Dissecting the effect of long-term exposure to air pollution on risk of dementia in UK Biobank Authors: Ensor Rafael Palacios, Chin Yang Shapland, Levi John Wolf, Liv Tybjærg Nordestgaard, Emma Anderson, Chloe Slaney, Dan Bernie, Dann Mitchell, Patrick Gavin Kehoe, Gareth James Griffith, Kate Tilling Published: 2025 ABSTRACT Mounting evidence links air pollution to dementia, the most prevalent cause of cognitive impairment in older people. Here we investigated individual and compound effects of particulate matters (PM10, PM2.5, PMcoarse, PMabs) and nitric oxides (NO2, NO) on risk of all-cause dementia, and its most common subtypes, Alzheimer’s disease (AD) and vascular dementia (VAD), using data from UK Biobank. We addressed factors that hinder causal interpretation of associations previously shown in the literature and their translation into clear public health policies. Specifically: 1) spatial confounding by area-level covariates, 2) collinearity among and identification of the most relevant air pollutants, and 3) the time window for pollution exposure. Furthermore, we used chronic obstructive pulmonary disease (COPD) and frequency of oily fish intake in positive and negative control analyses. We found NO2 to be the strongest risk factor for dementia, especially when considering participants with longer permanence at residential address as proxy for longer periods (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 5$$\end{document} years) of exposure (all-cause dementia hazard ratio HR=1.06, 1.02-1.11 per 9.86 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document} interquartile range). There was stronger evidence of an effect on risk for AD than VAD. Positive control analysis did not provide any evidence against causality, although the analyses of spatial confounding and negative control analyses revealed the presence of some residual bias, thus warranting care in the interpretation of the results. Together, our results highlight that targeting air pollution, in particular NO2 levels, could inform preventive public health policies for dementia. Introduction Dementia is a syndrome characterised by the progressive decline of cognitive, emotional, and motor abilities due to accruing brain damage. It eventually impairs a person’s everyday life to a point where extensive support is needed, thereby placing a huge burden on the caregivers and the health system [1]. Currently, there are over 50 million people suffering from dementia worldwide, a number that is predicted to triple by 2050 [1, 2], if the current absence of effective treatments and prevention strategies continues. One proposed modifiable risk factor for dementia is air pollution, carrying a smaller risk for individuals but with higher population prevalence than other factors (e.g., diabetes, obesity, hearing loss) [3]; this, together with the fact that it should be relatively easy to reduce via environmental policies, makes it one of the health hazards with highest public relevance. Air pollution encompasses chemicals, gases, and particles from various natural and anthropogenic sources, including particulate matter (PM) and nitrogen oxides (NOx) [4]. PMs are classified based on their composition (e.g., black carbon and soot, PMabs) and aerodynamic diameter (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\le 10 \mu m$$\end{document}, PM10, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2.5-10\mu m$$\end{document}, PMcoarse, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\le 2.5 \mu m$$\end{document}, PM2.5), with smaller particles able to penetrate deeper into human tissues (e.g., by crossing the blood brain barrier) [5, 6]. NOx includes nitric oxide (NO) and dioxide (NO2), which are highly reactive species that interact primarily with the components of the respiratory system [5]. The main pathways through which air pollution can affect the brain involve increased oxidative stress and inflammatory responses, which trigger cascade effects leading to tissue damage [7–9]. Notably, air pollution may unequally affect dementia subtypes, including Alzheimer’s disease (AD) and vascular dementia (VAD), because their partially distinct underlying neurodegenerative mechanisms (protein structure vs neurovascular alterations) could be differentially susceptible to pollutants toxicity. Growing epidemiological evidence supports an adverse effect of air pollution on risk of dementia [10–13]. In the UK, much work has used the UK Biobank cohort [14–18], an ongoing prospective cohort involving over 500,000 people since 2006 [19, 20]. These studies indicate that PM2.5, PM10, NO2, NOx may be risk factors for dementia, including AD and VAD. However, there remain unanswered questions that need addressing to allow effective translation of findings into public health policies. First, it is unclear the extent of spatial confounding of reported associations due to correlations between air pollution and area-level covariates, including socioeconomic position and noise pollution. Second, given the extended etiological time frame for developing dementia (years to decades) [10, 11], reported associations may be sensitive to the length of the exposure time window considered, but few studies have considered longer windows (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 5$$\end{document} years) [17, 21]. Third, findings are inconsistent with respect to which air pollutant is more relevant, partly due to their high spatial and temporal correlation. Finally, the specific effect of NO has been largely ignored, as studies have usually used NOx instead, combining NO2 and NO. In this study, we aimed to rigorously investigate both individual and compound effects of PM10, PM2.5, PMcoarse, PMabs, NO2, and NO on the risk of all-cause dementia, AD, and VAD. To do so, we triangulated several complementary approaches. We controlled for area-level covariates to minimise bias originating from spatially correlated confounders at a local level. We used variability of detected associations across UK Biobank recruitment centres to assess plausible causality: because neurotoxic effects of air pollution should be fairly constant across locations, high variability would likely indicate the presence of residual bias due to spatially correlated confounders at regional level. We then investigated which pollutants may be more harmful by using variable selection and mixture model methods with multi-pollutant models. With regards to temporal considerations, we explored the potential bias due to the length of exposure to air pollution, by comparing results using 1 or 5 years as minimum exposure time. We further evaluated the risk of bias in our analyses by triangulation with positive and negative control outcome analyses. Lastly we investigated possible interactions between air pollution exposure and socioeconomic position on risk of dementia, to identify potential vulnerable subgroups of people. Methods Study design and population The UK Biobank is an ongoing