Cluster vs. Robust Estimation of Risk Ratio using Expanded Logistic Regression

In their article, Dr. Janani et al. discussed some methods to obtain adjusted risk ratio (RR).1 Among options, authors mentioned the method named “expanded logistic regression”, which consists in changing the original dataset by duplicating data of each individual that developed the outcome.2,3 In this new GDWDVHW GXSOLFDWHG REVHUYDWLRQVDUHLGHQWL¿HGDV QRQRXWFRPH The probability of success in the original dataset will be equal to WKHRGGVRIVXFFHVVLQWKHPRGL¿HGGDWDVHWWKHUHIRUHDORJLVWLF UHJUHVVLRQ PRGHO ¿WWLQJWRWKH QHZ GDWDVHW UHVXOWVLQ ULVN UDWLR instead of an odds ratio. This simple tool could be useful for calculating adjusted RRs even using not sophisticated software. The main problem with WKLVPHWKRGLVWKDWWKHFRQ¿GHQFHLQWHUYDOVDUHZLGHUWKDQWKRVH observed with the reference methods.4 It was suggested that robust standard errors (SE) are needed to account for the within-subject correlation resulted from the duplicated observations.1 However, robust estimation of SE does not solve that problem because the dependence of duplicate observations persists. Recently, Dwivedia et al. proposed the cluster option to correct Thus, each case and its duplicate would be considered within a cluster, which allows estimating RRs considering the dependence of these observations. In order to represent the differences between robust estimation of SE and cluster option for logistic regression, this communication present an analysis comparing these two methods against logbinomial regression