Pollutants emitted by motorized traffic are also emitted by other (combustion) sources. A novel framework to formalize the process of determining whether the air pollution exposure contrast in a study was dominated by traffic, we developed a novel framework [18]. In brief, the framework combined three aspects of TRAP assessment and results from a study had to entail all three aspects to be included: 1) Included studies used measures of defined traffic-related pollutants and/or indirect traffic measures, such as distance to major roads or traffic density. Eligible pollutants were NO₂, NO_x, NO, carbon monoxide (CO), EC (including related metrics such as black carbon, black smoke, and PM absorbance), ultrafine particles (UFP), non-tailpipe PM trace metals [e.g., copper (Cu), iron (Fe) and Zinc (Zn)], polycyclic aromatic hydrocarbons (PAHs), benzene, PM₁₀, PM_2.5 and PM_coarse (Supplementary Table S1). 2) Both the pollution surface and participants’ addresses in the included studies had to meet the framework’s thresholds for spatial resolution (e.g., 5 km grid). 3) Eligible exposure assessment methods included appropriate models or surface monitoring at sufficient spatial resolutions (Supplementary Table S2).

Following this framework, we excluded studies on short-term (minutes to months) effects or self-reported exposures to TRAP. We included studies that assigned individual-level exposure based on models exploiting within-city (i.e., neighborhood) contrasts, that were considered to stem primarily from traffic. Studies that exclusively used between-city contrasts were excluded. In general, the larger the study area, the less likely a measured or modelled contrast in pollution stems primarily from traffic emissions. Therefore, epidemiological studies in larger regions (e.g., state- or country-wide studies) were only included when they adjusted for area in their analysis. PM is generally not specific to traffic. We included results pertaining to PM measures (aerodynamic diameter ≤10 µm [PM₁₀] or ≤2.5 µm [PM_2.5]) in certain settings, e.g., urban areas, so long as they met more stringent requirements for inclusion. For example, PM studies based exclusively on surface monitoring were excluded, but studies using chemical transport models, dispersion models or land-use regression models with a resolution finer or equal to 5 km were included.

To specify how well the studies met the multiple criteria of the exposure framework, we defined an indicator for high traffic specificity based on even stricter criteria. We used this indicator for sensitivity analyses. High traffic specificity was mainly assigned to models with finer resolution (<1 km) or PM models considering only traffic-specific sources/emissions also with a resolution <1 km.

We converted effect estimates for pollutants expressed as ppb or ppm to μg/m³, or mg/m³ using standard WHO scaling factors (standardization of units). For example, 1 ppb NO₂ = 1.88 μg/m³, assuming an ambient pressure of 1 atm and a temperature of 25°C [19]. Effect estimates for black carbon (BC), black smoke (BS) and PM_2.5 absorption (soot) were converted into EC-equivalent estimates [20, 21].

We performed a systematic literature search in PubMed and the specialized LUDOK (Literature database and services on Health Effects of Ambient Air Pollution https://www.swisstph.ch/en/projects/ludok/datenbanksuche/) database matching the PECOS (Population, Exposure, Comparator, Outcome and Study) question [15] for epidemiologic studies:

“In the adult population (P), what is the increase in risk of prevalence and incidence of diabetes (O) per unit increase (C) of long-term exposure to traffic-related air pollution (E), observed in studies relevant for the health outcome and exposure duration of interest (S).”

We searched the databases from 1 January 1980 through 31 July 2019. This end date was chosen a priori for the comprehensive HEI special report comprising dozens of exposures and health outcomes. The search strategy was based on a review protocol developed by the NIEHS OHAT (OHAT) and further refined using a combination of medical subheadings (MeSH) and keywords (Supplementary Table S3). The search strategy was supplemented with hand-searches of references in recent reviews. These were identified by the original search, an additional search in the LUDOK database or individual bibliographic databases curated by HEI and Panel members.

We applied the following inclusion and exclusion criteria according to the predefined PECOS statement. Studies needed to be published in English in a peer-reviewed journal.

We used DistillerSR, a web–based, systematic review software program version 2.29.8 [22], for screening, data extraction and risk of bias assessment. Initial screening based on title and abstract was done by two independent reviewers. Secondary screenings of study eligibility, especially regarding the exposure criterion, were conducted by two independent reviewers based on the full-text, supplements and related exposure assessment papers. At this full-text review stage, the reviewers documented reasons for excluding any given study (Supplementary Table S4). Any disagreement on inclusion was resolved by discussion.

