Random effects modelling vs logistic regression for the inclusion of cluster level covariates in propensity scores for medical device and surgical epidemiology

PurposeSurgeon and hospital related features such as surgeries volume can be associated with treatment choices and treatment outcomes. Accounting for these covariates with propensity score (PS) analysis can be challenging due to clustered nature of the data. Previous studies have not focused solely on the PS estimation strategy when treatment effects are estimated using random effects model(REM). We studied PS estimation for clustered data using REM compared with logistic regression. MethodsSix different PS estimation strategies were tested using simulations with variable cluster-level confounding intensity (odds ratio(OR)=1.01 to OR=2.5): i) logistic regression PS excluding cluster- level confounders; ii) logistic regression PS including cluster-level confounders; iii) same as ii) but including cross-level interactions; iv), v) and vi), similar to i), ii) and iii) respectively but using REM instead of logistic regression PS. Same analysis were tested in a randomised controlled trial emulation of partial vs total knee replacement surgery. Simulation metrics included bias and mean square error (MSE). For trial emulation, we compared observational vs trial-based treatment effect estimates. ResultsIn most simulated scenarios, logistic regression including cluster-level confounders gave more accurate estimates with the lowest bias and MSE. E.g. with 50 clusters x 200 individuals and confounding intensity OR=1.5, the relative bias= 10% and MSE= 0.003 for (i), compared to 21% and, 0.010 for (iv). In the Trial emulation, all 6 PS strategies gave similar treatment effect estimates. ConclusionsLogistic regression including patient and surgeon/hospital-level confounders appears to be the preferred strategy for PS estimation. Further investigation with more complex clustered structure is suggested. Competing interestsProf. Prieto-Alhambras research group has received grant support from Amgen, Chesi-Taylor, Novartis, and UCB Biopharma. His department has received advisory or consultancy fees from Amgen, Astellas, AstraZeneca, Johnson, and Johnson, and UCB Biopharma and fees for speaker services from Amgen and UCB Biopharma. Janssen, on behalf of IMI-funded EHDEN and EMIF consortiums, and Synapse Management Partners have supported training programs organised by DPAs department and open for external participants organized by his department outside submitted work. Ethics Approval and Informed ConsentThis study was approved by the secretary of state, having considered the recommendation from the Confidentiality Advisory Group (CAG reference: 17/CAG/0174). Informed ethical approval was given on the use of pseudonymised patients data included in the study.