Association between ambient and household air pollution with carotid intima-media thickness in peri-urban South India: CHAI-Project

Abstract Background Evidence linking ambient air pollution with atherosclerosis is lacking from low- and middle-income countries. Additionally, evidence regarding the association between household air pollution and atherosclerosis is limited. We evaluated the association between ambient fine particulate matter [particulate matter with an aerodynamic diameter of ≤2.5 µm (PM2.5)] and biomass fuel use on carotid intima-media thickness (CIMT), a surrogate of atherosclerosis, in India. Methods We analysed the third follow-up of the Andhra Pradesh Children and Parent Study cohort (2010–2012), which recruited participants from 28 peri-urban villages. Our primary outcome was mean CIMT, measured using a standardized protocol. We estimated annual average PM2.5 outdoors at residence using land-use regression. Biomass cooking fuel was self-reported. We fitted a within-between linear-mixed model adjusting for potential confounders. Results Among 3278 participants (48% women, mean age 38 years), mean PM2.5 was 32.7 [range 24.4–38.2] µg/m3, and 60% used biomass. After confounder adjustment, we observed positive associations between within-village variation in PM2.5 and CIMT in all participants [1.79%, 95% confidence interval (CI), −0.31 to 3.90 per 1  µg/m3 of PM2.5] and in men (2.98%, 95% CI, 0.23–5.72, per 1  µg/m3 of PM2.5). Use of biomass cooking fuel was associated with CIMT in all participants (1.60%, 95% CI, −0.46 to 3.65), especially in women with an unvented stove (6.14%, 95% CI, 1.40–10.89). The point-estimate for the PM2.5 association was larger in sub-groups with higher cardiometabolic risk profile. Conclusions Ambient and household air pollution were positively associated with CIMT in a peri-urban population of India, although with limited precision for some estimates. We observed differences in the association between ambient and household air pollution and CIMT by gender.

What is the main source of cooking fuel? Options: a) Electricity, b) Gas, c) Kerosene, d) Oil, e) Other.
Note: For multiple sources, field team recorded the predominant one.
Other includes wood, crop residues or dung.

Stove ventilation
We kept the same categories as the questionnaire Those with stove located outside were considered vented.
Was your stove or fire vented to the outside? Options: a) Yes, b) No.
Main source of lighting fuel We categorized those using biomass (c, d, and e) vs no biomass (a and b) What is the main source of lighting for your household? Options: a) Electricity, b) Gas, c) Kerosene, d) Oil, e) Other.
Note: For multiple sources, field team recorded the predominant one. Other includes wood, crop residues or dung. Covariates Primary occupation: a) At home doing housework, b) Unemployed, not seeking work; retired/ disabled, c) Unemployed, seeking work, d) Student/ training, e) Unskilled manual, f) Semi-skilled manual, g) Skilled manual, h) Skilled non-manual, i) Semi-Professional, j) Professional

Standard Living Index
We used the this standard index for socioeconomic status assessment validated for use in India. 1

Metabolic syndrome
At least 3 of the following criteria: 3 a) abdominal obesity, b) systolic blood pressure ≥130 mmHG or diastolic blood pressure ≥85mmHg or antihypertensive intake, c) impaired fasting glucose, d) HDL cholesterol <50 mg/dL for women and <40 mg/dL for men, e) Triglycerides ≥150 md/dL Tobacco smoke We kept the same categories as from the questionnaire.
Have you ever used tobacco on a regular basis (at least weekly)? Options: a) Never, b) Former (stopped over 6 months ago), c) Current (anytime in the last 6 months).
Environmental tobacco smoking We categorized participants in those exposed (a) or not (b).
Is there someone in your household who smokes tobacco at home? Options: a) Yes, b) No.

Alcohol use
We categorized participants in those who drank most days (a and b) and Rarely/Never (c, d and e) Would you describe your present alcohol intake as? (Included locally made spirits / Branded made spirits / Beer / Wine) Options: a) Daily/Most days, b) Weekends, c) 1-2 times/month, d) Special occasions, e) Never.

Diet
The percentage of energy from each food category, as well as fruits and vegetable amount, were derived from a validated questionnaire. 4 Semi-quantitative food frequency questionnaire (FFQ) Physical activity We used the metabolic equivalent unit values (MET) derived from the validated questionnaire. 5 Interviewer-administered quantitative physical activity questionnaire

