Long-term noise exposure and the risk for type 2 diabetes: a meta-analysis.

Diabetes mellitus is one of the leading causes for disability and mortality in modern societies. Apart from personal factors its incidence might be influenced by environmental risks such as air pollution and noise. This paper reports a systematic review and meta-analysis on the risk for type 2 diabetes due to long-term noise exposure. Electronic searches in MEDLINE, EMBASE and the Internet yielded 9 relevant studies (5 for residential and 4 for occupational exposure). They were checked against a predefined list of safeguards against bias producing individual quality scores, which were then fed to MetaXL to conduct a quality effects meta-analysis. People exposed at their homes to roughly L(den) > 60 dB had 22% higher risk (95% confidence interval [CI]: 1.09-1.37) for type 2 diabetes in comparison to those exposed to L(den) < 64 dB; when studies reporting contentious exposure categories were excluded, there was still 19% risk (95% CI: 1.05-1.35) for L(den) = 60-70 dB versus L(den) < 60 dB. In occupational environment there was not significant risk (relative risk [RR] = 0.91, 95% CI: 0.78-1.06) for < 85 dB versus >85 dB. There was no heterogeneity in the two groups (I² = 0.00). The results should be interpreted with caution due to methodological discrepancies across the studies; however, they are indicative of the close links that noise pollution might have not only to cardiovascular diseases but to endocrine dysfunction as well.

Introduction

Diabetes mellitus (ICD-10: E08-E13) is one of the heavy burdens of societies internationally. It has been estimated that about 347 million people worldwide have diabetes.[1] In 2010 there were 285 million newly diagnosed cases,[2] and by 2030 it is projected to become the 7th leading mortality cause.[3] Diabetes is also associated with considerable economic costs — more than 306 billion dollars in direct and 69 billion in indirect costs were spent in 2012 in the United States alone.[4] About 90 % of all diabetes cases can be attributed to type 2 diabetes (ICD-10: E11), which results from the ineffective utilization of insulin.[5] Nearly 25% of those >50 years of age have impaired glucose tolerance and each year 5-8% of them develop diabetes.[6] By 2025 the number of people with type 2 diabetes will be about 380 million.[7]

It is widely established that obesity, irregular and insufficient physical activity and genetic factors contribute to the onset of type 2 diabetes. However, modern lifestyle is characterized not only by the unhealthy personal choices that we make, but by severe environmental conditions that we have to endure and have little power over. Recently a meta-analysis suggested that air pollution might be a significant risk factor for type 2 diabetes.[8] There has been emerging evidence that noise pollution might increase the risk on its own.[9] The study of Sørensen et al.[9] circulated the Internet and other media, raising the concern for noise-induced type 2 diabetes, particularly so given the plague-spreading pattern of noise pollution.[1011] There is strong biological plausibility for this relationship. In summary, noise acts as an environmental stressor leading to activation of the hypothalamic-pituitary-adrenal axis, increases cortisol levels and can thus inhibit the β-cells insulin secretion and the peripheral insulin sensitivity. In addition, sleep disturbance, disruption of normal sleep patterns and chronic sleep deprivation can induce diabetes via increased fasting glucose, appetite modulation and general dysregulation of the metabolic and endocrine functions (for a review see Song,[2] Sørensen et al.[9] and Eriksson et al.).[12]

Recently new contradictory evidence emerged regarding the effects of noise exposure on diabetes.[1213] As far as the author is aware, there has not been a systematic review on this association. Therefore initially this research aimed at carrying out a systematic review, but after rigorous literature search, the identification of several other relevant papers allowed conducting a quantitative meta-analysis on the risk for type 2 diabetes due to long-term noise exposure.

Methods

Search strategy and protocol

The searches were carried out by the author and another researcher in order to minimize bias. The research question was: “What is the risk for developing type 2 diabetes among people long-term exposed to noise in comparison to those not exposed or exposed to lower levels?” There were no previous systematic reviews on the topic according to PROSPERO and the Internet in general. Electronic searches (through June 11, 2014) in English and Spanish were carried out in MEDLINE (PubMed with no filters) and EMBASE (ScienceDirect with relevant filters) using the free-term keyword combinations: “diabetes + noise” (n = 377 in PubMed, n = 991 in ScienceDirect); “diabetes + traffic” (n = 559 in PubMed, n = 823 in ScienceDirect); “diabetes + aircraft” (n = 43 in PubMed, n = 119 in ScienceDirect); “diabetes + railway” (n = 46 in PubMed, n = 159 in ScienceDirect); “ruido + diabetes” (n = 1 in PubMed, n = 81 in ScienceDirect). The Internet was widely searched as well, as were the reference lists of all included studies. Because the data of interest were not always obvious when the prime objective of the study was not the relationship noise — diabetes, all available papers were reviewed in full text. Experts in endocrinology and environmental hygiene, and authors of articles were contacted as well. All studies including epidemiological, experimental and review papers were considered. Those dealing with the effects of noise in people with already diagnosed diabetes and without control group were excluded (for example, traffic accidents in people with diabetes).