prospective cohort study involving over 500,000 participants aged 40–69 years recruited between 2006 and 2010 from one of the 22 recruitment centres across England, Wales, and Scotland at baseline [19, 20]. Participants’ sociodemographic, lifestyle, clinical, genetic, and biochemical information has been collected through questionnaires, physical measurements, sampling assays, genotyping, and linked health records at baseline and throughout follow-up time. For more details on this cohort, please see [19, 20]. Dementia cases Dementia cases were algorithmically defined based on the combination of information from baseline self-reports and linked data from hospital admissions and death registries [22]. Thus, dates of dementia incidence span a time period that goes from before baseline up to now. The diagnosis of dementia incidence from linked data was based on the International Classification of Diseases codes (ICD 9/10). The date of diagnosis was equal to the earliest recorded date of dementia occurrence, irrespective of the source. Participant’s outcome could be defined as either AD (n=2679) or VAD (n=1338); all-cause dementia (n=6141) included both, in addition to other dementia types (e.g., unspecified dementia). Because we considered air pollution exposure up to baseline, we excluded participants whose dementia was detected before the baseline assessment (2006-2010) (Fig. 1).Fig. 1Analysis flow chart. Horizontal lines indicate excluded participants Analysis flow chart. Horizontal lines indicate excluded participants Air pollution exposure UK Biobank provides \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$100 m^2$$\end{document} resolution estimates of air pollution exposure at the residential address of participants in 2010, for an area covering up to 400 km from Greater London. We assumed relatively stable pollution concentrations (spatial correlation matrices) over years and considered 2010 data a valid proxy for previous years’ exposures. Violations of this assumption can increase the exposure measurement noise and bias the results in either directions, depending on the underlying mechanism: for example, policies targetting more polluted areas before baseline could lead to underestimate exposure levels at baseline, and thus bias the results towards the null. Air pollutants for which information is present include PM10, PM2.5, PMcoarse, PMabs, NO2 and NO (see Supplementary Information for more detail). Here we encoded air pollution exposure as both continuous variables scaled by interquartile range (IQR), and categorical variables as quartiles (see Table 19 in Supplementary Information). In this study, we considered the individual and compound neurotoxic effects of PM10, PM2.5, PMcoarse, PMabs, NO2, and NO. Recent studies have estimated the effect of the joint pollutants’ exposure using a score defined as the subject-specific average of log hazard ratios (HRs) obtained from single-pollutant models [18, 23], which depends on the set of covariates adjusted for in the single-pollutant analysis. To avoid this dependence, we opted to use principal component analysis (PCA)[16]: specifically, we used the first principal component, accounting for 40% of overall variability in IQR-scaled air pollutant data, as a (unitless) score for the joint exposure to PM10, PM2.5, PMcoarse, PMabs, NO2, and NO across individuals, and called this the ‘global’ air pollution score. Thus, a participant with a low global score could still be greatly exposed to a single pollutant but have a low general level of exposure across pollutants. We did not include additional components because only the first one captured uniform exposure to all pollutants (please see Table 24 in Supplementary Information). As the results from single-pollutant models suggested a negligible effect of PM10 and PMcoarse on risk of dementia, we also performed PCA excluding these two pollutants, thus focusing on the most relevant variability in the data; this ‘restricted’ score was then preferred over the global one in all sensitivity analyses, and unless stated otherwise, pollution score refers to the ‘restricted’ score. The restricted score accounted for 53% of overall variability and was again the only component to capture uniform exposure to all pollutants (please see Table 25 in Supplementary Information). Confounding variables Potential confounders included age, sex, ethnicity, household income, educational attainment, and population density (urban or rural area); these were included in all analyses. We additionally included Index of Multiple Deprivation (IMD), noise pollution and recruitment centres in sensitivity analyses. See Supplementary Information for definition of these variables and Direct Acyclic Graphs (DAG, Figure 1 in Supplementary Information). As the IMD is not comparable across countries (England, Scotland and Wales), we restricted all analyses to participants with the English IMD (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 95\%$$\end{document} of all participants) (Fig. 1). All results are from complete case analyses, that is, participants with missing data in any variables were excluded; see Supplementary Information for more details. Statistical analysis We assumed that putative effects of air pollution exposure during the baseline period are constant over the future risk of dementia; we therefore conducted Cox regression analyses to investigate the association between air pollution and risk of all-cause dementia, AD, and VAD. In the primary analysis, we adjusted for age, sex, ethnicity, household income, educational attainment and population density. For the exposure, in separate models, we used single air pollutants (coded either continuously, with median as reference, or categorically, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1^{st}$$\end{document} quartile as reference) and air pollution score (both ’global’ score and ’restricted’ score). In all sensitivity analyses, we only used the ‘restricted’ air pollution score. In all Cox regression analyses, we used the time from baseline as a time unit, with follow-up period ending with dementia incidence, loss to follow-up, death or end of study, whichever came first. The end of study dates used in these analyses were \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$31^{st}$$\end{document} of October 2022 for England, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$31^{st}$$\end{document} of August 2022 for Scotland and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$31^{st}$$\end{document} of May 2022 for Wales1. We tested the proportionality hazard assumption using Schoenfeld residuals plots, and we found that age, sex, income, PM10, PM2.5, PMabs and NO2 appeared to violate the proportionality assumption (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p<0.05$$\end{document}); however, this was likely due to the large sample size, and the distribution of residuals appeared to be flat over time (Figure 14–17 in Supplementary