Further description of statistical analysis
Of the 6944 participants enrolled in the third follow-up of Andhra Pradesh Children and Parent Study (APCAPS) cohort (first clinic visit established within the villages), we included adults (age ≥ 18 years), men and nonpregnant women, with available CIMT measurement. Among those eligible for inclusion (n=6229), 3445 (50%) attended the second clinic visit at the National Institute of Nutrition (NIN) located in Hyderabad , for the cardiovascular risk profile measures (eFigure 1, Study flowchart). There were differences between those who attended the clinic visit at NIN and those who did not. Participants who attended the NIN clinic visit had a higher prevalence of known risk factors for cardiovascular diseases (eTable 2). To adjust for potential selection bias, we used inverse probability weighting (IPW) technique to account for the population representativeness. 6,7 The IPW were applied in two steps a) deriving the probability of attending the NIN for each individual and b) estimating the weight for each participant (1/probability of attending), which was subsequently used in the linear mixed model. Therefore, each participant contribution to the model is weighted, to achieve estimates that aim to be representative of the source population (i.e., participants with a profile that are more likely to attend receive lower weights compared with participants with a profile not attending).
Thus, the final model was fit in a pseudopopulation, generated with the derived weights. 6,7 We followed the published recommendations for the IPW generation. 6,8 The model to derive the attending probability to the NIN was build using a logistic regression model, where the outcome was binary Yes/No (Yesparticipants who attended NIN, NOparticipants who did not attend NIN). We used as covariates variables related to the fact of attending or not, but also related to our main outcome (CIMT), as recommended by the literature. 6,8 The covariates were all those included in the full adjustment model (model 3), with education and occupation as originally collected (education = 6 levels; occupation = 10 levels) and the auxiliary variables systolic blood pressure, diastolic blood pressure, impaired fasting glucose/diabetes, abdominal circumference, and metabolic syndrome criteria. The model also included an interaction term between village-ID, age, and gender, to account for the underlying potential process of not attending the NIN, which we hypothesized to be closely related to each village characteristics (e.g., distance to the NIN), and the participant age and gender accounted within villages. Sensitivity analysis using different auxiliary variables (e.g. distance to the NIN instead of village-ID) showed similar results (data not shown).
We assessed the IPW building process and met all suggested requirements: a) We observed a high degree of overlap between weights among participants who attended or not the clinic, b) We did not observe problems with large weights, 6,8 and c) we used the recommended additional tests to evaluate the model (Hosmer-Lemeshow, p=0.885; Hinkley´s method, p=0.791). 6,8 We followed the published recommendations for the multiple imputation (MI) followed the IPW generation. 6,8,9 We conducted multiple imputation for covariates using multivariate chained equation methods using the mice package in R. 9, 10 We investigated the missingness pattern and assumed a Missing at Random (MAR) mechanism.
We used all covariates used in the full adjustment model (model 3), the outcome, and auxiliary variables. We generated 10 imputed datasets, with 50 iterations, and the estimates were pooled following the Rubin´s rule. We checked the variables distribution and convergence. To allow for the hierarchical structure of the data, we used the predictive mean matching method for all covariates and entered the village-ID as dummy variables. 11,12 All analyses were conducted with R-3.4.2, 13 with the packages tidyverse, 14 mice, 10 miceadds, 15 lme4, 16,17 metafor, 18 ggplot2, 19 and forestplot. 20 eFigure 1. Association between village-average PM 2. 5

and CIMT
The blue line is the linear regression between village-average PM2.5 and village-average CIMT. The gray zone is the 95% CI. CIMT = carotid intima-media thickness; PM2.5 = particulate matter with an aerodynamic diameter of 2.5 micrometers or less.  4.91 (-0.11, 9.93) Analysis conducted in 10 multiple imputed datasets, using a linear mixed model accounting for within-between effects, with correction for selection bias through inverse probability weighting. CI = confidence interval; CIMT = carotid intima-media thickness; HAP = household air pollution; PM2.5 = particulate matter with an aerodynamic diameter of 2.5 micrometers or less.
* Model 3 was adjusted by age (modelled with natural spline, df=3), gender, occupation, education, standard living index, body-mass index, fruits and vegetables consumption, smoking status and environmental tobacco smoke, alcohol consumption, physical activity, biomass fuel use and whether stove was vented to the outside. The models for women did not include active smoking and have an interaction term between biomass fuel use and whether stove was vented to the outside. † Model 4: Model3 + blood pressure, impaired fasting glucose and non-HDL cholesterol. eTable 4. Association of within-village variation in PM2.5 and biomass fuel use with mean and maximum carotid intima-media thickness in all participants and stratified by gender (sensitivity analysis: multiple imputed data without correction for selection bias). .98) Analysis conducted in 10 multiple imputed datasets, using a linear mixed model accounting for within-between effects, without correction for selection bias. CI = confidence interval; CIMT = carotid intima-media thickness; HAP = household air pollution; PM2.5 = particulate matter with an aerodynamic diameter of 2.5 micrometers or less. * Model 1 was adjusted by age (modelled with natural spline, df=3) and gender. † Model 2 was additionally adjusted by occupation, education, standard living index, body-mass index, fruits and vegetables consumption, smoking status and environmental tobacco smoke, alcohol consumption, and physical activity. The models for women did not include active smoking. ‡ Model 3 was additionally adjusted by biomass fuel use and whether stove was vented to the outside. The models for women have an interaction term between biomass fuel use and whether stove was vented to the outside. eTable 5. Association of within-village variation in PM2.5 and biomass fuel use with mean and maximum carotid intima-media thickness in all participants and stratified by gender (sensitivity analysis: complete case analysis, with and without correction for selection bias). .05) Analysis conducted in complete case data, using a linear mixed model accounting for within-between effects, with and without correction for selection bias through inverse probability weighting. CI = confidence interval; CIMT = carotid intima-media thickness; HAP = household air pollution; PM2.5 = particulate matter with an aerodynamic diameter of 2.5 micrometers or less.
* Model 3 was adjusted by age (modelled with natural spline, df=3), gender, occupation, education, standard living index, body-mass index, fruits and vegetables consumption, smoking status and environmental tobacco smoke, alcohol consumption, physical activity, biomass fuel use and whether stove was vented to the outside. The models for women did not include active smoking and have an interaction term between biomass fuel use and whether stove was vented to the outside.