Types of Studies

This report includes studies that focus on the comparison of outcomes of different androgen suppression monotherapies, the comparison of outcomes of monotherapies versus combined androgen blockade, and the comparison of outcomes of immediate versus deferred initiation of androgen suppression. The primary criterion for selection of studies, with respect to efficacy outcomes, required that studies be designed as randomized controlled trials. For adverse events data, randomized trials were included, and nonrandomized phase II studies were included if they reported the frequency of patients withdrawing from therapy due to adverse events. For quality of life data, all reports that were identified by the search strategy were included.

Types of Participants

Trials that compared different options for monotherapy, or that compared monotherapy with combined androgen blockade, were included if they enrolled men with advanced prostate cancer who were not previously treated with hormonal therapy for prostate cancer. For this review, advanced prostate cancer is defined to include the following two groups:

1.

Those with disseminated and/or symptomatic metastases (defined as stage D1/D2, N+ or M1 disease); and

2.

Those with asymptomatic or minimally advanced disease (defined as stage C or T3-4NXM0 disease).

For trials that compared immediate versus deferred initiation of androgen suppression only, a third patient group was included:

3.

Those with a rising prostate-specific antigen (PSA), or other signs of progression, after surgery or radiotherapy for early stage disease.

Wherever possible, data for subgroups were analyzed separately; however, reports of outcome by subgroups were sparse in this literature. In addition, for purposes of the meta-analyses, sensitivity analyses were performed that restricted inclusion to studies that provided separate data for stage D2 patients. No other patient subgroup was represented by sufficient data to justify performing separate sensitivity analyses.

Trials that compared immediate with deferred treatment were included if they enrolled men in groups "2" and "3," as above.

Types of Interventions

Types of Outcome Measures

Trials were included if they reported at least one of the following outcomes, each of which were compared and analyzed separately:

• overall survival
• cancer-specific survival
• progression-free survival
• time to hormone refractory status
• time to treatment failure
• adverse effects of treatment
• quality of life
• patient preferences or satisfaction

Determining Study Eligibility

Procedures were followed with the intention of systematically screening references to include all studies relevant to this literature review. The title and abstract of each reference retrieved by the search initially were evaluated against the inclusion and exclusion criteria by one of two reviewers. Each reviewer was responsible for 50 percent of the retrieved citations. The reviewer sorted references into three categories: "include," "exclude," and "uncertain." All references sorted into the "include" category by either reviewer were slated for retrieval. All references sorted into the "exclude" category by one reviewer were evaluated by the other reviewer. Any reference sorted as "exclude" which the alternate reviewer sorted as "include" was slated for retrieval. All references sorted as "uncertain" were reconsidered by both reviewers with a bias toward being inclusive. After articles were retrieved, each reviewer evaluated all articles against the inclusion and exclusion criteria. Disagreements were resolved by consensus of the two reviewers.

Data Abstraction

Two reviewers independently collected data for each eligible study, recording it with electronic database software. The data elements that were abstracted, when available in published reports, are detailed in the data abstraction form (Appendix IV). Data elements were grouped into the following broad categories: trial identifiers, study methods (including enrollment and withdrawal numbers), patient characteristics, outcomes, and comments.

Among the data elements sought for these outcomes were the proportion of patients surviving at 1 year, 2 years, 5 years, and 10 years. The median survival durations were also sought for each outcome. When an article did not report these elements specifically, the reviewers estimated them from curves that were available in the published reports. After each reviewer completed data abstraction, the data were compared and disagreements were resolved by consensus of the two reviewers.

Assessing Study Quality

The reviewers assessed the quality of methods and reporting, with the intention of documenting the quality of individual trials and determining whether studies could be grouped into categories by grade of methodologic quality. Quality assessments included the following elements: method of random sequence generation, blinding of the randomization process during patient recruitment, blinding of the investigator and patient to treatment allocation prior to breaking the randomization code (except when the patient underwent surgical castration), reporting of withdrawals and handling of them in the analysis, use of intent-to-treat analysis, reporting of power analysis, whether compliance with treatment was monitored, and description of treatment protocols, including concurrent treatments.

Information on study quality was included in the data abstraction forms and in an evidence table. Decision rules for reviewer assessments on these elements are described in Table 6. Reviewers did not complete a formal quality rating scale for each trial. However, the decision rules that guided assessments of the different elements of methodologic quality are consistent with rules contained in many of the available scales (Chalmers, Smith, Blackburn et al., 1981; Moher, Jadad, Nichol et al., 1995; Mulrow and Oxman, 1997).

Table

Table 6. Decision Rules for Reviewer Quality Assessments.

During data abstraction, the reviewers discovered that information was commonly lacking about many of the above methodologic elements, making it impossible to make a definitive determination about many elements of methodologic quality. After consultation with members of the Technical Advisory Group, a simple method of evaluating study quality was adopted, using the two elements that were most often possible to determine from articles: blinding of intervention and use of intent-to-treat analysis. Studies of higher quality were those that used double-blinding or those where blinding was not applicable and those studies where an intent-to-treat analysis was performed on the outcome of interest. This method of classifying higher quality studies was used in the meta-analyses for the purposes of performing sensitivity analyses. Sensitivity analyses for those studies that were assessed as being of higher quality were performed after meta-analyses were performed on all eligible studies for a particular outcome.

