Evaluation of Methods to Simulate Correlated Binary Data

• Activity Number: 498
• Type: Topic Contributed
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Government Statistics
• Author(s): Robert Lew*+ and Hongsheng Wu and David Gagnon and Soe Soe Thwin
• Companies: VA Boston Healthcare and VA Boston Healthcare and Department of Veterans Affairs and Department of Veterans Affairs
• Keywords: Correlated binary data ; Bahadur's model ; Emirich/ Piedmonte algorithm

Abstract

Methods to analyze correlated binary response data have received increasing attention in recent years, resulting in a need for evaluation of these methods and hence a reliable tool to simulate correlated binary data. We present an evaluation of two popular methods to generate correlated binary data: one based on Bahadur's model (the Bahadur's approach) and the other based on Emirich and Piedmonte algorithm (the EP approach). To compare these two approaches, twelve simulated scenarios were designed using AR1 and exchangeable correlation structures, two time dimensions (3 and 5), and three association parameter beta values (0, 1, and 2). Five thousand data sets of fixed size (N = 1000) were generated for each scenario and estimated correlation coefficient were compared across the two methods. For both correlation structures considered and zero effect size, the Bahadur's approach and the EP approach performed equally well with respect to correlation coefficient estimation accuracy when correlated binary responses were taken at three time points. However, the EP approach outperformed the Bahbadur's approach for five outcome measures. Under other scenarios, we found lower correlation l