Supplemental Material: Limitations of Remotely-sensed Aerosol as a Spatial Proxy for Fine Particulate Matter

Here we consider spatial associations of monthly average PM and AOD, matched in space, for each of the 12 months of 2004, to supplement the daily and yearly analyses in the main paper. We focus on MODIS and GOES because of the extremely low retrieval availability of MISR, which would require interpolation of AOD values over very large areas with no retrievals. Even after averaging to the month, some locations have no retrievals, so we use a statistical smoothing model to estimate an AOD surface for each month, using a computationally efficient Markov random field representation of a thin-plate spline (Rue and Held 2005, Sec. 3.4.2; Yue and Speckman, unpublished manuscript) that readily fits the AOD retrieval data for each month (there are as many as 15,000 observations in a month). We use a heteroscedastic residual variance that accounts for the differing number of retrievals in different locations. While the model has the flexibility to do either substantial or little smoothing, in practice because the AOD values are fairly smooth at a fine spatial scale, the smoothed fields look similar to the raw fields, but with imputed values where no data are available. Correlations are higher in the warmer months, but are moderate at best (Supplemental Material, Figure 1). Results are similar when restricting to locations with at least three days with AOD retrievals in a given month. The poor correlations result in part because of the limited retrieval availability, particularly during winter, seen in red in Supplemental Material, Figure 1. Taking the twelve smoothed monthly values and averaging to the year at each location, Supplemental Material, Table 1 shows near-zero correlations of raw AOD and moderate correlations for calibrated AOD ∗email: [email protected]