Case-Crossover Spatial Lag Grid Differences Between Aerosol Optical Depth-PM2.5 and Respiratory-Cardiovascular Emergency Department Visits and Inpatient Hospitalizations in Baltimore, Maryland, USA: Identification of Homogeneous Spatial Areas

Optimal use of aerosol optical depth (AOD)-PM2.5 fused surfaces in epidemiologic studies requires homogeneous temporal and spatial fused surfaces. No analytic method is currently available to evaluate the spatial dimension. The temporal case-crossover design was modified to assess the association between Community Multiscale Air Quality (CMAQ) lag grids and four respiratory-cardiovascular hospital events. The maximum number of adjacent lag grids with the expo-sure-health outcome association determined the size of the homogeneous spatial area. The largest homogeneous spatial area included 5 grids (720 km2) and the smallest 2 grids (288 km2). PMC and PMCK analyses of ED asthma, IP asthma, IP MI, and IP HF were significantly higher in rural grids without air monitors than in urban with air monitors at lag grids 0, 1, and 01. Grids without air monitors had higher AOD-PM2.5 concentration levels, poverty percent, population density, and environmental hazards than grids with air monitors. ED asthma, IP MI, and HF PMCK ORs were significantly higher during the warm season than during the cold season at lag grids 0, 1, 01, and 04. The possibility of elevated fine PM and other demographic and environmental risk factors contributing to elevated respiratory-cardiovascular diseases in persons residing in rural areas was discussed.


Introduction
Published studies emphasize the detrimental effects of acute and chronic exposure to elevated ambient PM2.5 concentration levels to respiratory-cardiovascular chronic disease morbidity and mortality . Frequent, accurate, and timely measurements of ambient PM2.5 concentration levels, obtained from on-the-ground air monitors, are and PMB fused surface concentration levels after controlling for the confounding effects of apparent temperature (AT). Separate warm-cold season analyses will also evaluate the relationship between ambient temperature and AOD-PM2.5 concentration levels for the year and each season in the Baltimore study area and in grids with and without air monitors.

Materials and Methods
This study implemented the lag grid case-crossover design and utilized the same four experimental AOD-PM2.5 and PMB concentration level fused surfaces and the same four respiratory-cardiovascular chronic disease groups that were previously used in the lag day case-crossover studies completed for the Baltimore [3,84] and New York City [32] study areas. This section will include lag grid case-crossover study details. Other information about the lag day analyses are in earlier publications describing the Baltimore [3,84] and New York City [32] study areas.  99 12 km 2 CMAQ grids -these grids defined the Baltimore study area [3,84]. The 17 federal reference method (FRM) ambient PM2.5 air monitors, which were operational in 2004-2006, Table 1: Bidirectional Spatial Lag Grid Case-Crossover Analyses for the 99 12 km 2 CMAQ Grids With (Red) and Without (Clear) Ambient PM2.5 Air Monitors in the Baltimore Study Area a-c . (11,1) N (11, 5) (11,9)  are in the 15 grids with air monitors. The grids with ambient air monitors are identified with row-column grid coordinates, city/county abbreviations and are displayed in red hue in Table 1. There were 6 air monitors in BC, 4 in AA, 3 in PJ, 2 in B, and 1 each in M and in H Counties.

ED and IP Hospitalizations
The assessment included the four respiratory-cardiovascular chronic disease hospitalization end points. The 100]. Individual HSCRC electronic patient records did not include patient names, social security numbers, and residential addresses. However, each electronic patient record had temporal information about the ED visits and IP hospitalizations (year [2004-2006], quarter [winter, spring, summer, fall], day of the week [Sunday through Saturday], date of birth, gender, race, health insurance, spatial location for the residential address, five-digit U.S. Postal Service (USPS) residential Zone Improvement Plan (ZIP) code [101], and one primary and multiple secondary diagnostic fields, with International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) billing codes [102]. ICD-9-CM codes permitted the identification of electronic patient records with ED asthma visits and IP asthma (493), IP MI (410), and IP HF (428) hospitalizations in the primary diagnosis field. In addition, ICD-9-CM codes identified which electronic patient records had one or more of the three comorbid chronic diseases of atherosclerosis (414, 440), diabetes mellitus (250), and hypertension (401). The study's protocol, including the proposed data analyses, were approved by the Maryland Department of Health Institutional Review Board (IRB) [103] and the Maryland Health Services Cost Review Commission [100]. The MDH IRB concluded that since this study only utilized administrative hospital data, and because there was no contact with patients, it was approved through an exempted review.

Controls
There were three controls for each case. While the electronic patient records for the controls were the same as those for the cases, the cases had a different quarterly PM2.5 exposure value than the quarterly PM2.5 exposure values assigned to the controls. Three different monthly control exposure sampling strategies were utilized: first control was assigned the mean of the PM2. 5

Case-Control Strata
Each stratum included one case and three controls [3,32,84]. There were four different quarterly strata, each including a case with a different quarterly mean PM2.5 concentration level exposure value. Three controls were in each of the four quarterly strata but, as stated above, they differed from the cases on the assigned mean quarterly PM2.5 concentration level values [104]. The case and control means of the monthly ranks were used to confirm that the Braggio and associates [3,84] studies utilized a bidirectional temporal lag day case-crossover design [105]. Months were ordered in ascending order, from January (1) through December (12), and then the monthly ordinal ranks of 1-12 were assigned to each month, respectively. Cases with quarterly exposure values had mean monthly ranks of 2.0 for winter, 5.0 for spring, 8.0 for summer, and 11.0 for fall. The mean monthly ranks for the three controls were 5.5 (1 = January; 4 = April; 7 = July; 10 = October) for first control, 6.5 (2 = February; 5 = May; 8 = August; 11 = November) for the second and 7.5 (3 = March; 6 = June; 9 = September; 12 = December) for the third. For the first and second strata, mean quarterly ranks for the three controls were higher than the mean quarterly ranks for the two cases. In the last two strata, the mean quarterly ranks for the three controls were lower than the mean quarterly ranks for the two cases. Each case and the three associated controls were matched on age, gender, race, health insurance, residential ZIP code, year, and day of the week. PM2.5 concentration levels and effect modifiers varied temporally by year (3), quarter (4), day of the week (7), and spatially by CMAQ grid (99), thereby resulting in 8316 different possible variable combination values for the four AOD-PM2.5 and PMB fused surfaces. There were fewer CMAQ grids with fused surface concentration level values and health data: 15 with air monitors and 57 without air monitors, for a total of 72 CMAQ grids with heath data. Warm-cold season differences were preserved for subsequent analyses by using this bidirectional lag grid case-crossover design. 6 of 30 The comorbid health conditions included atherosclerosis [3,39,84], hypertension [3,32,39,84], and diabetes [3,10,32,84,[106][107][108]. Other confounders were apparent-temperature (AT, AT 2 ; [3,32,39,84,109,110]) and pollen [3,7,35,36,84,[111][112][113][114][115]. Major holidays (and the day after) were coded as dummy variables (1 = holiday or the day after, 0 = no holiday) in each annual ED visit and IP hospitalization file [3,32,84,116]. Dummy variables were used to code snowstorms (1 = yes; 0 = no) in each of the three annual files [3,32,84,117].

