Traffic-related air pollution and dementia incidence in the Adult Changes in Thought Study

Abstract

Background:

While epidemiologic evidence links higher levels of exposure to fine particulate matter (PM_2.5) to decreased cognitive function, fewer studies have investigated links with traffic-related air pollution (TRAP), and none have examined ultrafine particles (UFP, ≤100 nm) and late-life dementia incidence.

Objective:

To evaluate associations between TRAP exposures (UFP, black carbon [BC], and nitrogen dioxide [NO₂]) and late-life dementia incidence.

Methods:

We ascertained dementia incidence in the Seattle-based Adult Changes in Thought (ACT) prospective cohort study (beginning in 1994) and assessed ten-year average TRAP exposures for each participant based on prediction models derived from an extensive mobile monitoring campaign. We applied Cox proportional hazards models to investigate TRAP exposure and dementia incidence using age as the time axis and further adjusting for sex, self-reported race, calendar year, education, socioeconomic status, PM_2.5, and APOE genotype. We ran sensitivity analyses where we did not adjust for PM_2.5 and other sensitivity and secondary analyses where we adjusted for multiple pollutants, applied alternative exposure models (including total and size-specific UFP), modified the adjustment covariates, used calendar year as the time axis, assessed different exposure periods, dementia subtypes, and others.

Results:

We identified 1,041 incident all-cause dementia cases in 4,283 participants over 37,102 person-years of follow-up. We did not find evidence of a greater hazard of late-life dementia incidence with elevated levels of long-term TRAP exposures. The estimated hazard ratio of all-cause dementia was 0.98 (95 % CI: 0.92–1.05) for every 2000 pt/cm³ increment in UFP, 0.95 (0.89–1.01) for every 100 ng/m³ increment in BC, and 0.96 (0.91–1.02) for every 2 ppb increment in NO₂. These findings were consistent across sensitivity and secondary analyses.

Discussion:

We did not find evidence of a greater hazard of late-life dementia risk with elevated long-term TRAP exposures in this population-based prospective cohort study.

1. Introduction

The social, medical, and economic tolls of Alzheimer’s Disease (AD) and related dementias (ADRD) are a growing concern in our aging population (WHO, 2012). The identification of modifiable risk factors is critical. While extensive evidence links particulate matter air pollution from all sources to adverse health effects (Landrigan et al., 2018; Schraufnagel et al., 2019; US EPA, 2019a), more recent studies highlight the potential adverse cognitive impacts of exposures to traffic-related air pollution (TRAP) (Allen et al., 2017; Calderón-Garcidueñas & Ayala, 2022; Costa, 2017; HEI, 2022; Peters et al., 2019; Weuve et al., 2021). TRAP is a complex mixture of gases and particles that includes ultrafine particles (UFP, ≤100 nm diameter), black carbon (BC), nitrogen dioxide (NO₂), oxides of nitrogen (NOx), and other pollutants.

Toxicological studies have linked TRAP to ADRD hallmarks such as cerebral oxidative stress, inflammation, β-amyloid (Aβ) deposition, hyperphosphorylated tau (ptau), and cerebrovascular pathology (Huuskonen et al., 2021; Moulton & Yang, 2012; Patten et al., 2021; Sahu et al., 2021; Win-Shwe & Fujimaki, 2011). Autopsy, neuroimaging, and biomarker studies in humans have also reported associations between TRAP and Aβ, ptau, pre-tangle material, microvasculature damage, and other pathologies (Alemany et al., 2021; Calderón-Garcidueas et al., 2012; Calderón-Garcidueñas et al., 2004, 2008, 2013; Calderón-Garcidueñas & Ayala, 2022; Du et al., 2022; Hajat et al., 2023; Iaccarino et al., 2021).

Most epidemiologic studies of NO₂, NOx, and dementia incidence have reported hazard ratios (HR) consistent with neutral to positive effects around 0.97–1.20 for 5 ppb increments in NOx/NO₂ (Peters et al., 2019; Yu et al., 2020). A recent meta-analysis of longitudinal cohort studies reported a pooled relative risk (RR) of 1.02 (95 % CI: 0.99, 1.04) for a 5 ppb increment in NO₂ or NOx (Yu et al., 2020). More recent work has reported stronger associations between NO₂ and all-cause as well as vascular/AD dementia than with AD alone (Semmens et al., 2023). Only a few studies have investigated cognition and BC, a significant component of fine particulate matter (PM_2.5) that results from incomplete fuel combustion (US EPA, 2019b). These studies have reported both positive and negative effects (J. Li et al., 2022; Power et al., 2011; Yuchi et al., 2020).