Nine relevant studies were retrieved.[2912131415161718] The Internet search proved to be quite useful and yielded Song,[2] Rhee et al.[16] (missed with PubMed search) and Selander et al.[15] (also missed with PubMed search). Only the abstract of Heidemann et al.[13] was available, but the authors were kind to send us the full text. Therefore, the full texts of all papers were reviewed. During the peer-review of this paper two more studies were found.[1718]

Quality assessment of the studies

Based on the knowledge of the author, discussions with endocrinologist and environmental hygienist and extensive literature search a quality checklist including predefined safeguards against bias was prepared for the two reviewers. The checklist elements with the corresponding scoring protocol are given below:

Design (max 4)

Cohort (4)

Case-control (3)

Cross-sectional (2)

Using aggregated data (1)

Timeframe (max 0.5)

Reported (0.5)

Not reported (0)

Country where the study was carried out (max 2)

With good working and living conditions AND/OR high socio-economic standard (2)

Difficult conditions AND/OR lower socio-economic standard (1.5)

Very difficult conditions AND/OR very low socio-economic standard (1)

Not reported (0)

Mode of selection (max 2)

Random AND/OR considerable part OR the whole target population (2)

Non-random OR not specified (0)

Response rate (max 3)

≥ 80% AND/OR considerable part of the target population (3)

60-80% (1)

<60% OR not specified (0)

Final sample size (max 3)

Completely satisfactory AND/OR justified by power analysis (3)

Somewhat satisfactory AND no statistical justification (1)

Not satisfactory AND no statistical justification (0)

Participants (max 1)

Clearly described (1)

Ambiguous description OR not described (0)

Representativeness of the sample (max 3)

Completely representative of the population (3)

Somewhat representative (1)

Not representative (0)

Definition and assessment of diabetes mellitus type 2 (max 3)

Official WHO diagnostic criteria/official registry/database/records/physician-diagnosed (3)

Self-reported (1)

Assessment of noise exposure (max 3)

Geographic Information System geocoding AND/OR official sources AND/OR field measurements AND/OR noise dosimeters (3)

Job-exposure matrix AND/OR self-reported (1)

Duration of noise exposure considered (max 6)

Complete residential address history AND complete occupational history (6)

Incomplete residential address history AND complete occupational history OR complete residential address history AND incomplete occupational history (5)

Complete residential address history OR complete occupational history OR estimated/imputed residential address/occupational history (4)

Self-reported duration of exposure (2.5)

One-point-in-time assessment of exposure (2)

No information (0)

Noise metric (max 4)

Lden/Ldn/LAeq (4)

Need to raise the voice/estimated from job-exposure matrix (2)

Annoyance-based (1)

Biological plausibility for increased risk of type 2 diabetes mellitus given the overall noise exposure and the compared noise exposure categories (max 1.5)

Plausible (1.5)

Speculative (0.5)

Un-plausible (0)

Adjustments for personal covariates (max 5)

All of the relevant covariates (age, sex, socio-economic status, family history with diabetes, physical activity and sport, BMI/obesity, smoking, drinking, educational attainment) (5)

Some of the covariates (age, sex, family history with diabetes, BMI/obesity included) (4)

Some of the covariates (age, sex, family history with diabetes included partly with BMI/obesity included) (3)

Some of the covariates (age, sex, family history with diabetes, BMI/obesity not included) (0.5)

No adjustments OR none of the above included (0)

Adjustments for environmental covariates (max 3)

Air pollution only OR air pollution and other noise sources (3)

No adjustments (0)

Effect size calculation for meta-analysis (max 2)

No transformations AND no data imputation OR transformations not creating bias (2)

Transformations creating bias OR data imputation (1)

Transformations creating bias AND data imputation (0)

Additional transformations/imputation to the data AND/OR other source of bias associated with data extraction (max 3)

None (3)

Creating minor bias (1)

Major source of bias (0)

Overall (max 49)

As evident from the checklist, to each response option was assigned a different number of points in order to weight the corresponding element appropriately. For example, whether the timeframe (maximum of 0.5 points) was reported was not as important as the number of personal covariates included (maximum of 5 points) or the noise metrics (maximum of 4 points) utilized. Within each element the response options were also given different weights. The quality scoring was somewhat subjective, but was reasonable and yielded adequate quality weights for the studies (see below).

Data extraction

For each study the aim was to extract data for those comparisons that would ensure comparability with the other studies. When the studies did not have the relationship noise — diabetes as their main analysis,[1415161718] unadjusted estimates were computed and only those participants (cases and controls, exposed and non-exposed) relevant to these data were of interest.

Because Sørensen et al.[9] was the study of highest quality and following the recommendations for noise metrics, all residential exposure metrics were converted to Lden. They were not included directly in the analyses but allowed the data interpretation and were therefore included in the overall quality score. The categorical risk for Sørensen et al.[9] was extracted from Figure 1 in their article using GetData Graph Digitizer v.2.26 (http://www.getdata-graph-digitizer.com/), which allows for quite accurate digitalization of graph coordinates provided that the scale of each axis is accurate.

