The Azidothymidine Collaborative Working Group (AZTCG) [5] first showed that zidovudine reduced disease progression and mortality in patients with advanced acquired immunodeficiency syndrome (AIDS)-related complex or AIDS itself [5]. Nineteen patients in the placebo group and only 1 in the zidovudine group died (mean follow-up, 4 months). Subsequent analysis showed that patients originally given zidovudine retained a survival benefit for at least 18 months over patients who were switched from placebo to zidovudine [6]. The efficacy of the drug was subsequently evaluated in earlier stages of the infection [7]. Encouraging short-term results suggested that the indications for initiation of antiviral chemotherapy should include all HIV-infected patients with CD4 cell counts of less than 500 x 10⁶/L, regardless of symptoms [8]. The European-Australian Collaborative Group 020 (EACG-020) [9] even suggested a long-term benefit in asymptomatic patients with CD4 counts higher than 500 x 10⁶/L. The concept of early treatment of HIV-infected adults was challenged by the results of the Veterans Administration 298 (VA-298) [10] and Concorde trials [11], which showed no survival difference between immediate and deferred zidovudine therapy. Some questions remain. Does zidovudine provide a significant benefit for patients without advanced HIV disease? If so, what is the magnitude and duration of the benefit?

To help answer these questions, we did a meta-analysis of all the available data from double-blind, randomized, placebo-controlled trials that addressed the efficacy of zidovudine as monotherapy in HIV-infected adults without AIDS.

Methods

Top Methods Results Discussion Author & Article Info References

Identification, Eligibility, and Quality of Clinical Trials

We identified English-language reports of randomized trials through MEDLARS (Medical Literature Analysis and Retrieval System) of the National Library of Medicine. Specific databases searched included MEDLINE and MEDLINE backfiles, AIDSLINE, AIDSTRIALS, AIDSDRUGS, and CHEMID. We also searched Current Contents and the proceedings of the International Conferences on AIDS. We reviewed bibliographies of pertinent articles to ensure that all eligible trials were identified. The last search was done in mid-June 1994.

To be included in the meta-analysis, a clinical trial had to meet the following criteria: 1) a double-blind, randomized placebo-controlled design; 2) a patient population consisting of asymptomatic HIV-infected persons or symptomatic HIV-seropositive persons without a clinical diagnosis of AIDS (as defined by the 1987 Centers for Disease Control and Prevention [CDC] criteria); 3) a patient age of 12 years or older; 4) patients who had not previously or concurrently received antiretroviral agents other than zidovudine; and 5) the ability to distinguish patients with two or more disease events so that multiple counting was avoided. We excluded trials that addressed the efficacy of zidovudine given immediately after exposure to HIV or immediately after a diagnosis of HIV infection. Clinical trials were included regardless of zidovudine dosage or whether prophylaxis for Pneumocystis carinii infection was used. For trials that addressed symptomatic HIV-infected adults with and without AIDS, only the subgroup without an AIDS-defining illness was included in the meta-analysis.

We used a previously described method of assessing the quality of the included clinical trials for quality scoring [12]. We rated more than 30 elements of the study design, randomization, blinding, statistical analysis, and reporting of the studies using a score ranging from 0 (lowest quality) to 1 (highest quality).

Statistical Analysis

We pooled all the data using the DerSimonian and Laird random-effects model (13; see Appendix) that considers both within-study variance and variability among studies. When present, variability among studies increases the width of the confidence interval. We also used the Mantel-Haenszel (fixed-effects) model [14] for comparison calculations; this approach allowed us to verify the validity of the random-effects estimates in some analyses in which the numbers of events were small (the Mantel-Haenszel model works well even in this case). Unless otherwise specified, the results of the random-effects model are given. We used the Statistical Analysis System [15] to calculate the weighted regressions. Two-tailed P values and 95% CIs were used. Because there were few major comparisons, we did not adjust CIs and P values for multiple comparisons. All calculations were based on an intention-to-treat analysis, with patients analyzed in the manner in which they were randomly assigned.