Information). We conducted the following sensitivity and secondary analyses: Multiple deprivation index Due to correlations in spatial distribution of air pollution and socioeconomic position, it is possible that a backdoor (i.e., noncausal) path between pollution and dementia exists through socioeconomic position. To block this pathway, we additionally adjusted for area-level socioeconomic position, by adjusting for IMD; IMD was included in all subsequent analyses. We used IMD as opposed to the Townsend deprivation index (TDI), which is more widely used in UK Biobank studies. Whereas TDI is a score of material deprivation based on four household variables (unemployment, car and home ownership, household overcrowding) [24], IMD integrates area-level deprivation, at a Lower layer Super Output Area level, in seven distinct domains, including income, employment, health/disability, education/training, barriers to housing/services, environment and crime: thus, IMD can return a much more detailed picture of deprivation2. Importantly, its inclusion did not cause multicollinearity issues (please see Table 26 in Supplementary Information). Length of permanence at baseline residence As the effects of air pollution on the brain are likely cumulative, it is important to take into account the exposure (timeframe) window. We mentioned above that approximating long-term with baseline exposure can lead to bias due to increased measurement noise from changes in pollution levels over years. A second source of measurement noise is individual residence change; notably, baseline exposure differentially (better) captures long-term but not short-term exposure to pollution in participants who lived longer at baseline residence. Consequently, one can ask the question of whether associations differ among recent movers versus long-term residents to gain insights about the importance of the exposure window length. We thus examined whether associations between air pollution and dementia are sensitive to the length of stay at baseline residence by using 1 (1-year cutoff) versus 5 (5-year cutoff) years as minimum length of permanence cutoff. We chose 5 years because it gives a good tradeoff between minimal length of (reported) pollution exposure and number of participants additionally dropped from the analyses (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 15\%$$\end{document}). Exposure to pollutant mixture The PCA analysis could have discarded relevant information present in the exposure data by using only the first component. We thus used a mixture model that directly assesses the effect on dementia risk of simultaneously increasing by one quartile the exposure to all pollutants (see Supplementary Information for more details). For this analysis, we either included all pollutants in the mixture or excluded PM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{10}$$\end{document} and PM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{coarse}$$\end{document}; we also included IMD as a covariate and used 5-year cutoff for time at baseline residence. Variable selection analysis Because it is possible that only a few pollutants may be the driving risk factors for dementia, we conducted multi-pollutant analysis to disentangle individual pollutant associations with risk of developing dementia. We used three methods, namely survival analysis with either Lasso (L1) or elastic net penalisation and weighted quantile sum (WQS) logistic regression. These methods are differently affected by collinearity, which here arises due to spatial correlations between pollutants (see Supplementary Table 20). All methods used penalties or constraint selectively applied to the estimation of air pollutants coefficients, excluding covariates. For this analysis, we used both 1- and 5-years cutoff and adjusted for IMD. For more information, please see Supplementary Information. Recruitment centre analysis Putative causal effects of air pollution on dementia are expected to be fairly constant across locations; alternatively, variability across locations may be due to unaccounted bias. We thus assessed the spatial variability of the air pollution effect on dementia, by including random effects for UK Biobank recruitment centres in the Cox regression, with random effects encompassing a recruitment centre-specific intercept and air pollution slope; in this context, the variability in slope is of particular interest. This analysis was possible because participants appeared to be highly clustered around recruitment centres (Figures 18, 19 in Supplementary Information), which can therefore be used as proxy for geographical location. For this analysis we used both 1- and 5-years cutoff and adjusted for IMD. Positive control analysis We performed positive control analysis with time to COPD incidence as an outcome. This analysis aims at detecting bias in the data, by checking if an existing association could be detected in the population of interest. We chose COPD because there is strong evidence that air pollution is a risk factor for this condition [25], and the set of confounders is likely to be similar as for dementia (see [26]; excluding lifestyle factors as mentioned in our discussion). As for the dementia analysis, we performed Cox regression using the incidence of algorithmically defined COPD cases as an outcome, and excluded participants whose COPD incidence occurred before baseline assessment. For this analysis we used both 1- and 5-years cutoff and adjusted for IMD. Negative control analysis We performed negative control analysis with frequency of oily fish intake as a negative control outcome. This analysis aims to detect the presence of bias in our analysis after controlling for confounders, as any observed association between exposure and outcome would be likely due to selection bias or residual confounding. Fish intake frequency was chosen because it is unlikely to be directly caused by air pollution exposure, while the set of confounders are plausibly the same to those involved for dementia. Oily fish intake frequency was assessed through baseline self-report questionnaires (see Supplementary Information), and was encoded ordinally. We assessed the association between exposure to air pollution and frequency of oily fish intake using ordinal logistic regression, with and without adjustment for IMD. Noise pollution We additionally controlled for noise pollution exposure. Noise and air pollution are likely to be influenced by some common factors, for example vicinity to roads or topography of a city; on the other hand, noise pollution could be a risk factor for dementia incidence [27], and could thus confound the association between air pollution and dementia. For this analysis we used both 1- and 5-years cutoff and adjusted for IMD. Effect modification by IMD From a public health policy standpoint, it is relevant to know whether neurotoxic effects of air pollution change with socioeconomic position. We therefore estimated in a secondary