Long-term air pollution exposure and incident dementia in American elderly population: a national cohort study (2000-2018)

Epidemiological evidence suggests air pollution exposure may increase risk of Alzheimer’s 27 disease and related dementias (ADRD). However, previous U.S. studies have predominantly 28 focused on hospitalizations, which fails to fully capture ADRD. Here we constructed two national 29 population-based cohorts of those aged 65 and above from the Medicare Chronic Conditions 30 Warehouse (2000-2018), including doctor visits, to investigate the impact of long-term exposure 31 to ambient fine particulate matter (PM 2.5 ), nitrogen dioxide (NO 2 ), and ozone (O 3 ) on dementia 32 and AD incidence, respectively. We identified ~2.0 million incident dementia cases 33 ( N =12,233,371; dementia cohort) and ~0.8 million incident AD cases ( N =12,456,447; AD 34 cohort). Per interquartile range (IQR) increase in the 5-year average PM 2.5 (3.2 µg/m 3 ), NO 2 35 (11.6 ppb), and warm-season O 3 (5.3 ppb) over the past 5 years prior to diagnosis, the hazard 36 ratios (HRs) were 1.060 (95% confidence interval [CI]: 1.054, 1.066), 1.019 (95% CI: 1.012, 37 1.026), and 0.990 (95% CI: 0.987, 0.993) for incident dementias, and 1.078 (95% CI: 1.070, 38 1.086), 1.031 (95% CI: 1.023, 1.039), and 0.982 (95%CI: 0.977, 0.986) for incident AD, 39 respectively, for the three pollutants. For both outcomes there was strong evidence of linearity in 40 concentration-response relationships for PM 2.5 and NO 2 , suggesting the lack of a clear safe 41 threshold for these health-harmful pollutants. Our study suggests that exposures to PM 2.5 and 42 NO 2 , but not O 3 , may increase the incidence of dementia and AD. Improving air quality may 43 reduce the burden of ADRD and promote healthy aging.

thirds of dementia cases and is the sixth leading cause of death in the United States 2 . In 48 response to this devastating public health threat, the National Alzheimer's Project Act was 49 signed into law to overcome dementia, and the National Plan was launched with Goal 1 aiming 50 to prevent and effectively treat dementia (delay onset, slow progression) by 2025 3 . As there are 51 no disease-modifying treatments for the most common types of dementia, it is a top research 52 priority to identify modifiable risk factors for dementia that can be intervened on at the 53 population level. 54 There is growing evidence associating air pollution with neurodegenerative disease. Alzheimer's disease and related dementias (ADRD). Among them, 5 of 6 showed a positive 57 association between increased exposure to PM2.5 and dementia or AD; 4 of 4 showed an 58 association between NO2 and dementia or AD, while 1 of 3 did so for ozone (O3). Fu and Yung 59 (2020) 5 published a review and meta-analysis of AD and air pollution, and found a 2-fold excess 60 risk of AD for a 10 µg/m 3 increase of PM2.5 among 6 studies, and no increased risk for NO2 in 61 four studies, nor for O3 in three studies. There have been several longitudinal studies since 62 these reviews, with the majority finding positive associations between air pollutants and either 63 dementia or AD 6-14 . A few of these studies examine the associations in US populations, and 64 these studies have almost exclusively used hospitalization as a measure of morbidity 6,7,11,13 . The 65 diagnosis of ADRD, however, likely occurs in doctor visits, and ADRD does not generally result 66 in hospitalizations. Thus, hospitalization records may not well represent either disease incidence 67 or prevalence, and likely leads to an underestimation of the number of cases. In addition, neuropathologic changes are known to occur many years prior to the diagnosis 15 , and the 69 relevant time window in which air pollution might increase the risk of dementia or AD is unclear. 70 To address these knowledge gaps in studying ADRD incidence in the US, here we 71 constructed a national, population-based cohort study from Medicare data to investigate the 72 impact of long-term exposure to PM2.5, NO2, and O3 on dementia and AD incidence. To better 73 approximate disease incidence, we required a 5-year "clean" period without events of interest 74 and used all Medicare claims nationwide (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018), including Medicare inpatient and 75 outpatient claims, carrier file (primarily doctor visits), skilled nursing facility, and home health-76 care claims. We ascertained air pollution based on resident ZIP code, averaged over 5 years 77 prior to diagnosis, which was estimated from national spatiotemporal ensemble exposure 78 models. 79

