Effects of a Liquefied Petroleum Gas Stove Intervention on Gestational Blood Pressure: Intention-to-Treat and Exposure-Response Findings From the HAPIN Trial

Background: Approximately 3 to 4 billion people worldwide are exposed to household air pollution, which has been associated with increased blood pressure (BP) in pregnant women in some studies. Methods: We recruited 3195 pregnant women in Guatemala, India, Peru, and Rwanda and randomly assigned them to intervention or control groups. The intervention group received a gas stove and fuel during pregnancy, while the controls continued cooking with solid fuels. We measured BP and personal exposure to PM2.5, black carbon and carbon monoxide 3× during gestation. We conducted an intention-to-treat and exposure-response analysis to determine if household air pollution exposure was associated with increased gestational BP. Results: Median 24-hour PM2.5 dropped from 84 to 24 μg/m3 after the intervention; black carbon and carbon monoxide decreased similarly. Intention-to-treat analyses showed an increase in systolic BP and diastolic BP in both arms during gestation, as expected, but the increase was greater in intervention group for both systolic BP (0.69 mm Hg [0.03–1.35]; P=0.04) and diastolic BP (0.62 mm Hg [0.05–1.19]; P=0.03). The exposure-response analyses suggested that higher exposures to household air pollution were associated with moderately higher systolic BP and diastolic BP; however, none of these associations reached conventional statistical significance. Conclusions: In intention-to-treat, we found higher gestational BP in the intervention group compared with controls, contrary to expected. In exposure-response analyses, we found a slight increase in BP with higher exposure, but it was not statistically significant. Overall, an intervention with gas stoves did not markedly affect gestational BP.


Supplementary Information
. Summary of missing and invalid exposure measurements 8 Table S2. Results of ITT analyses testing for the difference between final and baseline SBP and DBP for intervention and control arms in each IRC 8 Table S3. Results of long-term exposure-response analyses for the difference between final and baseline MAP/PP across all IRCs 8 Table S4. Results of short-term exposure-response analyses for MAP/PP across all IRCs using all three BP measurements 9 Table S5. Results of long-term exposure-response analyses results for SBP and DBP in Guatemala IRC 10 Table S6. Results of long-term exposure-response analyses results for SBP and DBP in India IRC 11 Table S7. Results of long-term exposure-response analyses results for SBP and DBP in Peru IRC 12 Table S8. Results of long-term exposure-response analyses results for SBP and DBP in Rwanda IRC 13 Table S9. Results of short-term exposure-response analyses results for SBP and DBP in Guatemala IRC 14 Table S10. Results of short-term exposure-response analyses results for SBP and DBP in India IRC 15 Table S11. Results of short-term exposure-response analyses results for SBP and DBP in Peru IRC 16 Table S12. Results of short-term exposure-response analyses results for SBP and DBP in Rwanda IRC 17 Table S13. Effect modification by IRC, baseline gestational age, maternal age, baseline BMI for ITT analyses of SBP, DBP, PP, and MAP 18 Table S14. Effect modification by baseline gestational age, maternal age, baseline BMI for the association between PM2.5/BC/CO exposure and SBP based on log linear models in Guatemala 19 Table S15. Effect modification by baseline gestational age, maternal age, baseline BMI for the association between PM2.5/BC/CO exposure and SBP based on log linear models in India 20 Table S16. Effect modification by baseline gestational age, maternal age, baseline BMI for the association between PM2.5/BC/CO exposure and SBP based on log linear models in Peru 21 Table S17. Effect modification by baseline gestational age, maternal age, baseline BMI for the association between PM2.5/BC/CO exposure and SBP based on log linear models in Rwanda 22 Table S18. Results of ITT analyses testing for the difference between intervention and controls arms for repeated measures of BP after randomization, across IRCs 23 Table S19. Results of ITT analyses testing for the difference between intervention and controls arms for average postrandomization BP, across IRCs 23 Table S20. Personal 24-hour PM2.5 exposure (µg/m 3 ), BC exposure (µg/m 3 ) and CO (ppm) for mothers at baseline IRC (valid measurements only) 24 Table S21. Summary of SBP, DPB (mmHg) and gestational age (day) at baseline IRC 24 Table S22. Unadjusted long-term (a) and short-term (b) exposure-response analyses between PM2.5/BC/CO exposure and SBP/DBP 25 Figure S1. CONSORT flow chart showing HAPIN trial profile and analytical population of current analysis 26 Figure S2. Boxplots of personal exposure to PM2.5, BC and CO by intervention groups and visit 27 Figure S3. Systolic blood pressure by time since randomization (in days) and locally weighted scatterplot smoothing (LOWESS) curves in each IRC. 28 . Line plot of systolic/diastolic blood pressure change from baseline to follow-up 1 and from follow-up 2 by study arm (trial-wide) 32 Figure S8. Line plot of systolic/diastolic blood pressure change from baseline to follow-up 1 and from follow-up 2 by study arm and by IRC 33

Intention-to-Treat and Exposure-Response Analysis for Gestational Blood Pressure
This document contains the data analysis plan for gestational blood pressure of the HAPIN Study. Gestational blood pressure is one of the secondary outcomes. The goal is to avoid data-driven analyses during and at the end of the study to the extent possible.
Primary analyses include intention-to-treat (ITT) analysis and exposure-response (E-R) analysis. Secondary analyses include those that evaluate effect modification and sensitivity analyses -alternative health model specifications, alternative outcome definitions, consideration for missing data, and additional exclusion criteria).