Figure 1

Forest plot on the risk for type 2 diabetes due to longterm noise exposure and heterogeneity statistics. Note. Q, p and I2 — heterogeneity statistics

Because there is a 6 dB aircraft malus when comparing with road traffic noise (with respect to noise annoyance, assuming that the effects of noise are mainly mediated through annoyance),[19] the aircraft noise level categories in Eriksson et al.[12] (i.e., <50 dB, 50-54 dB, ≥55 dB) were transformed to road traffic noise equivalent (i.e., <56 dB, 56-60 dB, ≥61 dB) to ensure compatibility with the other studies reporting road traffic noise. The upper limit of the ≥61 dB category was fixed at 65 dB given the width of the adjacent category.

The subjective “heavy traffic” intensity that Heidemann et al.[13] reported was compared against the “no/very rare traffic” intensity, arbitrarily assuming exposures of <55-60 dB Lden versus >65-70 dB Lden. These categories were derived from consultations with a hygienist, inspection of local noise maps, of representative noise maps for Germany (Berlin) (http://www.stadtentwicklung.berlin.de/umwelt/umweltatlas/ei705.htm) and the corresponding noise levels at the façade of buildings along very busy streets.

For Selander et al.[15] LAeq,24h was converted to Lden according to the following algorithm:

IF: Lden ≈ Ldn,[20]

AND: Ldn = 1.21*LA10,18h − 14.7 dB (Brown, 1989 cited by Abbott and Nelson),[21]

AND: LA10,18h = LAeq,24h +3.5 dB (Brown, 1989 cited by Abbott and Nelson),[21]

THEN: Ldn = 1.21* (LAeq,24h +3.5 dB) — 14.7 dB ≈ Lden.

In this case: Lden ≈ 1.21* (50 dB + 3.5 dB) — 14.7 dB = 50.04 dB

For Rhee et al.[16] first LAeq,8h was converted to Lden and then to equivalent traffic noise (+6 dB malus):

IF: Lden ≈ Ldn,[20]

AND: Ldn = 1.21*LA10,18h − 14.7 dB, (Brown, 1989 cited by Abbott and Nelson),[21]

AND: LA10,18h = LAeq,8h +6.7 dB, (Huybregts and Samuels, 1998 cited by Abbott and Nelson),[21]

THEN: Ldn = 1.21* (LAeq,8h + 6.7 dB) − 14.7 dB ≈ Lden.

In this case: 53.5 dB to 58.14 dB to Lden = 64.14 dB; 71.5 dB to 79.92 dB to Lden = 85.92 dB

For the study of Bainbridge et al.[17] only data associated with occupational noise exposure and diabetes were extracted because it was impossible to reliably determine the actual noise levels in dB, equivalent to the reported leisure noise exposure.

Statistics and data processing

Meta-analysis was performed on 5 studies of residential and 4 studies of occupational noise exposure to estimate the risk for developing type 2 diabetes for people exposed to high noise levels. Only one effect size estimate per study was extracted, and when there was choice, the risk estimate adjusted for most of the relevant personal and environmental confounders was selected. Relative risks (RRs) with 95% CIs were chosen as estimates of effect size because they yield more conservative results than odds ratios (ORs). RRs were either calculated from the raw data when the study was not primarily focused on the relationship noise — diabetes,[1415161718] or were calculated from the reported adjusted ORs according to Zhang and Yu[22] in order to keep the important information.

The analyses were carried out using MetaXL (EpiGear International Pty Ltd., Version 1.4, 2011-2014, Brisbane, Australia) (www.epigear.com) add-in for Microsoft Excel (Version 2010). This software offers the advantage of applying the quality effects model in estimating pooled effect sizes.[2324] The quality effects model is argued to be superior to the commonly used random effects model, which does not apply if the studies included in the meta-analysis differ in systematic way from the possible range in the population. Essentially, it uses a quality index (Qi) (ranging 0.00-1.00) assigned to each study by the researcher according to a specially developed scoring protocol and then generates a study specific composite for each study. Qi represents higher probability the judgment of that study is credible as it approximates 1.00, and the composite takes into consideration study specific information and its relationship to other studies to re-distribute the inverse variance weights. The quality effects model requires only that quality scores rank the analyzed studies with respect to deficiencies rather than by fixed function of bias or being a function of quantitatively measured bias.[25] Quality indices were calculated by assigning Qi of 1.00 to Sørensen et al.[9] as the “best” study according to the list of predefined safeguards against bias and then dividing the score of each of the other studies by the score of Sørensen et al.[9] score (i.e., 37).[26]

Heterogeneity was explored using the Chi-square test and the quantity of heterogeneity across studies was measured by the I2 statistic[27] - I2 < 30% (mild), I2 = 31-50% (moderate) and I2 > 50% (high).[28] Sensitivity analysis was performed by assessing the contribution of each study to the summary effect estimate by excluding it and re-calculating pooled RRs for the remaining studies. Funnel plot depicted possible selective publishing.

