Marlene R. Miller, MD, MSc Robert L. McNamara, MD, MHS Jodi B. Segal, MD, MPH Nina Kim, MA Karen A. Robinson, MSc Steven N. Goodman, MD, PhD Neil R. Powe, MD, MPH, MBA Eric B. Bass, MD, MPH Baltimore, Maryland From the Division of Pediatric Cardiology (M.R.M.), Division of Cardiology (R.L.N.), Division of General Internal Medicine (J.B.S., N.R.P., E.B.B.), Oncology Center (S.N.G.), Division of Biostatistics at Johns Hopkins University School of Medicine; Graduate Training Program in Clinical Investigation (M.R.M.), Department of Epidemiology at Johns Hopkins University School of Hygiene and Public Health (R.L.M., N.R.P.); and the Baltimore Cochrane Center at University of Maryland (N.K., K.A.R.). This material was previously presented orally at the American College of Cardiology 48th Annual Scientific Session, March 1999. Dr Miller completed this study while she was a clinical fellow at Johns Hopkins University School of Medicine and a graduate student at Johns Hopkins University School of Hygiene and Public Health.
Reprint requests should be addressed to Marlene R. Miller, MD, MSc, Center for Quality, Measurement and Improvement, Agency for Healthcare Research and Quality, 2101 East Jefferson Street, Suite 502, Rockville, MD 20852. Email: mmiller@ahrq.gov.
References
Study Inclusion
Articles had to report original data on pharmacologic management of nonpostoperative AF in adults in the context of a randomized clinical trial to be eligible for inclusion in our review. Pairs of independent investigators reviewed all identified abstracts according to these inclusion criteria. All discrepancies about inclusion were resolved by consensus.
Study Quality Assessment
The Evidence-Based Practice Center team developed a data form for extracting information on study quality based on a review of forms used in other meta-analytic studies by study investigators,9-11 a literature review of the topic,12,13 and with the assistance of the Cochrane Collaboration. The form contained 22 questions assessing study quality in 5 areas: clarity of description of the study population; potential for bias and confounding; description of therapy, outcomes and follow-up; and statistical quality and interpretation. Each question included a 4 to 5–level subjective ranking of study quality with the resultant score for each of the 5 areas comprising the total points accumulated out of the maximum possible points for all relevant questions in that area. The overall study quality score consisted of the mean score of these 5 areas.
Teams of independent reviewers assessed the quality of each study with differences resolved by consensus. Given the difficult nature of assessing study quality based on article review, the team decided to collectively review and discuss any articles receiving an overall score less than 50% to reach decisions regarding study inclusion.
Data Extraction
Because of the large volume of articles for review, quantitative data were extracted by one reviewer and then checked for accuracy by a second reviewer with consensus resolution of differences. The reviewers were not blinded to the author, institution, and journal, because recent work has indicated that such masking makes little difference in the results.14 In trials involving both AF and atrial flutter patients, data were only extracted for the AF patients whenever possible.
Data Synthesis and Analysis
Before doing the meta-analysis we first performed both qualitative and quantitative assessments of heterogeneity between the trials to ensure appropriateness of subsequent data combination. The reviewers subjectively assessed qualitative heterogeneity on the basis of similarity between studies on age of subjects, type and duration of AF, comorbidities, therapeutic regimens, and follow-up times. We performed quantitative analysis of heterogeneity using the statistical test of data heterogeneity included in Review Manager (RevMan) version 3.1 (Cochrane Collaboration, Oxford, England).
For data synthesis we defined control treatment to include placebo, verapamil, diltiazem, or digoxin. An analysis of identified trials evaluating verapamil, diltiazem, or digoxin compared with placebo supported this definition, since all of these agents were found to have no efficacy compared with placebo for either conversion or MSR.15 We also combined treatment arms within a given study that used the same antiarrhythmic agent at different dosages. Analysis of these arms individually supported their consideration as one arm.15 When life table analysis was used, we extracted the resultant cumulative percentages of successful outcomes and applied them to the initial overall subject number in each trial arm to derive a proportion for meta-analysis inclusion.
We constructed evidence tables to present the data separately for the 2 main outcomes of conversion of AF and MSR and created scatter plots of the absolute rates of conversion and MSR.
For meta-analysis the primary effect measure chosen was the odds ratio (OR) with studies weighted based on the precision of the estimate within each study. A fixed-effects model was used. In cases of significant quantitative data heterogeneity, we explored the etiology of the heterogeneity and used random-effects modeling when appropriate.
We chose the following categorization of strength of evidence by noting the placement of the point estimate of the OR and the width of the confidence interval (CI) surrounding it: (1) strong evidence of efficacy: OR >1.0, 99% CI does not include 1.0 (P <.01); (2) moderate evidence of efficacy: OR >1.0, 95% CI does not include 1.0, but 99% CI includes 1.0 (.01 P .05); (3) suggestive evidence of efficacy: 95% CI includes 1.0 in the lower tail (.05< P <.25), and the OR is in a clinically meaningful range; (4) inconclusive evidence of efficacy: 95% CI is widely distributed around 1.0; and (5) strong evidence of lack of efficacy: OR near 1.0, 95% CI is narrow and does not include a clinically meaningful difference from an OR of 1.0. When the point estimate was less than 1.0, we called this negative efficacy and used the same categorization of strong, moderate, and suggestive evidence on the basis of the point estimate OR and CI. For clarity our reported CIs are at the 95% level.