analysis the interactions between air pollution exposure and IMD on the additive and multiplicative scale: as a measure for interactions on the additive scale we used the relative excess risk due to interaction (RERI), whereas for interactions on the multiplicative scale we used the product term in the linear model of the Cox regression [28]. Results Participant characteristics After exclusion criteria were applied, the total number of participants available for the main analysis was 349,809 (Fig. 1). For these participants, Table 1 shows the overall distribution of air pollution exposure quartiles, as well as the distributions for participants grouped by presence or absence of the different dementia outcomes. Similarly, Table 2 shows the characteristics of the same participants in terms of confounders used for the analyses.Table 1Distribution of air pollution exposure quartilesOverallAll-cause dementiaAlzheimer’sVascular dementiaExposureNoYesNoYesNoYes349,809343,6686,141347,1302,679348,4711,338pm2.5 1st25.4%25.5%23.4%25.4%23.4%25.4%23.7% 2nd25.2%25.2%25.5%25.2%25.3%25.2%27.0% 3rd25.3%25.3%26.3%25.3%25.9%25.3%26.0% 4th24.1%24.1%24.9%24.1%25.5%24.1%23.3%pmcoarse 1st25.5%25.5%25.6%25.5%26.7%25.5%24.9% 2nd25.4%25.4%26.4%25.4%25.9%25.4%28.0% 3rd24.9%24.9%24.8%24.9%24.1%24.9%23.2% 4th24.2%24.2%23.3%24.2%23.4%24.2%23.9%pmabs 1st27.0%27.0%25.5%27.0%25.5%27.0%24.6% 2nd24.4%24.4%26.0%24.4%26.2%24.4%26.5% 3rd24.0%24.0%25.5%24.0%25.2%24.0%25.1% 4th24.6%24.6%23.0%24.6%23.1%24.6%23.8%pm10 1st25.4%25.4%24.4%25.4%24.9%25.4%24.1% 2nd25.3%25.2%26.7%25.3%25.7%25.2%28.6% 3rd25.0%25.0%25.7%25.0%26.4%25.0%23.3% 4th24.4%24.4%23.2%24.4%23.1%24.4%23.9%NO2 1st25.8%25.8%23.4%25.8%24.2%25.8%23.0% 2nd25.1%25.1%26.9%25.1%26.6%25.1%27.4% 3rd24.5%24.5%25.5%24.5%24.4%24.5%25.8% 4th24.6%24.6%24.2%24.6%24.7%24.6%23.8%NO 1st25.7%25.8%23.7%25.8%23.4%25.8%21.2% 2nd25.3%25.3%25.2%25.3%25.2%25.3%27.4% 3rd24.6%24.6%25.7%24.6%25.5%24.6%26.0% 4th24.4%24.4%25.3%24.4%25.9%24.4%25.5%Percentage of participants assigned to the 1st, 2nd, 3rd and 4th air pollution exposure quartiles in UK Biobank and within groups with and without all-cause dementia, AD, and VAD. The percentages were computed after applying exclusion criteria shown in Fig. 1Table 2Characteristics of the confoundersOverallAll-cause dementiaAlzheimer’sVascular dementiaCharacteristicsnoyesnoyesnoyes349,809343,6686,141347,1302,679348,4711,338age56.3 (8.1)56.2 (8.0)64.1 (4.8)56.3 (8.1)64.5 (4.3)56.3 (8.1)64.7 (4.2)sex female52.6%52.7%43.9%52.6%48.4%52.6%36.7%ethnicity white95.0%94.9%96.5%95.0%96.9%95.0%96.7% other5.0%5.1%3.5%5.0%3.1%5.0%3.3%education14.2 (4.9)14.2 (4.9)12.4 (5.0)14.2 (4.9)12.2 (5.0)14.2 (4.9)12.0 (4.8)income inc. <18k22.7%22.3%44.7%22.6%44.7%22.6%48.1% inc. 18-31k25.6%25.6%30.7%25.6%32.0%25.6%30.3% inc. 31–5226.1%26.2%16.0%26.1%15.5%26.1%14.9% inc. 52-100k20.2%20.4%7.1%20.3%6.3%20.3%5.7% inc. >100k5.4%5.4%1.5%5.4%1.5%5.4%1.0%population density urban85.6%85.6%87.1%85.6%86.6%85.6%88.2% rural14.4%14.4%12.9%14.4%13.4%14.4%11.8%deprivation (q) 1st25.7%25.7%23.4%25.7%23.0%25.7%23.5% 2nd25.4%25.4%24.1%25.4%25.3%25.4%21.7% 3rd24.8%24.9%23.5%24.8%24.5%24.8%23.3% 4th24.1%24.0%29.0%24.1%27.3%24.1%31.5%noise pollution56.0 (4.3)56.0 (4.3)55.9 (4.2)56.0 (4.3)55.9 (4.2)56.0 (4.3)56.0 (4.4)lost follow-up no99.8%99.8%99.9%99.8%100.0%99.8%99.9%lost death no91.8%92.5%51.4%92.1%53.2%92.0%42.6%Categorical variables are described in terms of percentage, continuous variables are described with their mean and standard deviation, the latter in parenthesis. These characteristics are shown for the UK Biobank population and within groups with and without all-cause dementia, AD, and VAD. The characteristics were computed after applying restriction criteria shown in Fig. 1. The unit for education is years of education, for noise pollution is decibel. Please see Supplementary Information for more information on covariate coding Table 1Distribution of air pollution exposure quartilesOverallAll-cause dementiaAlzheimer’sVascular dementiaExposureNoYesNoYesNoYes349,809343,6686,141347,1302,679348,4711,338pm2.5 1st25.4%25.5%23.4%25.4%23.4%25.4%23.7% 2nd25.2%25.2%25.5%25.2%25.3%25.2%27.0% 3rd25.3%25.3%26.3%25.3%25.9%25.3%26.0% 4th24.1%24.1%24.9%24.1%25.5%24.1%23.3%pmcoarse 1st25.5%25.5%25.6%25.5%26.7%25.5%24.9% 2nd25.4%25.4%26.4%25.4%25.9%25.4%28.0% 3rd24.9%24.9%24.8%24.9%24.1%24.9%23.2% 4th24.2%24.2%23.3%24.2%23.4%24.2%23.9%pmabs 1st27.0%27.0%25.5%27.0%25.5%27.0%24.6% 2nd24.4%24.4%26.0%24.4%26.2%24.4%26.5% 3rd24.0%24.0%25.5%24.0%25.2%24.0%25.1% 4th24.6%24.6%23.0%24.6%23.1%24.6%23.8%pm10 1st25.4%25.4%24.4%25.4%24.9%25.4%24.1% 2nd25.3%25.2%26.7%25.3%25.7%25.2%28.6% 3rd25.0%25.0%25.7%25.0%26.4%25.0%23.3% 4th24.4%24.4%23.2%24.4%23.1%24.4%23.9%NO2 1st25.8%25.8%23.4%25.8%24.2%25.8%23.0% 2nd25.1%25.1%26.9%25.1%26.6%25.1%27.4% 3rd24.5%24.5%25.5%24.5%24.4%24.5%25.8% 4th24.6%24.6%24.2%24.6%24.7%24.6%23.8%NO 1st25.7%25.8%23.7%25.8%23.4%25.8%21.2% 2nd25.3%25.3%25.2%25.3%25.2%25.3%27.4% 3rd24.6%24.6%25.7%24.6%25.5%24.6%26.0% 4th24.4%24.4%25.3%24.4%25.9%24.4%25.5%Percentage of participants assigned to the 1st, 2nd, 3rd and 4th air pollution exposure quartiles in UK Biobank and within groups with and without all-cause dementia, AD, and VAD. The percentages were computed after applying exclusion criteria shown in Fig. 1 Distribution of air pollution exposure quartiles Percentage of participants assigned to the 1st, 2nd, 3rd and 4th air pollution exposure quartiles in UK Biobank and within groups with and without all-cause dementia, AD, and VAD. The percentages were computed after applying exclusion criteria shown in Fig. 1 Table 2Characteristics of the confoundersOverallAll-cause dementiaAlzheimer’sVascular dementiaCharacteristicsnoyesnoyesnoyes349,809343,6686,141347,1302,679348,4711,338age56.3 (8.1)56.2 (8.0)64.1 (4.8)56.3 (8.1)64.5 (4.3)56.3 (8.1)64.7 (4.2)sex female52.6%52.7%43.9%52.6%48.4%52.6%36.7%ethnicity white95.0%94.9%96.5%95.0%96.9%95.0%96.7% other5.0%5.1%3.5%5.0%3.1%5.0%3.3%education14.2 (4.9)14.2 (4.9)12.4 (5.0)14.2 (4.9)12.2 (5.0)14.2 (4.9)12.0 (4.8)income inc. <18k22.7%22.3%44.7%22.6%44.7%22.6%48.1% inc. 18-31k25.6%25.6%30.7%25.6%32.0%25.6%30.3% inc. 31–5226.1%26.2%16.0%26.1%15.5%26.1%14.9% inc. 52-100k20.2%20.4%7.1%20.3%6.3%20.3%5.7% inc. >100k5.4%5.4%1.5%5.4%1.5%5.4%1.0%population density urban85.6%85.6%87.1%85.6%86.6%85.6%88.2% rural14.4%14.4%12.9%14.4%13.4%14.4%11.8%deprivation (q) 1st25.7%25.7%23.4%25.7%23.0%25.7%23.5% 2nd25.4%25.4%24.1%25.4%25.3%25.4%21.7% 3rd24.8%24.9%23.5%24.8%24.5%24.8%23.3% 4th24.1%24.0%29.0%24.1%27.3%24.1%31.5%noise pollution56.0 (4.3)56.0 (4.3)55.9 (4.2)56.0 (4.3)55.9 (4.2)56.0 (4.3)56.0 (4.4)lost follow-up no99.8%99.8%99.9%99.8%100.0%99.8%99.9%lost death no91.8%92.5%51.4%92.1%53.2%92.0%42.6%Categorical variables are described in terms of percentage, continuous variables are described with their mean and standard deviation, the latter in parenthesis. These characteristics are shown for the UK Biobank population and within groups with and without all-cause dementia, AD, and