Effect Modifiers
Poverty [3, 35-37, 44, 84, 118-120] and population density [3,84] were two geographic-based demographic variables that came from the U.S. Census Bureau [121] and the Maryland Department of Planning, Maryland Data Center [122]. Maryland Zip Code Tabulation Area (ZCTA) spatial polygons for poverty and persons per square mile were obtained from the U.S. Census Bureau website [121].

AOD-PM2.5 and PMB Fused Surfaces
The same four experimental AOD-PM2.5 and PMB concentration level fused surfaces, previously described in detail in earlier publications [3,32,84] were used in this lag grid case-crossover data analysis study. The MODIS instrument on the two orbiting satellites, Terra and Aqua, obtained daily readings, the first in the morning (10:30 AM) and the other in the afternoon (1:30 PM), local time. The AOD measurement column was continuous from each satellite's trajectory in space to the earth's surface. PM2.5 particles within the AOD column changed the light refraction properties, thereby making it possible to obtain AOD unitless measurements. Thus, the AOD readings represented proxies for ambient PM2.5 concentration levels. The previously established relationship between the AOD unitless value and the ambient PM2.5 concentration level was used to assign a corresponding PM2.5 concentration level value to each AOD unitless value. Implementing this algorithm resulted in a continuous space-time AOD-PM2.5 fused surface [3,32,84,[94][95][96].
The updated HBM permitted the fusing of two or three different input surfaces with (not-Kriged) or without (Kriged) missing AOD-PM2.5 concentration level readings. PMC represented the inclusion of ambient PM2.5 monitor measurements with AOD-PM2.5 concentrations levels. Satellite recording failure or the presence of cloud cover interfered with obtaining unitless AOD column readings, thereby resulting in missing data values. To minimize the loss of daily AOD readings, a second AOD-PM2.5 fused surface was Kriged, resulting in the PMCK fused surface -the HBM fusion of monitor PM2.5 with Kriged PMC concentration levels. For the two remaining AOD-PM2.5 fused surfaces, the HBM was used to combine PMC (not Kriged) or PMCK (Kriged) with monitor PM2.5 readings and CMAQ PM2.5 model estimates to produce PMCQ (not-Kriged PMC) and PMCKQ (Kriged PMCK), respectively. Using the HBM, it was also possible to make PMB -the baseline fused surface, by combining PM2.5 monitor readings with CMAQ PM2.5 model estimates. The four experimental AOD-PM2.5 and PMB fused surfaces were displayed within the EPA CMAQ grid system by overlying the NASA native 10 km 2 AOD grid onto the EPA native 12 km 2 CMAQ grid.

File Linkage
The use of a previously developed polygon correspondence file minimized the spatial mismatch between the various irregular-shaped polygon files [123]. The polygon correspondence file made it possible to assign each USPS ZIP code polygon (health outcome), U.S. Census Bureau ZCTA polygon (poverty percent and population density), and CMAQ 12 km 2 grid template (four experimental AOD-PM2.5 and PMB concentration level fused surfaces) to one CMAQ grid [3,84]. Assembling the polygon correspondence file necessitated completing these steps: 1) Obtaining latitude-longitude centroid coordinates for each ZIP code and ZCTA polygons [83,121,122]. 2) Using a geographic information system (GIS) to assign each residential ZIP code polygon or ZCTA polygon to one CMAQ 12 km 2 grid, based on the spatial location of each latitude-longitude centroid coordinate of each polygon within a specific CMAQ grid [124]. By using the correspondence file, it was possible to implement the CMAQ 12 km 2 grid system as the standardized spatial polygon template for this data analysis study [3,83,84,123,124].
Each of the four linked AOD-PM2.5 files (one for ED asthma, a second for IP asthma, a third for IP MI, and a fourth for IP HF) was sorted on the spatial grid coordinates first and temporal values next. The sorting sequence for the Baltimore study area was columns first then rows, followed by the temporal variables of weekday, quarter, and year [3,84]. But, for this spatial lag grid analysis study, only the sorting sequences of columns and rows were implemented because they were of primary importance in this lag grid case-crossover study.

