Statistical model assessment in published dose-response meta-analyses is suboptimal: evidence from a methodological...

Dose-response meta-analysis (DRMA) is a crucial clinical and epidemiological research tool for synthesizing exposure-risk relationships. Despite its growing use, facilitated by the availability of statistical software, the appropriateness of the underlying statistical methods has not been thoroughly explored. This study aims to evaluate the reporting quality of key statistical measures in DRMA and compare their performance using empirical datasets.

We performed a systematic literature search to identify relevant studies, from which we extracted datasets and study characteristics. We fitted linear, quadratic polynomial and restricted cubic spline (RCS) models with fixed and non-fixed knots selection procedures. Key measures assessed included non-linearity, goodness-of-fit (GoF), model comparison, and the impact of outlying studies on statistical results. We compared P-values for non-linearity, GoF test for each model pair, and used the Akaike information criterion for model comparison. We evaluated the influence of individual studies on non-linearity and GoF using a leave-one-out (LOO) approach.

We included 146 unique DRMA studies, from which 242 datasets were extracted with median (interquartile range) of 9 (7, 14) individual studies. While the non-linearity test was conducted in 102/124 (82.3%) studies, other measures were infrequently reported. Only 13/146 (10.0%) studies assessed GoF, 10/79 (12.7%) provided model comparison results, and 46/146 (31.5%) examined the impact of outlying studies. Our reanalysis of 242 datasets demonstrated that the RCS model with a non-fixed knots selection procedure identified more non-linearity (110/242, 45.5%) and fitted well (205/242, 84.7%) than other models. The LOO approach showed that conclusions regarding non-linearity and GoF changed in approximately 50% of cases after excluding a single study, regardless of the model used.

Our analysis reveals suboptimal attention to key statistical issues in published DRMA studies. RCS with non-fixed knots selection have shown potential advantages than a few other alternative modeling approaches. To strengthen the credibility of meta-analytic findings, it is advisable for researchers to integrate both clinical judgment and rigorous statistical model evaluation in their analyses. The LOO assessment underscores the necessity for methods that can identify and accommodate outlying studies in DRMA.