Extraction of Data on Disease Progression

We extracted data for three types of disease events: progression to clinical AIDS (by the 1987 CDC criteria or by the 1982 criteria for studies done before 1988) or HIV-related death; any clinical progression (in which patients reached a primary or secondary clinical end point as defined in each study); and any progression to a primary end point (in which patients reached any primary clinical or CD4 cell count end point as defined in each study). For all end points, deaths probably related to treatment and deaths unrelated to HIV infection that occurred with preceding HIV progression (HIV-related symptoms but not necessarily AIDS) were included in the count of events. Additional analysis was done to examine the effect of counting all deaths, and results are reported if they differed. The control group was characterized as having "deferred" therapy because patients were usually unblinded when they reached a primary end point and the option of open zidovudine therapy could then be offered to placebo recipients (except for the first zidovudine trial [5], in which blinded follow-up continued for the duration of the study). When studies did not provide the total number of person-years of follow-up in each arm, we calculated this number by multiplying the number of patients in this arm by the mean or median duration of follow-up as given in each trial. The total number of person-years until the development of a primary end point was calculated similarly by multiplying the number of patients in this arm by the mean or median duration of follow-up until the development of primary end points as given in each trial.

Pooling of Efficacy Data and Regression Analysis

We estimated risk ratios for progression to any primary end point, any clinical end point, and AIDS or death using the total number of patients initially enrolled in each arm of each trial. We estimated incidence rate ratios and incidence rate differences [14] for progression to any primary end point using events per patient-years of follow-up (until the development of a primary end point). Because exact data on the number of person-years until the development of any clinical end point and AIDS or death were not consistently available, we could not do similar calculations for these two types of end points.

We separately pooled the clinical trials that had a mean follow-up of 4 to 14 months ("short-term") from those that had a mean follow-up of 21 to 37 months ("long-term"). No studies had a mean follow-up lasting between 14 and 21 months.

Univariate and multivariate regressions of the incidence rate differences for progression to any primary end point were done against three variables that characterized each trial: mean duration of follow-up (in months; we excluded follow-up after the development of a primary end point); presence or absence of symptoms at study entry; and the rate of progressions to a primary end point in the control (deferred zidovudine therapy) group. Each trial was weighted by the inverse of the variance of its incidence rate difference.

Subgroup Analysis of Progression to AIDS or Death

All trials gave data for progression to AIDS events or death and numbers of patients in each treatment arm for the subgroup of patients with CD4 cell counts greater than 500 x 10⁶/L and the subgroup with CD4 cell counts less than 500 x 10⁶/L. Data were not consistently available for separate stratification of patients with CD4 cell counts less than 200 x 10 (⁶/L) and patients with CD4 cell counts between 200 and 500 x 10⁶/L. In this analysis, we included in the count of events HIV-related deaths and deaths not related to HIV that were associated with previous progression to AIDS.

We estimated the risk ratio for disease progression for the subgroups with CD4 counts greater than 500 x 10⁶/L and the subgroups with CD4 counts less than 500 x 10⁶/L in each trial. Pooled risk ratios were estimated for different baseline disease stages. These stages included the following: asymptomatic patients with CD4 cell counts greater than 500 x 10⁶/L; asymptomatic patients with CD4 cell counts less than 500 x 10 (⁶/L); symptomatic patients with CD4 cell counts greater than 500 x 10⁶/L; and symptomatic patients with CD4 cell counts less than 500 x 10⁶/L. Similarly, we calculated pooled risk differences for progression to AIDS or death for each of the disease stages.

We did univariate and multivariate regressions of the natural logarithm of the risk ratio against four variables that characterized each trial subgroup: duration of total follow-up (short-term compared with long-term); presence of symptoms at entry; CD4 cell count (< 500 x 10⁶/L compared with greater than 500 x 10⁶/L); and the rate of events in the control group (deferred zidovudine therapy arm). In these regressions, each study was weighted by the inverse of the variance of the natural logarithm of the risk ratio.

Pooling of Toxicity Data

We did a separate analysis by pooling the data on the development of severe anemia, which was defined as a hemoglobin level of less than 80 g/L. We used data on events per 100 patient-years (total follow-up). The data were analyzed in four subgroups depending on the disease stage (asymptomatic or symptomatic) and the daily dose of zidovudine (500 to 1000 mg or 1200 to 1500 mg). Similarly, we analyzed the available data on the development of severe neutropenia (defined as an absolute neutrophil count of less than 0.75 x 10⁹/L).