Bidirectional Case-Crossover Spatial Lag Grid Analyses
The lag grid case-crossover analysis resembled the lag day case-crossover analysis, except that the spatial analysis evaluated lag grids, including the index grid, which always had a value of 0 (including all lag grids of 0-4, 01, 24, and 04), while the temporal analysis assessed lag days preceding the index day, which always had a value of 0 (including all lag days of 0-4, 01, 24 and 04). In the New York City [32] and Baltimore [3,84] studies, we utilized a bidirectional lag day case-crossover design to determine if there were differences due to asthma ED visits, and IP asthma, IP MI, and IP HF hospitalizations among the individual lag days (0-4) and summary lag days (01, 24, and 04) [3]. The lag day analysis determined if ambient PM2.5 concentration levels, evaluated by the four experimental AOD-PM2.5 and PMB concentration level fused surfaces, differentially contributed to the subsequent occurrence of one or more respiratory-cardiovascular ED visits or IP hospitalizations. In the lag day analysis, all outcomes were evaluated on different days but in the same grid. In the lag grid analysis, all outcomes were evaluated in different grids but on the same day.
The overall objective of the lag grid case-crossover assessment was to identify the number of 12 km-wide grid units from the index grid that demonstrated the same association between each of the four AOD-PM2.5 and PMB fused surface concentration levels and each of the four respiratory-cardiovascular hospital outcomes, ED visits or IP hospitalizations. The implementation of the lag grid case-crossover analysis is described in the three examples shown in Table 2. The first illustration is for grids without monitors, top of the table. The Spatial Lag Grid Analyses column displays lag grid values, both individual (0-4) and summary measures (01, 24, 04). The grid location for the individual grids with monitors is shown in bold red font for grids with air monitors and bold black font for grids without air monitors. The index grid location for the no monitor example is Row = 9 and Column = 7 (R9, C7). This index grid has a lag grid value of 0. Moving from the far-right side of the table (East) to the left side of the table (West), the grid adjacent to and immediately to the left of the index grid has a lag grid value of 1, and its grid location is R9, C6. The next three lag grids, with lag grid values 2-4, have sequential grid locations: R9, C5; R9, C4; and R9, C3, respectively. Summary lag grids 01, 24, and 04 were obtained by computing means for individual lag grid values. To illustrate, the summary lag grid 01 represents the mean for lag grids 0 and 1. Lag grid 24 includes the mean for lag grids 2-4. Lag grid 04 refers to the mean for lag grids 0-4.
The second example, shown in the middle of the table, is for lag grids with air monitors. Here, the index grid location is R6, C7. Again, the index grid has a lag grid value of 0. The first two lag grids with monitors are in the 6 th row, while the next three lag grids with monitors are in the 7 th row. Notice that column location decreases by 1 as the lag grid value increases by 1.
The third example, at the bottom of Table 2, is for both monitor grid conditions combined. The index grid location is R4, C7. All five individual lag grids, with lag grid values of 0-4, are in the 4 th row. Three of the five lag grids, 1, 2 and 4, include air monitors. The other two lag grids, 0 and 3, include grids without monitors.
Considering the spatial lag grid analysis descriptions summarized in Tables 1 and 2, it is possible to demonstrate that the three types of lag grid analyses presented above represent the inclusion of lag grids that either precede or follow the index grid, based on the index grid column location. Each complete lag grid sequence always includes five individual lag grid values, the index grid, and four additional lag grids. The four lag grids are selected from 1-4 lag grids that always precede the index grid. This prior selection of lag grids, relative to the location of the index grid, occurs for columns 9 E to 5 W. But, for the remaining four columns, 4 E to 1 W, the selection of the four preceding lag grids, relative to the location of the index grid, will invariably include 1-4 grids that always follow the index grid. To illustrate, when the index grid has a column location value of 4, the selection of the four preceding grids will result in the inclusion of column 3 for lag grid 1, column 2 for lag grid 2, column 1 for lag grid 3, and column 9, in one row above, for lag grid 4. When the index grid is in column 3, two grids precede the index grid, and two grids follow the index grid. The index grid for column 2 includes 1 grid that precedes the index grid and three grids that follow the index grid. For the fourth lag grid, when the index grid is in column 1, all four lag grids follow the index grid and are in the above row.
These are the results for all grids that preceded or followed the index grid: For grids with monitors, 62.0% of the grids had the same column value or a lower column value (preceded) than the index grids, and 38% of the grids had column values that were larger (followed) than the column values of the index grids. For grids without monitors, 69.3% preceded the index grids, and 30.7% followed the index grids. The combined no monitor grids with monitor grids included 68.4% of the grids preceded the index grids, and 31.6% followed the index grids. To summarize, these results support the conclusion that this spatial lag grid case-crossover design included lag grids that preceded or followed the index grid; thus, this outcome demonstrates that this study implemented a bidirectional lag grid case-crossover design [3,32,84,105].

Statistical Analyses
The four linked fine PM exposure-health outcome files, one each for ED asthma, IP asthma, IP MI, and IP HF were analyzed using CLR statistical software included in the SAS (Statistical Analysis System)/STAT proportional hazards ratio (PHREG) Procedure, version 14.3, along with Base SAS version 9.4 [3,32,84,[125][126][127][128][129]. PHREG performs CLR analyses on survival data, based on the Cox Proportional Hazards model, by quantifying the effects of explanatory variables on survival times. The Chi Square test evaluated group differences between categorical variables, such as case-control status, gender, and race [127]. Assessment of continuous variables, such as poverty percent and population density (converted to L10 values before analysis), involved comparing a case group mean to a control group mean and its associated 95% Confidence Interval (CI). Means and 95% CIs were obtained from Proc Means in Base SAS [127]. If the case mean was below or above the 95% CI lower or upper limits of the control group mean, this outcome was significantly different at p≤0.05 [125]. But, if the case mean was within the 95% CI lower and upper limits of the control mean, this outcome was not significant, p>0.05.

Final CLR Runs
Final CLR runs included the base model, effect modifiers for lag grids 0-4, and summary lag grids 01, 24, and 04. The three grid monitor conditions included grids without a monitor (No), grids with at least one monitor (Yes), and grids with or without monitors (Both). The null hypothesis was rejected if p≤0.05. The additional use of the Akaike Information Criterion (AIC) permitted the selection of CLR outcomes with lower values, thereby representing a better parameter fit [125,126]. Follow-up analyses evaluated the contribution of monitor (No, Yes) and season (Warm, Cold) in lag grids 0-4, 01, 24, and 04 [3].

