Predicting future weight status from measurements made in early childhood: a novel longitudinal approach applied to Millennium Cohort Study data

BACKGROUND/OBJECTIVE: There are reports that childhood obesity tracks into later life. Nevertheless, some tracking statistics such as correlations do not quantify individual agreement, whereas others such as diagnostic test statistics can be difficult to translate into practice. We aimed to employ a novel analytic approach, based on ordinal logistic regression, to predict weight status of 11-year-old children from measurements at age 5 years. SUBJECTS/METHODS: The UK 1990 growth references were used to generate clinical weight status categories of 12 076 children enrolled in the Millennium Cohort Study. Using ordinal regression, we derived the predicted probability (percent chances) of 11-year-old children becoming underweight, normal weight, overweight, obese and severely obese from their weight status category at age 5 years. RESULTS: The chances of becoming obese (including severely obese) at age 11 years were 5.7% (95% confidence interval: 5.2 to 6.2%) for a normal-weight 5-year-old child and 32.3% (29.8 to 34.8%) for an overweight 5-year-old child. An obese 5-year-old child had a 68.1% (63.8 to 72.5%) chance of remaining obese at 11 years. Severely obese 5-year-old children had a 50.3% (43.1 to 57.4%) chance of remaining severely obese. There were no substantial differences between sexes. Nondeprived obese 5-year-old boys had a lower probability of remaining obese than deprived obese boys: −21.8% (−40.4 to −3.2%). This association was not observed in obese 5-year-old girls, in whom the nondeprived group had a probability of remaining obese 7% higher (−15.2 to 29.2%). The sex difference in this interaction of deprivation and baseline weight status was therefore −28.8% (−59.3 to 1.6%). CONCLUSIONS: We have demonstrated that ordinal logistic regression can be an informative approach to predict the chances of a child changing to, or from, an unhealthy weight status. This approach is easy to interpret and could be applied to any longitudinal data set with an ordinal outcome.