VAD. The characteristics were computed after applying restriction criteria shown in Fig. 1. The unit for education is years of education, for noise pollution is decibel. Please see Supplementary Information for more information on covariate coding Characteristics of the confounders Categorical variables are described in terms of percentage, continuous variables are described with their mean and standard deviation, the latter in parenthesis. These characteristics are shown for the UK Biobank population and within groups with and without all-cause dementia, AD, and VAD. The characteristics were computed after applying restriction criteria shown in Fig. 1. The unit for education is years of education, for noise pollution is decibel. Please see Supplementary Information for more information on covariate coding Air pollution exposure and dementia association For continuously encoded exposures, we found some evidence of a negative effect of PM2.5 (HR=1.06, 1.03–1.10.03.10), PMabs (HR=1.04, 1.01–1.08.01.08), NO2 (HR=1.08, 1.04–1.12.04.12) and NO (HR=1.03, 1.01–1.06.01.06), but not of PMcoarse (HR=0.99, 0.97–1.02.97.02) and PM10 (HR=1.01, 0.99–1.04.99.04), on all-cause dementia risk (Fig. 2). We found similar results, but with stronger evidence, when considering AD as an outcome: PM2.5 (HR=1.10, 1.04–1.15.04.15), PMabs (HR=1.06, 1.02–1.11.02.11), NO2 (HR=1.11, 1.05–1.17.05.17) and NO (HR=1.05, 1.01–1.09.01.09) (Supplementary Figure 2). In contrast, there was little evidence of an effect of any air pollutants on VAD, as HR were closer to 1 and confidence intervals were wider (Supplementary Figure 3). Overall, PM2.5 and NO2 were the pollutants with stronger evidence of a neurotoxic effect: an increase of 1 unit in the IQR scale (1.27 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document} and 9.86 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document}) corresponded to a 6% and 8% increase of risk in all-cause dementia, and a 10% and 11% increase of risk in AD, respectively. For NO2, this is equivalent to a 16% and 22% higher risk of all-cause dementia and AD, respectively, for a person living in large and densely populated city like London compared to small coastal city like Plymouth in Devon (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 19 \mu g/m^3$$\end{document} difference in NO2 concentration3). Evidence was weaker for PMabs and NO, and little for PMcoarse and PM10.Fig. 2Effect of air pollution on all-cause dementia. From left to right, we show the pollution exposure with the associated HR and confidence interval, the forest plot and p-value of the effect estimate (Wald test); ‘iqr’ and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document}q’ refer to the continuous (scaled by IQR) and discrete exposures (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document} quartile), respectively. The models were adjusted for age, sex, ethnicity, educational attainment, income and population density Effect of air pollution on all-cause dementia. From left to right, we show the pollution exposure with the associated HR and confidence interval, the forest plot and p-value of the effect estimate (Wald test); ‘iqr’ and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document}q’ refer to the continuous (scaled by IQR) and discrete exposures (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document} quartile), respectively. The models were adjusted for age, sex, ethnicity, educational attainment, income and population density When using air pollution exposures encoded as quartiles, results were consistent with the above. In general, the effect sizes across pollutants for all-cause dementia and AD showed a dose-response trend, with the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4^{th}$$\end{document} quartile showing the highest concentration of air pollution has the strongest impact on dementia. As for the air pollution score, we found evidence of an adverse effect on all-cause dementia (HR=1.03, 1.02–1.05.02.05, Fig. 2), AD (HR=1.05, 1.02–1.08.02.08, Supplementary Figure 2) but not VAD (HR=1.02, 0.98–1.07.98.07, Supplementary Figure 3). We found similar results when using the global pollution score as the exposure (not shown); however, for both all-cause dementia (HR=1.006, 1.002–1.010.002.010) and AD outcomes (HR=1.009, 1.003–1.014.003.014) evidence was weaker, likely due to the inclusion of PMcoarse and PM10 score. In conclusion, higher joint exposure to (IQR-scaled) PM2.5, PMabs, NO2, and NO appeared to have a risk-inducing effect on all-cause dementia and AD risk, although evidence was weaker than that for individual PM2.5 and NO2 effects. Sensitivity analyses Inclusion of multiple deprivation index Adjusting for multiple deprivation index (IMD) resulted in less evidence of a neurotoxic effect for all pollutants (Fig. 3). Specifically, for continuous pollutant exposures and all-cause dementia, we found little evidence of an association for all air pollutants, except NO2 (HR=1.04, 1.00–1.08.00.08). With AD as an outcome (Supplementary Figure 4), evidence was also reduced (although it was still stronger than for all-cause dementia), such that there was some evidence of an association for PM2.5 (HR=1.07, 1.01–1.13.01.13), NO2 (HR=1.07, 1.01–1.14.01.14) and the pollution score (HR=1.04, 1.01–1.07.01.07). With VAD as an outcome (Supplementary Figure 5), we found little evidence of an association for all air pollutants. A similar pattern was observed using air pollution quartiles as exposure.Fig. 3Effect of air pollution on all-cause dementia. From left to right, we show the pollution exposure with the associated HR and confidence interval, the forest plot and p-value of the effect estimate (Wald test); ‘iqr’ and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document}q’ refer to the continuous (scaled by IQR) and discrete exposures (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document} quartile), respectively. The models were adjusted for age, sex, ethnicity, educational attainment, income, population density and deprivation index Effect of air pollution on all-cause dementia. From left to right, we show the pollution exposure with the associated HR and confidence interval, the forest plot and p-value of the effect estimate (Wald test); ‘iqr’ and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document}q’ refer to the continuous (scaled by IQR) and discrete exposures (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document} quartile), respectively. The models were adjusted for age, sex, ethnicity, educational attainment, income, population density and deprivation index Altogether, IMD seemed to strongly confound the association between air pollution and dementia, such that after adjustment for it only NO2 appeared consistent as a risk factor: an increase by 9.86 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document} in concentration corresponded to a 4% and 7% increase in risk of all-cause dementia and AD, respectively. Length