Results

Study characteristics

Out of the 9 included studies 5 reported data regarding residential[912131516] and 4 about occupational noise exposure.[2141718] Study characteristic and quality scores are given at Table 1. Song,[2] Sørensen et al.,[9] Heidemann et al.[13] and Eriksson et al.[12] designed and reported their studies with focus on the relationship noise — diabetes, while Suadicani et al.[14] studied the associating between occupational noise and all-cause mortality in Danish men, therefore relevant information was extracted without adjustments. The same refers to Selander et al.[15] and Rhee et al.[16] who were interested in noise effects on myocardial infarction and hypertension, respectively. Bainbridge et al.[17] and Jang et al.[18], on the other hand, looked at the associations between diabetes and noise-induced hearing loss. The type of residential noise source varied across studies as well. The participants of Eriksson et al.[12] and Rhee et al.[16] were exposed to aircraft noise, which is perceived differently than traffic noise. However, the noise sources were adequately transformed to ensure compatibility. Noise metrics were transformed as well to roughly correspond to Lden.

Table 1

Study characteristics and quality scoring

Study	Song[2]	Suadicani et al.[14]	Sørensen et al.[9]	Heidemann et al.[13]	Eriksson et al.[12]	Selander et al.[15]	Rhee et al.[16]	Bainbridge et al.[17]	Jang et al.[18]
D	Cohort (National Population Health Survey) (4)	Case-control** (Copenhagen Male Study) (3)	Cohort (Danish Diet, Cancer and Health cohort) (4)	Cohort (National Health Interview and Examination Survey and German Health Interview and Examination Survey for Adults) (4)	Cohort (Stockholm Diabetes Prevention Program) (4)	Case-control** (Stockholm Heart Epidemiology Program) (3)	Case-control** (3)	Case-control** (National Health and Nutrition Examination Survey) (3)	Case-control** (3)
TF	1994 — to date (0.5)	1970-1986 (0.5)	1993-1997 (0.5)	1997-1999, 2008-2011 (0.5)	1992-2006 (0.5)	1992-1994 (0.5)	Not specified (0)	1999-2004 (0.5)	2005-2009 (0.5)
C	Canada (2)	Denmark (2)	Denmark (2)	Germany (2)	Sweden (2)	Sweden (2)	South Korea (2)	US (2)	South Korea (2)
MS	Random (2)	Not specified (0)	Considerable part of the population (2)	Random (2)	Not specified (0)	Random (2)	Not specified (0)	Random (2)	Not specified (0)
SRR	Across all cycles >80% (3)	On average 81% (3)	Not specified (0)	56% (0)	On average 73% (1)	On average 72.5% (1)	46% for controls and 60% for cases (0)	89.5% for diabetes questionnaire (3)	Not specified (0)
FSS	n=1212 (1)	n=2998 (1)	n=57,053 (3)	n=3604 (1)	n=5156 (1)	n=2095 (1)	n=875 (0)	n=5140 (1)	n=1706 (1)
P	Household residents of the 10 Canadian provinces (1)	Caucasian middleaged men (1)	Residents of Copenhagen or Aarhus who were 50-64 years of age (1)	18-79-year-old residential population (1)	Residents of Stockholm County (35-56 years) (1)	Residents of Stockholm County (the controls are of interest) (1)	People within 5 km of a helicopter airbase, fighter-jet airbase, and not exposed to aircraft noise (nonexposed and helicopter airbase are of interest) (1)	Civilian, noninstitutionalized US population (20-69 years) (1)	Automobile manufacturing company workers (men) (1)
R	Completely (3)	Somewhat (men only) (1)	Completely (3)	Completely (3)	Completely (3)	Completely (3)	Somewhat (difficult to determine) (1)	Completely (3)	Somewhat (men only) (1)
DAD	Self-reported (1)	Self-reported (1)	National diabetes registry (3)	Physician-diagnosed/documented intake of antidiabetic medication (3)	WHO criteria (3)	Fasting glucose >6.7 mmol/L, self-reported, taking medication or physician-diagnosed (3)	Self-reported (1)	Self-reported (1)	American Diabetes Association criteria (3)
ANE	Job-exposure matrix (1)	Self-reported need to raise the voice (1)	Geocoding, official sources (3)	Self-reported traffic intensity (1)	Geographic Information System (3)	Nordic prediction method, official sources (3)	Cumulative noise dosimeter (3)	Self-reported need to raise the voice (1)	Noise dosimeters (3)
DE	Complete occupational history (4)	Self-reported duration of exposure (2.5)	Complete residential address history (4)	One-point-in-time assessment of exposure (2)	Calculated or imputed for part-time exposed or those who moved (4)	Complete residential address history + imputed for missing exposures (4)	One-point-in-time assessment of exposure (2)	Self-reported duration of exposure (2.5)	One-point-in-time assessment of exposure (2)
NM	dBA-year (2)	Need to raise the voice (>1-4 years + >5 years) (2)	Lden (4)	Annoyance-based (1)	Lden (4)	LAeq,24h (4)	LAeq,8h (4)	Need to raise the voice ≈ 85 dB (2)	dBA (2)
RDR*	85-95 dBA-year versus <85 dBA-year: OR=0.93 (0.61, 1.41)	n=1108 ≈ ≥85 dB (2.6% with diabetes) versus n=1 890 ≈ <85 dB (3.1% with diabetes)	64-67 dB versus <52 dB: IRR=1.18211 (1.01342, 1.35272)	“Heavy traffic” (≈65-70 dB) versus “no/very rare traffic” (≈ <55-60 dB): OR=1.41 (0.96, 2.08)	≥55 dB versus <50 dB: OR=0.94 (0.33, 2.70)	n=1397 <50 dB (92 with diabetes) versus n=698 ≥50 dB (61 with diabetes)	n=252 ≈ 53.5 dB (18 with diabetes) versus n=137 ≈ 71.5 dB (18 with diabetes)	n=1592 ≈ ≥85 dB (117 with diabetes) versus n=3302 ≈ <85 dB (254 with diabetes)	n=574≥85 dB (30 with diabetes) versus n=1132 <85 dB (69 with diabetes)
BP	Plausible (1.5)	Speculative (0.5)	Plausible (1.5)	Plausible (1.5)	Speculative (0.5)	Speculative (0.5)	Plausible (1.5)	Speculative (0.5)	Speculative (0.5)
APC	Obesity, physical activity, smoking, drinking, hypertension, education, income (3)	No adjustments (0)	Age, sex, BMI, waist circumference, smoking, diet, alcohol, sport, education, occupation, socioeconomic status (3)	Sex, age, smoking, heating of house/flat, education, BMI, waist circumference, sport, history of diabetes (4)	Sex, age, history of diabetes, socioeconomic status, physical activity, smoking, psychological distress, income, unemployment (3)	No adjustments (0)	No adjustments (0)	No adjustments (0)	No adjustments (0)
AEC	No adjustments (0)	No adjustments (0)	No adjustments (0)	Railway and airport noise, air pollution (3)	No adjustments (0)	No adjustments (0)	No adjustments (0)	No adjustments (0)	No adjustments (0)
ESC	OR → RR=0.94143 (0.65425, 1.31641) (1)	Raw data → RR=0.75 (0.48, 1.18) (2)	IRR ≈ RR=1.17732 (1.02013, 1.3345) (2)	OR → RR=1.37546 (0.96236, 1.95096) (1)	OR → RR=0.94 (0.34, 2.57) (1)	Raw data → RR=1.32704 (0.97149, 1.79626) (2)	Raw data → RR=1.83942 (0.98763, 3.24276) (2)	Raw data → RR=0.95540 (0.77244, 1.17752) (2)	Raw data → RR=0.85745 (0.56220, 1.29507) (2)
ATISB	No transformations (3)	No transformations (3)	Data digitalization and extraction from graph (1)	Heavy traffic ≈ 70-80 subjective dB (Lden) (0)	No transformations (3)	LAeq,24h → LA10,18h → Ldn ≈ Lden (0)	LAeq,8h → LA10,18h → Ldn ≈ Lden (0)	No transformations (3)	No transformations (3)
OQS	33	23.5	37	30	30	29.5	20.5	27.5	24