We assessed risk of bias (RoB) in the estimation of all exposure–outcome associations that were included in the meta-analyses. We used a modified version of the tool developed for the risk of bias assessment in systematic reviews for the WHO Air Quality Guidelines [16, 23]. In brief, the risk of bias tool guides the assessment of each study’s potential for bias from six domains and related subdomains of systematic error sources: 1) confounding; 2) selection bias; 3) exposure assessment; 4) outcome measurement; 5) missing data; and 6) selective reporting. Most domains have subdomains. The risk of bias for each subdomain and for each domain overall was given a rating of low, moderate or high. No summary classification was derived across the domains.

We conducted meta-analysis for each exposure-outcome pair where three or more studies reported results; we separately analysed findings from incidence and prevalence studies. Effect estimates from single-pollutant models were selected for the meta-analysis. For presenting results on each pollutant, we applied a uniform pollutant contrast to all contributing estimates and the resulting meta-analytic summary estimate (e.g., RR per 10 μg/m³ increment in NO₂), which necessitated converting some contributing estimates (see Supplementary Eq. S1). We chose the contrast of a given pollutant to reflect a realistic range of exposures in most studies, by using the pollutant concentration increments from a large European ESCAPE study [24]. Meta-analysis was not conducted for the exposure metrics related to distance and density of traffic, because the varying definitions across the studies precluded such analyses. We computed summary effect estimates with random effects models, using restricted maximum likelihood to estimate the between study variance [25]. Random effects models were chosen a priori because of the expected differences in effect estimates related to differences in populations and pollution mixtures. Statistical heterogeneity was assessed using primarily I², where I² values of <50% were interpreted as low; between 50% and 75% as moderate; and >75% as high degree of heterogeneity [26]. The risk estimates hazard ratio (HR), relative risk (RR), incidence rate ratio (IRR) and odds ratio (OR) were considered to approximate the risk ratio [27] and were therefore analysed together as done previously [28]. We use the general term RR to indicate any of the ratio measures.

If a sufficient number of studies were available, we performed additional meta-analyses to assess consistency of the association by: geographic regions; level of risk of bias (selection bias, missing data, confounding, exposure assessment, outcome assessment); smoking adjustment; traffic specificity; and adjustment for the co-exposure noise. All analyses and plots were done with the statistical program R (version 3.6.0), using the libraries “metafor” (v.2.4‐0), “meta,” (v. 4.16‐2), “forestplot” (v.1.10.1), “ggplot” (v. 3.3.3).

We assessed: 1) the quality of the body of evidence using a modified OHAT protocol [17], which itself is based on the GRADE (Grading of Recommendations Assessment, Development and Evaluation) approach; and 2) the confidence in an association between TRAP and diabetes in a “narrative” assessment. These complementary methods are described fully in the HEI Special Report, Additional Materials 5.3 [13]. We also reflect on the confidence assessment in a separate paper (under review).

For studies included in meta-analyses, we conducted the quality assessments separately for each pollutant and study design. Starting with a confidence rating depending on study design (moderate for cohort studies and low for cross-sectional studies), the rating was then downgraded for factors that decrease confidence (high RoB, unexplained inconsistency, imprecision, and publication bias) and upgraded for factors that increase confidence in the body of evidence (monotonic exposure-response, consistency across populations, and consideration of residual confounding). We did not consider the downgrading factor “indirectness” because we included only studies of human exposure to TRAP in direct association with diabetes. Furthermore, we did not use the upgrading factor “large magnitude of effect,” because this factor was unlikely to be meaningful. This a priori decision was based on experiences in the WHO systematic reviews of air pollution, where large or very large effect sizes (i.e., large RR > 2 or very large RR > 5 as defined in OHAT) never occurred [30, 31]. Large RRs were also not observed in our review (Supplementary Figure S1). Next, evaluations per pollutant were combined across study designs, and then across pollutants which was informed by the pollutant with the highest rating.

Since the OHAT assessment is geared toward studies entering a meta-analysis and focusses on the quality of the body of evidence rather than the presence of an association, the Panel also conducted a more inclusive “narrative” assessment. This additionally considered, e.g., pollutants with less than three studies reporting results or those studying indirect traffic measures. While many of the same aspects relevant to evidence synthesis were included in both assessments, there were some subtle differences, most notably regarding the magnitude and direction of the association, and the consistency across pollutants and indirect traffic measures.

In both assessments we rated the level of confidence as high, moderate, low or very low. The two approaches were considered complementary and combined into an overall confidence assessment.

