Exposures to road traffic, noise, and air pollution as risk factors for type 2 diabetes: A feasibility study in Bulgaria.

Type 2 diabetes mellitus (T2DM) is a growing public health problem in Bulgaria. While individual and lifestyle determinants have been researched; till date there has been no study on environmental risks such as road traffic, noise, and air pollution. As a first step toward designing a large-scale population-based survey, we aimed at exploring the overall associations of prevalent T2DM with exposures to road traffic, noise, and air pollution. A total of 513 residents of Plovdiv city, Bulgaria were recruited. Individual data on self-reported doctor-diagnosed T2DM and confounding factors were linked to objective and self-rated exposure indicators. Logistic and log-link Poisson regressions were conducted. In the fully adjusted logistic models, T2DM was positively associated with exposures to L(den) 71-80 dB (odds ratio (OR) = 4.49, 95% confidence interval (CI): 1.38, 14.68), fine particulate matter (PM) 2.5 25.0-66.8 μg/m 3 (OR = 1.32, 95% CI: 0.28, 6.24), benzo alpha pyrene 6.0-14.02 ng/m 3 (OR = 1.76, 95% CI: 0.52, 5.98) and high road traffic (OR = 1.40, 95% CI: 0.48, 4.07). L(den) remained a significant risk factor in the: Poisson regression model. Other covariates with consistently high multivariate effects were age, gender, body mass index, family history of T2DM, subjective sleep disturbance, and especially bedroom location. We concluded that residential noise exposure might be associated with elevated risk of prevalent T2DM. The inferences made by this research and the lessons learned from its limitations could guide the designing of a longitudinal epidemiological survey in Bulgaria.

Introduction

Diabetes mellitus is a growing public health problem in Bulgaria. According to the International Diabetes Federation, Bulgaria was one the European countries with relatively high prevalence of diabetes — 7.63% of those aged 20-79 years in 2013.[1] In 2012, 9.6% of the population had diabetes, with 26.1% of those still being undiagnosed.[2] That proportion had increased since 2006 when the prevalence was 7.9%.[3] Nearly 90% of all cases in 2012 could be attributed to type 2 diabetes mellitus (T2DM) (International Classification of Diseases (ICD) code — 10: E11). Additionally, 3.7% had prediabetes.[2] Given the low socioeconomic standard in the country and the dysfunctional healthcare system with limited resources hampering the access of patients to quality health services,[4] various preventive measures to reduce the burden of diabetes are gaining more and more prominence. However, lifestyle and genetic factors cannot fully explain the pandemic of cardiometabolic diseases worldwide.[5] While the high prevalence of insufficient physical activity, obesity, and metabolic syndrome in Bulgaria makes healthy lifestyle a logical target for preventive interventions,[6] health-enhancing behaviors will not tackle the risks associated with environmental determinants such as noise and air pollution.

These exposures are believed to have additive but distinct pathways of increasing the risk of T2DM. With respect to air pollution, sufficient data has been accumulated and has been synthetized in recently published meta-analysis.[7] Several biological mechanisms underlying the effect on T2DM have been discussed: Endothelial dysfunction, decreased peripheral glucose uptake, hepatic insulin resistance, systemic inflammation, and oxidative stress.

Conversely, only a few studies have explored the risk associated with noise pollution, their results are conflicting and their methodologies are largely incompatible.[8] Nevertheless, significantly higher risk of T2DM has been reported among those exposed to high residential noise.[8] The detrimental effects of noise are probably mediated through stress response, disruption of normal sleep patterns, and chronic sleep deprivation.[8] To this moment, there has been no previous study in Bulgaria on the environmental risk factors for T2DM. However, given the limited resources for environmental research in the country, and the fact that medical institutions give priority to clinical investigations, prior to designing a large scale longitudinal survey we believe such evidence is warranted, exploring its feasibility, supplementing basic effect size estimates that could be used to determine the required sample size and give us a feel of the anticipated practical importance of our findings; furthermore, possible methodological pitfalls could be highlighted and addressed at this preliminary stage in order to avoid additional costs and sources of bias in a population-based survey. Therefore, we aimed at exploring the overall associations of prevalent T2DM with objective and self-rated indicators of exposure to road traffic, noise, and air pollution in Plovdiv, Bulgaria.

Methods

Design and study population

This was a cross-sectional small scale observational survey. It was carried out in the period July-November, 2014 in the city of Plovdiv, Bulgaria. It is the second largest city in the country with a population of about 341 thousand people and a territory of 101.98 km2, [9] chosen for the present study as it is characterized by noise and air pollution exceeding the acceptable standards.[1011] Likewise, because more than 80% of diabetes-related mortality occurs in low- and middle-income countries,[12] this makes Plovdiv, Bulgaria an appropriate setting for our research.

A questionnaire comprising 59 closed and open-ended questions were administered in order to explore the associations between various neighborhood characteristics and several health endpoints. Only some of these questions were relevant to the objectives of the present study. In order to ensure sufficient sample size, two data collection procedures were employed — a nonprobability snowball sampling and a field sampling. Current residence in Plovdiv, Bulgaria and being ≥18 years of age were mandatory criteria for inclusion in the survey. People reporting hearing loss or uncorrected hearing impairment were excluded.

Snowball sampling started with a small group (n = 57) of acquaintances, colleagues, relatives, and students who were contacted either in person or via e-mail; they filled the questionnaire and were then asked to recruit members of their social network who became first wave participants. Each participant had to recruit up to 10 others to ensure continuity of the procedure. Because snowball sampling is a nonprobability approach and generates biased estimates due to the initial selection of seeds, no more than 50% of the predetermined number of participants were to be collected via this procedure. Simultaneously with it, the lead researcher carried out random field interviews among residents of Plovdiv, Bulgaria. As a first step, 41 neighborhood blocks were selected on an interactive city map. The researcher then visited those neighborhoods starting from the one closest to the center of the city. Participants were approached in local green spaces, at the street, in front of their homes or in local shopping centers, and they were invited to participate in the survey. They were informed that their residential satisfaction was surveyed. The interviews took about 10 min. If, for some reason, the interview was discontinued before completion of the questionnaire, the latter was still coded in the dataset if sufficient data on exposure or outcome variables were provided, or if other variables were available that could be used to impute the missing values (e.g., self-reported instead of objective exposure indicator). No financial incentives were offered.

