Ignoring non-significant factors without data may bias the results of meta-analysis of prognostic studies

Article type
Authors
Wang L1, Guyatt G2, Thabane L2, Busse J1
1Department of Anesthesia, McMaster University
2Department of Health Research Methods, Evidence and Impact, McMaster University
Abstract
Background:
Meta-analysis of prognostic studies usually pools reported measures of association for common factors across studies. Failure to consider non-significant factors that were excluded or not reported in regression models may overestimate pooled measures of association.
Objectives:
Using systematic reviews of predictors of persistent postsurgical pain or unemployment after breast cancer surgery, we explored the impact of imputing data for missing non-significant data on overall associations of risk factors.
Methods:
We pooled predictors to explore their association with persistent pain or unemployment after breast cancer surgery using random-effects models and the DerSimonian-Laird method. For our primary analyses, we imputed an odds ratio (OR) of "1" for predictors that were excluded or not reported in multivariable analyses due to non-significant association. We acquired the associated variance using the hot deck approach. We performed sensitivity analysis by excluding the imputed data for non-significant predictors (analysis of adjusted data reported). We calculated the ratio of odds ratio to estimate the difference between these two approaches.
Results:
Fifty-six studies with 66,740 patients were included. Most of the studies either excluded factors that were not significant in bivariate analysis (32%; 18 of 56) or failed to present data for non-significant predictors in their final regression models (68%; 38 of 56). 24 of 27 risk factors contained missing data for non-significant factors (Table 1).

The median ratio of odds ratio of pooling analyses using imputed data vs. only reported data in the final multivariable models was 1.07 (range: 1.01 to 1.15) for 9 poolable risk factors for persistent pain and 1.05 (range: 1.01 to 1.80) for 15 poolable predictors of unemployment (Figures 1 to 4). All pooled associations were larger in meta-analyses based on reported data only vs. when imputed missing data was considered, which exaggerated the magnitude of association by 1%-55%.
Conclusions:
Primary studies exploring prognostic factors often fail to report data for non-significant predictors. Failure to impute for missing non-significant predictors in meta-analyses systematically overestimates pooled measures of association. Systematic reviewers should acquire missing data for non-significant predictors from authors when possible, and impute data when not.
Patient or healthcare consumer involvement:
No