No epidemiologic studies have investigated the association between UFP exposures and late-life dementia incidence. UFPs are challenging to measure, and there are uncertainties as to how to best characterize long-term exposures to these given the lack of long-term or historical UFP data. Furthermore, the variability of instrumentation used across studies has made it challenging to synthesize findings (US EPA, 2019a; WHO, 2021) Still, experimental animal and short-term epidemiological studies indicate that the UFP fraction of PM_2.5 may play an important role in the adverse health effects associated with particulate matter (HEI, 2013; Ohlwein et al., 2019; Schraufnagel, 2020; US EPA, 2019a). Unlike larger particles, UFPs can directly enter the brain via the olfactory bulb (Kilian & Kitazawa, 2018; Win-Shwe & Fujimaki, 2011), they more readily adsorb to toxic substances, are less effectively phagocytized and cleared by alveolar macrophages, can cause a greater inflammatory response in both the lungs and other distal organs, and may induce more oxidative stress (Brown et al., 2001; Donaldson, 2001; Li et al., 2003; Lundborg et al., 2001; Oberdorster et al., 1994; Seaton et al., 1995; Stone et al., 2000). Smaller UFPs may be the most toxic (Han et al., 2023).

The Adult Changes in Thought (ACT) cohort is one of the largest, most extensive community-based brain aging cohorts in the world (Kukull et al., 2002). We previously showed that a 1 μg/m³ increment in the 10-year average PM_2.5 exposure was associated with a 16 % greater hazard of all-cause dementia in this cohort (HR [95 % CI]: 1.16 [1.03–1.31]) (Shaffer et al., 2021). Here, we leverage some of the largest, most spatially rich UFP/TRAP exposure surfaces, which were specifically developed for epidemiologic applications in this cohort, to conduct the first investigation of late-life ADRD incidence and UFP/TRAP exposures.

2. Methods

2.1. The ACT cohort

The ACT study began enrolling volunteers from a random sample of older (65+ yr), cognitively unimpaired individuals based in the greater Seattle area from what was Group Health and is now Kaiser Permanente Washington, an integrated healthcare delivery system. ACT enrollment has had several waves, with the original cohort of 2,581 enrolled from 1994 to 1996, an expansion cohort of 811 enrolled from 2000 to 2003, and continuous enrollment of 10–15 new enrollees per month from 2005 to 2020. Participants are assessed with in-person clinic cognitive testing every two years to identify cases of incident dementia and AD (see 2.2 Outcome Ascertainment) (Kukull et al., 2002; Wang et al., 2006). The study also collects participant demographics, medical history, lifestyle factors, and residential histories. Participants are followed until they meet the criteria for dementia or AD, at which point they are tested the following year to verify clinical diagnosis. Cases are given a diagnosis date associated with their last biennial visit and no longer followed. The same methodological and diagnostic criteria have been used since the cohort inception.

From 1994 through March of 2020, the full cohort included 5,763 individuals over 44,797 person-years. We restricted primary analyses to participants and person-years who were under observation after baseline, had available APOE genotype data, lived within the mobile monitoring modeling region (see 2.3 Exposure Ascertainment), and enrolled prior to 2018 (1994–2017). The last restriction was in response to the interval censoring nature of the study design since participants are only observed every two years. Administratively, this results in a smaller number of person-years at risk and artificially inflated dementia incidence rates during this time. Participants last seen mid 2018, for example, are not due for their cognitive assessment until mid 2020, and thus are censored in mid 2018. Our primary analyses thus consisted of 4,283 (74 %) participants over 37,102 (83 %) person-years that met our inclusion criteria (see Fig. S2 for participant retention details).

Study procedures were approved by the University of Washington and Kaiser Permanente institutional review boards. ACT participants signed informed consent forms.