The necessary sample size was calculated for a multivariate logistic regression model based on the most reliable exposure variable with lowest rate of missing values anticipated, i.e., self-reported traffic intensity (STI), which is considered a proxy for both road traffic air pollution and noise.[13] The sample size necessary to detect an OR of 2.11 for prevalent T2DM among people exposed to “extreme traffic” in comparison to those exposed to “none/very rare traffic” was 558.[14]

A total of 249 questionnaires were distributed via snowball sampling, and 213 were returned completed (response rate = 85.5%); 1,906 people were approached during the field sampling and of those only 368 agreed to participate (response rate = 19.3%). Main reasons for nonparticipation were lack of time and/or fear of revealing personal information.

Because of its noninvasive and observational nature and adult population the study did not undergo approval by an Ethics Committee.[15] All participants were assured of anonymity and participation was voluntary. Answering the questionnaire was taken to constitute informed consent.

Exposure assessment

Each participant answered a question about their address in Plovdiv, Bulgaria. For those reporting an exact address, it was geocoded using BatchGeocode (BatchGeo LLC, 2014)[16] and Google Earth Pro™ v. 7.1 (Google Inc., Mountain View, CA).[17] When, due to privacy and confidentiality, concerns participants had provided only the name of the street they lived at, the addresses were matched to the center point of the street, according to Agay-Shay et al.[18] Addresses not matched automatically or mismatched were coded manually. When possible, the building of residence was identified based on the street name and description of the building's characteristics (number of apartment floor, type of housing, presence of a backyard garden, etc.). Relevant exposure data were assigned after overlaying the air pollution and traffic noise maps onto the map of geocoded addresses.

Noise exposure

Road traffic is the dominant transportation source of noise in Plovdiv, Bulgaria with railway, aircraft, and industrial noise levels being less influential. Lden — defined as “average” noise levels during daytime, evening, and nighttime, applying a 5-dB penalty to noise in the evening and a 10-dB penalty to noise during the night — was chosen for a noise indicator. Combined Lden data from all traffic sources were elicited from the official strategic noise maps (10 m × 10 m grid, 4 m height) of Plovdiv, Bulgaria (http://www.webcitation.org/6biU8dDj0) created in 2009 in compliance with the Environmental Noise Directive 2002/49/EC.[19] Noise levels were assessed at the coordinates of the residential address unless the participant had indicated the orientation of the apartment. Lden was modelled with LimA v. 5 (Brüel and Kjær, Nærum, Denmark). Further description of the noise mapping process is available elsewhere.[20]

Noise annoyance was derived from an 11-point scale: “To what extent are you disturbed, annoyed, or irritated by noise when you are at home?” (“0, not at all” to “10, very much”). Lifetime exposure to loud noise was measured using two questions: “For how long during your lifetime have you lived in a place where noise was loud enough to disturb normal conversations?” and “For how long during your lifetime have you worked at a place where noise was loud enough to disturb normal conversations?” The need to speak in raised voices might be considered a proxy for exposure to at least 66 dB.[21]

Air pollution exposure

The objective indicators for air pollution in the neighborhood that we had information on, were average annual all-source fine particulate matter (PM)2.5 levels, a pollutant derived from fossil-fuel combustion, and average annual all-source benzo alpha pyrene (BaP) levels, a polycyclic aromatic hydrocarbon formed in the process of incomplete combustion of organic material. Due to insufficient resources, traffic count data could not be collected for each residential street, nor could we take field samples at residential addresses. Therefore, information on calculated levels was extracted from an official municipality source reporting PM2.5 and BaP pollution maps.[22] (http://www.webcitation.org/6biSbgjwd) Pollutant dispersion modelling was done for 2011 with SELMAGIS 9.28 (Lohmeyer GmbH and Co. KG). Available data on PM2.5 levels had been categorized: “0.0-17.5 μg/m3”, “17.5-20.3 μg/m3”, “20.3-25.0 μg/m3”, “25.0-40.0 μg/m3”, and “40.0-66.8 μg/m3”; and for BaP levels as follows: “0.0-1.0” ng/m3, “1.0-3.75” ng/m3, “3.75-6.0” ng/m3, “6.0-9.75” ng/m3, “9.75-14.02” ng/m3. [22] The spatial resolution of the visualized pollutant levels (≈400 m × 400 m) was lower than the noise map resolution, therefore, PM2.5 and BaP levels were still assigned to the participants even if they had reported only the name of their residential neighborhood, provided that exposure levels did not vary within the whole neighborhood.

The effect of subjective all-source air pollution in the neighborhood was assessed in order to capture a wider range of air pollutants and psychophysical aspects of air pollution. People's self-rated air pollution has its place in epidemiology when adequate assessment of long-term exposure is not possible.[13] Moreover, some of the deleterious effects of environmental pollution on health could be mediated through psychological mechanisms.[2324] Perceived air pollution was derived from three 11-point-scale-based questions: “How would you rate the air quality in your neighborhood?” (“0, very low” to “10, very high”, reverse coded question); “According to you, how severe is the air pollution in your neighborhood?” (“0, not at all” to “10, extremely”); and “To what extent are you annoyed by the air pollution in your neighborhood?” (“0, not at all” to “10, very much”).

Traffic exposure

We anticipated high percentage of missing values on the objective noise exposure and air pollution indicators, because Bulgarians often refuse to report their residential address.[13] Therefore, we introduced a variable representing STI as a proxy for traffic counts, air pollution, and noise.[13] STI question asked: “How would you describe the road that your home is located at and its traffic? Please, rate the traffic intensity based on comparisons with other streets of Plovdiv, Bulgaria.” (“very rare/no traffic”, “moderately busy street”, “considerably busy street”, “heavy traffic”, “extremely busy street/extreme traffic”).

Assessment of type 2 diabetes mellitus

T2DM status was determined by self-reported doctor diagnosis: “Have you been diagnosed with any of the following diseases/conditions?” with one of the response option being “Type 2 diabetes”.

