This paper presents a statistical evaluation of three highway air pollution models (CALINE3, HIWAY-2, and ROADWAY) using the tracer data from the General Motors Sulfate Dispersion Experiment. Since the models predict the ensemble mean whereas any given observation reflects a single realization or an event from a population, it should be recognized that the model predictions will almost always differ from the corresponding observations, even if the models and the input data for the models are perfect. The bootstrap resampling procedure is used to quantify the variability in the observed concentrations due to the stochastic nature of the atmosphere. The results suggest that the variability in the observations due to the random nature of the atmosphere is about 30%. Therefore, if the predicted values are within ± 30% of the measured concentrations, the differences between model predictions and observations should not be considered to be significant. Thus a ‘perfect’ air quality model should predict to within ± 30% of its corresponding observed concentrations. Comparisons of the model predictions paired and unpaired in time with measurements suggest that HIWAY-2 and ROADWAY perform best, but the performance of CALINE3 is acceptable. Application of the extreme value theory and the bootstrap resampling procedure to the modeled and measured data (unpaired) shows that all three models are capable of predicting the extreme concentrations within the model performance criteria set forth above.