Identifying an optimal strategy for converting pain as a continuous outcome to a responder analysis

Abstract

Background: In pain relief research, meta-analyses often combine continuous outcomes from various studies using mean differences. However, this approach can be difficult to interpret clinically. An alternative method involves aggregating the risk difference for patients who achieve a minimally important difference (MID) in pain reduction. The challenge is that many trials do not report responder analyses, necessitating continuous data conversion. Objective: To conduct a simulation study assessing the performance of four proposed methods for estimating the pooled risk difference (RD) of achieving the MID in meta-analyses of pain measured on a 10cm visual analogue scale (VAS). Methods: Individual patient data for VAS pain scores were simulated across 4,752 scenarios varying the treatment effect as change score in the intervention (-1.0 to 4.0) and control (-1.0 to 3.0) groups, study sample size (10-1000), number of studies per meta-analysis (3 to 30), shape of distribution (normal or skewed), and MID (1.0 or 1.5). The true pooled RD and 95% confidence interval (CI) were calculated from the simulated individual data. Four methods were evaluated: calculating RD based on pooled 1) median mean differences, 2) unweighted average differences, 3) weighted average differences, and 4) calculating RD for each individual study and then meta-analysing RDs. Bias, mean squared error, confidence interval (CI) coverage of true value, and empirical standard error (SE), and model-based SE were evaluated. Results: The median method showed the lowest bias (2.048; 95% CI: 1.759-2.338), while the individual method demonstrated the lowest RMSE (4.852; 95% CI: 4.661-5.044), empirical SE (0.148; 95% CI: 0.141-0.154), and model-based SE (2.198; 95% CI: 2.108-2.288), and highest CI coverage (55.717%; 95% CI: 53.185-58.250%). Differences between methods were minimal and not statistically significant. Performance was optimal when treatment effects were similar between groups and declined with increasing effect size differences. All methods performed poorly with skewed distributions. Conclusion: While the evaluated methods can provide useful estimates in many scenarios, they should be used cautiously, especially for large treatment effects or non-normal data. Researchers should prioritize conducting and reporting responder analyses in primary studies to reduce reliance on these estimation methods in meta-analyses.