of permanence at baseline residence We then repeated the analysis adjusted for IMD, but selectively including participants who lived \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 5$$\end{document} years (5-year cutoff) instead of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 1$$\end{document} year (1-year cutoff) at baseline residence. We found stronger evidence of a neurotoxic effect for most air pollutants. With all-cause dementia as an outcome (Fig. 4), there was some evidence of a positive association for continuously encoded PM2.5 (HR=1.04, 1.00–1.08.00.08), NO2 (HR=1.06, 1.02–1.11.02.11) and pollution score (HR=1.02, 1.00–1.04.00.04). As before, evidence was stronger with AD as an outcome (Supplementary Figure 6): PM2.5 (HR=1.08, 1.02–1.14.02.14), NO2 (HR=1.09, 1.02–1.16.02.16) and pollution score (HR=1.04, 1.01–1.07.01.07). Again, we found little evidence of an association between any air pollutant and VAD incidence (Supplementary Figure 7). We found analogous results when using air pollution exposure quartiles (Fig. 4, Supplementary Figures 6 and 7).Fig. 4Effect of air pollution on all-cause dementia. From left to right, we show the pollution exposure with the associated HR and confidence interval, the forest plot and p-value of the effect estimate (Wald test); ‘iqr’ and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document}q’ refer to the continuous (scaled by IQR) and discrete exposures (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document} quartile), respectively. The models were adjusted for age, sex, ethnicity, educational attainment, income, population density and deprivation index. Additionally, here we excluded participants who lived less that 5 years (instead of 1 year) at baseline address Effect of air pollution on all-cause dementia. From left to right, we show the pollution exposure with the associated HR and confidence interval, the forest plot and p-value of the effect estimate (Wald test); ‘iqr’ and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document}q’ refer to the continuous (scaled by IQR) and discrete exposures (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n^{th}$$\end{document} quartile), respectively. The models were adjusted for age, sex, ethnicity, educational attainment, income, population density and deprivation index. Additionally, here we excluded participants who lived less that 5 years (instead of 1 year) at baseline address Thus, after accounting for IMD as well as longer exposure time, 1 IQR increase in PM2.5 and NO2 was associated with a 4% and 6% increase in all-cause dementia risk (compared to 3% and 4% with 1-year cutoff), and 8% and 9% increase in AD (compared to 7% and 7% with 1-year cutoff). Exposure to pollutant mixture We used a mixture model to corroborate the results obtained from the air pollution score analysis. The results are displayed in Table 1–3 of the Supplementary Information for all-cause dementia, AD, and VAD, respectively. Overall, we found comparable evidence of an effect on risk of all-cause dementia to that found using single pollutant models. Similarly to above, there seemed to be stronger evidence of an effect when excluding PM10 and PMcoarse from the mixture, and with AD compared to VAD as an outcome. Variable selection analysis The results above point to NO2 and possibly PM2.5 as main risk factors for dementia among air pollutants. To better investigate which air pollutants are stronger risk factors for all-cause dementia, we applied Lasso, elastic net and WQS regression to a multi-pollutant model. The results of these analyses are shown in Table 4 of the Supplementary Information. When applying Lasso penalisation using 5-years cutoff, only NO2 (HR=1.002) appeared to have an effect. The same result (up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$3^{rd}$$\end{document} decimal place) was found with elastic net penalty, indicating that no other pollutant was as strongly associated with dementia as NO2; this result was additionally supported by WQS regression. Altogether, these analyses suggested NO2 as the main risk factor for dementia. Recruitment centre analysis We assessed the consistency of the association of air pollution exposure on all-cause dementia by including pollution exposure as a random effect varying across recruitment centres. The results of this analysis are shown in Figure 8 and Table 3 of the Supplementary Information. The largest average centre-specific HR was reported for NO2 using 5-years cutoff (mean=1.019): although in agreement with the above results, there was also considerable variability across recruitment centres (sd=0.076), suggesting little consistency of the association across England. Notably, only the HR’s for PMcoarse and PM10 were highly concentrated around 1, suggesting no effect on dementia risk. Therefore, this analysis highlighted some difference in the recruitment centre-specific HR’s for most pollutants, potentially indicating the presence of bias at a regional level. Positive control We used COPD as a positive control outcome. Accordingly, we found evidence of an association for all pollutants and COPD, with effect sizes in general much larger compared to those obtained with dementia as an outcome (Figure 9 of the Supplementary Information). Similarly to the dementia results above, adjusting for IMD (Figure 10 of the Supplementary Information) reduced the evidence of an association for all air pollutants; however, in contrast to the dementia analysis, the increase in evidence when using a baseline period \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 5$$\end{document} years was much smaller, if present (Figure 12 of the Supplementary Information). Overall, the positive control analysis lent support to our previous results, but also highlighted a possible difference in terms of etiological time window between dementia and COPD. Negative control We used oily fish intake frequency as a negative control outcome. For this analysis, we focused on the comparison between models adjusted for or not adjusted for IMD. Without adjustment, we found little evidence of an effect of air pollution, although with a perceptible general tendency to increase fish frequency intake (Supplementary 4). Adjusting for IMD slightly increased the association between air pollution and higher oily fish intake for all pollutants, suggesting that including IMD in the analysis may introduce some bias. Control for noise pollution We adjusted for noise pollution in addition to baseline covariates and IMD, using 5-year cutoff. Inclusion of noise pollution did not affect the results (13 in the Supplementary Information). In particular, we found some evidence of an effect of NO2 (HR=1.08, 1.03–1.13.03.13) on all-cause dementia: this effect is comparable to that previously obtained without inclusion of noise pollution (Fig. 4), suggesting no confounding by the latter. Effect modification by deprivation index We examined the modification of the effect of air