To account for intra-individual correlation across longitudinal observations, for counted data outcomes we used mixed modeling as implemented in SAS (SAS Institute, Inc., Cary, North Carolina) PROC MIXED (Bryk and Raudenbush, 1992; Singer, 1998) and generalized estimating equations (Liang and Zeger, 1986) for binary outcomes. Within each model comparison set, likelihood ratio testing (LRT) was used to compare the least complex model to each of the successive models that appended one or more additional terms. We identified the best fitting model within each model set as the one that minimized the Akaike’s information criterion (AIC; Akaike, 1973) for quantitative outcomes, or the comparable quasi-likelihood function under the independence model criterion (QICC; Pan, 2001) for binary outcomes. For binary outcomes, we also considered Wald tests to determine the significance of added terms, as more complex models were compared.
