Osmond and colleagues make several points about our use of their paper [1] as an example of survivor treatment selection bias. First, although they express concern about this bias in their paper, such a caveat, in our opinion, should not substitute for the use of appropriate statistical models that can address this bias.

Second, the authors state that their conclusions were based on survival comparisons between men with and those without diagnoses of P. carinii pneumonia rather than the (biased) proportional hazards model. In fact, survival in their subgroup with P. carinii pneumonia as an initial AIDS-defining diagnosis did not change significantly over the decade (P > 0.2). In addition, their comparisons of survival beyond CD4 counts less than 200 cells/mm³ according to type of initial AIDS-defining illness are also biased. Patients receiving prophylaxis for P. carinii pneumonia who nevertheless develop pneumonia tend to do so at very low CD4 counts [2] and therefore would have survived relatively longer (regardless of treatment) than would patients not receiving prophylaxis who develop pneumonia. Please note that Figure 1 of their paper [1] was mislabeled, and median survival beyond a CD4 count of 200 cells/mm³ in those with P. carinii pneumonia was less (not greater) in the later time period.

Third, Osmond and colleagues cite data on the use of antiretroviral agents and P. carinii pneumonia prophylaxis that were obtained from a cross-sectional (prevalence) survey [3] and extrapolate these data to their cohort study. This extrapolation may be misleading, given that cross-sectional samples reflect the use of treatment in patients with longer survival. Those who die quickly are under-represented in a sample that is taken at a single time point. The authors should instead investigate use of these drugs in the same cohort of men with incident CD4 counts less than 200 cells/mm³ on which they based their original analyses.

We stand by our contention that survivor treatment bias may have influenced the conclusions of the study by Osmond and associates [1]. The issue could be settled by reanalysis of their data using time-dependent covariates for treatment initiation, as recommended in our report.

We also note an error that occurred in our paper. The study of differences in survival between men and women after AIDS diagnosis, reported by Lemp and colleagues [4], was cited incorrectly as an example of how survivor treatment bias could lead to misleading conclusions, Lemp and coworkers restricted their analyses of the treatment effect of zidovudine to the time period after the availability of zidovudine (1987 to 1991). Thus, any differential survivor bias between men and women that may have been present in the analysis was greatly overstated in our paper and probably would not have affected the conclusions of Lemp and colleagues.