Assessment of confounding factors

Additional information was collected on:

Age, gender, ethnicity, highest educational level attained, marital status, occupation, self-rated socioeconomic status.

Body Mass Index (kg/meters2), sport activity (hours per week), number of wholegrain meals, alcohol beverages, meat/meat product meals, soda drinks, and sweet treats per week during the past year. When the participant had indicated the number of workouts per week rather than their overall duration, each workout was assumed to last 30 min on average.

Pack-years of smoking [(number of cigarettes a day*number of years of smoking)/20]

Family history of T2DM: “Have any of your first-degree relatives been diagnosed with any of the following diseases/conditions?”, with one of the response option being “Type 2 diabetes”.

Noise sensitivity was assessed with the Noise Sensitivity Questionnaire Short Form (NoiSeQSF). It measures multidimensional noise sensitivity on five subscales with three items in each.[15]

Duration of residence at the present address, number of nights per week sleeping with open bedroom windows, bedroom location.

Sleep disturbance was derived from an 11-point scale question: “How good is your sleep at home?” (“0, very good” to “10, cannot sleep at all”).

The type of sampling was coded as well.

Data analysis

The dataset was screened for missing rates. The variable with the highest percentage of missing values (40.9%) was Lden due to underreporting of residential addresses. Next, data were screened for univariate normality (graphical analysis, D’Agostino-Pearson K2 test, and Jarque-Bera test), and for univariate outliers (modified outlier labelling rule).[25] Outliers were to be retained if they were considered honest answers or if the cause for the outliers could not be determined. Otherwise they were winsorised.

Initial descriptive statistics were computed. Likert scale-like variables with ≥5 categories were included in parametric tests.[26] Welch's t-test and Pearson's chi-squared test/Fisher's Exact Test/Fisher-Freeman-Halton Test were conducted to assess the association between T2DM and other variables. The main analyses were based on multivariate logistic and log-linear Poisson regressions (distribution “Poisson”, link “Log”, method “Hybrid”, scale parameter “Fixed value”, covariance matrix “Robust estimator”).[2728] The latter is a preferable regression method when the aim is to approximate the relative risk (RR) in cross-sectional studies. However, because logistic regression is still widely used and formed the basis of our power analysis, all models are primarily reported in a logistic regression framework.

Three adjusted models were tested:

Model 1: Objectively measured noise and air pollution exposure (Lden + PM2.5 + BaP);

Model 2: Self-rated exposure to road traffic, noise, and air pollution (STI);

Model 3: Lifetime noise exposure (residential + occupational exposure to loud noise).

After inspection of the distribution of T2DM, cases across the 5-dB Lden categories in the dataset (with missing data), Lden was dichotomized (“71-80 dB” versus “51-70 dB”) in the regression models in order to enhance the contrast between the exposure categories, to minimize exposure misclassification bias rising from the limitations of the noise map, the fact that some residential addresses were matched to the center point of the street, and the unequal distribution of T2DM cases across the 5-dB categories. PM2.5 (“25.0-66.8 μg/m3” versus “0.0-25.0 μg/m3”) and BaP (“6.0-14.02 ng/m3” vs. “0.0-6.0 ng/m3”) were dichotomized as well in order to account for the small number of T2DM cases across their initial five categories. For the same reason, the categories of STI were collapsed (“very rare/no traffic” and “moderately busy street” → “low traffic exposure”; “considerably busy street” → “moderate traffic exposure”; “heavy traffic” and “extremely busy street/extreme traffic” → “high traffic exposure”).

The regression models were run on 50 imputed datasets in order to increase the precision and power of the analyses. Multiple imputation yields accurate estimates in <50% missing rates.[2930] This method has proven to be valid alternative for estimating missing rates in epidemiological surveys of environmental risk factors.[31] The exact adjustment set for the direct effects models was based on directed acyclic graphs (DAG) using DAGitty v. 2.3 (http://dagitty.net/). A recent simulation study showed that a priori theoretical identification of important confounders in logistic regression should always be preferred to empirical strategies.[32] Possible moderation effect of the type of sampling procedure was tested by computing interaction terms and evaluating them at the relaxed P < 0.2 level. Multicollinearity was tested. In sensitivity analyses, the results were stratified by participants’ characteristics when it was appropriate. Post hoc statistical power was computed for the logistic regression models using G*Power v. 3.1.9.2.[33]

Results were considered statistically significant at P < 0.05 (two-tailed) or when 95% confidence intervals did not overlap zero. Data were processed with Statistical Package for the Social Sciences (SPSS) [International Business Machines (IBM) Corporation released 2010. IBM SPSS Statistics for Windows, Version 19.0. Armonk, New York: IBM Corporation].

Results

A total of 581 questionnaires were available for processing, but after excluding 68 questionnaires due to an unacceptably high percentage of missing data (>50%) on all key variables, or due to meeting exclusion criteria, we were left with 513 individual cases for further analyses.

Participants’ mean age was 36.45 years (SD = 15.39, range: 18-83 years) and 185 participants (36.1%) were males. Most participants were ethnic Bulgarians (n = 436, 85.0%), had upper secondary (n = 257, 50.1%) or bachelor/master (n = 236, 46.0%) educational level, were married/had a spouse (n = 318, 62.0%), were employed (n = 286, 55.8%) and rated their socioeconomic status as “middle” (n = 352, 68.6%). There were 35 cases (6.8%) of T2DM in the sample.

According to official statistics for Plovdiv, Bulgaria[34] our sample was only partially representative of the studied population. That is, it was skewed toward women, younger people, those with middle socioeconomic status and higher education. The prevalence of diagnosed diabetes in the population is about 7.1% with 83% of those being >50 years of age and 56.7% being male.[2] In our sample, the prevalence of T2DM was similar, however, with a larger proportion of male cases and about 57.1% >50 years of age.

Participants differed between the two types of sampling [See supplementary Table S1]. People in the snowball sample were younger, more likely to be female, with a lower educational level, single, and studying. The two groups did not differ significantly on T2DM prevalence, but did on road traffic and BaP exposures. Most of these differences were due to the fact that we had recruited more young people at our university campus than in the field sample, and those students had younger peers in their social network. Therefore, the type of sampling was tested for moderation effects in the multivariate models.

