Background: Different analytical techniques have been used to study the determinants of overweight. However, certain commonly used techniques may be limited by the continuous nature and skewed distribution of body mass index (BMI) data. In this article, different regression models are compared to identify the best approach for analysing predictors of BMI. Methods: Data collected on 2270 nurses at 18 public hospitals in Rio de Janeiro, RJ (2010–2011) were analysed (80.6 % of the respondents). The explanatory variables considered were age, marital status, race/colour, mother’s schooling, domestic overload, years worked at night, Consumption of fried food, physical inactivity, self-rated health and BMI at age 20 years. In addition to gamma regression, regarded as the reference method for selecting the set of explanatory variables described here, other modelling strategies – including linear, quantile (for the 0.25, 0.50 and 0.75 quantiles), binary and multinomial logistic regression – were compared in terms of final results and measures of fit. Results: The variables age, marital status, race/colour, domestic overload, self-rated health, physical inactivity and BMI at age 20 years were significantly associated with BMI, independently of the method used. In the same way, consumption of fried food was significant in all the models, but a dose–response pattern was identified only in the gamma and normal models and the quantile model for the 0.75 quantile. Years worked at night was also associated with BMI in these three models only. The variable mother’s schooling returned significant results only for the category 12 or more years of schooling, except for overweight in the multinomial model and for the 0.50 quantile in the quantile model, in which the two categories were not significant. The results of the quantile regression showed that, generally, the effects of the variables investigated were greater in the upper quantiles of the BMI distribution. Of the models using BMI in its continuous form, the gamma model showed best fit, followed by the quantile models (0.25 and 0.5 quantiles). Conclusions: The different strategies used produced similar results for the factors associated with BMI, but differed in the magnitude of the associations and goodness of fit. We recommend using the different approaches in combination, because they furnish complementary information on the problem studied.