Attrition bias in rheumatoid arthritis databanks: a case study of 6346 patients in 11 databanks and 65,649 administrations of the Health Assessment Questionnaire.

OBJECTIVE: Patient dropout (attrition) can bias and threaten validity of databank-based studies. Although there are several databanks of rheumatoid arthritis (RA) in operation, this phenomenon has not been well studied.
METHODS: We studied the attrition patterns of patients with RA in 11 long-running databanks where patients were followed using semiannual Health Assessment Questionnaires. Attrition rates were calculated as the proportion of living patients who were in active followup at the cutoff date. Mantel-Haenszel methods and Weibull regression were used to model the relationship between attrition and age, sex, race, education, disease duration, functional disability, and other characteristics.
RESULTS: Overall, 6346 patients with RA were recruited into the study cohorts and followed for 32,823 person-years with 65,649 observations. The crude attrition rate was 3.8% per cycle. Rates were lowest in community-based databanks. Smaller size of the centers, inner-city location, and university clinic settings were associated with worse attrition. In multivariable analyses, younger age, lower levels of education, and non-Caucasian race predicted attrition. Level of disability and disease duration were not associated with attrition. Conclusion. In terms of person-years of followup and observation-points, this may be the largest study on attrition to date. While it is possible to have very high overall retention rates, certain types of databanks (smaller, inner-city-based, and university-based) are more likely to be biased due to selective retention of older, more educated Caucasian patients.