Table S1

Basic participants' characteristics by the type of sampling

Variable		No. cases	Snowball n = 208 (40.5%)	Field n = 305 (59.5%)	P-value (χ2/t-test)
Age, mean (SD)		510	32.48 (16.11)	39.19 (14.26)	<0.001
Gender: Male, n (%)		513	51 (24.5)	134 (43.9)	<0.001
Ethnicity: Bulgarian, n (%)		512	176 (85.0)	260 (85.2)	1.000
Education, n (%)		513			<0.001
	Basic		1 (0.5)	6 (2.0)
	Upper secondary		135 (64.9)	122 (40.0)
	Bachelor/master		63 (30.3)	173 (56.7)
	PhD, DSc		9 (4.3)	4 (1.3)
Marital status, n (%)		512			0.002
	Married/having a spouse		111 (53.6)	207 (67.9)
	Single		83 (40.1)	74 (24.3)
	Widowed		4 (1.9)	8 (2.6)
	Divorced		9 (4.3)	16 (5.2)
Occupation, n (%)		512			<0.001
	Employed		80 (38.5)	206 (67.8)
	Studying		115 (55.3)	59 (19.4)
	Unemployed		1 (0.5)	17 (5.6)
	Retired		12 (5.8)	22 (7.2)
Socioeconomic status, n (%)		511			0.437
	Lower		51 (24.6)	87 (28.6)
	Middle		149 (72.0)	203 (66.8)
	Upper		7 (3.4)	14 (4.6)
Type 2 diabetes: Yes, n (%)		508	11 (5.4)	24 (7.9)	0.291
Family history of diabetes: Yes, n (%)		497	67 (33.2)	90 (30.5)	0.556
BMI, mean (SD)		507	22.67 (4.55)	23.74 (4.10)	0.007
Pack years of smoking, mean (SD)		505	4.06 (8.12)	6.74 (14.47)	0.008
Years residential exposure to loud noise, mean (SD)		502	3.02 (8.00)	2.57 (6.76)	0.515
Years occupational exposure to loud noise, mean (SD)		504	2.12 (6.40)	4.23 (8.61)	0.002
Noise sensitivity, mean (SD)		478	3.35 (0.79)	3.35 (0.74)	0.925
Noise annoyance, mean (SD)		509	5.62 (2.36)	5.43 (2.13)	0.345
Perceived air quality, mean (SD)		485	5.43 (2.32)	5.55 (1.93)	0.546
Perceived severity of air pollution, mean (SD)		486	5.70 (2.43)	5.73 (1.96)	0.882
Air pollution annoyance, mean (SD)		485	5.83 (2.52)	5.79 (2.11)	0.870
Duration of residence, mean (SD)		485	13.49 (13.75)	17.88 (14.28)	0.001
Bedroom location: Noisy side, n (%)		481	62 (32.5)	71 (24.5)	0.061
Apartment floor, mean (SD)		483	4.50 (3.28)	3.62 (2.35)	0.001
Sport activity, mean (SD)		482	1.57 (2.08)	1.11 (2.41)	0.026
Wholegrain meals, mean (SD)		481	3.96 (3.22)	3.20 (2.98)	0.009
Alcohol beverages		474	1.31 (1.76)	1.20 (1.59)	0.485
Meat meals, mean (SD)		486	4.29 (3.28)	4.04 (2.83)	0.384
Soda beverages, mean (SD)		472	1.87 (2.40)	1.39 (1.97)	0.024
Sweet treats		480	4.09 (2.90)	2.99 (2.21)	<0.001
Sleep disturbance, mean (SD)		510	4.31 (3.49)	3.99 (2.75)	0.277
Self-reported traffic intensity, n (%)		496			0.598
	Very rare/no traffic		34 (17.3)	64 (21.4)
	Moderately busy street		76 (38.6)	96 (32.1)
	Considerably busy street		32 (16.2)	49 (16.4)
	Heavy traffic		29 (14.7)	45 (15.1)
	Extremely busy street/extreme traffic		26 (13.2)	45 (15.1)
Nights with open bedroom windows, mean (SD)		465	5.88 (2.11)	5.57 (2.34)	0.126
Lden [dB], n (%)		303			0.056
	51-55		15 (9.8)	3 (2.0)
	56-60		12 (7.8)	15 (10.0)
	61-65		32 (20.9)	29 (19.3)
	66-70		41 (26.8)	54 (36.0)
	71-75		49 (32.0)	45 (30.0)
	76-80		4 (2.6)	4 (2.7)
	>80		0 (0.0)	0 (0.0)
PM2.5 [mg/m3], n (%)		334			0.113
	0.0-17.5		2 (1.1)	0 (0.0)
	17.5-20.3		20 (11.4)	21 (13.3)
	20.3-25.0		26 (14.8)	16 (10.1)
	25.0-40.0		77 (43.8)	58 (36.7)
	40.0-66.8		51 (29.0)	63 (39.9)
BaP [ng/m3], n (%)		334			<0.001
	0.0-1.0		1 (0.6)	0 (0.0)
	1.0-3.75		36 (20.5)	23 (14.6)
	3.75-6.0		32 (18.2)	60 (38.0)
	6.0-9.75		106 (60.2)	75 (47.5)
	9.75-14.02		1 (0.6)	0 (0.0)

Note. Percentages are reported within the “type of sampling” variable. Calculations are based on the valid number of cases in the original dataset (i.e., excluding cases with missing values on “type of sampling” and/or respective participants’ characteristics)

According to Table 1, people with T2DM were more likely to be male and significantly older than those without. In addition, both groups differed on socioeconomic classes and occupations. Lifestyle factors such as dietary habits, alcohol consumption, and smoking were significantly associated with T2DM. Some environmental exposures were higher among those with T2DM. Correlations between the exposure indicators are given in Supplementary Table S2.

