Long-term Effect of Fine Particulate Matter on Hospitalization with Dementia

Abstract

Background

New evidence suggests that PM_2.5 is associated with dementia (LOD). However, epidemiological studies for the entire population have not been implemented yet.

Methods

We utilized approximately 94 million follow-up records from fee-for-service Medicare records for 13 million Medicare beneficiaries residing in the southeastern U.S. from 2000 to 2013. For exposure to PM_2.5, we used spatially and temporally continuous data that improved those from monitors. To account for time-varying PM_2.5 levels, the Cox model with the Andersen-Gill counting process was applied. Stratified analyses by subtype of dementia and levels of urbanization of residence were performed.

Results

During a median follow-up of 6 years, 1,409,599 hospitalizations for dementia occurred. The adjusted hazard ratio (HR) of hospitalization with dementia was 1.049 (95% confidence interval [CI], 1.048 to 1.051) per 1 μg/m³ increase in annual PM_2.5. In particular, the hazard ratio for vascular dementia (HR, 1.086; 95% CI, 1.082 to 1.090) was high. The magnitude of the effect grew as the level of urbanization increased (HR, 1.036; 95% CI, 1.031 to 1.041 in rural areas versus HR, 1.052; 95% CI, 1.050 to 1.054 in metropolitan areas).

Conclusions

Long-term exposure to higher PM_2.5 was associated with increased hospitalizations with dementia.

Introduction

Among the most potent characteristics of particulate matter with aerodynamic diameter ≤ 2.5 μm (PM_2.5) are its respirable size as well as toxic components from anthropogenic sources. Many studies have shown its harmful effects such as respiratory and cardiovascular diseases (CVD)^1–5.

New evidence suggests that PM_2.5 is also associated with dementia^6–12. The main hypothesis is that PM_2.5 directly or indirectly affects neurodegeneration by inducing inflammation and oxidative stress in the cerebrovascular system. An animal study found that air pollution may alter the blood brain barrier (BBB) permeability¹³ and interact with a genetic trait for Alzheimer’s disease¹⁴. Maher et al.¹¹ identified magnetite nanoparticles in the human brain that precisely match ambient air pollutants. They speculated that particles may enter the brain directly and generate damaging reactive oxygen species (ROS). Cacciottolo et al.¹⁴ reported an interaction between PM_2.5 and Apolipoprotein E (APOE) ε4/4 carriers among aged women. The APOE ε4 allele is known as a major genetic risk factor for Alzheimer’s disease¹⁵, the most common type of dementia¹⁶.

A recent large epidemiological study that includes metropolitan areas in the northeastern U.S has reported that the long-term effects of PM_2.5 increase hospitalization risks of Alzheimer’s disease, Parkinson’s disease, and dementia⁷. However, there have not been studies on the potential harmful effects of long-term exposure to PM_2.5 leading dementia hospitalization in much larger study populations that include medicare participants with a place of residence in un-monitored areas.

Because most older adults live in suburban areas¹⁷, another large-scale epidemiological study that also covers those areas is called for. With data acquistion at the population level for the eldery and spatio-temporal modeling based on satellite images, it has become possible to assess the relationship between PM_2.5 and late-onset dementia (LOD; onset of dementia at the age of 65 and older, hereinafter referred to as dementia).

The southeastern region has one of the highest percentages of aged population in the United States¹⁸. Also, it is one of the regions with the highest dementia prevalence standardized by population distribution, and many states in this region show the highest mortality due to dementia¹⁹.

Thus, we evaluated extensive hospital admission data from Medicare to determine the relationship between long-term exposure to PM_2.5 and the risk of dementia.

Method

Data

Medicare data

We obtained the Medicare Provider and Analysis Review (MedPAR) files for Medicare enrollment and fee-for-service claims for hospitalization from the Research Data Assistance Center (ResDAC). The study population consists of beneficiaries of the Medicare fee-for-service (FFS) plan who were aged 65 or older and resided in the southeastern part of the United States between 2000 and 2013. Specifically, the study area consisted of the seven states of Florida, North Carolina, Georgia, Tennessee, Alabama, South Carolina, and Mississippi.

The variables from data on Medicare beneficiaries that were used for the current study include hospitalization date, age, sex, race, state of residence, ZIP Code of residence, dual eligibility for both Medicare and Medicaid, and the date of death. We did not acquire any human identifiers such as name or Social Security number.