95% CIs in (), *Not included in the quality score, **In regards to the diabetes outcome this study was considered case-control, D = Design, TF = Timeframe, C = Country, MS = Mode of selection, SRR = Study response rate, FSS = Final sample size, P = Participants, R = Representativeness, DAD = Definition and assessment of diabetes, ANE = Assessment of noise exposure, DE = Duration of exposure, NM = Noise metric, RDR = Risk data reported, BP = Biological plausibility, APC = Adjustments for personal covariates, AEC = Adjustments for environmental covariates, ESC = Effect size calculation, ATISB = Additional transformation, imputation or source of bias, OQS = Overall quality score, RR = Relative risk, OR = Odds ratio, BMI = Body mass index, IRR = Incidence rate ratio

Generally all studies had adequate sample sizes and acceptable response rates, but Sørensen et al.[9] presented with the largest sample of over 57,000, which explains the highest weight for this study. Adequate assessment of diabetes was considered either diagnosis according to the standard WHO criteria, by a physician or personal health status obtained from official databases.[912131518] Self-reported diabetes, on the other hand,[2141617] was scored lower since it is not as reliable; furthermore, Rhee et al.[16], Bainbridge et al.[17] and Jang et al.[18] did not provide information regarding the type of diabetes or did not distinguish between type 1 and 2 diabetes mellitus, which should not be a considerable draw-back since 90% of all diabetes cases are attributable to type 2 diabetes[5]. As for noise exposure, higher preference was given to those studies which estimated complete lifetime residential and/or occupational noise exposure[291215] and used objective way to ascertain the exposure, i.e., Geographic Information System and/or valid methods for noise calculation.[9121516] One-point-in-time measurements in longitudinal designs are not quite so accurate.[1316] Heidemann et al.[13] proved problematic since they reported perceived traffic intensity. Noise exposure was heuristically imputed based on educated guess, although there are ways to estimate noise levels from the reported traffic intensity assuming that “reported annoyance is higher where the traffic intensity is higher”[29] and the frequency of extreme annoyance is overwhelmingly higher when extreme road traffic intensity is reported[30] — the percent of highly annoyed people can be converted to perceived decibels.[31] Because this study was somewhat contentiously interpreted, it received low score on this element. The job-exposure matrix method adopted by Song[2] was a time-consuming and sophisticated procedure, but was characterized by some limitations as the author admitted. It was actually not much more reliable than the self-reported frequency of raising the voice[1417] The occupational studies should be interpreted with extra caution when compared to the rest.