Table 1

Participants' characteristics by their type 2 diabetes mellitus status

Participants’ characteristics	No. cases	No T2DM n = 473 (93.1%)	T2DM n = 35 (6.9%)	P-value
Age, mean (SD)	505	35.27 (14.87)	51.09 (13.26)	<0.001
Gender: male, n (%)	508	158 (33.4)	25 (71.4)	<0.001
Ethnicity: Bulgarian, n (%)	507	403 (85.4)	31 (88.6)	0.804
Education, n (%)	508			0.148
Basic		5 (1.1)	2 (5.7)
Upper secondary		238 (50.3)	18 (51.4)
Bachelor/master		218 (46.1)	14 (40.0)
PhD, DSc		12 (2.5)	1 (2.9)
Marital status, n (%)	507			0.071
Married/having a spouse		292 (61.7)	23 (67.6)
Single		151 (31.9)	6 (17.6)
Widowed		9 (1.9)	2 (5.9)
Divorced		21 (4.4)	3 (8.8)
Occupation, n (%)	507			<0.001
Employed		262 (55.5)	21 (60.0)
Studying		171 (36.2)	2 (5.7)
Unemployed		15 (3.2)	3 (8.6)
Retired		24 (5.1)	9 (25.7)
Socioeconomic status, n (%)	506			0.007
Lower		122 (25.9)	16 (45.7)
Middle		331 (70.3)	16 (45.7)
Upper		18 (3.8)	3 (8.6)
BMI, mean (SD)	502	22.98 (4.08)	27.80 (4.95)	<0.001
Sport activity, mean (SD)	477	1.37 (2.35)	0.30 (0.89)	<0.001
Wholegrain meals, mean (SD)	476	3.53 (3.07)	2.77 (3.35)	0.200
Alcohol beverages, mean (SD)	469	1.22 (1.58)	1.60 (2.48)	0.381
Meat meals, mean (SD)	481	4.16 (2.86)	3.14 (1.07)	<0.001
Soda beverages, mean (SD)	468	1.63 (2.19)	0.63 (1.31)	<0.001
Sweet treats, mean (SD)	475	3.48 (2.42)	2.05 (2.44)	0.002
Pack years of smoking, mean (SD)	500	5.09 (11.67)	14.21 (18.66)	0.009
Family history of T2DM: Yes, n (%)	496	139 (30.2)	18 (51.4)	0.013
Years residential exposure to loud noise, mean (SD)	497	2.70 (6.93)	2.38 (7.87)	0.815
Years occupational exposure to loud noise, mean (SD)	501	2.98 (7.23)	8.93 (12.74)	0.010
Noise sensitivity, mean (SD)	473	3.32 (0.76)	3.57 (0.64)	0.042
Noise annoyance, mean (SD)	504	5.50 (2.27)	5.69 (1.43)	0.477
Sleep disturbance, mean (SD)	505	3.97 (3.09)	5.69 (2.26)	<0.001
Self-reported traffic intensity, n (%)	492			0.001
Very rare/no traffic		90 (19.5)	7 (22.6)
Moderately busy street		169 (36.7)	2 (6.5)
Considerably busy street		76 (16.5)	5 (16.1)
Heavy traffic		64 (13.9)	8 (25.8)
Extremely busy street/extreme traffic		62 (13.4)	9 (29.0)
Perceived air quality, mean (SD)	480	5.49 (2.10)	5.56 (2.23)	0.853
Perceived severity of air pollution, mean (SD)	481	5.71 (2.16)	6.00 (2.16)	0.475
Air pollution annoyance, mean (SD)	480	5.75 (2.29)	6.53 (2.11)	0.050
Duration of residence, mean (SD)	480	15.26 (13.67)	26.82 (17.09)	0.001
Bedroom location: Noisy façade side, n (%)	477	116 (26.0)	16 (53.3)	0.001
Nights with open bedroom windows, mean (SD)	460	5.68 (2.27)	6.05 (1.83)	0.286
Lden [dB], n (%)	299			<0.001
51-55		18 (6.4)	0 (0.0)
56-60		23 (8.2)	3 (15.8)
61-65		60 (21.4)	1 (5.3)
66-70		93 (33.2)	0 (0.0)
71-75		80 (26.8)	13 (68.4)
76-80		6 (2.1)	2 (10.5)
>80		0 (0.0)	0 (0.0)
PM2.5 [μg/m3], n (%)	330			0.267
0.0-17.5		2 (0.6)	0 (0.0)
17.5-20.3		38 (12.3)	3 (13.6)
20.3-25.0		41 (13.3)	1 (4.5)
25.0-40.0		127 (41.2)	6 (27.3)
40.0-66.8		100 (32.5)	12 (54.5)
BaP [ng/m3], n (%)	330			0.961
0.0-1.0		1 (0.3)	0 (0.0)
1.0-3.75		56 (18.2)	3 (13.6)
3.75-6.0		86 (27.9)	6 (27.3)
6.0-9.75		164 (53.2)	13 (59.1)
9.75-14.02		1 (0.3)	0 (0.0)

Note. T2DM = Type 2 diabetes mellitus, BaP = Benzo alpha pyrene, PM2.5 = Particulate matter pollutants of ≤2.5 μm in aerodynamic diameter, Lden = Day-evening-night equivalent sound level, BMI = Body mass index, Percentages are reported within columns. Calculations are based on the valid number of cases in the original dataset (i.e., excluding cases with missing values on “type 2 diabetes status” and/or respective participants’ characteristics)

Table S2

Correlations between objective and self-rated exposure indicators

Indicator	1	2	3	4	5	6	7	8	9	10
PM2.5	1.00	.71**	.28**	.01	–.04	.01	.04	.09	.13*	.09
BaP		1.00	.23**	.10	.02	–.02	.01	.12*	.13*	.05
Lden			1.00	.38**	.00	.03	.18**	.24**	.21**	.21**
Self-reported traffic intensity				1.00	.17**	.11*	.20**	.36**	.26**	.18**
Lifetime residential exposure to loud noise					1.00	.15**	.23**	.09*	–.03	.11*
Lifetime occupational exposure to loud noise						1.00	.06	.12**	.10*	.10*
Noise annoyance							1.00	.05	.07	.29**
Perceived air quality								1.00	.43**	.27**
Perceived severity of air pollution									1.00	.50**
Air pollution annoyance										1.00

Note. BaP = Benzo alpha pyrene, PM2.5 = Particulate matter pollutants of ≤2.5 μm in aerodynamic diameter, Lden = Day-evening-night equivalent sound level Model is based on the original dataset prior to replacement of missing values and dichotomizing variables. Spearman’s correlation coefficients are reported Correlation is *Significant at P < 0.05, **Significant at P < 0.01