pollution on dementia by IMD. We report both the effect modification on the additive and multiplicative scale in Tables 5–18 of the Supplementary Information. For all pollutants and both 1- and 5-year cutoff, there was little evidence of effect modification on both scales. Discussion We examined the individual and compound effect of PM2.5, PM10, PMcoarse, PMabs, NO2 and NO on the risk of developing all-cause dementia, AD, and VAD. We addressed several factors hindering both the causal interpretation and translation in public health policies of detected effects using the UK Biobank cohort. We found NO2 to be the strongest risk factor for dementia, especially when focusing on AD as an outcome. Sensitivity analyses highlighted the importance of considering longer exposure periods and accounting for spatial confounding. Together, our results highlight that targeting air pollution, in particular NO2 levels, could inform preventive public health policies for dementia. We found stronger evidence for an effect of PM2.5 and NO2 on all-cause dementia risk, whereas evidence was lacking for PMcoarse and PM10. Similar results were found when treating pollution exposures as IQR-scaled continuous or categorical variables; in the latter case, being in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4^{th}$$\end{document} quartile of the pollution exposure distribution represented generally the greatest risk. Lack of evidence for larger particles (PMcoarse and PM10) is congruent with some [14, 15, 29, 30] but not all [17, 18, 31, 32] UK Biobank studies. Mixed results may be explained by differences in the analyses used, including exposure derivation, outcome definitions (e.g., algorithmically defined vs specific ICD codes), model specifications (e.g., Cox vs accelerated failure time model), and the set of covariates adjusted for: in the present work, we excluded variables (e.g., lifestyle) whose effects on air pollution are uncertain (e.g., smoking, drinking, physical activity), as they may introduce selection bias. Our analyses using air pollution score suggested that combined exposure to PM2.5, PMabs, NO2 and NO does not increase the risk of dementia more than individual pollutant exposures. These results were corroborated by sensitivity analyses using mixture models. Although it is important to examine the health effect of mixtures rather than isolated pollutants [33], these analyses, in the context of dementia incidence, speak against possible super-additive interactions between different air pollutants, at least at a coarse, population level. Instead, it is simpler to explain our results by interpreting the compound effect of pollution as a sum of single pollutants effect. This leads to the question about which air pollutant may be a stronger risk factor for dementia. Variable selection indicated NO2 as the driving factor, suggesting that detected associations for PMabs and NO and possibly PM2.5, may reflect to some extent NO2 concentrations. This would in turn explain the smaller effect size of the pollution score compared to that of NO2: the score discarded some variability associated with NO2 exposure in order to accommodate that from PM2.5, PMabs, and NO. Notice that the same effect reduction should not be expected in the mixture model, which allows for pollutant-specific weights. Compared to all-cause dementia, there was more evidence of a neurotoxic effect for all pollutants, except PM10 and PMcoarse, with AD as an outcome. Notably, this was observed despite the fact that AD cases (n=2,679) were fewer than those for all-cause dementia (n=6,141). On the other hand, there was little evidence of an association with VAD for any pollutants. These results are in line with other studies that found stronger evidence for an effect on AD than on VAD [14, 16, 34], but are also at odds with known mechanistic pathways linking air pollution to cardiovascular disease [35], and cardiovascular disease to dementia [36]. Together, this could suggest that air pollution (or specific pollutants) unevenly promotes different neurodegenerative mechanisms underlying dementia subgroups: for example, it could affect protein structure and functions more strongly and/or faster than the neurovascular system via oxidative stress and inflammation, and this differential effect may be further promoted by co-occurrent health conditions (e.g., comorbidities, behavioural/dietary habits) unevenly distributed across the population. However, we also note that VAD had relatively few cases (n= 1,338) compared to AD, which could bias the analyses. When restricting analyses to participants who lived \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 5$$\end{document} years at baseline residence (as opposed to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ge 1$$\end{document} year), there was more evidence of a neurotoxic effect for pollutants except PM10 and PMcoarse, which is consistent with findings from few other studies testing this [17, 21]. Interestingly, we did not observe the same change when using COPD as a positive control (here serving as a negative control): this could be interpreted as an indication that specific consequences of air pollution on the nervous system may only become apparent on longer time scales, reflecting the extended aetiological time window for dementia. Previous studies, both in the UK Biobank [14, 16, 18, 29, 30, 32], and outside [10, 11] have included participants with shorter periods of known exposure (as extrapolated by time at baseline residence), and this could have biased the results. It may be difficult to identify the optimal temporal restriction criteria to use, as this should account for the trade off with reduced sample size (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sim 15\%$$\end{document} in our sample) and possible selection bias induced. The availability of datasets with increased sample size, information about past address history, and in general explicit measures of long term exposure to pollution are necessary to tackle these questions. Regarding spatial confounding, we tried to assess and account for it at both the local and regional level. When adjusting for IMD, there was less evidence of an association with both all-cause dementia and AD for all air pollutants previously found harmful, to the extent that only increase in NO2 exposure levels appeared consistently harmful. These results 1) hint at the importance of accounting for (local) area-level socioeconomic position to avoid overconfidence in the results, and 2) highlight the robustness of evidence of the role played by NO2 as a risk factor for dementia. When interpreting these results, however, it is important to note that negative control analyses suggested that including IMD, although likely to reduce spatial confounding, may also introduce some additional bias in the results. To check the nature of this bias, further analysis could address the validity of fish consumption as a negative outcome control (e.g., its