Rating the biological plausibility of a significant effect supporting the impact of noise on diabetes involved taking into account both the duration of exposure on which there were data and the thresholds or categories that were compared. If there was only one cut-point[14151718] with unclear upper and lower bounds, or if two lower noise levels with lacking lower bound for one of them were compared[12] the plausibility was scored as “speculative”. However, probably the most important determinant of the credibility of each effect size were the adjustments for relevant confounding factors. Except for Selander et al.,[15] Rhee et al.,[16] Suadicani et al.,[14] Bainbridge et al.[17] and Jang et al.[18] for which RRs were calculated from raw data (those studies did not test the association noise — diabetes directly), the other studies included most if not all of the important correlates of diabetes — mainly BMI/obesity, sporting habits/physical activity, family history of diabetes, demographics and unhealthy practices, i.e., smoking, drinking, poor diet. Conversely, only Heidemann et al.[13] adjusted their analyses for railway and aircraft noise, and air pollution.

Meta-analysis results

A forest plot is presented at Figure 1 and numeric results are reported at Table 2. For the purposes of the meta-analysis all studies were divided into two separate subgroups — Residential and Occupational noise exposure.

Table 2

Individual study effect seizes and pooled risk ratios for residential and occupational noise exposure groups

Residential noise exposure subgroup
Study	Comparosions (Lden [dB])*	RR	95% CI	Weight%
Rhee et al.[16]	64.14 versus 85.92	1.84	0.99-3.24	2.37
Selander et al.[15]	<50.04 versus ≥50.04	1.33	0.97-1.80	8.59
Sørensen et al.[9]	<52 versus 64-67	1.18	1.02-1.33	48.40
Heidemann et al.[13]	<55-60 versus 65-70	1.38	0.96-1.95	6.99
Eriksson et al.[12]	<56 versus ≥61 (61-65)	0.94	0.34-2.57	2.20
Pooled effect (subgroup)		1.22	1.09-1.37	68.55
Occupational noise exposure subgroup
Study	Comparosions (dB**)	RR	95% CI	Weight%
Suadicani et al.[14]	<85 versus >85	0.75	0.48-1.18	3.84
Song[2]	<85 versus 85-95	0.94	0.65-1.32	7.81
Bainbridge et al.[17]	<85 versus >85	0.96	0.77-1.18	15.44
Jang et al.[18]	<85 versus >85	0.86	0.56-1.30	4.36
Peeled effect (subgroup)		0.91	0.78-1.06	31.45
Pooled effect (overall)		1.12	0.95-1.31	100.00

*All exposures reported in different metrics are converted to Lden, **Perceived loudness or job-exposure matrix dB, CI = Confidence interval, RR = Relative risk

Results suggest that in comparison to Lden ≤ 64 dB, residential Lden > 60 dB (on average ranging 60-70 dB) is associated with about 22% higher risk for type 2 diabetes. As expected, the highest weight was distributed to Sørensen et al.[9] For occupational exposure (<85 dB versus >85 dB) the RR did not exceed 1.00 and was not statistically significant. Note that the overall pooled RR did not have much of a meaningful interpretation, but if interpreted it would indicate that there was 12% higher risk for diabetes in the higher exposure group versus the lower exposure group across all the studies. It should also be noted that despite the discrepancies between the studies in each subgroup, no heterogeneity was detected (I2 = 0.00), which might indirectly suggest that the comparison groups and effect size estimates that were extracted were appropriate and compatible.

Because Rhee et al.[16] and Selander et al.[15] had exposure groups difficult to compare to the rest, the pooled RR for the Residential subgroup was re-calculated after their exclusion. The pooled RR for residential exposure remained very close (RR = 1.19, 95% CI: 1.05-1.35), therefore the results were quite convincing, i.e. people exposed to Lden ≥ 60-70 dB have about 19% higher risk for type 2 diabetes when compared to those exposed to Lden ≤ 60 dB. There was a “grey area,” however. Sensitivity analysis indicated that when each individual study is excluded, the pooled RR remained above 1.00 (i.e., there was a risk for diabetes) and moreover statistically significant (95% confidence intervals [CIs] did not include 0.00). Noteworthy, Sørensen et al.[9] was not driving the effect on its own. In all scenarios heterogeneity was absent (I2 = 0.00) [Table 3].

Table 3

Sensitivity analysis for Residential noise exposure group

Excluded study	Pooled RR	95% CI	Heterogeneity
			Q	χ2, P	I2
Rhee et al.[16]	1.21	1.08-1.36	1.25	0.741	0.00
Selander et al.[15]	1.21	1.07-1.37	2.80	0.423	0.00
Sørensen et al.[9]	1.35	1.08-1.68	1.51	0.679	0.00
Heidemann et al.[13]	1.21	1.07-1.36	2.64	0.450	0.00
Eriksson et al.[12]	1.23	1.10-1.38	2.78	0.427	0.00

CI = Confidence interval, RR = Relative risk

Finally, a funnel plot depicted the possibility for selective publishing (regarding residential exposure) [Figure 2]. Due to the limited number of studies (n = 5) in this subgroup no statistical testing for asymmetry was considered, but visual inspection of the funnel plot does not raise immediate concerns. Regardless, this does not fully exclude publication bias.