Results
CLR runs evaluated the contribution of the four AOD-PM2.5 and PMB fused surfaces to the four health outcomes (ED asthma, IP asthma, IP MI, and IP HF) in grids (with air monitors, without air monitors, and both) at lag grid values of 0-4, 01, 24, and 04. Differences between experimental AOD-PM2.5 and PMB concentration level fused surfaces and demographics for all conditions are displayed in Table 3. The categorical analyses of PM2.5 concentration levels that were Below, Within, or Above the three-year fused surface mean's 95% CIs for each of the three grid conditions showed significant heterogeneity in all grids (Both), and in grids with (Yes), and without (No) air monitors. PMCK concentration level was significantly higher than PMB concentration level in the Both and No monitor grid groups, and significantly lower in grids with monitors (all p's≤0.05). PMC, PMCQ, and PMCKQ concentration levels were significantly lower than PMB concentration levels in all three monitor grid conditions (all p's≤0.05). Also, in all grids, PMB and the two AOD-PM2.5 fused surfaces that included CMAQ PM2.5 model estimates, PMCQ and PMCKQ, resembled each other in number (52)(53)(54)(55) and percentage (52%-56%) of grids with PM2.5 three-year concentration level mean values in the Above category. For the other two AOD-PM2.5 fused surfaces, PMC and PMCK, the number and percentage (30, 30% and 45, 45%, respectively) of grids with mean PM2.5 concentration levels in the Above category in all grids were lower than PMB (55, 56%). Monitor versus no monitor comparisons showed that PM2.5 concentration levels for PMB and three of the four AOD-PM2.5 fused surfaces -PMC, PMCQ, and PMCKQ -were significantly higher in grids with air monitors than in grids without air monitors (all p's≤0.05). Only PMCK had PM2.5 concentration levels that were not significantly different in grids with (14.27 μg/m 3 ) and without (14.39 μg/m 3 ) air monitors, both p's>0.05.
As seen at the bottom of Table 3, categorical analyses yielded new information about differences in grids with and without air monitors, which were not evident in the continuous measures for poverty percent and population density. Results for continuous measures supported the anticipated outcome of higher poverty percent and elevated population density in grids with air monitors compared to grids without air monitors (both p's≤0.05). For the demographic measures of poverty percent and population density, evaluated as categorical variables, there were more grids in the Above category in grids PMCK 14 Table 4 shows totals and percentages for the four respiratory-cardiovascular ED visit and IP hospitalization groups stratified on case-control status and patients' race. As seen in the Both column, second from left, ED asthma had the most observations (cases and  MI (19201) and IP HF (27518). Totals and percentages for each of the four health outcome groups were higher in the 57 grids with no monitors than in the 15 grids with monitors. For each of the four health outcome groups, approximately 25% of observations were cases and about 75% were controls. The ratio of one case to three controls was also evident in grids with and without air monitors for ED asthma, IP asthma, IP MI, and IP HF (all p's>0.05). Only for ED asthma was the total (%) for Black patients (22696, 48%) higher than the totals (%) for White patients (21294, 45%) and Other patients (3060, 6%). For IP asthma, IP MI, and IP HF the totals (%) were highest for White patients, intermediate for Black patients, and lowest for Other patients. The Race by CMAQ grid condition analyses were significant for all four health outcomes (all p's≤0.05). ED asthma and IP asthma Black patients' totals (%) were higher in grids with air monitors than in grids

CLR Analyses
CLR results showed significant differences between AOD-PM2.5 and PMB fused surfaces, health outcome groups, lag grids, monitor conditions, and warm-cold season differences. For all CLR runs, ORs were significant at lag grids 0, 1, 01 and 04 (all p's≤0.05). Effect modifiers were significant only for ED asthma in all grids (PMCQ population density, lag grid 0, protective OR, p≤0.05) and in grids with monitors (PMB, PMC, PMCK, PMCQ, PMCKQ fused surfaces for season, lag grids 2-4, 04, all p's≤0.05). By using the mean (95% CI) it was possible to evaluate OR differences for the four health outcomes, four AOD-PM2.5 and PMB fused surfaces at lag grids 0, 1, 01 and 04. ED Asthma Figure 1 shows ED asthma ORs (95% CIs) for the four AOD-PM2.5 and PMB fused surfaces. There are separate panels in Figure 1, one for each of the four lag grid values Figure 1: ED asthma ORs (95% CIs) for the four AOD-PM2.5 and PMB fused surfaces, in all grids (Both) and in grids with (Yes) and without (No) air monitors, and in separate panels for lag grids 0 (1a), 1 (1b), 01 (1c), and 04 (1d). of 0, 1, 01 and 04. In the Both grid condition, for lag grids 0, 1 and 01, ORs for each of the four AOD-PM2.5 fused surfaces were significantly higher than the PMB ORs (all p's≤0.05). In the Both grid condition, for lag grid 04, the PMC, PMCK and PMCKQ ORs were significantly higher than the PMB ORs (all p's≤0.05). In lag grids 0, 1, and 01, the no monitor PMC and PMCK ORs were significantly higher than monitor PMC and PMCK ORs (all p's≤0.05).  Figure 2 shows IP asthma ORs (95% CIs) for the four AOD-PM2.5 and PMB fused surfaces, and in separate panels for lag grids 0, 1, 01 and 04. In the Both grid condition, for lag grids 0, 1 and 01, the PMC, PMCK, and PMCKQ ORs were significantly higher than Figure 2: IP asthma ORs (95% CIs) for the four AOD-PM2.5 and PMB fused surfaces, in all grids (Both), and in grids with (Yes) and without (No) air monitors, and in separate panels for lag grids 0 (2a), 1 (2b), 01 (2c) and 04 (2d).
the PMB ORs (all p's≤0.05). In lag grids 0 and 01, only the no monitor the PMCK ORs were significantly higher than the monitor PMCK ORs (both p's≤0.05).

IP MI
IP MI ORs for the four AOD-PM2.5 and PMB fused surfaces, for lag grids 0, 1, 01 and 04, are displayed in Figure 3. In lag grids 0, 1 and 01, the PMC, PMCK, and PMCKQ ORs were significantly higher than PMB ORs (all p's≤0.05). Also in lag grids 0, 1 and 01, the no monitor PMC and PMCK ORs were significantly higher than the monitor PMC and PMCK ORs (all p's≤0.05).   Figure 4. In lag grids 0, 1 and 01, the PMC, PMCK, and PMCKQ ORs were significantly higher than PMB ORs (all p's≤0.05). For lag grid 04, only the PMCK OR was significantly higher than the PMB OR (p≤0.05). In lag grids 0, 1 and 01, the no monitor PMC and PMCK ORs were significantly higher than monitor PMC and PMCK ORs (all p's≤0.05).