2.2. Outcome ascertainment

ACT assesses participants during biennial visits using the Cognitive Abilities Screening Instrument (CASI), which combines common screening tests including the Mini-Mental State Examination and the Hasegawa Dementia Rating Scale to quantitatively assess attention, concentration, orientation, short- and long-term memory, language abilities and judgement, among other functions (Teng et al., 2004). Participants with CASI scores lower than 86/100 or those for whom there is clinical concern trigger a full dementia diagnosis evaluation that includes a physical and neurological examination with laboratory testing, imaging and a battery of neuropsychological tests (Teng et al., 2004). Our primary outcome of interest was clinical all-cause dementia, which was based on the Diagnostic and Statistical Manual of Mental Disorders (DSM) – IV criteria (American Psychiatric Association, 1994). Our secondary outcomes of interest were non-AD (vascular, mixed, and other dementias; as determined by DSM-IV) and probable or possible AD based on the National Institute of Neurological and Communicative Disorders and Stroke – Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRA) criteria (McKhann et al., 1984). Since DSM-IV and NINCDS-ADRA diagnostic criteria do not fully line up, non-AD (DSM-IV) and AD (NINCDS-ADRA) cases do not add up to all-cause dementia cases (DSM-IV).

2.3. Exposure Ascertainment

For our primary exposure assessment, we used spatial prediction models derived from a mobile monitoring campaign that involved collecting repeated samples at 309 locations near participant homes (Blanco et al., 2022). Briefly, between March 2019 and March 2020, we sampled for BC (AethLabs MA200), NO₂ (Aerodyne Research Inc. CAPS), and UFP (TSI NanoScan 3910, reported as a total particle number concentration [PNC] for 10–420 nm particles). We used these data with numerous geographic covariates to develop annual average land use regression models. The models performed well with cross-validated MSE-based R² ($R_{CV}^2$) of 0.65 for PNC, 0.60 for BC, and 0.77 for NO₂.

In secondary and sensitivity analyses of UFP exposure, we used UFP models derived from other size fractions and instruments (NanoScan: size-specific 10–420 nm PNC [$R_{CV}^2$: 0.51–0.66]; Testo DiSCmini: 10–700 nm PNC [$R_{CV}^2$: 0.65] and median particle size [$R_{CV}^2$: 0.75]; TSI P-TRAK 8525 with and without a diffusion screen: 36–1000 nm [$R_{CV}^2$: 0.73] and 20–1000 nm [$R_{CV}^2$: 0.76] PNC, respectively), and from on-road data (P-TRAK: 20–1000 nm PNC) that removed the on-road plumes ($R_{CV}^2$: 0.71) (Doubleday et al., 2023).

In sensitivity analyses of NO₂ exposure, we used: a) an alternative, time-varying spatiotemporal (ST) NO₂ model, and b) 2019 (non-time-varying spatial) predictions from this model (Zuidema et al., Under Review). The model is derived from a rich set of data sources, including long-term regulatory monitoring, supplemental monitoring, and low-cost sensor locations, and it features long-term time trends and dimension-reduced land use regression ($R_{CV}^2$: 0.85). A comparable model for PM_2.5 has been extensively used to reliably predict air pollution levels in the Puget Sound (mentioned below) (Keller et al., 2015; Lindström et al., 2014; Shaffer et al., 2021; Szpiro et al., 2010). Predictions from the 2019 spatial model were restricted to the mobile monitoring region so that they would be in alignment with both the spatial and temporal coverage of other TRAP models. The time-varying ST predictions changed over time and were available for a slightly larger area, although the larger spatial coverage had a minimal impact since the majority of ACT residences were within the smaller mobile monitoring region (see Supplemental Material Fig. S1).

We assessed PM_2.5 exposures based on the 2019 (spatial) predictions from a similarly developed ST PM_2.5 model (Bi et al., Under Review; Keller et al., 2015; Lindström et al., 2014; Shaffer et al., 2021; Szpiro et al., 2010). In sensitivity analyses, we used the time-varying ST PM_2.5 exposure predictions from this model, which were available for a slightly larger area (see Fig. S1).