Validity of Disease Progression Data

To assess the validity of data, we examined the percentage of patients in each trial arm that was unblinded before the termination of the study for any reason other than reaching a primary end point or HIV-related death. We noted any deviations from the original double-blinded protocol. The effect of noncompliance was estimated by evaluation of levels of zidovudine and its glucuronide in serum or by surrogate markers (such as an increase in the mean corpuscular volume of erythrocytes that was noted in some studies). In addition, for studies that provided P. carinii prophylaxis to less than 20% of the enrolled patients, we calculated the weighted risk ratio for AIDS or death being caused by P. carinii in the deferred compared with the early zidovudine therapy group. Finally, we noted the race and sex percentages and the predisposing risk factors for HIV of patients in each study.

Results

Top Methods Results Discussion Author & Article Info References

Study Characteristics and Quality Scoring

We identified 10 double-blind, randomized, placebo-controlled trials on the efficacy of zidovudine in HIV-infected adults without AIDS [5, 7, 9-11, 16-20]. Although one study [16] met all other criteria and provided data on opportunistic infections, Kaposi sarcoma, lymphoma, and death events, it did not discriminate among patients with two or more events; it was therefore excluded. This trial included less than 100 person-years of follow-up. The remaining nine trials combined had more than 10 000 person-years of follow-up. One trial [7] included two separate treatment groups that received different doses of zidovudine; this allowed two comparisons. Altogether, 10 comparisons were pooled in the meta-analysis.

The studies had a mean quality score of 0.75 (range, 0.57 to 0.94), which is considered to be high compared with scores of trials in other clinical domains that we have scored in the past.

Selected characteristics of the studies are listed in Table 1. These trials enrolled patients with differing disease severity at study entry; the predefined primary end points also varied. Generally, trials in patients with more advanced disease at entry also used more advanced primary end points. With allowance for slight variations in definitions among studies, studies of patients with similar disease severity at entry used similar end points.

Data on Disease Progression

The data on disease progression as defined by three different end points are presented in Table 2. There were 459 progressions to any primary end point in the early zidovudine therapy groups compared with 652 in the deferred therapy groups. A total of 454 and 626 clinical end points were reached in each group, respectively. Only 8% of patients in the early zidovudine therapy arm and 11.8% of the patients in the deferred therapy arm progressed to AIDS or death (as defined in the Methods section). Fewer than 3% of the enrolled patients were followed until they died of HIV infection.

The EACG-020 investigators censored data on disease progression 3 months after treatment stopped. However, a preliminary analysis [21] suggests that the results of this study are similar if late events in patients withdrawn early from the study are counted. Similarly, in ACTG-016, AIDS developed in five patients after the development of severe AIDS-related complex. However, because the investigators did not mention to which arm of the trial these patients belonged, they were not included in the analysis. Their number is too small to affect the calculations.

Total follow-up was 5641 person-years in the early zidovudine therapy arm of the comparisons and 5629 in the deferred therapy arm. Follow-up to a primary end point differed substantially from the total follow-up in two trials. In VA-298, the follow-up to a primary end point was 211 person-years in the early therapy arm and 193 person-years in the control arm (total follow-up, 385 and 395 person-years, respectively). In the Concorde study [11], the follow-up to a primary end point was 2363 person-years in the early therapy arm and 2236 person-years in the control arm (total follow-up, 2717 and 2702 person-years, respectively).

Pooling of Efficacy Data

Progression to Any Primary and Clinical End Point

Figure 1 shows the incidence rate ratios for disease progression to any primary end point. All studies except that of the European-Australian Haemophilia Collaborative Study Group [19] gave an incidence rate ratio estimate that favored early zidovudine therapy. Overall, early zidovudine use was associated with a 41% reduction in the number of events (95% CI, 26% to 53%). The benefit was more obvious in the short-term trials, which had a 49% (CI, 36% to 59%) reduction in the number of disease-progression events. Some beneficial effect continued in the long-term trials but did not attain statistical significance (27% reduction; CI, –3% to 48%). Similar risk ratios were obtained when calculations were based on the number of patients rather than on the number of patient-years: The pooled risk ratio for all trials was 0.58 (CI, 0.46 to 0.75) by the random-effects model and 0.70 (CI, 0.63 to 0.77) by the fixed-effects model; the pooled risk ratios were 0.53 (CI, 0.44 to 0.64) for short-term trials and 0.77 (CI, 0.52 to 1.13) for long-term trials.

View larger version (28K): [in this window] [in a new window]   Figure 1. Incidence rate ratios for progression to any primary end point. Calculations are based on events per patient-years. ACTG = AIDS Clinical Trials Group; AZTCG = Azidothymidine Collaborative Working Group; EACG = European-Australian Collaborative Group; VA = Veterans Administration; ZDV = zidovudine.

