Estimating dose-response relationship between ethanol and risk of cancer using regression spline models.

Cancers of the upper aero-digestive tract (i.e., oral cavity, pharynx, larynx and oesophagus) are largely attributable to smoking and drinking habits, but the correct estimation of the dose-response relationship between alcohol and cancer risk is challenging. Step functions are widely used to estimate risks and to evaluate trends of continuous exposure. However, results are influenced by the selection of the reference category and cutpoints. More flexible models, like spline regression and fractional polynomial models, may be an attractive alternative for avoiding strict assumptions about the dose-response relationship. Data from a large series of hospital-based case-control studies conducted in Italy and the Swiss Canton of Vaud in the last 2 decades were reassessed to compare findings from logistic regression spline models and standard step function analysis. For all examined cancers, the risk increased to the consumption of 150 grams of ethanol per day (1.5 litre/day of wine), with a possible threshold effect emerging for cancer of the pharynx and larynx (<50 grams of ethanol per day) only. For higher consumptions, the risks flattened. Our study suggests that regression spline models can be useful to estimate the pattern of risk of a continuous exposure variable, such as alcohol consumption, and provide more accurate estimates than categorical analysis when ORs within each interval, especially in the reference category, are not homogeneous. 2004 Wiley-Liss, Inc.