To interpret results of our original review (indicated in tables and figures as “Global 2022”) in light of evidence published after the ending date of this review’s literature search, we repeated the search for eligible studies, starting from June 2019 up to May 2022. Studies identified in this new search were not incorporated into the risk of bias and confidence assessment. However, we incorporated their findings into supplemental meta-analyses to investigate the robustness of our original meta-analytic results to the inclusion of recently published evidence (indicated in tables and figures as “Global 2023”).

The search strategy for all health outcomes considered for the comprehensive review yielded 13,660 unique articles. After initial screening, exclusion of studies not meeting the inclusion criteria, and restricting to articles on diabetes outcomes, we identified 45 studies, 21 of which entered this review after full-text assessment (Table 1, Supplementary Figure S2: PRISMA flow chart). Most studies were excluded, because the spatial scale of the pollution surface or participants’ address did not meet the criteria (Supplementary Table S4).

All studies were published after 2010. Nine studies estimated the association of TRAP with incidence of diabetes, 10 with diabetes prevalence, and two with both incidence and prevalence (the Rome Longitudinal [32] and the SAPALDIA study [33, 34]). The majority of the studies were conducted in Europe (10) or North America (8), followed by China (2) and Australia (1). Three studies were exclusively of women (BWHS [35, 36], SALIA [37], ALSWH [38]). NO₂ or NO_x were the most commonly studied pollutants (17), 11 studies investigated at least one particle metric, and seven included proximity metrics. Exposure levels ranged from very low (e.g., Australia, Canada) to high (e.g., Rome, Italy, China), with ranges in annual means of 5–42 µg NO₂/m³ and 4–25 µg PM_2.5/m³. The 11 cohort studies, all conducted in Europe or North America, included 2,931 to over 1 million participants with a range of follow-up of 4–16 years. The ten cross-sectional studies had 513 up to 2.5 million participants.

Diabetes definitions varied, and included self-report of physician-diagnosed diabetes (five studies), disease registers (two studies), administrative data (e.g., insurance claims) indicating diabetes diagnosis or prescription of hypoglycemic medications (three studies), clinical examinations at study centers, measuring blood glucose (five studies), or using a combination of different data sources (blood glucose measurements, questionnaire, medication, data linkage, six studies). Most smaller cohort studies (n < 10,000 participants) used clinical examinations (SAPALDIA, HNR, MESA, CHAMPIONS) or self-reported physician-diagnosed diabetes, whereas larger administrative cohort or cross-sectional studies typically relied on linkage to administrative databases or registers (e.g., ONPHEC, Rome longitudinal, Table 1).

Meta-analyses indicated positive associations of all traffic-related air pollutants with diabetes incidence and prevalence, though estimates were imprecise (Figure 1). For example, higher exposure to NO₂, the TRAP for which there were the most studies (seven studies), corresponded to higher diabetes prevalence (RR 1.09; 95% CI: 1.02; 1.17 per 10 μg/m³); the individual estimates were highly heterogeneous, especially for the NO₂ results (Figure 2). The association was less pronounced for diabetes incidence (RR 1.04; 95% CI: 0.96; 1.13 per 10 μg/m³; Figure 3). The summary estimates for EC, PM_2.5 and PM₁₀ were also positive but even less precise and based on fewer individual studies.

For pollutants not included in the meta-analyses (such as ultrafine particles PNC or NO, marked in Table 1 without ^c ) elevated risks were observed for measures of NO_x but not the various measures of PM in the prevalence analyses. The incidence analyses showed elevated risks for diabetes with NO and PNC. Notably, the traffic-specific PM_2.5 in the HNR cohort [39] yielded a substantially larger association compared to the total PM_2.5 mass estimates (RR 1.36 vs. 1.03 or 1.05 per 1 μg/m³). All but one study (MESA) showed positive (though imprecise) associations with distance and density of traffic (Table 1, Supplementary Figures S3, S4).

The ONPHEC [40], British Columbia Diabetes Cohort [41], CANHEART [42], and Rome Longitudinal study [32] were considered to have high RoB due to incomplete confounder control (missing adjustment for smoking or socioeconomic status). The SAPALDIA cohort [33, 34] was considered to have high potential for selection bias due to long survival in a cohort before inclusion into the analysis and the 33 CCHS study had extensive missing data [43] (Supplementary Table S5).

In subgroup analyses excluding these studies, association magnitudes were similar or larger (Supplementary Tables S6, S7). For example, restricting to prevalence studies with smoking adjustment eliminated heterogeneity entirely and yielded meta-analytic estimates for NO₂ of 1.09 [95% CI: 1.02; 1.17] (from 1.17 [1.09; 1.25]), and for PM₁₀ of 1.19 [0.87; 1.63] (from 1.43 [1.28; 1.59]).