Monitor Versus No Monitor OR Percent
The no monitor -monitor OR percent (∆OR%) values for the four AOD-PM2.5 and PMB fused surfaces and four lag grids 0, 1, 01 and 04, are displayed, in four separate panels, one for each of the four health outcomes, in Figure 5. For the four health outcomes, ∆OR% values were larger for PMC and PMCK than the other three fused surfaces. These differences occurred for each one of the four lag grid values. The ∆OR% values at lag 04 was smaller than the ∆OR% values at lag grids 0, 1 and 01. For the four health outcomes, PMB ∆OR% values were negative at lag grids 0, 1, 01 and 04. Results were mixed for PMCQ and PMCKQ. IP asthma PMCQ had negative ∆OR% values as did IP asthma PMB. PMCKQ ∆OR% at lag grids 0, 1 and 01 were positive, resembling PMC and PMCK. One difference, however, was the PMCKQ ∆OR% values were lower than the PMC and PMCK ∆OR% values at lag grid 04. PMCQ ∆OR% values were negative for ED asthma and IP asthma, resembling PMB ED asthma and IP asthma ∆OR% values. Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 10 January 2022

Size of Homogeneous Spatial Area
The largest size of the homogeneous spatial area was identified by the presence of a significantly higher AOD-PM2.5 fused surface OR than the PMB OR for each chronic disease group (ED asthma, IP Asthma, IP MI, and IP HF) at lag grids 0, 1, 01 and 04. Table  5 summarizes the homogeneous spatial area findings. For both monitor grid conditions combined (Both, top of Table 5) analyses of ED asthma and IP HF with the PMCK fused surface produced ORs that were significantly higher than the PMB ORs at lag grid values of 0, 1, 01 and 04 (all p's≤0.05). ORs for ED asthma with the PMC and PMCKQ fused surfaces were significantly higher than the PMB ORs at lag grids 0, 1, 01 and 04 (all p's≤0.05). Based on these significant outcomes at lag grid 04 for these three AOD-PM2.5 fused surfaces, it was concluded that the largest homogeneous spatial area for PMC, PMCK and PMCKQ fused surfaces was (1 grid 12 km wide x 5 grids) 60 km wide. In grids with air monitors, only the ED asthma PMCK ORs were significantly higher than the PMB ORs for lag grids 0, 1 and 01 (all p's≤0.05). The width of the largest homogeneous spatial area for PMCK was 2 grids wide, or 24 km. In grids without air monitors the analyses of ED asthma, IP MI, and IP HF with the PMC and PMCK fused surfaces produced ORs that were significantly higher than PMB ORs at lag grids 0, 1, 01 and 04 (all p's ≤0.05). The analyses of ED asthma with PMCKQ resulted in ORs that were significantly higher than the PMB ORs at lag grids 0, 1, 01 and 04 (all p's≤0.05). The maximum size of the homogeneous spatial area for PMC, PMCK and PMCKQ was 5 grids wide.    1 Each AOD-PM2.5 fused surface mean is compared with PMB fused surface mean (95% CI). Only significant differences are shown, p≤0.05. 2 AOD-PM2.5 and PMB fused surface means in grids without air monitors are compared to means (95% CI) in grids with air monitors. Significant outcome at p≤0.05 is displayed with symbol of "<". Not significant outcome for p>0.05 is shown with symbol of "=".

Monitor-No Monitor Differences
The purpose of these analyses was to determine if the AOD-PM2.5 and PMB fused surfaces had concentration levels that differed between grids with and without air monitors. Results are presented at the bottom of Table 5. The ORs for ED asthma, IP asthma, IP MI, and IP HF with PMC and PMCK fused surfaces were significantly higher in rural grids without air monitors than in urban grids with air monitors (all p's≤0.05).

Warm-Cold Season
There were warm-cold season differences for the four the AOD-PM2.5 fused surfaces and the four health outcomes (Online Resource includes electronic supplementary material consisting of Figures S1 [ED asthma: a-d, lag grids 0, 1, 01, 04, respectively] though S4 [IP HF: a-d, lag grids 0, 1, 01, 04, respectively] for ED asthma (referenced), IP asthma, IP MI, and IP HF (referenced); EMS_1,pdf file). These outcomes were expressed as larger ORs during the warm season and smaller ORs during the cold season. Specifics for each health outcome will be presented below.
ED asthma ( Figure S1), IP asthma ( Figure S2), and IP MI ( Figure S3) PMB, PMC, PMCK, PMCQ, and PMCKQ warm season ORs were significantly higher than cold season ORs for lag grids 0, 1 and 01 (all p's≤0.05). ED asthma, IP asthma, and IP MI PMCK warm season ORs were significantly higher than cold season ORs for lag grid 04 (all p's≤0.05). IP HF PMB, PMC, PMCK, PMCQ and PMCKQ warm season ORs were significantly higher than cold season ORs for lag grids 0, 1, 01 and 04 (all p's≤0.05). Table 6 shows the three-year mean (95% CI) PM2.5 concentration levels for the four AOD-PM2.5 and PMB fused surfaces and three-year mean ambient temperature values in F° by season (Both, Warm, Cold). The Both condition shows the significant correlations between each fused surface PM2.5 concentration level and ambient temperature (all p's≤0.05). The measure of association shown is r 2% , the square of the correlation (r) expressed as a percent. The r 2% value quantifies the percentage of the variance that is shared by each fused surface PM2.5 concentration level and ambient temperature. The r 2% measures are largest for PMCK, smallest for PMB and PMCQ, and intermediate for PMC and PMCKQ, in all three monitor grid conditions (Both, No, Yes). PMB r 2% was higher in grids with monitors (11.2%) and lower in grids without monitors (3.6%). PMCQ's r 2% was also higher in grids with monitors (16.5%) than in grids without monitors (9.3%). Differences in the r 2% measure between grids without and with air monitors were positive for PMC (41.3%-37.1% = 4.2%), PMCK (61.2%-50.2% = 11%), and PMCKQ (37.8%-34.0% = 3.8%), and negative for PMB (3.6%-11.2% = -7.6%) and PMCQ (9.3%-16.5% =-7.2%) . Concentration levels for the five PM2.5 fused surfaces were significantly higher during the warm season than during the cold season, under all three grid conditions (all p's≤0.05).
As expected, ambient temperature during the warm season was significantly higher than during the cold season for the three grid conditions (all p's≤0.05). Figure 6 displays the warm -cold season ∆OR% values. The ∆OR% measures were positive for each of the four AOD-PM2.5 and PMB fused surfaces, for ED asthma (panel a), IP asthma (b), IP MI (c), and IP HF (d) in lag grids 0, 1, 01, and 04. All four panels in Figure 6 demonstrate that PMC and PMCK had consistently higher ∆OR% values than the other two AOD-PM2.5 (PMCQ, PMCKQ) and the baseline PMB fused surfaces at lag grids of 0, 1, and 01. ED asthma ∆OR% values were smaller than the ∆OR% values for IP asthma, IP MI, and IP HF. Also, for all four health outcomes and the four AOD-PM2.5 and PMB fused surfaces, ∆OR% values were smaller at lag grid 04 than at the other three lag grid values of 0, 1, and 01.