We estimated running 10-year average exposures for each participant based on these prediction models and participant residential histories (see Note S1 for additional details on the mobile monitoring exposure models). While there is no consensus on an ideal exposure assessment period for capturing the effects of air pollution on dementia, we used a ten-year window to capture the long etiologically relevant period for dementia (Jack et al., 2018). We assessed shorter, 1- and 5-year exposure windows in secondary analyses. Residential histories were derived from: 1) billing records that were available for periods of time during which participants were enrolled in Group Health or Kaiser Permanente Washington from 1989 onwards (~57 % of person-months with address histories); 2) ACT study records (33 %); 3) a mailed residential history survey that was distributed to participants that were still enrolled as of 2019 and had significant gaps in their address history (8 %); and 3) a LexisNexus search for participants who exited the study (death, dementia, or another reason) prior to 2019 (2 %). We thus had full address histories for a large fraction of the population dating back to 1989 or earlier. Note S2 details how remaining residential gaps were imputed. Only person-years with addresses within the mobile monitoring region at least 50 % of the time (high exposure coverage) were included (see Fig. S2).

2.4. Statistical analyses

We used the epiR package in R (v. 3.6.2) to estimate all-cause dementia incidence rates (R Core Team, 2023; Stevenson et al., 2022). We fitted Cox proportional hazards regression models using the survival package in R (R Core Team, 2023; Therneau, 2015; Therneau & Grambsch, 2000) to investigate the association between ten-year average TRAP exposure and late-life dementia incidence. We ran separate models for each TRAP exposure (PNC, BC, and NO₂), used age as the time axis, and adjusted for APOE genotype (≥1 ε4 alleles) through stratification of baseline hazard functions since subgroups may have non-proportional hazards for dementia incidence as a function of age (Deary et al., 2002). We adjusted for the following variables in the parametric part of the model (i.e., assumed proportional hazards): sex (male vs. female), self-reported race (White vs. People of Color), ten-year average PM_2.5, calendar year (two-year categories; to account for changes that could have occurred in both air pollution and dementia incidence over time), education, and Neighborhood Disadvantage Index (NDI). NDI was based on a participant’s longest-lived address at or prior to baseline and is a validated indicator composed of six census tract-level variables from the year 2000 American Community Survey including: a) percent of adults 25 years and over with less than a high school education; b) percent of unemployed males; c) percent of households with income below poverty level; d) percent of households receiving public assistance; e) percent of households with children that are headed by an unmarried female; and f) median household income (Miles et al., 2016). NDI is a standardized measure, with values below 0 indicating less disadvantage, and values above 0 indicating greater disadvantage than the national average. Table 1 details how categorical adjustment variables were modeled. The inferential procedures we used accounted for delayed entry and right censoring, relied on the Efron approximation method to handle event ties, and used robust standard error estimates to guard against possible model mis-specification (Kleinbaum & Klein, 2012).

Table 1.

Descriptive statistics for ACT cohort participants included in these analyses. The analytic cohort includes 37,025 person-years from 1994 to 2017 that met our inclusion criteria (see 2 Methods).^a.

	Overall (N = 4,283)
Age at Baseline (Years)
Mean (SD)	75 (6)
Median [Min, Max]	73 [65, 101]
Follow-Up Duration
Mean (SD)	9 (5)
Median [Min, Max]	8 [>0, 24]
Sex
Male	1796 (42 %)
Female	2487 (58 %)
Race
White	3866 (90 %)
People of Color	417 (10 %)
Degree
None	351 (8 %)
GED/High School	1636 (38 %)
Bachelor’s	991 (23 %)
Master’s	660 (15 %)
Doctorate	251 (6 %)
Other	394 (9 %)
Neighborhood Disadvantage Index (NDI)
Mean (SD)	−0.71 (0.70)
Median [Min, Max]	−0.81 [−2.66, 2.64]
APOE Genotype
≥ 1 e4 Allele	1,143 (27 %)
No e4 Allele	3,140 (73 %)
IPW for Modeling Cohort
Mean (SD)	1.00 (0.12)
Median [Min, Max]	0.98 [0.87, 3.79]
All-Cause Dementia Cases
Yes	1,041 (24 %)
No	3,242 (76 %)
Alzheimer’s Disease Dementia Cases
Yes	842 (20 %)
No	3,441 (80 %)
Non-Alzheimer’s Disease Dementia Cases
Yes	388 (9 %)
No	3,895 (91 %)

^aRace is dichotomized due to the small number of People of Color. People of Color includes Black, Asian, American Indian or Alaskan Native, Asian, Native Hawaiian or Pacific Islander, and other/mixed. The sum of AD and non-AD dementia is greater than all-cause dementia due to different ascertainment criteria (see 2 Methods).

