Background: Poisson regression is routinely used in occupational and environmental epidemiology. For typical Poisson regression analyses, person-time and events are tabulated by categorising predictor variables that were originally measured on a continuous scale. In order to estimate a dose-response trend, a researcher must decide how to categorise exposures and how to assign scores to exposure groups.

Aims: To investigate the impact on regression results of decisions about exposure categorisation and score assignment.

Methods: Cohort data were generated by Monte Carlo simulation methods. Exposure categories were defined by quintiles or deciles of the exposure distribution. Scores were assigned to exposure groups based on category midpoint and mean exposure levels. Estimated exposure-disease trends derived via Poisson regression were compared to the "true" association specified for the simulation.

Results: Under the assumption that exposures conform to a lognormal or exponential distribution, trend estimates tend to be negatively biased when scores are assigned based on category midpoints and positively biased when scores are assigned based on cell specific mean values. The degree of bias was greater when exposure categories were defined by quintiles of the exposure distribution than when categories were defined by deciles of the exposure distribution.

Conclusions: The routine practice of exposure categorisation and score assignment introduces exposure misclassification that may be differential with respect to disease status and, consequently, lead to biased exposure-disease trend estimates. When using the Poisson regression method to evaluate exposure-disease trends, such problems can be minimised (but not necessarily eliminated) by forming relatively refined exposure categories based on percentiles of the exposure distribution among cases, and by assigning scores to exposure categories that reflect person-time weighted mean exposure levels.