Five studies evaluated confounding by concurrent noise exposure (British Columbia Diabetes Cohort, Plovdiv Diabetes Survey, both SAPALDIA analyses, Rome longitudinal [32–34, 41, 44], Supplementary Table S8). Most TRAP effect estimates were attenuated upon noise adjustment, but still showed elevated risks. For example, the NO₂ prevalence results in the SAPALDIA study were reduced from 1.21 [1.05; 1.39] to 1.19 [1.03, 1.38] when adjusting for noise [34].

The modified OHAT assessment was conducted for the 16 studies entering meta-analyses (Table 2). Among factors reducing the quality of the evidence, the most common factor was imprecision (wide CI and including unity despite sufficient sample size). For NO₂ and diabetes incidence, the confidence was upgraded due to monotonic exposure-response functions reported in two studies [40, 45]. We upgraded the evidence on NO₂ and prevalence due to potential downward bias. We arrived at a moderate confidence assessment for overall TRAP based on the moderate confidence for NO₂. While the confidence was low for the other pollutants, the associations for these pollutants were suggestive of an association, though imprecise.

A confidence rating of moderate was also reached in the narrative assessment that considered all studies. This rating was based on the meta-analytical evidence of an association of NO₂ with diabetes prevalence and suggestive evidence of an association of NO₂, NO_x, traffic-related PM with incident and prevalent diabetes. The confidence in the evidence was further supported by the monotonic exposure-response relationships reported in two studies, positive albeit imprecise associations involving indirect traffic measures, and numerous positive associations from studies that adjusted for likely confounders. Further, associations generally remained positive after adjustment for noise exposure (Supplementary Table S8). Finally, effect estimates were larger among the subgroup of studies with more extensive confounder adjustment, and among studies that used comprehensive outcome ascertainment methods (versus self-report and administrative data) (Supplementary Tables S6, S7).

Since our systematic search ending in July 2019, new studies have been published on TRAP and diabetes. We extended our search to May 2022 resulting in 304 hits. Five studies met the inclusion criteria (Table 3) adding estimates to all meta-analyses on diabetes incidence and the PM_2.5 prevalence analyses (Supplementary Figures S5–S7). While the pooled estimates did not change dramatically, risk estimates were still elevated and confidence intervals became narrower; especially for the PM_2.5-incidence analyses that was borderline significant (Supplementary Figure S5). Additionally, the Danish study [46] with traffic-specific pollutant estimates and the HNR analysis from 2020 [47] with longer follow-up and refined source-specific exposure assessment as compared to the 2015 analysis [39] showed significantly elevated risks related to traffic-specific NO₂, EC, and PM_2.5. Both also add to the evidence on ultrafine particles. However, measures were not comparable and thus meta-analysis was not possible for the different metrics of UFP. Overall, the results of the HEI 2022 review were strengthened by supplemental analyses of the studies identified in the updated search.

Recent reviews of ambient air pollution—as opposed to our focus on traffic-related air pollution—in association with diabetes found similar results (Supplementary Table S9). With a larger study base, Lui et al. [6] and Yang et al. [5] not only reported significantly elevated risks for diabetes prevalence with NO₂, but also with PM₁₀, and PM_2.5 (for example, including 11 studies vs. 3 studies in the PM_2.5 prevalence analyses). Diabetes incidence risk was significantly elevated with PM_2.5 in both reviews, and additionally with PM₁₀ in the analysis by [5] considering two more studies. As in our analysis, the reviews did not find a significantly elevated risk with NO₂ and diabetes incidence. Effect estimates seemed slightly larger in our prevalence analysis, though more imprecise (for example, 1.09 [1.02; 1.17] vs. 1.05 [1.03; 1.08] and 1.07 [1.04; 1.11]) in the NO₂ prevalence analysis. Another review reported elevated diabetes risks in association with living close to major roads [48].

