Meta-analysis of survival curve data using distributed health data networks: application to hip arthroplasty studies of the International Consortium of Orthopaedic Registries.

Overview

The motivating example for this paper comes from a distributed health data network, the International Consortium of Orthopaedic Registries (ICOR), which aims to examine risk factors for orthopedic device failure for registries around the world. Unfortunately, regulatory, privacy, and propriety concerns made sharing of raw data impossible, even if de-identified. Therefore, this article describes an approach to extraction and analysis of aggregate time-to-event data from ICOR. Data extraction is based on obtaining a survival probability and variance estimate for each unique combination of the explanatory variables at each distinct event time for each registry. The extraction procedure allows for a great deal of flexibility; models can be specified after the data have been collected, for example, modeling of interaction effects and selection of subgroups of patients based on their values on the explanatory variables. Our analysis models are adapted from models presented elsewhere--but allowing for censoring in the calculation of the correlation between serial survival probabilities and using the square root of the covariance matrix to transform the data to avoid computational problems in model estimation. Simulations and a real-data example are provided with strengths and limitations of the approach discussed.