Analysis of the data on progression to any clinical end point yielded risk ratios similar to those obtained in the analysis of progressions to any primary end point both for individual trials and for pooled estimates. The estimated pooled risk ratios were 0.61 (CI, 0.48 to 0.77) for all trials, 0.51 (CI, 0.39 to 0.66) for short-term trials, and 0.78 (CI, 0.60 to 1.01) for long-term trials.

The pooled estimates did not change when all deaths unrelated to HIV infection were considered in the calculations.

Progression to AIDS or Death

Figure 2 shows the risk ratios for progression to AIDS or death. With the exception of the European-Australian Haemophilia Collaborative Study Group trial, each study had an estimated risk ratio that favored early zidovudine therapy. For several trials, however, the benefit was small and not statistically significant. For the Concorde trial, the trend is reversed if the few deaths unrelated to HIV infection are included in the analysis, as was done in the original article. The pooled risk ratio favored early zidovudine use with a 45% (CI, 21% to 62%) reduction in the number of events. The short-term benefit was similar to the benefit seen for progression to any primary or clinical end point. A nonsignificant trend for long-term efficacy was seen (risk ratio, 0.88; CI, 0.73 to 1.08), but it disappeared when all deaths unrelated to HIV infection were included in the calculations (risk ratio, 0.96; CI, 0.81 to 1.13). Inclusion of all the deaths unrelated to HIV infection did not affect the other pooled risk ratio estimates.

View larger version (28K): [in this window] [in a new window]   Figure 2. Risk ratios for progression to the acquired immunodeficiency syndrome or death. Calculations are based on events per number of patients. ACTG = AIDS Clinical Trials Group; AZTCG = Azidothymidine Collaborative Working Group; EACG = European-Australian Collaborative Group; VA = Veterans Administration; ZDV = zidovudine.

Regression of Incidence Rate Difference for Progression to Any Primary End Point

Although incidence rates and risk ratios provide an estimate of the relative benefit, the incidence rate difference measures the absolute benefit. In a univariate regression, the incidence rate difference for progression to any primary end point correlated significantly with the presence of symptoms (regression coefficient [ß] = 8.2; P = 0.001) and the rate of events in the control (deferred therapy) group (ß = 0.42; P = 0.001) as well as with the duration of follow-up (ß = –0.16;P = 0.03). In a multivariate regression, the adjusted regression coefficients were significant for both duration of follow-up (ß = –0.22;P = 0.01) and for the rate of events in the control group (ß = 0.49; P = 0.001). The adjusted coefficient for symptoms was not significant because symptoms were highly correlated with the rate of events in the control group.

This regression analysis signifies that a larger absolute benefit could be realized when patients are expected to have more progression events and that the absolute benefit of early zidovudine therapy decreases over time. The model had a good fit to the data (adjusted coefficient of determination, 0.89).

Analysis of Progression to AIDS or Death

Subgroup Data

The 14 available comparisons could be grouped into 6 strata depending on the presence of symptoms, CD4 cell count, and the trial's mean duration of follow-up (Table 3). Three trials provided data for two separate CD4 subgroups. In addition, patients from a short-term study (AZTCG) [6] were followed after the study's completion in an observational, open-label zidovudine study in which all patients wishing to take zidovudine (both from the original therapy group and the original placebo group) were observed for disease progressions for an additional 18 months. This continuation study was not blinded or placebo-controlled and therefore was not included in our original analysis. However, because it provided long-term follow-up data on the original randomized, double-blind trial, we included it in our analysis of stratified data. Patients participating in the VA-298 study were also offered open-label zidovudine at the conclusion of the blinded trial and had continued follow-up. Although data from an abstract [22] were available on progression to AIDS and on deaths during a total follow-up of 36 months (9 more months than that reported in the published article), we could not extract data on AIDS or death. We therefore excluded these results so that they were not counted twice.

