) algorithm, using a hierarchical model for covariate information, is used to handle the missing data. It was found that the varied patterns of missingness of survey responses among the heterogeneous subgroups within the practices had large effects on estimated effects of practice chaos on patient care. c 2007 John Wiley & Sons, Ltd.