We imputed missing covariate values for NDI (~3%), race (~0.16 %), and education (~0.02 %) as the mean of the existing values. For missing APOE genotype data, we used inverse probability weighting (IPW) to address possible selection bias since over 10 % of participants had missing values. IPW assumes that observations are missing at random and worked by upweighting observations that were more likely to have missing APOE genotyping based on other available covariates (e. g., education) and downweighing observations less likely to have missing APOE genotyping. We calculated participant weights by fitting a logistic regression model for APOE genotype availability based on the strongest covariate predictors. Selected covariates were guided by Lasso logistic regression, which used a tuning parameter selected through cross-validation. Model weights were calculated by taking the inverse predicted probability of APOE availability from this procedure and stabilized (to reduce their variability) by multiplying them by the probability of APOE genotype availability given sex status. As a sensitivity analysis, we fitted Cox proportional hazards regression models without IPW.

We ran several other sensitivity analyses. We ran single pollutant models without PM_2.5 adjustment and a four-pollutant model that simultaneously adjusted for PNC, BC, NO₂, and PM_2.5. We ran models that additionally adjusted for age at study entry (five-year age bins) to adjust for potentially different cohort characteristics associated with entering the study dementia-free at younger versus older ages. Instead of using age as the time axis with two-year calendar bin adjustment, we ran models with calendar year as the time axis with two-year age bins. We also ran a restricted cohort analysis with person-years between 2010 and 2017 to limit the temporal extrapolation of our 2019 exposure surfaces. Finally, we ran a non-restricted analysis that included all study years (1994–2020) and did not account for artificially inflated dementia incidence rates two years prior to the data freeze.

3. Results

Table 1 summarizes the characteristics of participants included in our analyses. On average, participants were 75 years old at baseline and followed for 9 years. The majority self-reported White race (90 %) and close to half held at least a bachelor’s degree. The average NDI was –0.71 (less than 0), indicating less disadvantage than the national average. Roughly one quarter had ≥1 APOE ε 4 alleles. One quarter of participants developed all-cause dementia during the follow-up period, with the majority of these cases being due to AD. The overall all-cause dementia incidence rate was 28 (95 % CI: 26–30) cases per 1000 person-years (1,041 cases/37,102 person-years). Participants with higher TRAP exposures, those who were older, female, lower educational attainment, people with ≥1 APOE ε4 allele, and people living in more disadvantaged neighborhoods had higher unadjusted incidence rates (see Fig. S3 for details).

The median (IQR) ten-year average TRAP exposures throughout the study period were 10,227 (9394–11,269) pt/cm³ for PNC, 566 (506–630) ng/m³ for BC, 9.4 (8.5–11) ppb for NO₂ (Table 2). Fig. S4 depicts these exposures as well as those assessed in sensitivity analyses.

Table 2.

Distribution of 10-year average pollutant exposure levels for primary analyses. Exposures are for 37,025 person-years. UFP is defined as the total PNC for 10–420 nm particles from the TSI NanoScan instrument.

Pollutant	Min	Q25	Mean	Median	Q75	Max
PNC (pt/cm3)	4,358	9,394	10,537	10,227	11,269	20,850
BC (ng/m3)	235	506	575	566	630	1,118
NO2 (ppb)	4.1	8.5	9.9	9.4	11	21

We did not find evidence of an association between ten-year average exposure to TRAP and late-life dementia incidence (Fig. 1). The primary, two pollutant model estimated that the hazard of all-cause dementia did not significantly differ for every 2000pt/cm³ increment in PNC (HR, 95 % CI: 0.98, 0.92–1.05), 100 ng/m³ increment in BC (0.95, 0.89–1.01), or 2 ppb increment in NO₂ (0.96, 0.91–1.02). These findings were consistent in single pollutant models (no PM_2.5 adjustment) and the four-pollutant model that simultaneously adjusted for PNC, BC, NO₂, and PM_2.5. We saw similar results in sensitivity analyses when: 1) we used different PNC exposure models (Fig. 2); 2) the cohort was restricted to 2010–2017; 3) we did not temporally restrict the cohort in any way (1994–2020); 4) we adjusted for baseline age; 5) we used calendar year as the time axis with a two-year age adjustment; 6) we did not use IPW; and 7) we assessed NO₂ exposures using 2019 and time-varying spatiotemporal model predictions (Fig. S5) as well as PM_2.5 (Fig. S6). Secondary analyses similarly showed no evidence of an association between: a) all-cause dementia and prior one- or five-year TRAP exposures; or 2) ten-year TRAP exposures and dementia incidence from AD or non-AD (Fig. S7).