Pooling of Stratified Subgroup Data

As shown in Table 3, no subgroup of patients had a statistically significant long-term benefit as expressed by the pooled risk ratio for disease progression. However, some trend of efficacy was seen both for the asymptomatic patients with CD4 counts greater than 500 x 10⁶/L (risk ratio, 0.76; CI, 0.48 to 1.19) and for symptomatic patients with CD4 counts less than 500 x 10⁶/L (risk ratio, 0.77; CI, 0.45 to 1.32). When we included in our calculations the few deaths unrelated to HIV infection, this trend was further reduced (risk ratios, 0.90 [CI, 0.70 to 1.15] and 0.91 [CI, 0.60 to 1.39], respectively). Even if real, the absolute benefit for asymptomatic patients with CD4 counts greater than 500 x 10⁶/L was small: Only 1.2 (CI, –0.7 to 3.1) of 100 patients avoided AIDS or HIV-related death (Table 4).

In contrast Table 3, a statistically significant benefit was shown for both asymptomatic and symptomatic patients in short-term trials who had CD4 counts less than 500 x 10⁶/L (risk ratios, 0.43 [CI, 0.30 to 0.64] and 0.26 [CI, 0.13 to 0.56], respectively). The absolute magnitude of the benefit was more substantial in symptomatic patients and was the highest in trials that enrolled patients with more advanced symptomatic disease (Table 4). Inclusion of the few deaths unrelated to HIV infection did not affect the short-term estimates.

Regression of the Risk Ratio for Progression to AIDS or Death

In univariate regressions, the relative benefit from early zidovudine therapy in reducing progression to AIDS or death was greater in trials with short-term follow-up (ß = –0.85;P = 0.001) but did not correlate with the presence of symptoms at entry, the CD4 count, or the rate of events in the control group. However, in multivariate regression, the relative benefit was significantly associated with both short-term follow-up (ß = –0.56;P = 0.03) and the presence of symptoms at entry (ß = –0.62;P = 0.02). A marginal inverse association was seen with the rate of events in the control group (ß = 0.019; P = 0.05), but the regression coefficient was small (risk-ratio estimate, 1.01; CI, 1.00 to 1.03). The adjusted coefficient of determination for the model was 0.72.

These regression coefficients further support the notion that the relative benefit of early zidovudine therapy is greater in the short term. For a given duration of follow-up, CD4 category, and rate of events in the control group, patients with symptoms experience a larger relative benefit.

Pooling of Toxicity Data

The increased risk for development of severe anemia in asymptomatic patients receiving early zidovudine at doses similar to the currently recommended dose of 600 mg (range, 500 to 1000 mg) was statistically significant (incidence rate ratio, 2.1; CI, 1.1 to 4.1), but it was small in magnitude (0.4 events per 100 person-years) (Table 5). For symptomatic patients, the excess incidence of severe anemia probably reflected the high doses of zidovudine that were used, coupled with increased vulnerability to bone marrow toxicity in the symptomatic stages of the disease. The excess incidence of severe neutropenia in asymptomatic recipients of 500 to 1000 mg of zidovudine daily was small (1.1 events per 100 person-years; P = 0.07). Unfortunately, no studies used this dose range in symptomatic patients for a comparison. Overall, the toxicity results should be interpreted cautiously because the data vary considerably.

Validity of Data on Disease Progression

The percentage of patients not receiving randomized, blinded therapy at the termination date of the study for reasons other than reaching an end point or dying was similar in all trials, ranging from 16% to 39% with the exception of the Concorde study. The corresponding percentage for the Concorde study was 54%, but only 19% and 16% of the trial time for each arm was spent receiving therapy that deviated from the original randomized treatment. The percentage of patients not adhering to the blinded, randomized treatment in each trial was of the same magnitude in the zidovudine and control arms. An absolute difference of more than 12% was seen only in the ACTG-019 study (Table 1).

Because of modifications of trial protocol, 39 of the 168 patients assigned to receive placebo in the VA-298 study [10] switched to open-label zidovudine without having reached an end point at a mean of 10.2 months before the end of the study. Similarly, in the Concorde trial, a protocol amendment allowed 204 of the 872 placebo recipients to start open-label zidovudine on the basis of clinical discretion and a low CD4 count (<500 x 10⁶/L); nevertheless, the placebo group spent only 16% of their trial time receiving zidovudine before reaching an end point [11]. In the EACG-017 trial, 18 of 162 patients (11%) receiving placebo took advantage of a protocol amendment and changed to open-label zidovudine when their CD4 cell count decreased to less than 200 x 10⁶/L [20].

