Time‐Series Analysis of US Obesity Trends during 1995–2009 and Future Projection

The trends in obesity in the US over the past three decades have been widely reported, but it remains surprisingly rare that the temporal association of the time series prevalence data of overweight (25 kg/m 2 ≤BMI≤29.9 kg/m 2 ) and obesity (BMI≥30 kg/m 2 ) is properly accounted in previous research. Commonly regression analysis is conducted contingent on the implicit independent and identically distributed (i.i.d.) assumption, which indicates that overweight and obese rates in one year are independent from those in other year(s), an unrealistic assumption. We fit conditional autoregressive (AR) time series models to investigate the possibility of model misspecification and its impacts on prevalence analysis and forecast. A simple linear model without accounting for time series dynamics and a conditional AR model are fitted using the Behavioral Risk Factor Surveillance System (BRFSS) 1995–2009 data. The estimates indicate that AR model produces better with‐in sample predications, reducing about 30% root mean squared errors (RMSE). Moreover, the AR model generates better out‐of sample forecasts with more precise estimates (ie, narrower, by about 50% of the 95% confidence intervals) compared to simple linear models. Thus, to incorporate time series dynamics in longitudinal obesity trend analysis can help generate better forecast including its variation. Grant Funding Source : NIH/NIDDK (R01DK81335‐01A1), NIH/NICHD (1R01HD064685‐01A1).