576Is the method of longitudinal data analysis equally relevant for common versus uncommon endpoints?

Abstract

Abstract Background We explored the effects of predominant breastfeeding on infection-related hospitalization (uncommon outcome), over the first year of life, using the Mother Infant Health cohort study (MIHS), and the effect of early antiretroviral treatment (ART) on viral suppression (< 400 copies/mL) (common outcome), in children aged 7 to 12 weeks, using a subset of the Children with HIV Early AntiRetroviral Treatment (CHER) clinical trial data. We assessed the sensitivity of findings to different models to account for dependency of uncommon and common binary outcome. Methods We fitted generalized linear mixed model with (1) random intercept and (2) random slope, generalized estimating equations (GEE) with 3) an exchangeable correlation structure; 4) autoregressive correlation structure of order 1 (AR1) and 5) unstructured correlation structure and 6) logistic regression model. Results Eighty four and 119 children from MIHS were non-predominantly and predominantly breastfed, respectively. There were 34 infection-related hospitalizations overall. Most infants were hospitalized once, except for four with two hospitalizations. We analysed 88 HIV-infected children from the CHER trial. On average, a child achieved viral suppression twice, range of one to four. The effect of predominant breastfeeding on infection-related hospitalization was similar across all models, except for the GEE with AR1 that had a high estimate (wider confidence intervals). The effect of early ART exposure on viral suppression varied across models. Conclusions The sensitivity of estimates to the method of analysis was driven by frequency of the outcome.