The labels low, intermediate, and high are labels of convenience for the presentation that follows. They are not to be construed as qualitative labels for the presumed actual underlying mortality rates for the populations of interest. The true underlying rates may, in reality, be higher than the levels indicated for the set carrying that label. For example, the median life expectancy for the population enrolled into R89 (AIDS patients with CMV) is, in all probability, less than the 3.5 years listed for the ''high'' rate. The same is likely to be true for the population studied in R90. The estimated median life expectancy is around 10 years from conversion. The population studied is likely to have been positive for some time and, hence, has a life expectancy considerably less than that.

The data used for the reanalyses are as given in Table H-2, as derived from the FDA report. The closing date for recruitment into the individual trials is assumed to correspond to the date of the last dose for the trials, as recorded in the report. The period of recruitment is assumed to be that interval from enrollment of the first patient into a trial and the date of the last dose. The rate of enrollment was assumed to be constant over that interval. The actual period is less than that used herein by a matter of days or a few weeks, depending on the trial.

For the analyses presented below to be informative on a real time basis, one would have to have close surveillance of each study person so as to learn of deaths as they occur or within a matter of days after they occur. Lags in detection or reporting of deaths to the analysis site would degrade the value of the analyses as a real time decisionmaking tool.

Similarly, to be useful, one has to have 100% surveillance of all persons enrolled for mortality, regardless of their status in the trial (i.e., such surveillance including dropouts). We have no way of knowing the extent to which deaths noted in the FDA are the result of such surveillance. The report does not indicate the methods used to identify deaths occurring after the defined period of treatment and followup for the trials discussed. Hence, we have no way of knowing the extent to which the surveillance can be viewed as complete. There is, in fact, intrinsic evidence to suggest it is incomplete simply from the counts of deaths provided. For example, one would expect most of the patients in R89 to have died by mid 1993. The median life expectancy for the population enrolled into that trial was not likely to be much more than a year, if that.

The caveats regarding followup need to be kept in mind when reviewing the results below. An under-reporting of deaths would mean that the actual p-values noted in the tables below would be smaller than given and that the powers would be larger than indicated. Since the ability to trace people diminishes as a function of the time of the last contact, we surmise that the quality of followup was better for the 26 week period of followup than for followup through 27 June 1993. Hence, the emphasis is on analyses relating to 26 weeks of followup.

The results for each trial analyzed separately is given in Tables H-3, -4, and -5 for the low, intermediate, and high rates. The first panel of each table is at the close of each trial assuming 26 weeks of followup for each person enrolled. Note that for this method of analysis only 4 of the 9 deaths observed in the first three trials are counted. The other 5 deaths occurred beyond the closing date for followup and are censored. The second panel is for followup through 27 June 1993.

The sample size, assuming a uniform rate of enrollment over the interval defined by enrollment of the first patient to the last administration of treatment in a given trial is given in the column labeled N.