Worked example of how model choice for a meta-analysis can affect the overall pooled results and subsequent conclusions
| Model | Difference in means (95% CI) | P value |
| FE | –0.55 (–0.70 to –0.40) | <0.001 |
| RED&L | –0.90 (–1.36 to –0.44) | <0.001 |
| REHKSJ | –0.90 (–1.45 to –0.35) | 0.003 |
| IVhet | –0.55 (–1.27 to 0.17) | 0.13 |
| QE | –0.38 (–1.12 to 0.36) | 0.31 |
Data are from the work of Park et al 50 on preemptive epidural analgesia for acute and chronic postthoracotomy pain in adults. Negative values indicate improvement, that is, reductions, in pain intensity in the preemptive vs control group.
p, α value; Q (p), Cochran Q statistic and α value for Q; QE, quality effects model with studies weighted by Cochrane Risk of Bias Results reported in the original meta-analysis; RED&L, traditional random effects model of DerSimonian and Laird; REHKSJ, random effects model with Hartung-Knapp-Sidik-Jonkman adjustment. Results are similar for Q (p) [(116.3 (<0.001)], I2 (95% CI) [85.4% (78.3–90.2)] andgiven that the same approach is historically used for analysis across all models.
FE, fixed effects model; IVhet, inverse variance heterogeneity model;