Lden, PM2.5, BaP, and type 2 diabetes mellitus

In the univariate analyses, there were higher odds for T2DM associated with Lden, PM2.5, and BaP—7.19 (95% CI: 2.79, 18.53), 1.25 (95% CI: 0.51, 3.09), and 1.15 (95% CI: 0.50, 2.67), respectively. After these three exposure indicators were adjusted for each other and additionally controlled for important a priori confounders, the elevated odds of T2DM persisted in the multivariate model, but only the effect of Lden reached statistical significance [See Table 2]. The model explained about 45% of the variance in T2DM, but the observed power of the test was high only for Lden (100%), whereas for PM2.5 (11%) and BaP (33%), it was low enough to account for the nonsignificant results. Interaction terms between the type of sampling and Lden, PM2.5, or BaP were nonsignificant. The ORs associated with Lden in men and women were 5.55 (95% CI: 0.93, 33.19) and 6.42 (95% CI: 0.64, 64.67), respectively. When the analysis was restricted to participants living at the address for ≥10 years the OR associated with Lden remained similar (OR = 4.54, 95% CI: 1.05, 19.65) but increased for PM2.5 (OR = 2.08, 95% CI: 0.27, 16.26) and BaP (OR = 2.27, 95% CI: 0.54, 9.46). In the Poisson regression model, the risks were relative risk (RR) = 3.08 (95% CI: 1.24, 7.62), RR = 1.29 (95% CI: 0.49, 3.41) and 1.39 (95% CI: 0.60, 3.18) for Lden, PM2.5, and BaP, respectively.

Table 2

Multivariate associations between Lden, PM2.5 and BaP and type 2 diabetes mellitus (logistic regression)

Sample size (n = 513)	P-value	OR	95% CI
Lden 71-80 dB (ref. “51-70 dB”)	0.013	4.49	(1.38, 14.68)
PM2.5 25.0-66.8 mg/m3 (ref. “0.0-25.0 μg/m3”)	0.726	1.32	(0.28, 6.24)
BaP 6.0-14.02 ng/m3 (ref. “0.0-6.0 ng/m3”)	0.362	1.76	(0.52, 5.98)
Age (continuous)	<0.001	1.06	(1.03, 1.10)
Gender: Female (ref. “male”)	0.024	0.30	(0.10, 0.85)
Ethnicity: Non-Bulgarian (ref. “Bulgarian”)	0.507	1.61	(0.40, 6.49)
Body mass index (continuous)	0.022	1.13	(1.02, 1.24)
Family history of T2DM: Yes (ref. “no”)	0.019	3.12	(1.20, 8.08)
Noise sensitivity (continuous)	0.274	1.56	(0.70, 3.43)
Air pollution annoyance (continuous)	0.699	1.05	(0.81, 1.36)
Noise annoyance (continuous)	0.335	0.88	(0.67, 1.15)
Bedroom location: Noisy façade (ref. “quiet side”)	0.031	3.11	(1.11, 8.70)
Sleep disturbance (continuous)	0.032	1.22	(1.02, 1.47)
Pack-years of smoking (continuous)	0.691	1.00	(0.97, 1.02)

Note. ref. = Reference category, T2DM = Type 2 diabetes mellitus, BaP = Benzo alpha pyrene, PM2.5 = Particulate matter pollutants of ≤2.5 μm in aerodynamic diameter, Lden = Day-evening-night equivalent sound level, Model is based on 50 imputed datasets, Average Nagelkerke R2 = 0.45, Post-hoc power is 1.00 for “Lden”, 0.11 for “PM2.5” and 0.33 for “BaP” (specifications: Two tails, α = 0.05, binomial distribution, 6.9 % prevalence of T2DM, n = 513, R2 with other covariates = 0.1/0.27/0.27)

Road traffic exposure and type 2 diabetes mellitus

People exposed to high road traffic had a crude OR = 3.60 (95 % CI: 1.57, 8.29). In the adjusted logistic model, however, high road traffic exposure was associated with nonsignificant effect (OR = 1.40, 95% CI: 0.48, 1.07) [See Table 3]. The model explained about 39% of the variance in T2DM and had low statistical power (15%). The odds associated with high traffic exposure among men and women were OR = 1.09 (95% CI: 0.25, 4.75) and OR = 2.18 (95% CI: 0.26, 18.55), respectively. Poisson regression model yielded RR = 1.19 (95% CI: 0.54, 2.62) associated with high traffic exposure.

Table 3

Multivariate associations between self-reported traffic intensity and type 2 diabetes mellitus (logistic regression)

Sample size (n = 513)	P-value	OR	95% CI
Self-reported traffic intensity
low	ref.	1.00	.
moderate	0.836	1.15	(0.30, 4.45)
high	0.533	1.40	(0.48, 4.07)
Age (continuous)	<0.001	1.06	(1.03, 1.09)
Gender: Female (ref. “male”)	0.012	0.29	(0.11, 0.76)
Ethnicity: Non-Bulgarian (ref. “Bulgarian”)	0.566	1.48	(0.39, 5.70)
Body mass index (continuous)	0.003	1.15	(1.05, 1.25)
Family history of T2DM: Yes (ref. “no”)	0.054	2.39	(0.98, 5.80)
Noise sensitivity (continuous)	0.324	1.48	(0.68, 3.22)
Air pollution annoyance (continuous)	0.522	1.08	(0.85, 1.39)
Noise annoyance (continuous)	0.340	0.89	(0.69, 1.14)
Bedroom location: Noisy façade (ref. “quiet side”)	0.028	3.35	(1.14, 9.87)
Sleep disturbance (continuous)	0.019	1.22	(1.03, 1.44)
Pack-years of smoking (continuous)	0.550	0.99	(0.97, 1.02)

Note. ref. = Reference category, T2DM = Type 2 diabetes mellitus, Model is based on 50 imputed datasets, Average Nagelkerke R2 = 0.39, Post-hoc power is 0.15 for “high traffic exposure” (specifications: Two tails, α = 0.05, binomial distribution, 6.9% prevalence of T2DM, n = 513, R2 with other covariates = 0.2)

Lifetime exposure to loud noise and type 2 diabetes mellitus

Only the univariate effect of lifetime occupational exposure was significant (OR = 2.13, 95% CI: 1.45, 3.12). Neither of the two lifetime exposures was associated with elevated multivariate odds of T2DM [See Table 4]. Stratification by gender yielded ORs of 0.63 (95% CI: 0.20, 1.94) and 1.03 (95% CI: 0.51, 2.07) in men and 0.38 (95% CI: 0.04, 3.51) and 0.29 (0.02, 5.11) in women. In the Poisson regression, the RR was 0.69 (95% CI: 0.34, 1.41) for residential and 0.91 (95% CI: 0.66, 1.27) for occupational exposure.