Beneficiaries were identified as having dementia if they claimed their first hospitalization with the diagnosis codes for dementia. All types of diagnosis codes in ICD-9 (the primary, secondary, and other concurrent diagnosis codes up to the 10^th order) were considered. For the list of ICD-9 codes for dementia, we referenced the Chronic Condition Data Warehouse (CCW)²⁰, Centers for Medicare & Medicaid Services (CMS). The included ICD-9 CM codes were as follows: 290.0, 290.20, 290.21, 290.3, 290.40, 290.41, 290.42, 290.43, 291.2, 294.0, 294.1, 294.10, 294.11, 331.0, 331.1, 331.11, 331.19, 331.2, 331.7, and 331.82. It is worth noting that early-onset dementia (EOD; dementia onset before 65 years old) was not of interest in our study²¹. The codes for EOD (290.1 and 290.1x) were also not found in the original dataset

The variables from the above Medicare claims for hospitalization with dementia included patient’s date of admission and diagnosis codes up to the tenth diagnosis. All codes were determined at discharge and in International Classification of Diseases, 9^th Revision, Clinical Modification (ICD-9-CM).

Members who did not have information on sex, race, or locatable Zip Code or had erroneous entries such as hospitalization records after death or who were hospitalized with dementia at study entry were excluded from statistical analysis.

PM_2.5 data

The vast majority of studies that have investigated the effect of PM_2.5 on health have used PM_2.5 measurements from ground monitoring stations as a surrogate for individual exposure in their location. However, PM_2.5 data from monitoring stations have many limitations in terms of both spatial and temporal coverage. Also the absolute number of monitoring stations is few and thus ecological assignment of exposure is hindered by the remote distance between the study subjects and the monitors. Furthermore, the monitors are mainly located in urban areas. Often this has led to rural areas being isolated from the environmental epidemiological studies on air pollution. Another disadvantage is that many ground monitors operate only on every a few days. Therefore, daily readings are not available for many regions. As a result, the use of discrete PM_2.5 data from the exiting monitoring stations has suffered from less statistical power and bias towards the null²².

To overcome these limitations, Lee et al.²³ have produced spatially and temporally continuous PM_2.5 concentrations at 1-km resolution on a daily basis. They calibrated observations from a satellite with a 1-km resolution to PM_2.5 data from ground monitors. Aerosol Optical Depth (AOD) values measured by the Moderate Resolution Imaging Spectroradiometer (MODIS) device aboard the Aqua satellite operated by NASA are considered to be indirect measurements of the ground levels of PM_2.5. Those values were fitted against the ground PM_2.5 levels, incorporating other factors such as meteorological observations and land use that affect the relationship between the AOD values and ground surface level of PM_2.5 from monitoring sites. The modeling has shown high performance (mean from 10-fold cross-validated R² = 0.76 and root- mean-squared prediction errors = 2.74; slopes for the observations and predictions were almost 1, indicating no bias). Further details are described elsewhere²³. We obtained additional data for extended years and reproduced the updated PM_2.5 estimates.

To assign PM_2.5 exposure to the Medicare members, we used the Zip Codes of their mailing addresses. The closest PM_2.5 estimate was linked to the centroid of the Zip Code in the Euclidian geodesic distance. Then the annual averages were calculated for each Zip Code.

We also acquired PM_2.5 data from the monitoring stations operated by the Environmental Protection Agency (EPA). Each Zip Code area was given PM_2.5 observations measured by the closest ground monitor from the zip centroid in the Euclidian distance. The calculation of the annual average PM_2.5 by Zip Code followed.

Geographical Boundaries and Community-level Variables

To locate the beneficiaries’ Zip Codes, the shapefile of Zip Codes from the Esri Data and Maps dataset were used. To determine the level of urbanization of the Zip Code, we acquired the Current Metropolitan Statistical Area/Micropolitan Statistical Area (CBSA) shapefile for the year 2007²⁴. This is one of the TIGER/Line Shapefiles from the Master Address File/Topologically Integrated Geographic Encoding and Referencing (MAF/TIGER) Database (MTDB) provided by the U.S. Census Bureau. It delineates the boundaries of metropolitan statistical areas and micropolitan statistical areas defined by the Office of Management and Budget (OMB). Metropolitan statistical areas are defined as urbanized areas of 50,000 or more population. Micropolitan statistical area are termed urban clusters with populations between 10,000 and 50,000²⁵.

The level of urbanization of the Zip Code was determined according to whether or not the centroid of the Zip Code fell inside the boundary of the metropolitan/micropolitan statistical area.

The Medicare dataset does not provide a variable on individual-level educational attainment. To complement it, we used educational attainment data at the Zip Code level in 2007 through 2011 from the American Community Survey (ACS). The ACS is administered by the U.S. Census Bureau’s Population Estimates Program to provide up-to-date information about ongoing changes in the U.S. population in between the decennial census. It is conducted on a yearly basis to selected households asking different questions from the census. It is the largest household survey in the nation and provides information for policy making by the federal and state governments.

