This article presents alternatives for modeling body mass index (BMI) as a continuous variable and the role of residual analysis. We sought strategies for the application of generalized linear models with appropriate statistical adjustment and easy interpretation of results. The analysis included 2,060 participants in Phase 1 of a longitudinal study (Pró-Saúde Study) with complete data on weight, height, age, race, family income, and schooling. In our study, the residual analysis of models estimated by maximum likelihood methods yielded inadequate adjustment. The transformed response variable resulted in a good fit but did not lead to estimates with straightforward interpretation. The best alternative was to apply quasi-likelihood as the estimation method, presenting a better adjustment and constant variance. In epidemiological data modeling, researchers should always take trade-offs into account between adequate statistical techniques and interpretability of results.
Body Mass Index; Linear Models; Statistical Data Interpretation
