Spatio-temporal modeling of particulate air pollution in the conterminous United States using geographic and meteorological predictors

Table 1 Bias and precision statistics from cross-validation (CV) of PM _2.5 , PM ₁₀ , and PM _2.5–10 models from 1999-2007

Pollutant	Region^A	Monthly values									Spatial CV R² ^G
Pollutant	Region^A	N^B	N excluded^C	Model R^2D	CV R²	Intercept^E	Slope^E	NMBF (%)^F	CVMAE^F	NMEF (%)^F	Spatial CV R² ^G
PM_2.5	All	108,718	4	0.84	0.77	0.3	0.87	−1.6	1.61	14.3	0.89
	Northeast	24,318	0	0.85	0.81	0.2	0.92	−1.4	1.44	11.4	0.88
	Midwest	15,767	0	0.85	0.81	0.2	0.91	−0.7	1.31	10.6	0.89
	Southeast	24,201	1	0.88	0.83	0.2	0.92	−0.4	1.31	9.7	0.82
	Southcentral	12,762	0	0.79	0.72	0.2	0.89	−0.6	1.44	14.1	0.83
	Southwest	13,448	2	0.79	0.69	0.4	0.81	−5.5	2.65	26.8	0.83
	Northwest	9,052	0	0.65	0.50	0.7	0.62	−4.6	2.07	28.9	0.62
	Central Plains	9,170	1	0.72	0.60	0.4	0.81	−2.8	1.66	23.2	0.81
PM₁₀	All	104,509	22	0.71	0.58	0.7	0.77	−5.1	5.21	24.4	0.69
	Northeast	16,982	0	0.67	0.57	0.7	0.76	−4.7	4.17	19.8	0.68
	Midwest	10,088	0	0.63	0.48	0.9	0.71	−6.0	4.82	21.2	0.56
	Southeast	20,316	0	0.62	0.49	0.7	0.76	−4.0	3.89	17.6	0.46
	Southcentral	8,092	0	0.61	0.45	0.8	0.74	−6.0	6.24	27.4	0.44
	Southwest	24,050	19	0.76	0.62	0.7	0.79	−4.7	6.92	27.8	0.72
	Northwest	5,943	1	0.59	0.49	0.8	0.71	−1.6	5.33	30.2	0.72
	Central Plains	19,038	2	0.61	0.50	0.8	0.71	−7.6	5.11	31.3	0.66
PM_2.5-10 ^H	All	41,098	1,936	0.67	0.52	0.6	0.76	−3.2	4.18	38.9	0.61
	Northeast	8,375	423	0.49	0.35	0.9	0.58	−8.9	3.46	42.7	0.53
	Midwest	4,567	233	0.61	0.43	0.7	0.70	−1.5	3.72	34.4	0.49
	Southeast	7,178	359	0.45	0.28	0.5	0.75	−4.2	3.02	38.0	0.36
	Southcentral	3,614	23	0.61	0.40	0.9	0.62	−11.4	5.63	44.6	0.33
	Southwest	9,237	296	0.74	0.56	0.5	0.81	−1.6	5.64	36.6	0.64
	Northwest	2,579	340	0.55	0.47	0.3	0.92	18.1	3.97	48.2	0.58
	Central Plains	5,548	262	0.56	0.41	0.6	0.74	−2.3	3.87	39.8	0.61

^ACorresponds to regions shown in Figure 1.
^BIncludes data from CV sets one through nine; see text for details.
^CThree PM_2.5 values above 70 μg/m³ (>99.99th percentile) and one low value, as well as 22 PM₁₀ values above 150 μg/m³ (>99.99th percentile) were excluded from CV statistics as outliers. Extreme values may have been due to local events such as wildland or other fires, dust storms, etc.
^DCalculated on the native rather than natural-log scale and among observations used for CV for comparison to the CV R². For PM_2.5–10, predicted levels<=0 were removed.
^EFrom major axis regression of predictions on measurements (both are natural-log transformed monthly means); see text for details.
^FNMBF is normalized mean bias factor; CVMAE is cross-validation mean absolute error; NMEF is normalized mean error factor; see text for details.
^GSpatial CV R² calculated at 1,245, 1,192, and 512 sites with >35 valid monthly-average measurements for PM_2.5, PM₁₀, and PM_2.5–10, respectively.
^HCalculated as the difference between monthly PM₁₀ and PM_2.5 measurements and, separately, monthly PM₁₀ and PM_2.5 model predictions. Of the 1,936 values excluded as outliers, 11 were removed due to extreme PM₁₀ or PM_2.5 measurements; an additional 1,925 were due to measured or predicted PM_2.5–10 below the limit of detection of 0.57 μg/m³ (<3.4th percentile of measured and <1.6th of predicted PM_2.5–10).

ISSN: 1476-069X