Plausible pathways regarding how TRAP could lead to diabetes are discussed in the literature. Important mechanisms include oxidative stress induced inflammation leading to endothelial and mitochondrial dysfunction, resulting in impaired insulin signalling and insulin resistance [10]. Animal studies provide evidence that exposure to high concentrations of traffic particles may be a risk factor in the development of diabetes [49–51]. Studies evaluating mechanistic pathways underlying such metabolic perturbations induced by urban PM and near roadway air pollution have identified possible contributory roles played by inflammation and altered fatty acid metabolism. Indeed, Lucht et al. [47] observed that diabetes incidence in an adult population was mediated by markers of inflammation (adiponectin and C-reactive protein). While our results build on evidence found especially for the association with NO₂, mechanistic studies on NO₂ are scarce [52] and NO₂ could be an indicator for other highly correlated pollutants from the same source. However, a recent study on Witstar rats was able to demonstrate reactive oxygen species formation and mitochondrial and endothelial dysfunction after 3 weeks of repeated high NO₂ exposure [53]. Epidemiologic studies also found TRAP-associated higher risks for glucose homeostasis dysregulation measured as insulin concentration in cord blood, fasting blood glucose, insulin sensitivity, HOMA-IR, HbA1c in newborns [54], children [55, 56], adolescents [57], and adults [58] indicating a role of early-life exposure.

The systematic approach to study selection and evaluation using an a priori specified framework for exposure assessment and for a systematic evaluation of the epidemiological evidence are major strengths of this review. Even though none of the pollutants are uniquely traffic-specific, the use of several indicators of TRAP allowed the evaluation of consistency across pollutants and enabled the Panel to base its conclusions on a larger number of studies with diverse exposure metrics. Additionally, the application of two complementary methods (the modified OHAT assessment for studies entering meta-analyses and the narrative assessment considering all studies for the evaluation of the epidemiological evidence maximizes what can be learned from the epidemiologic studies, including evidence from less studied pollutants like UFP and traffic-specific PM fractions.

The overall number of studies per pollutant was small, limiting our ability to conduct meta-analysis or subgroup analysis for some exposure-outcome pairs, and to investigate publication bias.

It has been proposed that effects of air pollutants on the metabolic system commence at an early age [54, 55]. Studies entering this review, including the newest available studies, comprised older adult populations (mean age >50 years) and have excluded persons with already manifest pollutant-dependent diabetes at baseline from the incidence analyses. Thus, a selection bias toward a healthier population might have compromised the ability to study associations with diabetes incidence. The subgroup analysis showed more robust results for studies with low risk of selection bias (Supplementary Table S6).

Another limitation refers to the possible underestimation and misclassification of diabetes. This may depend on the age of the study participants regarding results on incidence of diabetes or on study design and available data sources. Cohort studies with individual data or smaller cross-sectional studies show more rigorous outcome ascertainment with less risk of bias as opposed to the larger studies based on administrative data. Reliance on self-report or documented disease would miss 24% up to 50% of cases depending on the region, while in-depth study center examinations will have a much higher sensitivity due to the long oligosymptomatic prediagnostic phase of diabetes [2]. Non-differential outcome misclassification (independent from exposure status) related to incomplete case ascertainment might bias the results to the null [59, 60]. This was seen for prevalence studies in the sub-group analysis regarding risk of bias due to outcome ascertainment, but not incidence studies (Supplementary Tables S6, S7).

We were not able to distinguish between type 1 and type 2 diabetes. Since 90% of adult diabetes cases are type 2, and the vast majority of incident diabetes cases in adults are type 2 diabetes, we conclude that our results primarily refer to type 2 diabetes.

In cities, where the majority of the world´s population resides, traffic remains an important source of air pollution. The majority of studies were from high-income countries in Europe and North America with generally lower levels of air pollution than in other world regions. However, the one study from China with mean exposure at the higher end of the exposure range (35.3 μg/m³ NO₂) also showed increased risk of diabetes. The available evidence provides overall moderate evidence that TRAP increase diabetes risk. Large studies with rigorous case ascertainment are needed, including in low and middle income countries and other locations with higher exposures. Studies are also needed to assess the change in composition of TRAP due to diesel and gasoline fleet turnover to lower-emission vehicles with a rising share of non-tailpipe emissions in the overall share of traffic-related particulate matter (e.g., from SO₂ emissions). The interplay of TRAP with co-exposures in polluted spaces, most notably noise and green space, needs to be better understood for effective intervention [61].

Studies assessing critical windows of exposure, e.g., in younger populations and preclinical outcomes along the mechanistic path to clinically manifest disease are warranted. Evidence suggests that underlying pathology may be underway as early as childhood and adolescence [62]. Future experimental studies should provide more mechanistic evidence for a better understanding of the molecular and cellular actions of long-term exposure to NO_x and other TRAP on the cardiometabolic system.

In conclusion, we found moderate confidence in the evidence for an association of long-term exposure to traffic-related air pollution and diabetes, with higher effect estimates observed in prevalence studies. We observed increased risks in populations in various geographical regions and contexts and conclude, that TRAP is a risk factor for diabetes.