Discussion
This data analysis study implemented, for the first time, a spatial lag grid case-crossover design to identify homogeneous spatial areas in rural grids without air monitors that displayed a similar association between elevated AOD-PM2.5 and PMB concentration levels and increases in respiratory-cardiovascular ED visits and IP hospitalizations as has been found in urban grids with air monitors. Analyses of ED asthma, IP MI, and HP HF with PMC and PCMCK attained a maximum homogeneous spatial area (12 km * 12 km * 5 grids [0-4] ) of 720 km 2 . The largest homogeneous spatial area identified by the analysis of ED asthma with PMCKQ was also 720 km 2 .
The homogeneous spatial area analysis results for the Baltimore study area, urban grids with air monitors and rural grids without air monitors combined, were the same as those obtained for the no monitor grid analyses. For the urban area air monitor grid condition, the comparisons between each of the four AOD-PM2.5 fused surfaces and PMB were not significant, except for the analysis of ED asthma with PMCK at lag grids 0, 1 and 01. These results indicate that the benefit of utilizing AOD-PM2.5 is only detected when the analyses are completed in grids without air monitors. Other analyses demonstrated what combinations of the four AOD-PM2.5 and PMB fused surfaces with the four respiratory-cardiovascular hospital events resulted in larger ORs in rural grids without air monitors and smaller ORs in urban grids with air monitors. PMC and PMCK ED asthma, IP asthma, IP MI, and IP HF ORs in grids without air monitors were significantly higher than ORs in grids with air monitors at lag grids 0, 1 and 01. There were differences in the pattern of significant ORs obtained with the spatial lag grid case-crossover design used in this study and the temporal lag day case-crossover design reported by Braggio and associates in 2020 [3]: For the temporal case-crossover analyses only lag days 0, 1 and 01 were significant. The implementation of the spatial case-crossover analyses in this study obtained significant lag grids of 0, 1, 01 and 04. Categorical analyses of Baltimore study area demographics identified, for the first time, a subset of homogeneous spatial grids in rural areas without ambient air monitors that resembled urban grids with ambient air monitors in poverty percent and population density: In rural areas there were 8 grids with poverty percent values in the Above category, and 7 grids in urban areas with poverty percent values in the Above category. There were 19 rural grids with population density in the Above category and 11 urban grids in the above category. In addition, there were higher totals and percentages of IP MI (12016, 62.6%) and IP HF (15684, 57.0%) in rural areas than in urban areas (7185, 37.4% and 11834, 43.0%, respectively). But the totals and the percentages of Black ED asthma patients (11844, 25.2%) and IP asthma patients (2358, 17.5%) were higher in urban areas than in rural areas (10852, 23.1% and 2152, 16.0%, respectively). These results support the conclusion that within the Baltimore study area there is a subset of rural grids without air monitors that resembles the set of 15 urban grids with air monitors in demographic risk factors and in the association between elevated AOD-PM2.5 concentration levels and increases in respiratory-cardiovascular ED visits and IP hospitalizations.
In an epidemiologic study Hirshon and associates [40] evaluated PM2.5 concentration and zinc level readings from the Baltimore PM2.5 Supersite and children's asthma ED visits and IP hospitalizations. These authors reported that PM2.5 zinc concentrations resulted in increases in asthma hospital events. The publication did not identify the zinc source, however. One explanation could be that the ambient PM2.5 zinc levels recorded at the Baltimore PM2.5 Supersite [130] may have come, in part, from a nearby TRI site [131].
The Baltimore PM2.5 Supersite location can be remapped onto one of the 99 CMAQ grids utilized in the Baltimore study area [3]. As a result of this remapping, we discovered that the location of the Baltimore PM2.5 Supersite is in CMAQ grid R6, C6. This same Baltimore study area grid also had three FRM PM2.5 air monitors and one EPA-identified TRI facility that emitted ambient zinc fumes and dust. An adjacent Baltimore study area, grid R6, C7, had one TRI facility that released zinc fumes and dust in the air.
As a follow-up to the Hirshon and associates [40] study, we also looked at the number of TRI facilities that released zinc fumes or dust in the Baltimore study area between 2004-2006. There were five different businesses operating during this three-year timeframe. One company had facilities in two distinct locations, contributing a total of 11 zinc point sources: 7 zinc fumes or dust point sources in grids with air monitors (urban grids) and four in grids without air monitors (rural grids). This analysis suggests that the ambient PM2.5 zinc measured at the Baltimore PM2.5 Supersite in 2002 could have originated, at least in part, from a nearby TRI facility that emitted zinc fumes and dust. Although the FRM PM2.5 air monitors data we utilized did not include PM2.5 zinc, it is possible that ambient zinc from TRI zinc emitting facilities in selected at-risk grids (with or without air monitors) in the Baltimore study area could have contributed to children's asthma ED visits and IP hospitalizations.
Although this study did not evaluate the environmental hazard associated with living close to brownfields [132,133] or EPA-designated Toxic Release Inventory (TRI) facilities [131], several published studies have described the environmental contamination from manufacturing efforts in Maryland that have included the EPA TRI facilities in the State [134][135][136][137] and brownfields in South Baltimore [132,133]. Perlin, Sexton, and Wong [135] found that there were 122 TRI sites in Maryland. About half of the TRI facilities were in Baltimore City, Howard, Anne Arundel, and Baltimore Counties. Maryland residents near a TRI site were medically underserved [137]. South Baltimore brownfields have higher respiratory and heart disease mortality rates among White working-class residents than the rest of the City and State [132]. Litt, Tran, and Burke [133] described a variety of environmental hazards in Southeast Baltimore that included heavy metals, solvents, and insecticides. Living in these environmentally-compromised areas for long durations could increase residents' adverse responsiveness to lower ambient PM2.5 concentration levels and their enhanced contribution to respiratory-cardiovascular ED visits or IP hospitalizations [5,24,132,133].
The suggestion introduced here is that the accuracy of AOD-PM2.5 concentration levels in estimating actual ambient PM2.5 concentration levels determined by on-the-ground fine PM ambient air monitors is partly contingent on spatial scale [44,68,[138][139][140]. Remote sensing studies that have used AOD to estimate ambient PM2.5 concentration levels have concluded that smaller grids provide greater accuracy than larger grids. We utilized the NASA 10 km 2 grid when we accessed the AOD unitless readings for the Baltimore and New York City study areas. Since our objective has been to evaluate how accurately the four experimental AOD-PM2.5 concentration level fused surfaces, in comparison to baseline PMB, estimated ambient PM2.5 concentration levels and their contribution to the occurrence of respiratory-cardiovascular chronic disease ED visits and IP hospitalization in all grids and grids with and without ambient air monitors, the NASA 10 km 2 grid was mapped to the EPA CMAQ 12 km 2 grid before the Baltimore and New York City epidemiologic studies were undertaken [3,32,84]. In the Braggio and associates' publications [3,84] and the results of this study, we have controlled for scale effects on the accuracy of AOD-PM2.5 fused surfaces to represent actual ambient PM2.5 concentration levels by utilizing the EPA CMAQ 12 km 2 grid as the smallest spatial area of analysis. The larger sites, those that defined grid multiples with air monitors, grid groups without air monitors, and both air monitor grid conditions combined, are not less accurate than the individual 12 km 2 grid since the larger areas represent the inclusion of multiple homogeneous 12 km 2 grids into larger spatial grid areas. Likewise, the accuracy of the identified homogeneous spatial areas, in rural grids without air monitors and in urban grids with air monitors, is equivalent to the accuracy which characterizes each CMAQ grid.
