Semiparametric Regression Approach to Adjusting for Meteorological Variables in Air Pollution Trends

A semiparametric regression model is used to analyze relationships between ambient ozone, related meteorological predictor variables, and time trends. This model enjoys a number of advantages over alternatives. The model produces outputs that parallel the outputs of linear regression but without forcing the structure of linear regression upon the data. Instead, the abundant data are allowed to determine the form of the model as in other nonparametric regression models. Four Houston, TX, sites with relatively complete ozone and meteorological data from 1983 to 1995 are analyzed. For these four sites, the model explains over 60% of the variation in log (ozone) and estimates declines in meteorologically adjusted ozone of 13−26%. These estimated declines appear to be real and are consistent with other evidence, such as declines in NO x , declines in exceedance days, and implementation of control policies. Because the estimated declines are interpreted, as in ordinary regression, as having been adjusted for year-to-year differences in the meteorological variables, one can infer that factors other than changes in the weather are responsible for the declines.