Dispersion Model or Statistical Model?

There is a need to know the space and time distributions of pollutant concentrations to estimate the health effects of air pollutants. Dispersion models and statistical models are the basis for the two different methods used to estimate these distributions. The dispersion model is based on the underlying processes and requires inputs of source emissions, meteorology, and terrain. The statistical model is based on empirical relations that relate observed concentrations to land-use in the region of interest; such models are commonly referred to as land-use regression (LUR) models. Dispersion models are criticized because they require knowledge of inputs, such as emissions, that are often poorly known, and because they have an inherent uncertainty of as much as plus and minus a factor of two. However, dispersion models allow predictions to be made at a site with no air quality monitors or for future scenarios. LUR models are adequate when there are sufficient air quality monitoring data available at the site, because they reflect the actual scenario. However, they apply only to the specific scenario and space and time domain of the available observations, they cannot be used to forecast the future when emissions may change, and the relationship between concentrations and land-use cannot be justified a priori.

This paper reviews some case studies where the pros and cons of the two approaches are further revealed. A hybrid approach is recommended, where the observations are used to guide the dispersion model solutions, and/or the dispersion model solutions are used to guide the statistical analysis. For example, for the scenario of dispersion downwind from a busy street, it is well-known from dispersion theory that concentrations are inversely proportional to distance from the street, and that simple relation can be incorporated in the statistical analysis.