Fig. 1.

Estimated hazard ratios for the estimated hazard ratio of all-cause dementia incidence associated with an increment in ten-year average TRAP exposure. Single pollutant Cox proportional hazards models use age as the time axis, are stratified by APOE genotype, and further adjust for sex, race, calendar year, education, and NDI. Two pollutant (primary) models further adjust single pollutant models for PM_2.5. The four-pollutant model includes PNC, BC, NO₂, and PM_2.5 in a single model.

Fig. 2.

UFP sensitivity (total PNC) and secondary (size-specific PNC and size) analyses for the estimated hazard ratio of all-cause dementia incidence associated with an increment in ten-year average UFP (1000pt/cm³) exposure. Cox proportional hazard models are similar to the primary UFP two pollutant models but use exposures from different UFP models.

4. Discussion

We leveraged the long-standing ACT cohort study with spatially rich TRAP exposure surfaces specifically designed for this cohort to evaluate associations between TRAP and incident dementia. We did not observe an association between elevated dementia risk and higher levels of the TRAP measures we assessed (Fig. 1). Sensitivity and secondary analyses produced similar findings, indicating that our results were robust to numerous modeling choices. Further, our TRAP results were consistent when we adjusted our primary TRAP exposure models for spatial (2019) or time-varying (ST) PM_2.5 (Fig. S6), as well as when we additionally adjusted for the remaining two TRAPs in a four-pollutant model (Fig. 1). Unlike these TRAP findings, we estimated a 17 % greater hazard of all-cause dementia (HR [95 % CI]: 1.17 [1.03–1.32]) for a 1 μg/m³ increment in the 10-year average ST PM_2.5 exposure adjusted for PNC. Although not the focus of this analysis, this result is consistent with our previously reported PM_2.5 estimate in a single pollutant model (HR [95 % CI]: 1.16 [1.03–1.31]) (Shaffer et al., 2021), indicating that this estimate is robust and not likely to be confounded by TRAP exposures.

To our knowledge, this is the first study to investigate the association between late-life ADRD incidence and UFP exposures (total and size-specific). UFPs are challenging to measure, and there are uncertainties as to how to best characterize these exposures. In this analysis, our primary UFP model (10–420 nm PNC) followed the WHO’s air quality recommendations to include particles sizes starting at 10 nm or less (WHO, 2021). We conducted sensitivity and secondary analyses, however, to address the large variation of monitoring instruments used across studies (i.e., the technology and its measured particle sizes, e.g., <100 nm, 10 + nm, and 20+ nm particles). We found no meaningful differences in the subsequent health inferences including no strong evidence of a size-dependent effect of UFP on dementia incidence. In fact, there was a slight trend towards higher risk with larger particle sizes (although confidence intervals were overlapping), which is in conflict with evidence suggesting that smaller UFPs may be more toxic (Han et al., 2023). Studies assessing UFP source and composition may be useful as these data become available (Moreno-Ríos et al., 2021).

Our BC findings are in line with the few other existing dementia studies that have reported mixed findings. Yuchi et al. (2020) reported a small elevated risk of AD and non-AD dementia incidence that was consistent with a wide range of effects from protective to harmful for increments in BC exposures (Yuchi et al., 2020). Their study was based on the administrative health data of approximately 678,000 Canadians ages 45–84 years, and exposure was assessed from land use regression models of ground observations. Mortamais et al. (2021) reported similar findings for all-cause, AD, and vascular or mixed dementia incidence in a large French population-based cohort that was evaluated for cognition every two years and assessed for time-varying air pollution exposures using land use regression models (Mortamais et al., 2021). Li et al. (2022), on the other hand, reported a higher risk of dementia (HR: 1.10 [1.09–1.11]) per IQR (0.3 μg/m³) increment in BC (J. Li et al., 2022). Their study used the Medicare data of two million participants ages 65 years or older residing in the northeastern United States and BC models based on satellite imagery, chemical transport modeling, and ground-based observations.

