The risk ratio reported by Bucher and Griffith for progression to AIDS or death in long-term trials is identical to that which we obtained in our paper [1] (0.96 [CI, 0.81 to 1.13]; 0.88 [CI, 0.73 to 1.08]) if deaths not related to HIV infection are excluded. We believe, however, that their subgroup analysis is not sufficiently robust. Despite further stratification according to CD4 count, studies were pooled regardless of the presence or absence of symptoms (an important covariate), and no meta-regression was done. Overall, we claimed no important long-term benefit from early zidovudine treatment for any subgroup in terms of progression to the acquired immunodeficiency syndrome or death. A borderline trend of long-term benefit was seen only for progression to any primary end point (risk ratio, 0.73; CI, 0.52 to 1.03), an outcome not addressed by Bucher and Griffith. Second, because early treatment was compared with deferred treatment rather than with placebo, we believe that the inclusion of the follow-up studies of AZTCG and AIDS Clinical Trials Group (ACTG)-019 in the subgroup analysis was legitimate. Regardless, the exclusion of these studies would not alter the results substantially. Several trials (VA-298, Concorde, and European-Australian Collaborative Group [EACG]-017) allowed the use of open-label zidovudine for a substantial portion of their patients who had not reached a primary end point. According to the approach of Bucher and Griffith, these patients should also have been excluded. Such exclusion, however, would have introduced overt bias. Finally, our regression analysis predicts that asymptomatic patients with the highest control rates of progression are likely to have the worst long-term outcome from early initiation of zidovudine treatment. This finding is consistent with the risk ratio of 1.29 (CI, 0.95 to 1.75) obtained by Bucher and Griffith for the group of asymptomatic patients with CD4 counts less than 200 cells/mm³ (that is, the asymptomatic group with the highest control rate of progression). Overall, we are pleased that, despite their differing approach, Bucher and Griffith reached results so closely akin to ours.

Drs. Costagliola and Mary-Krause make a valid comment about the unknown but important effect of patients lost to follow-up. We also expressed our concern about this effect in our report. We disagree, however, that the number of lost patients was particularly high in these studies. In six studies Table 1, only 363 of 5449 patients were lost to follow-up (3.4 patients per 100 patient-years), with a maximum loss of 8%. Such rates are not unusual; in fact, high withdrawal rates are common in longitudinal trials that enroll asymptomatic patients who do not have imminent disease [2]. The ratio of "end points-to-patients lost to follow-up" proposed by Costagliola and Mary-Krause is arbitrary and depends on their choice of end point in trials with multiple primary end points. For example, EACG-020 had four primary end points. For this study, this ratio would be 2.6 for progression to the earliest endpoint but only 0.2 for the most advanced endpoint. The P value for the comparison of early and deferred zidovudine treatment is statistically significant (P < 0.001) for the former end point but not for the latter (P > 0.2). Moreover, any cutoff value (2 or other) for the ratio of "end points-to-patients lost to follow-up" to segregate trials of "good" from those of "poor" quality would be equally arbitrary. Such analyses may be further biased if patients are more likely to be lost to follow-up early (rather than late) in a trial. Patients who stay in a trial for a year have already shown committed interest and may be less likely to subsequently withdraw from the study without informing their physicians. This may explain why the dropout rates per 100 patient-years are slightly higher in short-term than in long-term trials.