Fixed-effect or random-effect models? A methodological reappraisal of subgroup analyses in mesenchymal stem cell therapy for knee osteoarthritis.
Systematic Review
We commend Cao et al. for their systematic review demonstrating the efficacy of intra-articular mesenchymal stem cell (MSC) therapy in alleviating pain and improving function in patients with non-surgical knee osteoarthritis (OA). However, we reanalyzed their subgroup analyses to evaluate the methodological implications of statistical model selection (fixed-effect vs. random-effect models) on result reliability. In dose-stratified analyses, Cao et al. applied fixed-effect models to low-dose (I = 0%) and high-dose (I = 80%) MSC subgroups. Upon reanalysis using random-effect models, the high-dose group showed no statistically significant differences in Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) total scores compared to the control group at 6 months [MD = 8.75; 95% CI (-2.10, 19.61); P = 0.11] or 12 months [MD = 12.68; 95% CI (-4.96, 30.32); P = 0.16], contrasting with Cao et al.’s original findings. The low-dose subgroup, with no heterogeneity, yielded identical results across both models. Similarly, in cell-source stratification (adipose-derived MSCs [ADMSCs] vs. bone marrow-derived MSCs [BM-MSCs]), reanalysis of ADMSCs using random-effect models demonstrated significant 6-month WOMAC improvement [MD = 9.32; 95% CI (3.73, 14.92); P = 0.001] but non-significant 12-month differences [MD = 12.90; 95% CI (-1.76, 27.55); P = 0.08], diverging from Cao et al.’s conclusions. BM-MSCs results remained consistent due to negligible heterogeneity (I = 0%). These findings underscore that fixed-effect models artificially narrow confidence intervals in heterogeneous populations, overestimating clinical significance. Our results align with Cochrane guidelines, emphasizing that random-effect models better accommodate inter-study diversity, yielding conservative and clinically generalizable estimates. This critique reinforces the necessity of transparent statistical model selection in meta-analyses, particularly when subgroup heterogeneity may influence therapeutic interpretations.
