Methods: Monitoring of PNC began in April 2001 using condensation particle counters (3022A, TSI) in Augsburg, Barcelona, Helsinki, Rome, and Stockholm. Concurrent measurements of air pollutants and weather were used, as well as selected interactions between the two, to fit a regularized linear model (also called ridge regression). This technique is robust with respect to inclusion of irrelevant explanatory variables and can be modified to be highly tolerant of missing data, two highly beneficial features when there are many explanatory variables.
Results: The most important predictor variables were the nitrogen oxides. The models appear to fit PNC data relatively well, with R2 of 0.77, 0.80, 0.58, 0.84, 0.81 respectively for the five cities. Split-halves analysis (modelling on half of the data with validation on the other half) indicates that the modelling process was fairly reliable.
Conclusion: A statistical model can be applied to existing data on traffic-related air pollutants and weather variables in order to predict PNC levels. The retrospective prediction of PNC levels appears to be sufficiently reliable for use in epidemiological research.