OBJECTIVES: In economic and epidemiologic studies utilizing longitudinal data, researchers often require a fixed follow-up period. By allowing differential follow-up, researchers can eliminate one source of selection bias. Our objective was to compare analyses of two study populations from the General Practice Research Database (GPRD) differing only in length of follow-up time. METHODS: Study population 1 (SP1; n=28,643) included patients with a given condition first diagnosed between June 1988 and February 1999. Study population 2 (SP2; n=21,289) included only those patients in SP1 with at least a year of follow-up. Comorbid conditions, resource use, and reasons for loss to follow-up were examined. RESULTS: Approximately 26% (n=7,354) of patients in SP1 were excluded from SP2. Reasons for exclusion were death (22%), loss to follow-up (34%), and right-censoring (44%). Age, gender, and calendar year of the index event were not dramatically different between the two populations. Both crude and age-adjusted prevalence rates of most comorbid conditions at index date were lower in SP2. Incidence rates showed a similar pattern. Mortality was significantly higher in SP1 (5.7 vs. 3.8/100 PY; p<0.05). There was a decrease in the proportion of patients never treated with drugs for the selected condition from 42% in SP1 to 36% in SP2. Mean and median values for annualized rates were lower in SP2, possibly indicating that some patients with higher resource use were selectively excluded. CONCLUSIONS: The impact of selection bias associated with defining patient cohorts based on minimum follow-up time affects both economic and epidemiologic analyses. Overall, the results of this study suggest that researchers should define a cohort without regard to follow-up whenever possible. The appropriate and necessary follow-up period will clearly vary by condition and perhaps by local practice patterns. The data themselves can help inform the appropriate time frame for follow-up.