Results 80
Study population characteristics. Table 1 provides descriptive information on the dementia 81 cohort and AD cohort. Both cohorts were followed after requiring a 5-year period without events 82 of interest to better capture disease incidence. There were 12.2 and 12.4 million people in the 83 dementia and AD cohorts, respectively (Table 1). Most of the studied subjects (78.5% and 84 78.1% for dementia and AD, respectively) entered the cohorts between ages 65 and 74. The 85 median follow-up was 7 years in both cohorts. 16.6% developed dementia (~2.0 million cases), 86 and 6.5% developed AD (~0.8 million cases). More than 90% were not eligible for Medicaid, 87 indicating that most were above the poverty level. A majority of the study population had a 88 comorbidity at some point during follow-up. 89 90 Table 1. Descriptive statistics for the study population   92   93 Note: a means none of the above comorbidities; b presented as mean concentration (interquartile range).

94
Air pollution levels. The average annual level of PM2.5 of cohort participants during the study 95 period, 9.3 µg/m 3 , was below the US EPA standard of 12 µg/m 3 ; NO2 levels were considerably 96 below the EPA annual standard of NO2 of 53 ppb. The annual warm-season average O3 was 97 42.6 ppb. EPA does not have a standard for annual warm-season O3. As a reference, the EPA 98 standard for daily maximum of 8-hour average O3 is 75 ppb. We examined warm-season O3, 99 because O3 is more readily formed in the warm season 16 , and this metric is often used in long-100 term epidemiological studies 17 . Figure 1 shows the distribution of the three pollutants across the 101  Effect modifications. We examined 5 potential effect modifiers (gender, race 154 (white/black/other), Medicaid eligibility, population density (quartiles), and age (<75, ≥75). Figure  155 4 shows hazard ratios in each subgroup, based on the interaction term between PM2.5 or NO2, 156 and the potential effect modifier. Most marked results were seen for an increased hazard of 157 dementia and AD for blacks vs. whites in relation to both PM2.5 and NO2; a similar pattern was 158 found for those eligible for Medicaid. At the same time, those living in the rural areas (i.e. lowest 159 quartile of population density) were found to have a notably lower association between both 160 dementia and AD and both PM2.5 and NO2. All three of these effect modifiers may reflect a 161 general pattern of increased susceptibility to the effects of PM2.5 and NO2 among those of lower 162 Regarding age, those less than 75 had a markedly stronger association between dementia and 165 both PM2.5 and NO2, while the association was stronger between AD and NO2 among those 166 older than 75. Finally, we found little evidence of an interaction between pollution and gender in 167 relation to dementia or AD. Figure S1 shows the results for O3, with relatively few hints of effect 168 modification and all subgroup-specific estimated hazard ratios below one. 169

175
Attributable fraction. The strongest relationship we found with both endpoints was for PM2.5 176 among the three pollutants. If the US PM2.5 levels could be lowered by 3.2 μg/m 3 , which is the 177 IQR, then the attributable fraction of dementia and AD due to current exposure levels, based on 178 our main results from tri-pollutant models assuming a linear relationship, would be 6% and 7% 179 respectively. Namely, an estimated 6% of dementia cases and 7% of AD cases would be 180 avoided if PM2.5 levels decreased by 3.2 μg/m 3 , which is approximately the difference between 181 our large cities like New York and Chicago and smaller cities like Portland, Buffalo, or 182 Baltimore 21 . 183 Sensitivity analysis. Associations between long-term exposure to PM2.5, NO2, O3 and 184 dementia or AD were robust to a series of sensitivity analyses. First, a more strict "clean period" 185 by excluding anyone who had a diagnosis for dementia or AD in their first 10 years of follow-up yielded results similar to the main analyses (Supplementary Table S2). Second, based on this 187 new subcohort (with 10-year clean period), the use of alternative exposure windows (annual 188 exposure 10, 5, 1, or 0 years prior to disease diagnosis, i.e., lags 10, 5, 1, or 0) all support a 189 positive association with PM2.5 and NO2, but not O3, though HRs varied in magnitude 190 (Supplementary Table S2). For both outcomes, associations with PM2.5 and NO2 were 191 attenuated with increasing lag periods. Third, the observed associations with dementia or AD 192 were not mediated by nor modified by comorbidities, such as diabetes, hypertension, stroke, 193 and heart failure (Supplementary Table S3 long-term exposure to PM2.5 and NO2, but not O3, were significantly associated with an 198 increased incidence of dementia and AD; both also suggest that misclassification has somewhat 199 biased our findings to the null (Supplementary Tables S4 and S5). 200