Statistical analyses

We fit extended Cox models to regress the first hospitalization related with dementia against the annual PM_2.5 concentration level while controlling for other covariates.

Our study cohort was defined as Medicare membership between 2000 and 2013 in the southeastern U.S. Therefore, the follow-up started January 1^st, 2000 and a subject entered the cohort during the study period when he or she lived in the area and reached the age of 65 or if they moved into the area after the age of 65. We assumed the start day was the first day of each year (January 1^st). Follow-up was terminated when the first hospitalization with dementia or death occurred, the study period reached the end point (December 31^st, 2013), or beneficiary was lost to follow-up, whichever came first.

Since the nature of exposure of interest varies with time, the dataset was prepared according to the Andersen-Gill counting process²⁶. That is, if the follow-up period was several years, multiple yearly records were created for those subjects until the outcome of interest occurred. Each follow-up record (year) was defined in the interval format that has elapsed days since entry. Annual PM_2.5 levels were matched according to mailing Zip Code and calendar year of follow-up.

In the model, we controlled for age, sex, race, dual eligibility for Medicaid, educational attainment data at the Zip Code level, and U.S. state of address. We stratified analyses by age since it is a most potent risk factor for dementia and is also correlated with the length of the exposure period. Other factors were controlled by adding those variables in the model statement as covariate terms. Medicaid eligibility was used as a surrogate measurement for income, as the eligibility is related mostly to low income. The zip code level education was included as a surrogate measurement for individual-level educational attainment. Meanwhile, it is established that PM_2.5 causes CVD^27–29, which is also associated with an increase in the incidence of dementia^30–32. Thus we did not control for cardiovascular diseases in our model since it is an intermediate variable lying on the causal pathway from PM_2.5 and dementia to avoid overadjustment bias³³.

We assessed the relation between PM_2.5 and dementia by level of urbanization of residence and quantified the effects in rural areas. Using the CBSA map data mentioned in the earlier section, we classified Zip Code numbers into three categories: metropolitan area, micropolitan area, and rural area; and performed separate analyses for each area. In doing so, we also compared the use of monitoring stations and the use of satellite data as an exposure metric to examine how different exposure assessment affect the effect estimates by urbanization.

Various sensitivity analyses were performed in terms of exposure assessment and outcome ascertainment. In the first sensitivity analysis, we limited study subjects only to those who had entered the cohort at the age of 65 not at a later age. Our intention was to prevent left censoring, because we were interested in dementia in the elderly population. This also kept the exposure history constant by preventing entry into the study such as people moving into the study area. In the second sensitivity analysis, we evaluated how results changed when the exposure metric was conventional, namely the use of ground monitoring stations. Lastly, a longer exposure-period of 5 years was examined. We confined the data only for those who appears to the Medicare records at the age of 65 in the study area since 2005. For instance, the hospitalization records in 2005 were given the 5-years average of PM_2.5 since 2000 in the corresponding zip code area.

Data construction and statistical analyses were conducted using ArcGIS 10.5 (Esri, Inc., Redlands, CA, USA), R (R Foundation for Statistical Computing, Vienna, Austria) and SAS 9.4 (SAS Institute, Inc., Cary, NC, USA).

Results

There were 13,374,404 Medicare beneficiaries who resided in the seven states in the Southeastern U.S. (AL, FL, GA, MS, NC, SC, and, TN) from 2000 through 2013. After data refinement, 64,503 people (0.5 % of the original dataset) were excluded from the analysis, and thus follow-up records from 13,309,901 members were used. The median follow-up period was 6 years and the characteristics of the study population are shown in Table 1. The average PM_2.5 concentration was 11.02 μg/m³ with a standard deviation of 2.5 μg/m³ (25^th to 75^th percentile range=8.40 to 14.01 μg/m³). The spatial distribution of PM_2.5 is described in Figure 1.

Table 1.

Characteristics of the Medicare Fee-For-Service Beneficiaries in the Southeastern U.S. 2000–2013.

Characteristic	Statistic
Mean Age, yr (SD)	70.4 (7.3)
Female, No. (%)	7,438,243 (55.9)
Race, No. (%)
White	11,022,085 (82.8)
Black	1,833,340 (13.8)
Other	454,476 (3.4)
Medicaid, No. (%)	1,814,448 (13.6)
Level of Urbanization, No. (%)
Metropolitan	10,243,823 (77.0)
Micropolitan	1,974,696 (14.8)
Rural	1,091,382 (8.2)
State, No. (%)
Florida	5,307,795 (39.9)
North Carolina	2,033,958 (15.3)
Georgia	1,693,701 (12.7)
Tennessee	1,445,879 (10.9)
Alabama	1,118,188 (8.4)
South Carolina	1,044,485 (7.8)
Mississippi	665,895 (5.0)

Abbreviations: yr, year; SD, standard deviation; N, number; m, meter; wk, week

Figure 1.

