Alternative regression models to assess increase in childhood BMI

Abstract

Background: Body mass index (BMI) data usually have skewed distributions, for which common statistical modeling approaches such as simple linear or logistic regression have limitations.

Methods: Different regression approaches to predict childhood BMI by goodness-of-fit measures and means of interpretation were compared including generalized linear models (GLMs), quantile regression and Generalized Additive Models for Location, Scale and Shape (GAMLSS). We analyzed data of 4967 children participating in the school entry health examination in Bavaria, Germany, from 2001 to 2002. TV watching, meal frequency, breastfeeding, smoking in pregnancy, maternal obesity, parental social class and weight gain in the first 2 years of life were considered as risk factors for obesity.

Results: GAMLSS showed a much better fit regarding the estimation of risk factors effects on transformed and untransformed BMI data than common GLMs with respect to the generalized Akaike information criterion. In comparison with GAMLSS, quantile regression allowed for additional interpretation of prespecified distribution quantiles, such as quantiles referring to overweight or obesity. The variables TV watching, maternal BMI and weight gain in the first 2 years were directly, and meal frequency was inversely significantly associated with body composition in any model type examined. In contrast, smoking in pregnancy was not directly, and breastfeeding and parental social class were not inversely significantly associated with body composition in GLM models, but in GAMLSS and partly in quantile regression models. Risk factor specific BMI percentile curves could be estimated from GAMLSS and quantile regression models.

Conclusion: GAMLSS and quantile regression seem to be more appropriate than common GLMs for risk factor modeling of BMI data.

Figure 1

Univariate density distributions of children's BMI with regard to underlying risk factors. Maternal BMI and weight gain in the first two years were divided up into two categories. The risk factors seem to produce a slightly right-skewed distribution for exposed in comparison to non-exposed children, whereas the confounder variable sex does not.

Figure 2

Values for the 90^th and 97^th BMI percentiles in respect to weight gain in the first two years (in kg), estimated by GAMLSS (dark lines) and quantile regression (grey lines), with fixed values for all other covariates. The dashed lines denote the estimated values for the 97^th percentiles for GAMLSS and quantile regression (QR), respectively. The dots represent observed values in the dataset.