Towards Improving Transparency of Count Data Regression Models for Health Impacts of Air Pollution

In studies on the health impacts of air pollution, regression analysis continues to advance far beyond classical linear regression, which many scientists may have become familiar with in an introductory statistics course. With each new level of complexity, regression analysis may become less transparent, even to the analyst working with the data. This may be especially true in count data regression models, where the response variable (typically given the symbol y) is count data (i.e., takes on values of 0, 1, 2, …). In such models, the normal distribution (the familiar bell-shaped curve) for the residuals (i.e., the differences between the observed values and the values predicted by the regression model) no longer applies. Unless care is taken to correctly specify just how those residuals are distributed, the tendency to accept untrue hypotheses may be greatly increased. The aim of this paper is to present a simple histogram of predicted and observed count values (POCH), which, while rarely found in the environmental literature but presented in authoritative statistical texts, can dramatically reduce the risk of accepting untrue hypotheses. POCH can also increase the transparency of count data regression models to analysts themselves and to the scientific community in general.