Figure 2

Funnel plot for studies in Residential noise exposure subgroup (effect size against precision)

Discussion

Key results

Nine relevant studies were identified by the systematic review; they were allocated into two subgroups (Residential and Occupation noise exposure) and included in the meta-analysis. Despite the methodological differences, using a series of transformations, the noise metrics and sources in each of the subgroups were arbitrarily converted to Lden traffic noise in order to facilitate the interpretation of the quantitative results. Similarly, applying the quality effects model accounted for different sources of bias and redistributed the study weights accordingly.

Overall the results suggested that people exposed at their homes to roughly Lden > 60 dB had 22% higher risk for type 2 diabetes in comparison to those exposed to Lden < 64 dB (4 dB “grey area”); when studies reporting contentious exposure categories were excluded, there was still 19% risk for Lden = 60-70 dB versus Lden < 60 dB. Sensitivity analyses did not change the results and no heterogeneity was detected which gave another level of credibility to the selection, transformation and interpretation of the effect sizes from the examined categories. These risks refer to residential noise exposure. In occupational environment there was not significant risk (RR < 1.00) for < 85 dB versus >85 dB.

Strengths and limitations

This was the first study to attempt summarizing the results from published peer-reviewed literature regarding the risk for type 2 diabetes due to noise exposure. In addition, it identified non-peer reviewed literature and some studies which contained useful information but might have been omitted if only MeSH terms were searched. Best efforts were made to extract effect sizes comparable to each other while adhering to the conservative principle. Rigorous quality scoring protocol was developed and each study was given an adequate appraisal. Applying the quality effect meta-analytical model allowed combining methodologically different studies and interpreting the pooled statistics in a meaningful way. As suggested by Doi and Thalib[232425] this model might offer additional benefits over the random effects model making the results more precise especially given the unsatisfactory reporting and unstandardized designs in public health research.

There are some issues which might be sources of bias. On one hand, Zhang and Yu's approach gives precise pooled estimate for the RRs but somewhat narrower 95% CIs.[32] Conversely, in epidemiology, as outlined by Babisch in regards to noise hygiene, the actual statistical significance does not always correspond to scientific and practical importance of the findings (for example, due to insufficient statistical power); therefore “a low relative risk with a small confidence interval that only just fails statistical significance” is of importance for environmental research and future noise policy.[33] That is, we should be more concerned with the pooled point estimate rather than the statistical significance of the 95% CIs around it (to a certain degree of reasonability, of course), so this should not be a considerable limitation.

The fact that no exposure-response relationship could be derived questions whether a quantitative meta-analysis is even necessary as an addition to the systematic review. Although the exposure-response slopes of different studies can be pooled disregarding comparability of noise metrics,[34] the author did not feel confident in the feasibility of this approach given the available studies, and especially given the fact that the relationship is very unlikely to be linear. The other option, that is the categorical approach, requires comparable noise indicators and the same exposure categories across the studies.[34] Babisch has argued that “studies that are not suitable with respect to issues of exposure misclassification, selection bias and observation bias or confounding should be excluded from the meta-analysis, which is the basis for a quantitative risk assessment”.[34] Although the author tends to agree, the inferences made by this meta-analysis are still of particular viability because:

They do not refer to a construction of dose-response relationship or pooling regression slopes;

The discrepancies between the studies and major sources of bias were to a great extend accounted for by the adopted quality effects model, which is still underutilized in epidemiology (there are only few published meta-analyses using this method and none in noise research), although it makes possible meta-analyzing heterogeneous studies;[35]

It suggests that despite exposure assessment, noise metrics and study designs higher residential noise is associated with higher odds for type 2 diabetes in comparison to people exposed to lower residential noise levels.

Important information can often be disregarded or consciously excluded from meta-analyses due to statistical or methodological issues in pooling regression slopes or generating exposure-response functions.

Noteworthy, the transformations of aircraft to road traffic noise were based on annoyance functions and assumed that the main effects of noise are mediated through annoyance. Also it is argued that when a research field is still underexplored (such as the relationship between noise exposure and diabetes) it is reasonable to publish meta-analysis comparing relative quantitative results, i.e., higher exposure versus lower exposure conditions with somewhat fluctuating and contentious ranges of the exposure categories. Ndrepepa and Twardella, for example, published a meta-analysis on the risk for cardiovascular diseases due to noise annoyance comparing higher versus lower perceived exposure categories,[36] which at some instances were considerably more incompatible than the reported in the present study; nevertheless, their study received international recognition and contributed to the state of current knowledge.

