Because considerable information about progression of human immunodeficiency virus (HIV) infection has been provided by studies of cohorts of individuals with prevalent HIV infection, this study was designed to investigate bias due to onset confounding (differential time-since-infection distributions) and differential length-biased sampling in epidemiologic analyses of data from such cohorts. Subjects were participants in the Italian Seroconverters Study, a seroincident cohort of more than 1,200 adults seen at ambulatory care clinics in Italy, with observed HIV seroconversion in 1980-1988. Acquired immunodeficiency syndrome (AIDS) diagnoses, based on the 1987 Centers for Disease Control case definition, and mortality were ascertained through Italian national registries through 1994. To estimate bias in prevalent cohorts, a series of pseudoseroprevalent (PSP) cohorts were drawn by sampling, from among the total seroincident cohort, prevalent AIDS-free subjects in each calendar year. The relative AIDS risk associated with a given covariate was calculated in each PSP cohort and compared with the relative AIDS risk for that covariate in the seroincident cohort. Relative risks were estimated by both the ratio of AIDS incidence densities and the relative AIDS hazards from proportional hazards regression. Differential length bias was not evident, as assessed in the following way: Among 338 individuals with seroconversion dates in 1983-1986, the relative risk of AIDS for subjects born before 1951 compared with those born more recently was 1.67 (95% confidence interval (CI) 1.30-2.14). Although differential length-biased sampling was expected to bias this relative risk toward 1.0, the observed relative risk for earlier birth ranged from 1.79 to 2.86 in 1987-1992 PSP cohorts. Onset bias was observed: Among 644 subjects with seroconversion in 1980-1988, the AIDS relative risk for 1980-1985 seroconverters compared with 1986-1988 seroconverters was 1.09 (95% CI 0.76-1.55). Onset bias was seen in 1988-1990 PSP cohorts (relative risks for early seroconversion = 1.47, 1.46, and 1.34, respectively); in 1991-1992, relative risks were close to the expected value of 1.09, and CIs on relative risks from all PSP cohorts after 1989 included 1.0. Confounding attributable to differential length-biased sampling in prevalent cohorts does not necessarily bias estimates of the impact of covariates on rate of progression to AIDS. Bias can arise when a covariate suspected of affecting AIDS risk is closely linked to date of acquisition of HIV infection. However, onset bias appears to wane as subjects' dates of infection become more remote.