Multilevel Temporal Analysis of Repeated Cross-sectional Survey Data: Diabetes and the County Environment

Society for Epidemiologic Research (SER) Annual Meeting, June 18-21, 2019 Minneapolis, MN

Justin M. Feldman, MPH, ScD, David C. Lee, MD

INTRODUCTION: Random-effects within-between (REWB) models can incorporate temporality in analyses of repeated cross-sectional surveys, but have rarely been used in epidemiology. In an REWB model, within-group coefficients are equivalent to an econometric fixed-effect estimator ­­– both estimate the effect of a change in a group-level variable on growth in a dependent variable. These within-group estimates are less vulnerable to time-stable confounding.

METHODS: Using data on 3.4 million respondents to the CDC Behavioral Risk Factor Surveillance Survey (2003-2012), we employ an REWB logistic model to estimate associations between diabetes and county measures of active commuting (% who walk, cycle, or take transit to work), unemployment, and food environment (fast food / total restaurants; convenience stores / all food stores). We compare results to a standard mixed model that pools across county-years. All models controlled for a national time trend, smoking, and socio-demographic variables.

RESULTS: Associations for odds of diabetes estimated by the pooled model were in the hypothesized directions (positive for unemployment, fast food, and convenience stores; inverse for active commuting). Results from the REWB model were more ambiguous. A within-county increase of 1 standard deviation in active commuting was associated with an 8% reduction in odds of diabetes, but this estimate was imprecise (OR: 0.92; 95% CI: 0.82, 1.02). We found no evidence that within-county change in food environment predicted growth in county diabetes. The within-county effect of unemployment was inestimable due to its collinearity with time (there was a strong national trend of increasing unemployment over the study period).

CONCLUSION: The REWB model, which provides for stronger causal inference vs. a pooled model, failed provide evidence supporting the effects of county food environment on diabetes, but offered marginal evidence for a beneficial effect of increased active commuting.