Combining distributed regression and propensity scores: a doubly privacy-protecting analytic method for multicenter research

Abstract

PurposeSharing of detailed individual-level data continues to pose challenges in multi-center studies. This issue can be addressed in part by using analytic methods that require only summary-level information to perform the desired multivariable-adjusted analysis. We examined the feasibility and empirical validity of 1) conducting multivariable-adjusted distributed linear regression and 2) combining distributed linear regression with propensity scores, in a large distributed data network.Patients and methodsWe compared percent total weight loss 1-year postsurgery between Roux-en-Y gastric bypass and sleeve gastrectomy procedure among 43,110 patients from 36 health systems in the National Patient-Centered Clinical Research Network. We adjusted for baseline demographic and clinical variables as individual covariates, deciles of propensity scores, or both, in three separate outcome regression models. We used distributed linear regression, a method that requires only summary-level information (specifically, sums of squares and cross products matrix) from sites, to fit the three ordinary least squares linear regression models. A comparison set of analyses that used pooled deidentified individual-level data from sites served as the reference.ResultsDistributed linear regression produced results identical to those from the corresponding pooled individual-level data analysis for all variables in all three models. The maximum numerical difference in the parameter estimate or standard error for all the variables was 3×10−11 across three models.ConclusionDistributed linear regression analysis is a feasible and valid analytic method in multicenter studies for one-time continuous outcomes. Combining distributed regression with propensity scores via modeling offers more privacy protection and analytic flexibility.

Estimating the marginal hazard ratio by simultaneously using a set of propensity score models: A multiply robust approach

“…For example, the Sentinel System is a national electronic system funded by the US Food and Drug Administration to monitor the safety of approved medical products using data from more than a dozen health plans and delivery systems 44 . The IPW Cox model stratified on data‐contributing site provides one approach to estimating marginal hazard ratios in multisite studies, where each site fits a site‐specific propensity score model 45‐47 . In this section, we extend the proposed multiply robust method in Section 4 to enable each participating site to postulate multiple site‐specific propensity score models.…”

Inverse probability weighted Cox model in multi-site studies without sharing individual-level data

“…9,29 A recent study employed distributed regression technique to adjust for propensity score via modeling. 30 The methods proposed here, together with those used in prior studies, offer a suite of privacy-protecting methods that allow researchers to use the propensity score in several ways – matching, stratification, regression modeling, and now weighting – to adjust for confounding in multi-site studies of time-to-event outcomes.…”

“…In practice, local estimation of propensity scores is more flexible and easier to conduct than global estimation. 30…”

Recovery of original individual person data (IPD) inferences from empirical IPD summaries only: Applications to distributed computing under disclosure constraints

“…A more systematic investigation of the methodological challenges in privacy‐constrained distributed computing has focused on multivariable confounding adjustment . Here, among other methods, inverse‐weighting meta‐analysis remains an available standard approach.…”

“…A more systematic investigation of the methodological challenges in privacy-constrained distributed computing has focused on multivariable confounding adjustment. [16][17][18][19] Here, among other methods, inverse-weighting meta-analysis remains an available standard approach. However, more focus has been given to regression methods based on reducing confounders to a single propensity score, also resulting in IPD aggregation through matching and stratification.…”