Discussion

Key findings

To our knowledge, this was the first exploration of the feasibility of a large scale epidemiological study on traffic, noise, and air pollution as environmental risk factors for T2DM in Bulgaria. Overall, although our results are somewhat weak due to methodological limitations, they suggested that future research in the country is feasible. People exposed to Lden 71-80 dB might have significantly higher odds of prevalent T2DM. Exposures to high road traffic, PM2.5 25.0-66.8 μg/m3, and BaP 6.0-14.02 ng/m3 were suspected risk factors as well, although they failed statistical significance. Lifetime exposure to loud noise was not associated with T2DM. These effects were higher among women, although nonsignificant. Given the systematic overestimation of the effects when ORs are applied as if they were RRs and the ongoing debate on the appropriateness of the two estimates;[3536] we additionally looked at Poisson regressions, which estimated statistically significantly higher risk of T2DM associated with Lden and nonsignificant risks associated with PM2.5 and BaP. A nonsignificantly elevated risk was seen as well for participants exposed to high traffic.

The fact that our self-rated exposure indicators were not significant risk factors can partially be attributed to the low power of some tests or to a temporal instability of participants’ perception of exposure. When we tested noise sensitivity and noise annoyance in a multivariate context, they were not significant or influential covariates. If noise sensitivity that has been linked to cardiovascular health[37] does not independently rise or modify the risk of T2DM, this would mean that highly sensitive groups will benefit just as much from physical reduction of noise exposure. On the contrary, if personal sensitivity increases the risk on its own and is not contingent on the actual exposure, then even reducing noise levels might not be enough to sufficiently decrease morbidity rates, while additional psychosocial approaches might help. These speculations prompt further scientific inquiry into the subjective exposure indicators and individual environmental sensitivity.

The RR of T2DM associated with exposure to Lden 71-80 dB is lower than that estimated in a recent meta-analysis,[8] which might be due to the weak conclusions of that meta-analysis (due to methodological discrepancies in the primary studies), the limitations of our survey and/or the different social and environmental context in a middle-income country such as Bulgaria. The effect of PM2.5 was positive, even given the low power of our test and it is likely to become significant in a further study as is the case in the literature.[7] Moreover, significant effects might have been missed due to the low spatial resolution of the air pollution maps we used. As for BaP, the mechanisms underlying its alleged effects are still unclear, although increase in adipose mass, production of reactive oxygen species, and inflammation are suspected pathways.[38] Other authors have found links between urinary concentrations of polycyclic aromatic hydrocarbons and diabetes, glucose homeostasis, and metabolic syndrome.[383940] We cannot compare our results to previous findings because, as far as we are aware, the direct relationship between atmospheric BaP levels and T2DM has not been established. The significant and strong effect of sleeping in a bedroom with noisy façade is in line with previous work. Recently, Pirrera et al.[41] concluded that “averaged sound pressure level as noise indicator in a field context is insufficient to draw conclusions with respect to noise-induced sleep disturbances” and that “bedroom location is not only a major determinant for noise assessment, but also for sleep outcomes”. Nighttime noise, however, was not tested in our models due to its almost perfect linear correlation with Lden.

Methodological considerations and future research

Our study has several noteworthy limitations. First, its cross-sectional design precludes us from inferring causal links between exposure and outcome variables. However, as it has been argued with respect to cardiovascular effects of noise,[42] it is unlikely that people with T2DM will purposefully move to noisy or air polluted locations. Moreover, the effect of Lden did not change in long-term residents and, due to their poor socio-economic standard, most Bulgarians are unlikely to have moved to a different address simply because of the traffic. Some of the covariates were assessed only for the last year (e.g., dietary habits), whereas others reflected lifetime exposure (e.g., pack-years of smoking). Thus, the diet, for example, might have been modified in compliance with the prescribed nonpharmacological treatment of T2DM. This would explain why people with T2DM consumed less sweet treats and soda beverages. As an attractive alternative design requiring fewer resources, we propose the cross-sectional cohort study, where a representative population is sampled cross-sectionally and then thorough exposure and outcome history is collected over a specified time period.[43] Regardless of the study design, however, random sampling will be essential.

Another set of inherent methodological pitfalls are sociological. While we fail to provide sufficient evidence for definitively confirming or rejecting some risk factors due to the smallish sample size, even in representative epidemiological studies, one should refrain from interpreting the 95% CIs solely in terms of statistical significance.[44] Furthermore, the power was sufficient for the effect of Lden. Regarding missing data, imputing 50 datasets should have addressed those adequately.[2930] The low response rate alone, on the other hand, cannot be held as an evidence of nonparticipation bias:[45] Bulgarians have poor survey culture and might be reluctant to participate in surveys on environmental pollutants.[13] As for the nonrepresentativeness of our sample, it is a common misconception that random sampling is mandatory in risk assessment; if the sole purpose is to estimate the relative risk, it is unnecessary unless the factor making the sample nonrandom acts as an effect modifier.[46] In the future, offering some form of financial incentives will probably increase participation rates, while public popularization of environmental health concerns might get Bulgarians more involved.