Discussion 201
We found elevated hazard ratios for both dementia and AD in relation to PM2.5, and less 202 markedly to NO2, while hazard ratios for warm-season O3 were not elevated. We did this study 203 in a large US cohort (12 million), with national coverage, and including non-urban areas. For 204 both PM2.5 and NO2, we found a larger effect on AD compared to dementia, which may reflect 205 that fact that dementia includes a wide range of diseases with distinct etiologies, some of which 206 may be unrelated to air pollution, while AD is a subset of dementia and a single disease, for 207 which we found a stronger association. We also found that shorter time windows between 208 exposure (PM2.5 or NO2) and disease showed higher effect estimates, and we posit that it 209 implies an acceleration of an existing process (i.e., accelerating cognitive decline which was 210 already well developed). Moreover, our diagnosis free period requirement provides reasonable assurance that we are looking at incidence, and use of physician's visits, nursing home data, 212 etc. to ascertain diagnosis avoids missing large numbers of cases, possibly not missing at 213 random, which likely occurs in studies using diagnoses based on hospital admission records. 214 Some of our models showed a protective effect of O3. However, when we compare 215 results in Figure 2, we see that in single pollutant models the effect estimate for O3 was null, 216 while in bi-pollutant models with either PM2.5 or NO2, the effect size for O3 was pushed below 217 the null (albeit not significantly) and only in the tri-pollutant model was it protective at the 218 conventional 0.05 level. Moreover, in the bi-pollutant models with O3, the effect sizes for PM2.5 219 and NO2 increased from their level in the single pollutant models. We interpret this as evidence 220 that there is no effect of O3, and the protective effect seen in the tri-pollutant model may be due 221 to collinearity. 222 Our results are broadly consistent with developing literature, which shows relatively 223 consistent effects for PM2.5 and NO2, but less consistent for O3. We observed an HR of 1. in the human brain, indicating that particles from urban air pollution can reach the blood-brain 245 barrier (e.g. through interacting with dysfunctional cell) 31 . 246 Our data suggest that lowering air pollution would have a meaningful reduction on AD and 247 dementia that, when applied to the US population, would be an important tool in the fight against 248 dementia and AD. Assuming these associations are causal, our findings suggest that about 6% 249 of dementia cases and 7% of AD cases would be avoided if PM2.5 levels decreased by 3.2 μg/m 3 . 250 It should be noted thatassuming our findings are generalizable to other parts of the worldthe 251 potential decrease in the burden of AD with lowered air pollution could be greater, considering 252 that the average annual PM2.5 level worldwide in 2015 was estimated at 42 μg/m 3 . 32 253 Our study has several strengths. To our knowledge, this is the first nationwide, 254 population-based cohort study that focuses on the simultaneous health effects of PM2.5, NO2, 255 and O3 on dementia and AD. The large sample size gives us ample power to detect effects even 256 though they are small, which is often the case in environmental studies. Second, the use of 257 Medicare claims data that include doctor's visits rather than restricting the data to 258 hospitalizations is likely to include many more cases, given that many cases are never 259 hospitalized, and also cases which are diagnosed earlier and hence better reflect incidence. 260 Evidence can be found by comparing recent data in another paper about dementia and AD 261 hospitalization in Medicare data, 7 to the data in the current paper. To allow for a fair comparison, 262 we used the same inclusion/exclusion criteria and restricted to the same time period (2000-263 2016) and geographic region (i.e. the lower 48 states), and we found that using just 264 hospitalization missed nearly 90% of dementia cases and 60% of AD cases, compared to using 265 our current data including doctor's visits (Supplementary Table S6). Third, we used a 266 conservative method by requiring a 5-year "clean" period and restricting analysis to subjects 267 with continuous enrollment in Medicare FFS, and Part A (hospital insurance) and Part B 268 (medical insurance) programs throughout the study period, which can ensure that cases were 269 newly diagnosed and thus better approximate incidence. Lastly, we were able to control for a compared Medicare data to clinical diagnoses considered as the gold standard, and found that 284 the sensitivity of dementia was 0.85 but was considerably lower, 0.65, for AD.
We have assumed that outcome misclassification is non-differential (conditionally 286 independent of exposure to air pollutants, conditional on confounders); there are no data 287 indicating otherwise. We have conducted two types of sensitivity analyses to adjust for such 288 misclassification of classifying dementia or AD cases as without dementia or AD (false negative, 289 or 1-sensitivity), and the misclassification of non-dementia, non-AD subjects to one of the 290 diseases (false positive, or 1-specificity). Both these methods of adjustment for false negative 291 and false positives were in agreement that our results were likely to under-estimate the true 292 hazard ratios for PM2.5 and NO2 for both dementia and AD. 293 Another limitation of our study is the potential for exposure error, although the exposure 294 prediction model we used has excellent predictive accuracy 34-36 . Using larger scale ambient air 295 pollutions assigned to individuals has been shown to have a net bias towards the null, 296 consistent with non-differential measurement error, which reflects to some degree classical type 297 of error 37-39 . In addition, our study is subject to unmeasured and residual confounding. While we 298 were able to control for a number of potential confounders at the neighborhood level, we had no 299 individual-level data on SES and education, a limitation implying some mismeasurement of 300 confounders, which may have biased our results (moderately, given that these unmeasured 301 confounders are not likely to act as very strong risk factors for dementia), in an unknown 302 direction. Furthermore, we only studied the Medicare FFS population who enrolled in both Part A 303 and Part B programs, precluding generalizability to the entire US elderly population. 304