As for the number of included studies, from statistical point of view there is no minimum number of studies for a meta-analysis as long as it is ≥2. Heuristic justification can also be found in Davey et al.[37] who reviewed 22 453 meta-analyses from Cochrane Database of Systematic Reviews and found that in the categories applicable to the present study (dichotomous outcome, endocrinology, onset of chronic disease) half of the meta-analyses included only 3 studies and 75% included 6. If we adhere to Horton's Precautionary principle “we must act on facts, and on the most accurate interpretation of them, using the best scientific information. That does not mean that we must sit back until we have 100% evidence about everything” (Horton, 1998 cited by WHO[38]).

Future research

Clearly the inferences made by this meta-analysis are too preliminary to allow adequate estimation of the burden associated with noise-induced type 2 diabetes. Future research should consider not only more robust designs (for example, full noise exposure history, air pollution adjustments, etc.) but also adequate reporting and statistical analyses in order to facilitate data extraction and to enhance the compatibility with previous studies. Even papers dealing with noise or diabetes only as covariates for which the main association has been adjusted for will benefit meta-analysts if the coefficients for those covariates are actually reported. Others who have already published research on air pollution and diabetes, for example, might provide invaluable information regarding the effects of noise if they mine their datasets or link the health outcome to noise exposure.

A more person-centered approach is suggested as well. Currently environmental hygienists and epidemiologists tend to adopt a more or less toxicological interpretation of exposure — response relationship when it comes to noise effects. However, unlike other pollutants noise is a phenomenal entity, that is, the extent to which it affects us is likely deeply rooted in our perceptions, attitudes and judgments. It is speculated that if variables representing psychoacoustical phenomena such as noise sensitivity, perceived acoustical environment, etc. are included in the analyses, some of the unexplained variance could be accounted for by. Besides, if personality and attitude towards noise and noise policy, perceived scientific consensus, etc. actually contribute to the cumulative risk for diabetes, a person-centered prophylaxis (i.e., behavioral counseling, public debates, and educational campaigns) might supplement noise control engineering the same way general stress management protocols are applied to people with diabetes.[39]

Valuable knowledge will be gained if the associations of noise exposure and various indicators of metabolic control of diabetes (fasting glucose, glycated hemoglobin, blood lipid profile, microalbuminuria, etc.) are studied. After all, patients’ lives are not over with the diagnosis or onset of diabetes. Environmental noise might not only increase the risk for this group of metabolic disorders, but affect their course, patients’ quality of life and prognosis.

Similarly, it might be feasible to study the effects of environmental noise on testosterone levels among men with diabetes. There is compelling evidence that in animal models chronic noise exposure, being a prototypal environmental stressor, is associated with significant decrease in plasma testosterone.[40] On the other hand, testosterone deficiency is common in men with diabetes and is associated with increased cardiovascular risk.[41] Therefore testing the complex relationships between noise, cardiovascular diseases, diabetes and hypogonadism might yield some very interesting and important results. In fact, a network meta-analysis might aid addressing the feasibility of this theory.

Another important aspect of future research should be the mutual enhancement of diabetes and noise exposure with respect to sensorineural hearing loss (SNHL).[1842] SNHL results from pathologic changes in structures of the inner ear[43] and diabetes is likely increasing inner ear's susceptibility to noise damage. The high prevalence of hearing impairment among diabetic patients might be mediated by several biological mechanisms grouped into neuropathic, angiopathic, and a combination of the two.[44] In particular, those refer to sclerosis of arteria auditiva interna, capillary thickening of stria vascularis, atrophy of ganglion spirale, demyelination of nervus vestibulocochlearis and impaired brainstem auditory reactivity (for a brief overview see Bainbridge et al.[17] and Siddiqi et al.).[45] Unfortunately, there is little if any evidence to suggest the extent to which diabetes and noise exposure are mutually confounded with respect to SNHL, are moderator effects occurring, etc.

From methodological point of view, the quality effects model will not only benefit future “noise meta-analyses”, but recently it was modified in order to improve risk assessment.[46] Given that current methodology for estimating burden of disease from noise is characterized by a significant degree of uncertainty due to a fair amount of assumptions, data imputations, extrapolations, etc.,[47] adding another level of credibility to the calculations might improve the accuracy of predictions and respective political actions. However, in order to ensure reproducible analyses under the quality effects model, we would need standardized quality scoring checklists designed specifically for the assessment of primary studies in noise hygiene. Furthermore, the development of such checklists under the supervision of expert scientific groups (e.g., in international workshops or conferences) might become a separate research field with important practical applications.

Finally, computing RR in primary research by applying log binomial or log Poisson regressions (although the latter might yield somewhat conservative results) as alternatives to OR and logistic regression,[48] an issue which is already addressed with respect to cardiovascular effects of noise,[49] will facilitate the interpretations of the risk.

Conclusions

People exposed to high residential noise (Lden > 60 dB versus Lden < 60-64 dB) might be at significantly higher risk (19-22%) for developing type 2 diabetes. Occupational noise exposure (>85 dB versus <85 dB) was not associated with higher risk. The results should be interpreted with caution due to methodological discrepancies across the studies; however, they are indicative of the close links that noise might have not only to the cardiovascular diseases but to endocrine dysfunction as well.