A1167
Title: Comparing performance of three propensity score weighting methods for continuous exposures in nutritional epidemiology
Authors: Yuriko Muramatsu - Kyoto University (Japan) [presenting]
Tosiya Sato - Institute of Statistical Mathematics and Shiga University (Japan)
Shiro Tanaka - Kyushu University Hospital (Japan)
Abstract: Inverse probability of treatment weighting (IPTW) estimates causal effects in entire study population by eliminating the bias due to observed confounders. For continuous exposures, it is necessary to deal with the large weights caused by a lack of positivity, and to identify the exposure distribution to estimate generalized propensity score (GPS), the conditional probability of receiving a certain level of exposure. Generalized overlap weighting (GOW) and generalized matching weighting (GMW) mitigate the influence of a lack of positivity, so they may estimate exposure effects with less bias and higher precision. Quantile binning approach estimates GPS by dividing exposures into bins, thereby preventing model misspecification. The primary and secondary aims are to compare the performance of three weighting methods and to evaluate the favorable number of bins. Using a dataset from an epidemiologic study of patients with diabetes, the associations between continuous nutrient intake and diabetes complications risk were estimated. The performance metrics were covariate balance, weight variability, and the precision of estimates. Results showed that the performance of the three methods was favorable with 2 or 5 bins and depended on the degree of overlap in the GPS. With reduced overlap, GOW was effective for confounding adjustment for continuous exposures.
