Imputation of missing adjusted data for non-significant factors in meta-analysis of prognostic studies

Article type
Authors
Wang L1, Kennedy S2, Romerosa B3, Kwon H4, Kaushal A2, Craigie S1, Chang Y2, Almeida C5, Izhar Z2, Couban R1, Parascandalo S6, Guyatt G2, Reid S2, Khan J2, McGillion M2, Busse J1
1Michael G DeGroote Institute for Pain Research and Care, McMaster University, Canada
2McMaster University, Canada
3Complejo Hospitalario de Toledo, Spain
4Wayne State University, USA
5Federal University of Rio Grande do Sul -UFRGS, Brazil
6University College Cork, Republic of Ireland
Abstract
Background: Many prognostic studies use data-driven models: each independent factor is tested in a bi-variable analysis, and only those that show evidence of association (e.g. P value ≤ 0.05) are entered into an adjusted model. Others report data only for those independent factors in their adjusted model that show a significant association with the dependent variable. Systematic exclusion of these data presents a risk of overestimation by only pooling estimates of association from predictors that appear in adjusted regression models, and for whom data is provided.
Objectives: To investigate if imputation of missing non-significant data in final regression model will avoid overestimation of the predictive power of the risk factors, using predictors for persistent pain after breast cancer surgery as an example.
Methods: We pooled nine predictors to explore their association with the development of persistent pain following breast cancer surgery using random-effects models. We imputed an odds ratio (OR) of '1' for predictors that were excluded from adjusted analyses due to non-significant bi-variable analyses, or that were reported, but with no data due to lack of significance in the final regression model. We acquired the associated variance for all such imputations using the hot deck approach. We performed sensitivity analysis to examine the impact of imputing data for non-significant predictors excluded from adjusted analyses by re-running our analyses and excluding the imputed data.
Results: Fifty-nine (51.3%) out of 115 study-sets for nine predictors failed to reported the adjusted data for non-significant predictors. Our sensitivity analyses found no significant differences in results whether or not we incorporated missing data for non-significant predictors (Table 1). However, the associations of the nice predictors with persistent pain were consistently larger in meta-analyses based on the adjusted data only than in the full analyses including imputed missing data.
Conclusions: Imputation of missing data for non-significant predictors did not cause any significant associations to lose significance, but the magnitude of association was reduced.