Deviations from the original protocol allowing open-label zidovudine use before an end point was reached could have affected the magnitude of the observed zidovudine efficacy. The most likely effect of cross-overs from placebo to active treatment is a diminution in the observed treatment effect and rates of events in the control groups. Patients withdrawn because of rapidly decreasing or very low CD4 cell counts could have progressed to a clinical end point soon after the observation period [23]. Similarly, patients lost to follow-up might have been closer to reaching an end point or might even have reached one. Although these deviations and withdrawals do not appear large enough to markedly influence the results, they may affect the precision of the data. This is a limitation that can be partially overcome only by analysis of individual patient data, which were not provided in the published articles. Similarly, the effect of an on-treatment approach, as opposed to the intention-to-treat approach used in these trials, could be investigated with such data.

Noncompliance was generally minimal and probably even less than what is expected in clinical practice: All studies except the ACTG-019 study gave estimates of noncompliance of less than 10%. The ACTG-019 study detected zidovudine in 79% to 84% of the serum samples of patients randomly assigned to receive zidovudine (although only 8% of the zidovudine group specimens had an erythrocyte mean corpuscular volume of less than 95 fL). Moreover, 9% and 3% of the serum samples from the placebo groups of the ACTG-019 and the Concorde trials, respectively, contained zidovudine.

Pneumocystis carinii prophylaxis was given to fewer than 20% of patients in each study except the VA-298 and Concorde studies. Six of the eight trials that did use substantial prophylaxis provided reasons for progression to AIDS or death: Fifty-four of the 111 progressions in the placebo group were caused by P. carinii infection compared with 21 of the 49 progressions in the zidovudine group. The corresponding risk ratio was 1.02 (CI, 0.71 to 1.47). Thus, there is little chance that the absence of P. carinii prophylaxis disproportionately increased the number of progressions to AIDS caused by Pneumocystis carinii pneumonia in the arm that showed higher rates of progression.

In the Concorde and EACG-020 studies, women composed 15% of the patient population; in all other studies, the percentage ranged from 1% to 8%. Fewer than 600 women were studied in all the trials combined. Only the VA-298 study had a substantial number of black patients (37% and 34% in each arm). In all other studies that provided data on race, white patients composed 91% to 99% of all patients. Most patients were homosexual or bisexual men. The percentage of intravenous drug users varied from 9% to 16%. Heterosexual contact was the mode of HIV transmission for 6% to 14% of all patients.

Separate data on disease progression events among female patients were available only in the ACTG-016 and ACTG-019 studies. A previous analysis of these data from individual patients [24] showed that women probably experience the same benefit from zidovudine as men. The beneficial effects of zidovudine reported for the entire studied population also applied to subpopulations of blacks, Hispanics, and intravenous drug users [23]. We combined the subpopulation data at the treatment group level from these two trials and the VA-298 trial. These trials cumulatively studied 244 black patients, 221 Hispanic patients, and 313 intravenous drug users. The risk ratio estimates for progression to AIDS were 0.56 (CI, 0.26 to 1.21) for black patients, 0.51 (CI, 0.14 to 1.91) for Hispanic patients, and 0.71 (CI, 0.38 to 1.32) for intravenous drug users. These pooled estimates favor early zidovudine use in these subpopulations and probably lack statistical significance only because of the few patients available for pooling.

Discussion

Top Methods Results Discussion Author & Article Info References

Our meta-analysis included more than 6000 HIV-infected patients without AIDS and more than 11 000 person-years of follow-up. Pooling of all data according to the mean duration of follow-up in each trial showed that early initiation of zidovudine was efficacious in the short term (for all studies lasting as long as 14 months) for all the end points that were addressed. This benefit decreased in trials that had longer follow-up. A trend for long-term efficacy (longer than 21 months) was noted when all primary end points were considered. This beneficial trend was smaller for progression to AIDS or death. The sparse data on survival suggest no evidence of a long-term survival benefit for either asymptomatic (the Concorde trial) [11] or symptomatic patients (the VA-298 trial) [10, 22] who had early therapy. A meta-analysis of all the available survival data showed no benefit from early zidovudine [25]. Yet, only slightly more than 3% of all patients enrolled in the analyzed studies died of HIV infection during the trial follow-up periods. Therefore, the statistical power of the mortality data is low [26].