causal structure), use a different outcome, possibly with time-to-event information, and use a different predictor (i.e., negative exposure control), possibly without such a strong spatial component as air pollution. Furthermore, quantitative bias analysis could also be used to adjust for this bias, although it would require external data to make assumptions about the bias. Finally, we did not find much evidence in support of increased vulnerability to pollution based on socioeconomic position; although these results are plausible, effect modifications may still be present for subgroups of people at particular extremes of the socioeconomic position scale, which may not be well represented in the UK Biobank cohort [37]. At a regional level, we assessed the spatial variability of detected associations across UK Biobank recruitment centres. Using random effect models, we detected negligible variability in the recruitment centre-specific HRs for PMcoarse and PM10, which were tightly clustered around 1, thus providing further evidence of their negligible association with dementia. On the other hand, for all other pollutants and pollution score, variability was an order of magnitude higher, with standard deviations on the HR scale between 3–7%, and several centre-specific HR’s below 1. Whilst we again found the greatest average effect for NO2, this heterogeneity hints at the presence of unmeasured, systematic differences in participants attending various recruitment centres, for example due to regional stratification of socioeconomic and healthcare related variables, or due to different selection bias mechanisms operating across England. There are several limitations that may undermine our conclusions and whose impact could not be tested here. First, outdoor measures of air pollution levels used in this study do not comprehensively capture individual exposure, neither in space nor in time. Indeed, exposure computed at residential address does not account for indoor levels as well as outdoor levels experienced during commuting or at work. Similarly, concentrations for 2010 can only return noisy estimates of long-term exposure because of changes in pollutant concentrations over time, for example, due to the introduction of the clean air zones. Thus, our analysis is possibly biased, with the direction of the bias depending on the mechanisms underlying measurement error. Previous work has found the temporal approximation valid when comparing results using time-varying exposure levels [16]; nonetheless, more precise measures (e.g., using consecutive yearly averages) may improve estimates of the cumulative effects of air pollution on dementia and enable investigation of relevant exposure time windows, as mentioned above. Second, measurement of time of dementia onset can be biased, as it depends on factors such as frequency of use of the healthcare system and presence of comorbidities, which in turn may cause spurious associations. Third, the classification of dementia, even in its major subtypes, is highly problematic and debated; thus, results indicating a differential effect of air pollution on dementia subtypes need to consider these issues. Fourth, our results may have limited generalisability on settings outside high-income countries; this is a common problem of studies on air pollution and dementia, which mostly include people from high-income countries who live in urban areas [38–44], as in our case (see Figures 18–19 in Supplementary Information). Generalisability is problematic if there are unmeasured (or measured with limited support) effect modifiers with different distribution between the analysis sample and the population of interest. Here, living in proximity of urban centres is plausibly associated with the inclusion process as well as socioeconomic position and related lifestyle patterns; then, if these characteristics (e.g. house insulation, mobility and commuting, or healthcare and social infrastructure) modify the effect of ambient air pollution on dementia, results from high socioeconomic settings will not generalise directly to settings with low socioeconomic levels. Fifth, despite efforts to account for possible confounding, caution is required when interpreting pollution data, and in particular NO2 signal, as it may mark a broader urban and traffic mixture rather than a sole causal agent. Despite the above limitations, this work clearly strengthens evidence supporting the role of air pollution, in particular NO2, as a key risk factor for dementia, and advocates for actions aimed at improving air quality management. The estimated NO2 and PM2.5 annual air pollution exposure levels for participants included in this study well exceeded annual air quality guideline levels set by the World Health Organisation (NO2 10 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document}, PM2.5 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document}) [45], with even minimum levels surpassing these thresholds (NO2 12.93\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document}, PM2.5 8.170\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu g/m^3$$\end{document}) (see Table 20 in Supplementary Information); this highlights the importance of decreasing emissions, concentrations and exposure to air pollution [46]. Because the main anthropogenic source of NO2 is fossil fuel combustion for heating, power generation, and vehicle exhaust [47], addressing emissions requires policies promoting the use of cleaner energy sources. At the same time, as most participants lived in or around urban centres, which are the areas with highest pollution concentration (see Figures 18 in Supplementary Information), these results also emphasise the importance of a type of urban planning that, within a general framework of increased quality of life and healthcare, aims at reduced the concentration of and exposure to air pollution, especially NO2. Conclusion The present work adds evidence in support of the role of air pollution in increasing the risk for dementia, while addressing some key questions and issues in the field. We reported neurotoxic effects for various air pollutants, and in particular PM2.5 and NO2, even after comprehensively adjusting for socioeconomic deprivation. Notably, there was strongest evidence for an involvement of NO2, which we propose as a key target for public health policies. Additionally, we found that air pollution was a stronger risk factor for AD than VAD; this differential effect, in turn, warrants more research on the mechanisms underlying neurotoxic effects of air pollution. Given the smaller sample size of VAD (compared to AD and all-cause dementia) studies with larger samples of VAD are also needed. Finally, our results also highlight the strong necessity to consider appropriately extended time windows for air pollution exposure, at least within the range of 5 years, in studies of dementia. Supplementary Information Supplementary Material 1. Supplementary Material 1.