Article type
Year
Abstract
Introduction:
Objectives: The acceptable approach in meta-analysis is to conduct a comprehensive search for all eligible trials. However, the effort may not be warranted given the impact on the conclusions when the principal goal is to estimate a summary effect size as opposed to investigating study heterogeneity or analyzing subgroups. In a meta-analysis assessing the dyspepsia risk associated with nonsteroidal antiinflammatory drugs (NSAIDs), we compared the estimated relative risks of dyspepsia, and the precision of those estimates, based on different methods of sampling trials.
Methods: All published randomized controlled trials (RCT's) were identified by an unrestricted search of MEDLINE, EMBASE, HEALTHSTAR and BIOSIS from 1966-1997. Titles, abstracts and/or reports were reviewed to identify studies that included a placebo group, studied adults, used oral NSAIDs for more than 4 days, and reported dyspepsia by subject. Data extraction was performed in duplicate with consensus resolution of discrepancies. Four methods for sampling studies to analyze were compared: (1) all eligible studies; (2) all large (sample size n > 200) studies identified at the abstract screening stage; (3) a 50% simple random sample of titles drawn at the title screening stage; and (4) all large studies and a 25% simple random sample of small studies identified at the abstract screening stage. For each sample of studies, the relative risk of dyspepsia and a 95% confidence interval were estimated using a random effects model. For methods (1) and (2) respectively, only one subgroup of trials is possible and the pooled result for those trials is reported. If method (3) or (4) is applied, different subgroups of studies might be sampled, and the median pooled relative risk, and median lower and upper confidence interval bounds are reported based on 100 simulated samples.
Results: 4849 published titles were identified, from which 1768 articles were selected, 96% of which were retrieved and reviewed, yielding 37 eligible trials, of which 12 had n > 200. The analysis produced the following results for the different sampling methods:
Sampling Method Number of Studies Relative Risk 95% CI
(1) 37 1.46 (1.16,1.84)
(2) 12 1.58 (1.14,2.19)
(3) 18 1.45 (1.05,2.01)
(4) 18 1.53 (1.15,2.02)
Discussion: For meta-analyses whose primary purpose is to estimate an overall treatment effect, and for which many RCT's are available, sampling studies may provide a more efficient use of resources to conduct the analysis with only a small loss of precision. However, this gain in resource efficiency must be balanced against a possible loss in ability to investigate study heterogeneity and treatment differences between subgroups.
Objectives: The acceptable approach in meta-analysis is to conduct a comprehensive search for all eligible trials. However, the effort may not be warranted given the impact on the conclusions when the principal goal is to estimate a summary effect size as opposed to investigating study heterogeneity or analyzing subgroups. In a meta-analysis assessing the dyspepsia risk associated with nonsteroidal antiinflammatory drugs (NSAIDs), we compared the estimated relative risks of dyspepsia, and the precision of those estimates, based on different methods of sampling trials.
Methods: All published randomized controlled trials (RCT's) were identified by an unrestricted search of MEDLINE, EMBASE, HEALTHSTAR and BIOSIS from 1966-1997. Titles, abstracts and/or reports were reviewed to identify studies that included a placebo group, studied adults, used oral NSAIDs for more than 4 days, and reported dyspepsia by subject. Data extraction was performed in duplicate with consensus resolution of discrepancies. Four methods for sampling studies to analyze were compared: (1) all eligible studies; (2) all large (sample size n > 200) studies identified at the abstract screening stage; (3) a 50% simple random sample of titles drawn at the title screening stage; and (4) all large studies and a 25% simple random sample of small studies identified at the abstract screening stage. For each sample of studies, the relative risk of dyspepsia and a 95% confidence interval were estimated using a random effects model. For methods (1) and (2) respectively, only one subgroup of trials is possible and the pooled result for those trials is reported. If method (3) or (4) is applied, different subgroups of studies might be sampled, and the median pooled relative risk, and median lower and upper confidence interval bounds are reported based on 100 simulated samples.
Results: 4849 published titles were identified, from which 1768 articles were selected, 96% of which were retrieved and reviewed, yielding 37 eligible trials, of which 12 had n > 200. The analysis produced the following results for the different sampling methods:
Sampling Method Number of Studies Relative Risk 95% CI
(1) 37 1.46 (1.16,1.84)
(2) 12 1.58 (1.14,2.19)
(3) 18 1.45 (1.05,2.01)
(4) 18 1.53 (1.15,2.02)
Discussion: For meta-analyses whose primary purpose is to estimate an overall treatment effect, and for which many RCT's are available, sampling studies may provide a more efficient use of resources to conduct the analysis with only a small loss of precision. However, this gain in resource efficiency must be balanced against a possible loss in ability to investigate study heterogeneity and treatment differences between subgroups.