A delay in the progression of a devastating illness may be an important benefit even if overall survival is not prolonged. To put the results of this analysis in practical terms Table 4, on the average, for every 100 asymptomatic patients with CD4 counts greater than 500 x 10⁶/L, fewer than 12 are expected to progress to AIDS or death within 3 years; fewer than 3 (if any) would be expected to avoid this progression by receiving early zidovudine therapy. Of 100 asymptomatic patients with CD4 counts less than 500 x 10⁶/L, 6 to 9 are expected to progress to AIDS or death in 1 year, and between 2 and 6 will avoid this progression by receiving early zidovudine therapy. This benefit probably disappears by the third year of therapy. At the other end of the spectrum, of 100 patients with advanced AIDS-related complex who have a 22% to 46% chance of developing an AIDS-defining illness or dying within 4 months, more than 8 are expected to evade these events by starting zidovudine therapy rather than deferring therapy [5]. This benefit lasts for about 2 years [6].

In clinical practice, initiation of antiretroviral therapy depends not only on issues of efficacy but also on side effects from the medication [27] and the overall quality of life. In the ACTG-019 study, the loss in quality of life from drug side effects was equal to the gain caused by delay in disease progression [28]. The cost of treatment should also be considered [29].

The pooled data suggest that zidovudine efficacy decreases over time. Resistance to zidovudine has been widely shown at the molecular level [30] and further supports the idea that zidovudine monotherapy has time-limited efficacy. If so, then the drug might be reserved for patients with AIDS, the only group that has been shown to experience a survival benefit from zidovudine for at least 21 months [6] despite more common side effects [27]. The Multicenter AIDS Cohort Study [31] showed that patients who start treatment with antiretroviral agents within 3 months of the onset of AIDS have longer survival after AIDS onset than patients who receive antiretroviral treatment before their AIDS-defining illness. However, a large observational study [32] showed that even in patients with AIDS, the survival benefit disappears after the second or third year of treatment.

The approach of deferring antiretroviral therapy until late stages of the disease does not consider several potentially important factors. Asymptomatic patients with high CD4 cell counts have few major side effects, whereas very sick patients may experience a high morbidity from zidovudine side effects, with a larger reduction in the quality of life and a larger cost of treatment. The correlation of in vitro resistance with clinical failure has not been proved, and viral load, development of syncytia-inducing phenotypes, and other unknown variables may affect the clinical outcome regardless of zidovudine susceptibility. Viral load, as measured by messenger RNA expression in peripheral blood cells, has been shown to predict disease progression independently of surrogate markers such as the CD4 cell count [33]. The continuous presence of substantial levels of HIV in the plasma during all stages of infection [34] may require continuous antiviral treatment starting at an early stage of illness. Loss of immune function may be irreversible after the viral burden progresses beyond a certain point. Therefore, initiating treatment too late may not allow full exploitation of the potential benefits of current antiretroviral therapy.

Perhaps most importantly, the availability of alternative antiretroviral agents and combination therapy may influence the decision on when to initiate zidovudine therapy. We did not consider the current practice of switching to other antiretroviral agents or using combinations of antiretroviral agents after zidovudine monotherapy loses efficacy [2]. In the future, it may be possible for the effectiveness of antiretroviral treatment to last more than 3 years by using different medications sequentially or in combination.

In summary, early initiation of zidovudine delays disease progression, but this benefit has a limited duration. Symptomatic patients experience a larger benefit than asymptomatic patients. Analysis of individual patient data may improve the precision of our findings. More importantly, longer follow-up of the analyzed trials and additional data on sequential or combination therapy with other drugs is critical for defining the proper role of zidovudine in the management of HIV-seropositive patients.

Addendum

Since this paper was submitted, long-term data for progression to AIDS or death were published for the ACTG-019 trial [35] (mean follow-up, 2.6 years). The study had been unblinded in August 1989, and open-label zidovudine had been offered to all participants. According to our protocol, we incorporated these long-term data in the subgroup analyses. Their inclusion in the respective subgroup did not change our results. Asymptomatic patients with CD4 counts less than 500 x 10⁶/L did not experience any long-term (>21 months) benefit from early initiation of zidovudine therapy: The pooled risk ratio for progression to AIDS or death was 0.98 (CI, 0.85 to 1.14), and the corresponding pooled risk difference was 0.3 (CI, –1.2 to 1.7) progressions spared per 100 person-years. Similarly, for trial subgroups, the results of the regressions of the relative benefit (natural logarithm of the risk ratio of progression to AIDS or death) did not change. In multivariate regression, the relative benefit from early initiation of zidovudine was still significantly larger in short-term trials (ß = –0.59;P = 0.006) and in symptomatic patients (ß = –0.61;P = 0.01). This relative benefit was slightly decreased in trials with very high rates of progression in the control group (ß = 0.018; P =0.03). The latter finding may be consistent with the ACTG-019 analysis [35] showing that the observed long-term benefit was driven by the subgroup of patients with CD4 counts higher than 300 x 10⁶/L, whereas patients with lower CD4 counts (200 to 300 x 10⁶/L) had higher event rates and experienced no long-term benefit.

