Alcohol and liver cirrhosis mortality in the United States: comparison of methods for the analyses of time-series panel data models.

BACKGROUND: To explore various model specifications in estimating relationships between liver cirrhosis mortality rates and per capita alcohol consumption in aggregate-level cross-section time-series data.
METHODS: Using a series of liver cirrhosis mortality rates from 1950 to 2002 for 47 U.S. states, the effects of alcohol consumption were estimated from pooled autoregressive integrated moving average (ARIMA) models and 4 types of panel data models: generalized estimating equation, generalized least square, fixed effect, and multilevel models. Various specifications of error term structure under each type of model were also examined. Different approaches controlling for time trends and for using concurrent or accumulated consumption as predictors were also evaluated.
RESULTS: When cirrhosis mortality was predicted by total alcohol, highly consistent estimates were found between ARIMA and panel data analyses, with an average overall effect of 0.07 to 0.09. Less consistent estimates were derived using spirits, beer, and wine consumption as predictors.
CONCLUSIONS: When multiple geographic time series are combined as panel data, none of existent models could accommodate all sources of heterogeneity such that any type of panel model must employ some form of generalization. Different types of panel data models should thus be estimated to examine the robustness of findings. We also suggest cautious interpretation when beverage-specific volumes are used as predictors.