Spatial distribution of PM_2.5 by Zip Code

In the total follow-period of 94,045,403 person-years, 1,409,599 hospitalizations related to dementia occurred, at an average age of 82.2 years. A total of 3,564,780 deaths occurred from 2000 through 2013 including 374,201 deaths after or upon hospitalization with dementia. Table 2 shows the associated increase in hospitalization with dementia with increased annual PM_2.5 level. The risk of dementia was increased by about 5% per 1 μg/m³ increase in annual PM_2.5 level, (hazard ratio [HR], 1.049; 95% confidence interval [CI], 1.048 to 1.051). The response differed by the sub-category of dementia. The HR of Alzheimer’s disease was 1.060 (95% CI, 1.057 to 1.062) and that of vascular dementia was 1.086 (95% CI, 1.082 to 1.090).

Table 2.

Association between the PM_2.5 and hospitalization with dementia.

	N (%)	Hazard Ratioa (95% CI)
All Dementia	1,409,599 (100)	1.049 (1.048, 1.051)
Alzheimer’s	571,652 (40.6)	1.060 (1.057, 1.062)
Vascular	157,926 (11.2)	1.086 (1.082, 1.090)
Others	680,021 (48.2)	1.045 (1.043, 1.046)

Abbreviations: CI, confidence interval

^aper 1 μg/m³ increase in annual PM_2.5 concentration

Adjusted for age, sex, race, Medicaid eligibility, educational attainment at the Zip Code level, and U.S. state of address

There were differences in relation between PM_2.5 and dementia by level of urbanization (Fig. 2). The hazard ratio of dementia related to annual PM_2.5 increased as the Zip Code area fell in a more urbanized area. The risk was lower in rural Zip Code areas (HR, 1.036; 95% CI, 1.031 to 1.041) than those in metropolitan areas (HR, 1.052; 95% CI, 1.050 to 1.054). Compared to results in a dashed line that used discrete PM_2.5 data from the monitoring stations, hazard ratios showed stronger effects in all regions.

Figure 2.

Difference in PM_2.5 effects by level of urbanization.

Table 3 shows our analyses were robust to the change in exposure metric and more strict definition of study population, and exposure time window. When the records from the ground monitoring stations were used as exposure, the size of the effect decreased but was still statistically significant, without any change in the direction of estimates. When the study population was limited only to those who entered the study at the age of 65 (no late entry), the hazard ratios increased, but the order of magnitude remained the same. When data were confined only to those who entered the study at the age of 65 since 2005 to examine the much longer exposure time window of 5 years, the results rather strengthened. All of the results were statistically significant at the α=0.05 level.

Table 3.

Hazard Ratio^a for hospitalization with dementia by outcome ascertainment, exposure metric, and time window.

	Entry at Age 65	Monitor-based	5 Yearsb
All dementia	1.074 (1.071, 1.077)	1.028 (1.027, 1.029)	1.091 (1.083, 1.099)
Alzheimer’s	1.114 (1.108, 1.120)	1.037 (1.036, 1.039)	1.108 (1.093, 1.124)
Vascular	1.106 (1.098, 1.114)	1.044 (1.041, 1.047)	1.154 (1.134, 1.175)
Others	1.066 (1.063, 1.070)	1.025 (1.024, 1.026)	1.082 (1.074, 1.091)

Abbreviations: CI, confidence interval

^aper 1 μg/m³ increase in annual PM_2.5 concentration

^bFive-years of moving average

Adjusted for age, sex, race, Medicaid eligibility, educational attainment at the Zip Code level, and U.S. state of address

Discussion

In this study, we assessed an association between long term exposure to PM_2.5 and dementia hospitalizations in the Southeastern United States from 2000 through 2013. The extensive and population-level study was enabled by the improved exposure metric, spatiotemporally and temporally continuous data on the PM_2.5 concentration.