Previously published studies have also found higher PM2.5 concentration levels and worse health outcomes during the warm season than during the cold season [3,10,18,31], while other reports have reported larger cold season effects and smaller warm season effects [7,13,28,32,41,97,98]. Cheng and colleagues [8] found significant increases in asthma hospital admissions on warm (>23° C) and cool (<23° C) days in Taipei, Taiwan. Our analyses found significant positive associations between AOD-PM2.5 concentration levels and ambient temperature, suggesting that higher temperatures during the warm season, relative to the cold season temperatures, contributed to the increased OR of respiratory-cardiovascular ED visits and IP hospitalizations. In grids with and without air monitors, the r 2% values were higher for PMCK than the other three AOD-PM2.5 and PMB fused surfaces. This measure of association between PMCK and ambient temperature explained 59% of the shared variance in all grids, 61% in grids without air monitors, and 50% in grids with air monitors. Results for PMC-ambient temperature shared variance were 41% in the Baltimore study area, 41% in grids without air monitors, and 37% in grids with air monitors. Percentages of fused surface-ambient temperature shared variance measured by r 2% were smaller for PMB: 4.6% in the Both condition, 3.6% in grids without air monitors, and 11.2% in grids with air monitors. The association between AOD-PM2.5 concentration level and ambient temperature could be a proxy for another factor that could contribute to increases in ambient fine PM concentration levels: Possible fine PM sources could include transportation and manufacturing air pollution particles [141,142]. Therefore, these warm-cold season results also support using the PMC or the PMCK fused surface as a candidate to replace PMB, the currently available baseline, as a reliable and accurate estimate of ambient PM2.5 concentration level.
These spatial lag grid results provide additional support for the use of PMC and PMCK fused surfaces in epidemiologic investigations of the association between ambient fine PM and respiratory-cardiovascular hospital events. Both PMC and PMCK consistently identified OR differences between air monitor and no air monitor grids relative to the mixed results for the other two fused surfaces that included PM2.5 CMAQ model estimates, PMCQ and PMCKQ. PMC and PMCK showed larger changes between warm and cold season ORs. Published studies have contributed to the continued improvement in the use of remote sensing methodology by showing how AOD can be utilized to more accurately estimate ambient PM2.5 concentration levels in areas with and without air monitors [3, 6, 9, 13-16, 20, 22, 30, 32, 48, 51, 52, 54, 57-74, 77, 80-82, 84, 89, 92, 143]. Based on the results of this study, spatial attributes favor the use of PMCK over PMC as the replacement for PMB.
There are strengths in the analytical methods utilized in this spatial lag grid casecrossover study. To our knowledge, this is the first time the case-crossover design has been modified to evaluate lag grids. The selection of controls that preceded or followed the cases makes the lag grid case-crossover design spatially bidirectional [105]. The use of the lag grid case-crossover design to quantify spatial heterogeneity provides an assessment tool to identify homogeneous spatial areas. Study results demonstrate, also for the first time, the presence of grids without air monitors (rural grids) that resembled grids with air monitors (urban grids) in exposing rural residents to increased fine PM levels and seeking medical care for a respiratory-cardiovascular chronic disease hospital events. A final contribution of this study was the identification of benefits derived by utilizing AOD-PM2.5 concentration level fused surfaces, especially PMCK, as estimates of ambient fine PM levels in urban and rural areas.
There are unresolved methodological issues that limit the generalizability of study results. The 99 CMAQ grids that defined the Baltimore study area included 15 urban grids with air monitors and 84 rural grids without air monitors. All 15 urban grids with monitors had associated respiratory-cardiovascular hospital event data. Only 57 of the 84 rural grids without air monitors had respiratory-cardiovascular hospital event data. Grids without air monitors that lacked health data could have been those grids over the Chesapeake Bay. The Chesapeake Bay CMAQ grids, located in the south-east corner of the Baltimore study area, included more water than land mass. Residents live on part of Maryland's irregular coastline and islands. Underestimates for total patients with respiratory-cardiovascular ED visits and IP hospitalizations could have occurred because Maryland residents can and do obtain medical treatment out of state, e.g., Washington, DC, Virginia, Pennsylvania. To be consistent with the way the linear boundaries for the New York City study area and the Baltimore study area were established for the purpose of developing the AOD-PM2.5 and PMB fused surfaces, it was necessary to include all 99 CMAQ grids in the Baltimore study area. Another study limitation was the inability to identify which grids were included in the homogeneous spatial areas. Based on the way this study implemented the spatial lag grid case-crossover design, only the size of the homogeneous spatial area could be determined as CMAQ grid multiples, with each one 12 km wide. There was no independent confirmation of actual ambient PM2.5 concentration levels in rural grids without air monitors. There was also (relative) spatial heterogeneity in the 15 urban grids with air monitors because there were only 17 ambient air monitors for an area of (12 km * 12 km * 15 grids) 2160 km 2 . If the 17 fine PM ambient air monitors were equally distributed among the 15 CMAQ grids, each monitor would occupy 127.1 km 2 . Monitor accuracy for the fine PM measurements is highest at the monitor's location. Fine PM measurement accuracy decreases as the distance from the monitor increases. It is not clear to what extent health care access impacted the results. Available evidence indicated that this may not have been a bias since there were more patients with the four respiratory-cardiovascular chronic diseases in rural grids without air monitors than in urban grids with air monitors.
Future research efforts should involve the identification of criteria that will lead to the replacement of the currently used PMB baseline fused surface with another improved AOD-PM2.5 fused surface. Relevant attributes that could be helpful in the selection of an updated AOD-PM2.5 baseline can include grid resolution below 10 km 2 , absence of missing AOD unitless readings, and improved accuracy of PM2.5 concentration level estimates in areas without ambient air monitors [3,84]. In addition, to increase the reliability and validity of PM2.5 readings in grids without air monitors [84], it will be necessary to have available, independent on-the-ground ambient PM2.5 readings in grids without ambient air monitors. This more ambitious goal could be reached by using portable and accurate PM2.5 monitors to supplement fine PM readings available from the EPA AQS network [50]. The overall goal of these proposed improvements will be to protect the respiratorycardiovascular health of residents in rural areas, as has been accomplished thus far for residents in urban areas.

