Bias and Efficiency in a Matched Observational Study with Varying Cluster Size

• Activity Number: 88 - SPEED: Causal Inference and Related Methodology Part 2
• Type: Contributed
• Date/Time: Sunday, July 28, 2019 : 5:05 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Eric KH Chow* and Rajani Kaimal and Vedant Pargaonkar and Sara Bouajila and Katharine Sears-Edwards and Jennifer Tremmel and Manisha Desai
• Companies: Quantitative Sciences Unit, Stanford University School of Medicine and Quantitative Sciences Unit, Stanford University School of Medicine and Interventional Cardiology, Stanford University School of Medicine and Stanford University School of Medicine and Cardiovascular Medicine, Stanford University School of Medicine and Interventional Cardiology, Stanford University School of Medicine and Stanford University Quantitative Sciences Unit
• Keywords: matching; case-control; retrospective; efficiency; cluster size; bias

Abstract

Matching has shown promise in mitigating confounding in observational studies when RCTs are not feasible; however, it can be challenging to find sufficient controls in rare diseases such as in spontaneous coronary artery dissection (SCAD) among peripartum women. We tested a “greedy match” algorithm (GM) that matches peripartum women to similarly aged non-peripartum controls with varying cluster sizes (1-4) versus an optimal match algorithm (OptMatch in R) with fixed cluster sizes of 1:2 and 1:3. Match quality was evaluated using the standardized bias (< 0.25) and variance ratio (0.5-2.0). The 1:3 optimal match selected the most controls, but they were biased towards being older (123 controls, std bias = 0.54). The 1:2 optimal match selected the best-matched controls, but selected the fewest controls (82 controls, std bias = 0.09). We created match sets in-between these with the GM algorithm. The GM selected 101, 107, and 116 controls (with increasing std bias = 0.34, 0.40, 0.48) by setting the age caliper to 1, 2, and 4yrs. Allowing for varying cluster size increased the efficiency of control selection but at the penalty of more bias in the match clusters with more controls.