As a result, we found that long-term exposure to PM_2.5 is associated with an increase in hospitalization with dementia. Our finding is consistent with previous studies^{7,8,14,34–36}, especially with large-scale studies. In a similar study from the same population in the New England region, Kioumourtzoglou et al.⁷ reported a hazard ratio of dementia as 1.08 (95% CI, 1.05 to 1.11) per 1 μg/m³ increase in annual average PM_2.5. It would be reasonable to compare that result with one for the metropolitan areas that used monitoring stations as an exposure metric to make the results more comparable, which is 1.053 (95% CI, 1.052 to 1.053). Chen et al.³⁴ also analyzed 6.6 million adults and reported the hazard ratio of dementia was 1.07 (95% CI, 1.06 to 1.08) for living within 50 m of a major road compared to more than 300 m distant.

PM_2.5 is thought affect neurodegeneration through various pathways, directly or indirectly. PM_2.5 may penetrate and be deposited in the brain through the olfactory nerves³⁷ and cause defects in the brain directly causing inflmamation³⁸. Also, it is well established that PM_2.5 has detrimental effects on cardiovascular systems^1,39. Thus, it can elevate the risk of dementia by increasing cardiovascular disease^39–41. Vascular dementia showed the higher hazard ratio, implying an effect modified by cardiovascular conditions.

The adverse health association with PM_2.5 increased with level of urbanization. We were also able to quantify this effect in rural areas that have been given less attention. As a Zip Code area is more urban, a stronger PM_2.5 effect was observed. The use of discrete PM_2.5 data from the monitors underestimated the effect of PM_2.5. When they were applied, all of the effect estimates decreased and the difference was most striking in the rural areas. Indeed, the use of continuous PM_2.5 data for exposure outperformed the use of discrete PM_2.5 data and produced stronger effect estimates in all cases. We speculate that this is because the uncertainties between the monitoring stations and the actual residences were considerably resolved by using the satellite data. Thus, subjects were given a more heterogeneous distribution of PM_2.5 than would come from monitoring stations. It is well known that the use of discrete data from ground monitoring stations for exposure in epidemiological studies leads to underestimation of the effect estimates in many cases²². The results from the sensitivity analysis in our study reinforce this theory.

This study has limitations. Medicare records are based on the insurance claims of beneficiaries’ hospitalization, rather than on dementia incidences. That is, patients who were hospitalized for any reason were identified as having dementia once they had any type of ICD-9 codes of dementia. Therefore, the outcome records may not represent all incidences but only capture progressed dementia cases.

Our study population is an open cohort by definition since its membership status may vary by time. Therefore, two possibilities arise. The first is the case where a patient developed dementia before the age of 65. We were limited in our ability to handle issues related with EOD during analyses due to the absence of such data. However, EOD was not our outcome of interest aforementioned since it is perceived separately from LOD⁴².

The second possibility is where a beneficiary older than 65 developed dementia in other areas than the southeast and then joined Medicare. This is expected to more common than the first possibility, since our study cohort allows staggered entry. In that case, our study would be more of a study about prevalence rather than incidence, which would make our inference weaker. This is also an issue in the perspective of exposure assessment, since late enterers’ exposure history and experience may be different from those who were followed in the study area from the start. During the sensitivity analysis restricted to the subset of population that was followed since the age of 65, we tried to minimize these issues.

Medicare membership status was recorded on a yearly basis, which involves tracking to a new Zip Code if a beneficiary had moved within the study area. Therefore, we could update any change in exposure to PM_2.5 accordingly as long as they stayed in the southeastern area.

Another limitation is that exposure was based on Zip Code, not on street address. However, we do not see this as a major concern in our study for two reasons. First, the spatial variation of PM_2.5 is moderate at the local level^43–45. Moreover, urban areas in the southeastern U.S. have spatial uniformity in PM_2.5 concentrations⁴³. Thus the magnitude of error is expected to be low. We are more concerned that exposure to ambient PM_2.5 may be affected as the indoor air pollution level can be modified by ventilation systems or the use of air purifiers. However, non-differential exposure measurement error in this study typically leads to bias towards the null, which attenuates the association^46–48.

Individual-level data on risk factors for dementia such as smoking or alcohol consumption were not available. Although smoking and alcohol consumption themselves are less likely to be directly related to the ambient PM_2.5 levels, they can be associated through other socio-economic status (SES) variables such as income or education. Studies have reported an association between SES and smoking/drinking^49–51 and income may affect residence and neighborhood. We tried to controlled for income using a surrogate measurement, Medicaid eligibility. Higher educational attainment is related with a lower risk of dementia as well⁵². However, we were only able to adjust for it at the community-level due to the absence of such variable at the individual-level.

To the best of our knowledge, this is the first study on the relationship between PM_2.5 and dementia that covers the entire population in the extensive region. In doing so, we confirmed an adverse association and achieved estimation for rural areas with less bias. Further research is needed to identify which components of PM_2.5 are most associated with dementia.