Conclusions
The spatial lag grid case-crossover results provided support for the use of this analytical method to identify homogeneous spatial areas that demonstrated the same relationship between elevated AOD-PM2.5 concentration levels and increased respiratory-cardiovascular hospital events in the entire Baltimore study area and in rural grids without air monitors. With the PMC and PMCK fused surfaces, the largest homogeneous spatial areas, showing the same relationship between elevated AOD-PM2.5 and ED asthma and IP HF, were 720 km 2 for the Baltimore study area (Both) and in rural grids without air monitors. These results suggest that some rural grids without air monitors were not necessarily different from urban grids with air monitors in the contribution of increased ambient fine PM to the occurrence of respiratory-cardiovascular hospital events. While limitations in the spatial case-crossover analyses could not identify specific grids included in the homogeneous spatial areas, other results from categorical data analyses confirmed that there was a subset of rural grids without air monitors that demonstrated higher poverty percent, increased population density and elevated AOD-PM2.5 concentration levels as has been found for urban grids with air monitors. Warm-cold season analyses confirmed that elevated AOD-PM2.5 concentration levels during the warm season contributed to increases in respiratory-cardiovascular ED visits and IP hospitalizations. New information confirmed the association between elevated AOD-PM2.5 concentration levels and ambient temperature. Higher temperatures could be proxies for other sources of fine PM that are more prevalent during the warm season than during the cold season. PMC and PMCK fused surfaces consistently demonstrated larger differences between warm and cold seasons than the other two AOD-PM2.5 fused surfaces that included CMAQ PM2.5 estimates (PMCQ and PMCKQ) or the currently used baseline, PMB. Future research efforts should reevaluate the contribution of increased ambient fine PM levels and areaspecific demographic and environmental hazards to increased susceptibility of developing respiratory-chronic diseases especially in Maryland's rural areas, and rural areas in other locations.
Supplementary Materials: The following are available online at www.mdpi.com/xxx/s1, Four figures display warm-cold season odds ratios (ORs) at spatial lags 0, 1, 01 and 04, for each of the four AOD-PM2.5 and PMB concentration level fused surfaces, four respiratory-cardiovascular emergency department and inpatient hospitalizations in all CMAQ 12 km 2 grids, and in grids with and without air monitors: Figure S1, ED asthma ORs and 95% CIs for the four AOD-PM2.5 and PMB fused surfaces during the warm and cold seasons at lag grids 0 (S1a), 1 (S1b), 01 (S1c) and 04 (S1d); Figure S2, IP asthma ORs and 95% CIs for the four AOD-PM2.5 and PMB fused surfaces during the warm and cold seasons at lag grids 0 (S2a), 1 (S2b), 01 (S2c) and 04 (S2d); Figure S3, IP MI ORs and 95% CIs for the four AOD-PM2.5 and PMB fused surfaces during the warm and cold seasons at lag grids 0 (S3a), 1 (S3b), 01 (S3c) and 04 (S3d); Figure S4, IP HF ORs and 95% CIs for the four AOD-PM2.5 and PMB fused surfaces during the warm and cold seasons at lag grids 0 (S4a), 1 (S4b), 01 (S4c) and 04 (S4d).  Institutional Review Board Statement: The Maryland Department of Health Institutional Review Board and the Maryland Health Services Cost Review Commission approved the study's protocol to utilize the State's electronic hospital patient records, ED visits and IP hospitalizations, to complete the analyses described in this paper.
Informed Consent Statement: Not applicable.
Data Availability Statement: The Maryland Health Services Cost Review Commission provided one year's access to confidential hospital data, after payment of the data usage fee. It is not possible to share the electronic patient records since the electronic patient records contain information that could be used to identify the names and addresses of patients in the electronic ED and IP files.
Acknowledgments: Judy Qualters of the U.S. Centers for Disease Control and Prevention (CDC) endorsed and supported this data linkage and analysis project. Fred Dimmick, U.S. Environmental Protection Agency (EPA), facilitated the initial 'proof-of-concept' project for this inter-agency collaboration on this data linkage and analysis project, as well as two prior data linkage studies. Eric S. Hall, EPA, continued the formal interagency partnership between CDC and EPA on this data linkage and analysis project and was the co-developer of the HBM used in the New York City and Baltimore data analysis studies. Michelle Morara, Battelle Memorial Institute, completed the HBM model update from two to three input surfaces.

Conflicts of Interest:
The authors declare no conflict of interest.
Disclaimer: The EPA Office of Research and Development was an active participant in this study. Although the EPA reviewed and approved the final manuscript for publication, the content of this paper, nor the conclusions reached by the authors, may not necessarily reflect the official Agency policy. Mention of trade names or commercial products in this manuscript does not constitute endorsement or recommendation for use by the EPA or the authors.