Findings in the literature have been stronger overall for NO₂ and NOx. Two recent meta-analyses of longitudinal cohort studies investigating NO₂, NOx, and dementia incidence have reported a pooled relative risk ratio (RR [95 % CI]) for dementia incidence of 1.02 (0.99, 1.04) for a 5 ppb increment in NO₂ or NOx (Yu et al., 2020), and a HRs of 1.05 (0.99–1.13) and 1.03 (1.00–1.07) for a 10 μg/m³ increment in NOx and NO₂, respectively (Abolhasani et al., 2022). A few large studies, however, have reported more dramatic positive associations. In two population-based cohort studies using the health care data of over two million older Ontario adults, Chen et al. (2017) assessed NO₂ exposure using LUR-satellite models and reported dementia incidence HRs of approximately 1.07–1.10 for a 11.3–14.2 ppb increment in NO₂ (Chen et al., 2017a; Chen et al., 2017b). Carey et al. (2018) also reported a higher risk of dementia based on primary care data of 130,978 adults ages 50–79 living in the Greater London area and NO₂ dispersion models to assess exposure (Carey et al., 2018). They reported a HR of 1.16 (1.05–1.28) for every 7.5 μg/m³ increment in NO₂ exposure. More recently, researchers observed a positive association between NO₂ and all-cause as well as vascular/AD dementia but no association between NO₂ and AD alone in a population of older US adults (Semmens et al., 2023). Differences in population ages, air pollution sources and levels, geographic domains, dementia ascertainment, and modeling might explain these mixed conclusions.

An important strength of this study is that ACT ascertains dementia incidence based on regular in-person assessments, and the study has followed consistent protocols over time. Most studies ascertain dementia from administrative records, which have the potential to introduce a substantial amount of outcome misclassification since dementia is generally underdiagnosed (National Institute of Aging, 2014; Power et al., 2016; Weuve et al., 2021; Wu et al., 2016). Changes in the dementia diagnostic criteria, increases in the number of patients seeking medical attention, and changes in education levels over time further complicate administrative record studies (Wu et al., 2016). Moreover, while there will always be losses due to competing risks in aging cohort (Power et al., 2016), this is less of a concern in ACT. ACT has achieved an impressive Completeness of Follow-up Index (the number of completed participant visits, divided by the number of possible visits) of 94.5 % by giving participants the option of completing clinic, home or, as a last alternative, phone follow-ups (Clark et al., 2002; Crane et al., 2016). Further, the interval censoring nature of this study (participants are observed every two years) likely had a minimal impact on our results. Dementia is generally a gradually progressive syndrome, and that exact timing of incidence is challenging to detect.

We had excellent residential histories, with most being of exact quality, and few had residential gaps. Nonetheless, several methods were used to complete residential histories (see 2.3 Exposure Ascertainment), with the residential history survey and a LexisNexus search possibly being more error prone. Since these may have been linked to participants who were the healthiest (remained in the study to complete the survey) or sickest (had exited the study therefore LexisNexus was used), this may have resulted in some degree of differential misclassification, although these methods only accounted for ~10 % of the available address histories. Overall, these methods allowed us to capture the movement of ACT participants, most of whom moved at least once during the study period, with some moving as many as 17 times. Many studies assess exposure based on baseline address and with the use of cruder spatial models (e.g., from regulatory monitoring sites) (Power et al., 2016; Weuve et al., 2021). Regulatory monitoring sites commonly used in air pollution studies have limited spatial coverage of BC and NO₂, and they do not measure UFPs. We had higher than average spatial resolution exposure surfaces from an innovative mobile monitoring campaign that was specifically designed for this epidemiologic application (Blanco et al., 2022). While we only included person-years with 10-year address histories within this region at least 50 % during any given time, the possibility of selection bias or this resulting in a systematically different populations was small since this only marginally changed the analytic sample (Fig. S2).

