Relative Risk Regression for Clustered Data with Application to Oral Health Research
Yingren Luo, Aya Mitani · University of Toronto Journal of Public Health · 2023
Introduction: Cluster-weighted generalized estimating equations (CWGEE) and doubly-weighted GEE (DWGEE) are used to produce unbiased estimates when informative cluster size (ICS) exists. However, their performance in estimating the relative risk (RR) from a Poisson regression is unknown. Methods: Using the dental data from the San Juan Overweight Adults Longitudinal Study (SOALS), we estimated the association between patient-level (sex, education level, smoking status, age) and tooth-level (bleeding upon probing) predictors and two types of outcomes with high and low prevalence each. We compared the odds ratio (OR) and RR estimates from logistic and Poisson CWGEE/DWGEE, respectively. Results: For patient-level covariates, the ORs estimated from logistic CWGEE/DWGEE and the RRs estimated from Poisson CWGEE/DWGEE were similar to the low-prevalence outcome (tooth loss). With the high-prevalence outcome (tooth loss or increase in attachment loss or pocket depth), the ORs were further from the null compared to the RRs. For example, within CWGEE, the OR of smoking was 1.301 (95% CI: 1.063-1.592), whereas the RR of smoking was 1.235 (95% CI: 1.052-1.450). For the tooth-level covariate (bleeding), there was a considerable difference between OR/RR on tooth loss estimated from CWGEE vs DWGEE. For example, the RR estimated from CWGEE was 1.692 (95% CI: 1.386-2.067), whereas the RR estimated from DWGEE was 1.354 (95% CI: 1.072-1.710). Discussion: CWGEE and DWGEE may produce different estimates for tooth-level (sub-cluster) covariates, especially when the prevalence of the outcome is low. In general, RRs estimated from Poisson CWGEE/DWGEE are closer to the null compared to ORs estimated from logistic CWGEE/DWGEE.