Adverse Events

Abstraction and analysis of data on adverse events present particular difficulties. The difficulties encountered in this project are representative of the general problem of the limitations of clinical trials as a source of data on adverse events. One well-recognized problem is that some adverse events may be so infrequent that clinical trials are not large enough to capture events that may be of concern when the treatment is used in the general population of patients. A second problem is inconsistency in which adverse events are reported and how they are measured. Efforts to improve standards in reporting of randomized trials have emphasized the need for more thorough and systematic reporting of the spectrum of effects for an intervention (Mosteller, Gilber, and McPeek, 1980).

The literature on androgen suppression for the treatment of prostate cancer lacks a consistent framework for reporting adverse events. Studies vary substantially with respect to which adverse events are reported and how these are measured. Reports are often ambiguous with respect to whether failure to report on a specific adverse event means that the event did not occur or that data on the event was not collected. The methods for collecting, pooling, and reporting adverse events data used in this report attempt to systematize, to the extent feasible, a body of evidence that has serious limitations.

There are three types of evidence relating to adverse events: adverse events by category (e.g., cardiovascular, endocrine), adverse events leading to withdrawal from therapy, and adverse events reported by degree of severity. It is generally recognized that each of these methods of androgen suppression has characteristic adverse event profiles, in which certain categories of adverse events are more prominent or of greater clinical importance. Taken together, the primary categories of adverse events for all androgen suppression methods include the following: cardiovascular, endocrine, gastrointestinal, hepatic, and ophthalmologic. This review focused on these categories of adverse events.

Evidence on the severity of adverse events is likely to be of most use to patients and clinicians for the purposes of choosing interventions. However, there are very few data of this type. Only three studies have reported data on adverse events by degree of severity, which is insufficient to compare interventions. Therefore, adverse events leading to withdrawal from therapy serve as indicator of severity.

Supplemental Analysis: Meta-analysis

Meta-analysis describes a set of analytic techniques that allows for synthesis of data across a number of similarly designed studies. One of the rationales for performing meta-analysis is to increase the power by which to detect an effect when individual studies do not obtain statistically significant results. An excellent approach to the meta-analysis of trials in prostate cancer was provided in the analysis by (Caubet, Tosteson, Dong, and coworkers,1997). This report used that general approach with some additional guidance from the paper of (Whitehead and Whitehead,1991). One interesting methodologic problem resulted from the variety of monotherapies available and the fact that they are not all compared against the same control. A general methodology for handling this problem was given by (Hasselblad,1998) and was applied to this particular problem.

Survival rates for prostate cancer are poor, which implies a large value for the hazard rate (the rate of death across time). Although these values may not be exactly constant over time, in general, it was assumed that the hazard is constant. The actual survival curve often is not available and, as a result, those assumptions cannot be checked.

Based on the above assumptions, the object was to obtain estimates of the hazard rate for each arm of each study or to obtain the estimate of the proportional hazards term and its standard error. For studies that do not provide this, estimates from other statistics may be obtained. (Caubet, Tosteson, Dong, and coworkers,1997) suggested that the log-hazard ratio Beta can be estimated from the chi-squared value of the log-rank test:
(1)
where e is the number of events. In those cases where Kaplan-Meier curves are given, it is generally possible to estimate the individual hazards. Finally, in those cases where the survival at a given point in time, S , is given as S = X/N (where X is the number surviving and N is the number of subjects followed, the hazard can be estimated as:
(2)
where t is the time at which the survival was measured.

The hazard rates can be combined using a fixed or random effects model. The fixed effects model assumes that the hazard rates are the same in each study. In this case, the measures are often combined using an inverse variance weighting formula. This assumption is probably unreasonable given the variety of studies. The populations, initial stage, treatment, and length of followup are slightly different in each study.

The model used for the meta-analyses in this document is a random effects model (see, e.g., DerSimonian and Laird, 1986, and Hedges and Olkin, 1985). Random effects models differ from fixed effects models in that a measure of the variation between studies is included in computation of the total uncertainty used to compute weights for each estimate. The random effects mean and variance are used to compute confidence intervals in a manner that corresponds exactly to the fixed effects case.

To use this methodology, a table of hazard rates for each arm of each study must first be constructed. For example, Evidence Table I.13 (Appendix I) shows the 2-year hazard rates for all the monotherapies used in the treatment of prostate cancer. In addition, dummy variables were created to indicate the type of treatment. Finally, dummy variables were created for each individual study (not shown) in order to pool across hazard ratios (similar to that done for odds ratios).

Meta-analyses were performed where appropriate using random effects models. Sensitivity analyses were used to test for heterogeneity of methods (including the effect of including studies of lower methodologic quality), participants, and interventions. For the meta-analysis of combined androgen blockade, sensitivity analysis was also used to test whether studies that reported 5-year survival differed from studies reporting 2-year survival. An initial analysis was performed to determine whether the results of castration and administration of diethylstilbestrol are comparable and thus whether it is valid to pool studies in which the control arms used either of these monotherapies. Separate analyses will also be used to compare the available monotherapies, compare monotherapy with combined androgen blockade, and compare the outcomes of immediate androgen suppression with those of deferred treatment.

Supplemental Analysis: Cost-Effectiveness Analysis

The cost-effectiveness of androgen suppression strategies for patients with advanced prostate cancer also was evaluated for this report. The decision analysis model incorporates benefits and harms of the interventions, captures a broad range of costs, and accounts for quality of life effects. The model is conducted from a societal perspective, with all costs and benefits expressed in present value terms with a rate of time discount of 3 percent, in accordance with the guidelines of the Panel on Cost-Effectiveness in Health and Medicine ( Gold, Seigel, Russell et al., 1996 ). The analysis was conducted with DATA software version 3.0.16 (TreeAge). Further details about the cost-effectiveness analysis methods will be described in the cost-effectiveness analysis chapter.