Appendix: Pooled Analysis of the Risk Ratio, Incidence Rate Ratio, and Incidence Rate Difference Using the DerSimonian and Laird Random-Effects Model

Pooled Risk Ratio

Let t_i and c_i be the number of treated and control patients, respectively, with the event of interest in the ith study. For n studies to be pooled, let n_ti and n_ci represent the number of persons in the treatment and control groups, respectively, in the ith study (i = 1, 2,., n). The estimated natural logarithm of the risk (cumulative incidence) ratio in the ith study is

(t_i/nti)

ln (RR_i) = ln—————————

(c_i/nci)

Its standard error is

(n_ti – t_I

n_ci – c_i)

se_I = (—————————+—————————)^–1/2

(t_i n_ti c_I n_ci)

and the factor by which the estimated ln (RR_i) is weighted in the classic fixed-effects analysis is

w_i = 1/(se_i)².

The overall fixed-effects risk ratio across all n studies is equal to

Sigmaw_i ln (RR_i)

ln (RRf) =—————————————.

Sigmaw_i

The hypothesis that the n underlying risk ratios are equal can be tested by comparing the statistic

Q = Sigmaw_i (ln [RR_i] –ln [RRf])²

to a given percentile (for example, the 95th) of a chi-square distribution with n –1 degrees of freedom. The DerSimonian-Laird analysis [13] differs from its fixed-effects counterpart when

Q –(n –1)

D =————————————————-

Sigmaw²_i

Sigmaw_i-—————————

Sigmaw_i

is greater than zero. Let L = D if D > 0; if not, L = 0.

The DerSimonian-Laird weight factor for study i is equal to

1

w^*_i = (L +———-)^–1

w_i

Let

Sigmaw^*_i ln (RR_i)

ln (RR^*) =————————————————.

Sigmaw^*_i

The overall estimate of the random-effects risk ratio becomes

RR^* = e^ln (RR^*)

and the corresponding 95% confidence interval becomes

e(ln [RR^*] ± 1.96/radical[Sigmaw^*_i]).

Pooled Incidence Rate Ratio

The pooled analysis of the incidence rate ratio (with person-time data) is similarly computed by replacing the estimate of ln (RR_i) with

(t_i/ptti)

ln (IRR_i) = ln—————————-,

(c_i/ptci)

where pt_ti and pt_ci denote person-time in the treatment and control groups, respectively, in the ith study, and by using

se_i = ([1/t_i] + [1/c_i])^–1/2.

Pooled Incidence Rate Difference

The estimated incidence rate difference in the ith study is

t_i c_i

IRD_i =————- -————-,

pt_ti..

TX.-

pt_ci

with standard error

t_i c_I)^–1/2

se_i = (———————- +————————)

(pt²_ti..

TX.-

pt²_ci)

and fixed-effects weight

w_i = 1/(se_i)².

The overall fixed-effects incidence rate difference is

Sigmaw_i (IRD_i)

IRDf =————————————

Sigmaw_i

The corresponding Q statistic is

Q = Sigma_i (IRD_i – IRDf)²

The values for D, L, and w^*_i are based on the same general formulas given above for the risk ratio. The overall estimate of the random-effects incidence rate difference becomes

Sigmaw^*_i (IRD_i)

IRD sup^* =———————————————

Sigmaw^*_i

and the 95% confidence interval becomes

IRD^* ± 1.96/radical(Sigmaw^*_i).

Grant support: By R01 MH45686 from the National Institute of Mental Health (Dr. Sacks and Dr. Lau); R01 HS07782 from the Agency for Health Care Policy and Research of the U.S. Public Health Service (Dr. Lau); and T32 AI07389 (Dr. Ioannidis), K08 AI01046 and R01 AI33290 (Dr. Skolnik), and U01 AI27667 (Dr. Sacks) from the National Institute of Allergy and Infectious Diseases.