Regarding our exposure and outcome variables, there is a possibility for misclassification bias, but it is likely to be nondifferential. Noise and air pollution levels were modelled for 2009 and 2011, respectively. An update of the strategic noise map of Plovdiv, Bulgaria by recalculating and redefining the acoustic model of the street network might be necessary as well because its accuracy is currently partially outside the acceptable range defined by the Environmental Noise Directive 2002/49/EC.[47] The fact that only major arterial streets were sampled to calibrate the model might be skewing the exposure. On the other hand, the pathogenesis of T2DM is long, therefore, exposure levels referring to a past period are arguably more informative than current exposure. Matching addresses to the center point of the street is problematic as well as is the fact that lifetime residential history was not collected. By dichotomizing Lden, PM2.5, and BaP, however, we should have reduced these imprecisions. Stratification of the analyses by duration of residence was not always feasible because we had very few cases in the subgroups (i.e., T2DM was reported mostly by long-term residents).

With respect to STI, it measured traffic intensity on the residential street, but for those living on small side streets close to a major road artery and vice versa adjusting for the distance to the closest major road will improve the accuracy of the models. Finally, our PM2.5 and BaP indicators were somewhat crude proxies for air pollution, because they were derived from air pollution maps with large grid size, rather than sampled and verified at each residential address. Using readily available noise and air pollution maps was associated with some degree of exposure misclassification and is not substantiated for future studies aiming at estimation of precise exposure-response relationships, because taking all exposure levels <60 dB as one reference category loses information. They were not intended for the purposes of epidemiological research, making them unsuitable for generating high quality evidence at lower exposure levels. Enrolment of acoustic engineers and officers from the Regional Health Inspections will facilitate field sampling of noise and air pollution, calculating, and calibrating land-use regression models. Extensive occupational and residential address history should be acquired as well, and the type of domestic heating should be assessed. A look at the effects of wider range of air pollutants and further study of polycyclic aromatic hydrocarbons, by assessing both their concentrations in air and urinary metabolites, are warranted.

The fact that lifetime exposure to loud noise did not raise the risk of T2DM but Lden 71-80 dB did, is probably due to an inherent limitation of our design and the inaccuracy in the self-reported exposure measure. The observed range of values on this variable and the few people ever exposed might be the causes of the flattened association with T2DM. Likewise, the question on speech disturbance made no distinction between transportation noise and other sources such as domestic equipment, toddlers, noisy neighbors, etc. As for occupational exposure to loud noise, the fact that previous studies failed to establish a link with T2DM[8] as well suggests that there might be a qualitative difference between noise exposure at work and at home. Provided that people typically spend more active time at work, nighttime noise mediated through sleep disturbance might, in fact, be driving the effect. Disruption of sleep patterns has been recognized as a risk factor for metabolic disorders and diabetes.[4849] Our data additionally supports the importance of bedroom location, sleeping with open bedroom windows, and sleep disturbance as risk factors for T2DM. Therefore, occupational noise might simply be less influential because people typically do not sleep at work. Moreover, they might have developed some psychological and behavioral coping strategies in occupational settings. Finally, because we did not specifically target industrial workers, the variability in this variable might have additionally been too low.

T2DM status was assessed via self-reported doctor diagnosis but about 26.1% of the diabetes cases remain undiagnosed.[2] Further, we failed to assess the date of diagnosis, therefore, the temporal links with noise and air pollution cannot be disentangled. An emphasis should be put on obtaining valid biomedical information about metabolic control indicators and accompanying diseases through clinical examinations or medical documentation. The use of more sophisticated and available questionnaires in Bulgaria for the assessment of some covariates such as sleep quality,[50] physical activity, and diet[51] is recommended as well.

Conclusions

Residential noise exposure was associated with significantly higher odds of prevalent T2DM. Although there were suspected effects of other self-rated and objective exposures, they failed statistical significance. Bedroom location, sleeping in a room with open windows, and self-rated sleep disturbance were associated with elevated risks as well.

The inferences made by this small scale research and the lessons learned from its limitations could guide the designing of a population-based epidemiological survey in Bulgaria. Future environmental health research calls for acquisition of valid exposure and biomedical data, involving the community and further investigating both objective and self-rated exposure indicators. Our results might be of interest not only to Bulgarian environmentalists but to the wider international research community as well. Despite the sources of bias in our study and its crude design, it adds to the limited body of evidence of the metabolic effects of environmental noise. By providing various exposure indicators and reporting both ORs and RRs, we facilitate comparisons with previous findings. Moreover, results from South-Eastern Europe are still scant.

Financial support and sponsorship

Nil.

Conflicts of interest

The authors declare that, with respect to this study, they do not have any potential conflict of interest.

Table 4

Multivariate associations between duration of lifetime residential and occupational exposure to loud noise (per 1 interquartile range increase) and type 2 diabetes mellitus (logistic regression)

Sample size (n = 513)	P-value	OR	95% CI
Lifetime residential exposure to loud noise (per 16 years)	0.257	0.58	(0.23, 1.49)
Lifetime occupational exposure to loud noise (per 13 years)	0.739	0.91	(0.53, 1.58)
Age (continuous)	<0.001	1.06	(1.03, 1.10)
Gender: Female (ref. “male”)	0.012	0.28	(0.11, 0.75)
Ethnicity: Non-Bulgarian (ref. “Bulgarian”)	0.475	1.64	(0.42, 6.37)
Body mass index (continuous)	0.003	1.15	(1.05, 1.26)
Family history of T2DM: Yes (ref. “no”)	0.042	2.51	(1.03, 6.09)
Noise sensitivity (continuous)	0.329	1.48	(0.68, 3.24)
Air pollution annoyance (continuous)	0.595	1.07	(0.84, 1.37)
Noise annoyance (continuous)	0.612	0.94	(0.72, 1.21)
Bedroom location: Noisy façade (ref. “quiet side”)	0.009	3.79	(1.40, 10.30)
Sleep disturbance (continuous)	0.016	1.23	(1.04, 1.45)
Pack-years of smoking (continuous)	0.610	0.99	(0.97, 1.02)

Note. ref. = Reference category, T2DM = Type 2 diabetes mellitus, Model is based on 50 imputed datasets, Average Nagelkerke R2 = 0.40, Post-hoc power is 0.38 for “Lifetime residential exposure to loud noise” and 0.05 for “Lifetime occupational exposure to loud noise” (specifications: Two tails, α = 0.05, lognormal distribution, 6.9% prevalence of T2DM, n = 497, R2 with other covariates = 0.06/0.25)