Implications for future research 305
Our study provides clear evidence that long-term exposure to PM2.5 mass is significantly 306 associated with increased ADRD incidence and lowering air pollution potentially has an 307 important public health effect. Future studies of air pollution and dementia in low-to-middle-308 income countries (LMIC) of which there are few, will be important. Understanding the potential bias and unmeasured confounding, given the limitations of observational studies, is 310 encouraged. Examining the role of specific pollutant components in ADRD may also be 311 important, because different components of PM2.5 (e.g., metals, elemental carbon, organic 312 carbon, sulfate, and nitrate) and different sources of PM2.5 (e.g., traffic, industrial, cooking, and 313  We created separate datasets for dementia and AD. For each cohort, we further required a "clean" period 332 of 5 years after enrollment, during which there were no diagnosis codes for dementia or AD. By removing 333 potentially prevalent cases in their first five years of follow-up, a diagnosis after that "clean" period more likely approximates "incidence". We considered that 5 years was a reasonable period to ensure that a 335 person truly was not demented prior to the Medicare diagnosis; however, we also explored a 10-year 336 clean period in sensitivity analyses. Therefore, study subjects entered the cohort on January 1st of the 337 year following the "clean" period and were followed until first diagnosis of the outcome of interest across 338 all Medicare claims, death, or end of follow-up. We excluded this 5-year clean period from follow-up time 339 to avoid immortal time bias. This study was approved by the Institutional Review Board of Emory 340 University and a waiver of informed consent was granted.   Individual-level age at entry, sex, race, and Medicaid eligibility were obtained from the Medicare 371 denominator file. We also obtained neighborhood-level covariates in our study. These included ZIP code-372 level SES variables (population density, % Black population, education, median household income, % 373 owner-occupied housing units, and % population above 65 years of age living below the poverty line), 374 county-level behavioral risk factors (smoking rate and body mass index) and health care capacity 375 variables (number of hospitals and active medical doctors), as well as a geographical region. Data were 376 also available for co-morbidities (diabetes, heart failure, stroke, hypertension) in CCW. These covariates 377 have been associated previously with ADRD and may be associated with air pollution, and hence were 378 candidate confounders to be included in models 41,42 .

380
We fit a series of stratified Cox proportional-hazards models with a generalized estimating equation 381 (GEE) 43 to estimate the associations between long-term exposure to PM2.5, NO2, and O3 on dementia or 382 AD among the elderly, where the coefficient for the exposure variable was the parameter of interest, and 383 years of follow-up was the time scale. Specifically, we fit single-pollutant, bi-pollutant, and tri-pollutant 384 models and estimated hazard ratios (HRs) per interquartile-range (IQR) increase in the 5-year average of 385 the annual PM2.5, NO2, and warm-season O3 concentrations in the 5 years prior to diagnosis. All three