We used ambient air pollution levels at residential locations as indicators of personal exposure. Individuals spend the majority of their time at home (Leech et al., 2002). Characterizing outdoor air pollution infiltration and dispersion, indoor sources, and individual behavior (e. g., time activity patterns, including past occupational exposures) has been shown to impact exposure levels to some degree (Allen et al., 2003; Jarvis et al., 2010; Jung et al., 2011; Klepeis et al., 2001; Park & Kwan, 2017; Vardoulakis et al., 2020; Zipprich et al., 2002), but personal exposure sampling to reflect long-term average exposures is unfeasible for large epidemiologic cohort studies. In addition, we extrapolated 2019 mobile monitoring exposure surfaces to estimate long-term exposures prior to enrollment (see Note S1) since historical data were unavailable for UFP, and data for BC and NO₂ had limited spatial resolution. While we adjusted for calendar time such that comparisons were made among people with similar characteristics (e.g., age, sex, SES) during the same calendar year, we assumed that the shape (rank order) of the exposure surfaces was constant over time. Our work and that of others indicates that this assumption may be reasonable since TRAP exposures have decreased over time, but the spatial contrasts have remained similar (Blanco, 2021; Kim et al., 2017; Levy et al., 2015; Meng et al., 2019; Molter et al., 2010; Wang et al., 2011). Moreover, New York State is one of the only places in the United States with continuous, long-term UFP measures since 2009 (NY DEC, 2023; US EPA, 2023). Our ancillary analyses showed stable UFP levels over time for any given site, and concentration differences that were greater across sites at any given year (even for sites separated by <1 km, Fig. S8). These data indicate that capturing the high spatial variability of UFP by measuring many locations (for example with mobile monitoring) is most critical and that our exposure surface can likely be used to characterize levels back to at least the 2000s with good confidence. While restricting our models to more recent time periods (2010–2017) produced similar findings (suggesting that this was not a critical limitation of this work), exposure misclassification was likely higher during earlier years. Moreover, analyses using time-varying NO₂ or PM_2.5 exposures from spatiotemporal models produced similar results to the time-fixed 2019 spatial or mobile monitoring exposure surfaces (Figs. S5, S6). Nonetheless, some degree of exposure measurement error was introduced into these analyses. Future studies should investigate longer-term UFP exposures as continuous, time-varying models become available, as this would be the most robust way to answer this research question. Finally, we noted earlier that there is no consensus on an ideal exposure assessment period for capturing the effects of air pollution on dementia (Jack et al., 2018). We evaluated exposures ten-year prior to dementia incidence given the available data, but it’s possible that earlier and cumulative, life-long exposures may also be relevant (Clifford et al., 2016).

There are a few features of the ACT cohort design that should be considered in the interpretation of these findings. For one, this study is best positioned to characterize the effects of TRAP exposure on late-life dementia incidence (65+). Moreover, although individuals are randomly invited to participate, this cohort may have some volunteer bias and possibly represent “healthy survivors” whose cognition and survival was not affected by air pollution earlier on. If true, this could have dampened the relationships between TRAP and dementia that we observed. Moreover, while we assessed dementia subtypes based on clinical evaluations in secondary analyses, subtype pathologies often overlap, and clinical distinction between pure and mixed pathologies can be challenging (Jack et al., 2018; Kovacs et al., 2008; Nichols et al., 2023; Schneider et al., 2007). Generalizability may be an issue since the demographic composition of the ACT cohort generally reflects the composition of the greater Seattle area (Kukull et al., 2002), which is mostly White, well-educated, and middle income (Table 1). Additional studies in more diverse populations are needed. Similarly, participants reside in the Puget Sound, which is generally characterized by low air pollution levels (Blanco et al., 2022). Combined with limited exposure variability after adjusting for calendar time, this could have widened our HR confidence intervals. Studies with greater exposure variation would be valuable for improved power (i.e., the ability to observe health effects). Finally, our models may have had some residual confounding since we adjusted for neighborhood-level SES but had limited additional adjustment for individual-level SES indicators such as income. Nonetheless, past work has demonstrated that neighborhood SES is often more tightly correlated with air pollution exposures than individual SES (Hajat et al., 2013).

Overall, we did not find evidence of an effect of TRAP exposure, including various UFP measures, on late-life dementia incidence. ACT’s long-standing design and extensive participant information along with our spatially rich exposure surfaces are a strength of this study. Consistency across our primary, sensitivity, and secondary analyses strengthens our conclusions. This is the first study of its kind, and therefore additional studies are needed to replicate this work – especially once longer-term TRAP data become available.

Data availability

The authors do not have permission to share data.