Table 4. Overall performance of exposure assessment methods.
Assessment of the various statistical and machine learning methods over space and time by averaging: Root Mean Squared Error (“RMSE”), Mean Absolute Deviation (“MAD”), correlation between predicted and observed values (“Corr”), and empirical coverage of the predictive 95% intervals. Both RMSE and MAD are in units of μg/m3. The methods considered are: raw CMAQ output, ordinary least squares (“OLS”), inverse distance weighting (“IDW”), universal Kriging(“UK”), downscaler, random forests (“RF”), support vector regression (“SVR”) and Neural networks (“NN”). Methods use either CMAQ and/or other geographic covariates (“Covs”).
| Method | RMSE | MAD | Corr | Coverage |
|---|---|---|---|---|
| CMAQ | 7.19 | 4.68 | 0.51 | -- |
| OLS (CMAQ) | 4.80 | 3.09 | 0.65 | 0.62 |
| OLS (Covs) | 4.63 | 2.97 | 0.67 | 0.79 |
| OLS (CMAQ + Covs) | 4.22 | 2.63 | 0.74 | 0.83 |
| IDW | 3.22 | 1.82 | 0.85 | -- |
| UK (CMAQ) | 3.08 | 1.70 | 0.87 | 0.95 |
| UK (Covs) | 3.25 | 1.79 | 0.85 | 0.93 |
| UK (CMAQ + Covs) | 3.15 | 1.76 | 0.86 | 0.93 |
| Downscaler (CMAQ) | 3.10 | 1.70 | 0.87 | 0.94 |
| RF (CMAQ + Covs) | 4.23 | 2.74 | 0.73 | 0.96 |
| SVR (CMAQ + Covs) | 3.83 | 2.22 | 0.79 | -- |
| NN (CMAQ + Covs) | 3.90 | 2.49 | 0.78 | -- |