Examining heterogeneity in meta-analysis - Comparing results of randomized trials and nonrandomized studies of interventions for low back pain

Study Design. Literature review. Objective. To assess the influence of various factors in statistical heterogeneity of meta-analyses of interventions for low back pain. One of these factors was study design: randomized controlled trial ( RCT) versus nonrandomized study ( NRS). Summary of Background Data. The presence of statistical heterogeneity poses a challenge to the conduct and interpretation of meta-analyses. Methods. We searched MEDLINE, EMBASE, and The Cochrane Library up to May 2005 for comparative studies of interventions for low back pain. The interventions with the highest number of NRSs were selected. All NRSs and RCTs of the same interventions were combined using meta-analysis. Subgroup analyses and meta-regression were performed according to study design and other factors that were selected by a panel of 20 experts. Results. NRSs frequently either agree with RCTs or underestimate the effects compared with RCTs. The interventions and the respective factors that explained statistical heterogeneity were a) surgery versus conservative treatments ( 17 NRSs and 8 RCTs): study design ( odds ratio, OR: 1.56 and 4.69 for nonrandomized and randomized studies, respectively), pain duration ( OR: 1.75 and 3.55 for chronic and acute, respectively), and involvement of workers' compensation ( OR: 1.85 and 5.07, with and without, respectively); b) surgery with fusion versus surgery without fusion ( 17 NRSs and 3 RCTs): spondylolisthesis ( OR: 2.15 and 1.22, with and without, respectively); c) Instrumented fusion versus noninstrumented fusion ( 15 NRSs and 8 RCTs): previous surgery ( OR: 2.89 and 1.36, with and without, respectively) and levels fused OR: 1.50 and 2.98, single and multilevel, respectively). Conclusion. Comparisons between RCTs and NRSs may be influenced by various factors, including study design. However, other factors were more powerful explanatory variables than study design. These factors included pain duration, involvement of workers' compensation, presence of spondylolisthesis, previous surgery, and levels fused.