Baseline Participant Characteristics
For the analysis, baseline characteristics will be summarized by intervention versus control arms. Means, standard deviations, and range will be calculated for continuous variables and frequencies and percentages will be calculated for categorical variables. Missing data will be reported as a separate category. We do not expect any imbalance between arms, and do not plan to include these variables in our ITT analysis, but we provide them as descriptive. We will also check that there is no imbalance in potential confounding variables (see below for potential confounders, Table 5). If there is imbalance (p<0.05 for a test between arms) for potential confounders we may consider inclusion of potential confounding variables in the ITT analysis.

Outcome Summary
Outcomes: Systolic Blood Pressure (SBP) and Diastolic Blood Pressure (DBP) in mmHg.
Gestational blood pressure was assessed on enrollment (baseline, <20 weeks' gestation), and two follow up visits at approximately 24-28 gestational weeks (follow-up 1) and 32-36 gestational weeks (follow-up 2). At each measurement period, seated and resting blood pressure was measured in triplicate with right arm, using an automatic digital blood pressure machine, and the average of the three readings was used in the analysis. Field workers confirmed that the pregnant women participants had not smoked, nor had alcohol or caffeinated drinks or cooked using biomass in the 30-minute period prior to the blood pressure measurement. If a participant was found to have a SBP >= 140 mmHg and/or a DBP >= 90 mmHg, she was referred to the nearest health center or hospital.
SBP values less than 70 and DBP values less than 35 were excluded as implausible. There were no implausible high vales. 14 participants on antihypertensive medication at any time of the pregnancy will be excluded from the analysis. SBP and DBP will be summarized by intervention status and visits in means, standard deviations. The number of measurements and summary of corresponding gestational age at each visit will also be reported in the table.
At the same time as the blood pressure measurements, we measured personal samples of fine particulate matter with aerodynamic diameter ≤ 2.5μm (PM2.5), black carbon (BC) and carbon monoxide (CO). Only valid exposure samples will be used in data analysis. PM2.5, BC, and CO will be summarized by visits in means, standard deviations, medians and IQRs (Table 3). The number of valid, and percent of all measurement (i.e., including invalid/missing samples) will also be reported in the supplementary table 1, where the total N will be the number of women with available blood pressure measurement. Correlations between person PM2.5, BC, and CO exposures will be reported as well.

Intention-to-Treat Analysis
The Table below summarizes the intention-to-treat (ITT) analysis methods for SBP/DBP. All analyses will adjust for 10 randomization strata using dummy variables. For the ITT analyses of any outcome, no baseline covariateadjusted effects will be estimated, except for centered baseline BP measurement. The change in BP from baseline to final measurement will be used as the outcome, adjusting for randomization strata and baseline BP. The model is below: ℎ = 0 + 1 1 + 2 2 + … + 11 10 + where for individual , ℎ is the difference between baseline and final (follow-up 2) blood pressure (either SBP or DBP), 1 is an indicator variable (0 for control and 1 for intervention), 2 through 11 are indicator variables for 10 randomization strata, and ~(0, 2 ) represents independent normal error. The parameter of interest 1 captures differences in the change of BP from baseline to follow-up 2 due to the intervention. The above ITT model assesses the effects of study arm on gestational BP over the gestational period under observation Additional secondary ITT analysis. We also conducted ITT analyses using 1) a mixed model with two repeated measures of post-randomization BP, controlling for baseline BP, and 2) a linear regression model with no repeated measures comparing the average of two post-randomization BPs between arms, again controlling for baseline BP.
Additional potential secondary ITT analysis. If imbalance between control and intervention groups for baseline covariates which are potential confounder (Table 1) suggests problems with randomization, and the covariate is a potential confounder (see below), covariate-adjusted effects will be evaluated as a sensitivity analysis.
Missing Data. Our primary approach to missing outcome data will be a complete-case analysis by excluding participants without a baseline BP measurement or without any post randomization BP measurements. It is anticipated that missing GBP data will be infrequent and will be balanced between intervention arms. Effect modification. We will carry out separate analyses by IRC, mother's age, mother's baseline BMI, and mother's baseline gestational age, the latter three variables dividing into 2 strata by their median.

Exposure-Response Analysis
For each household air pollutant (PM2.5, BC, and CO), 24-hour personal exposure measurements at baseline, first and second follow-up visit will be used. Only valid exposure samples will be included in the exposure-response analysis.
We plan to use two different models to assess the longitudinal exposure-response relationship between personal PM2.5/BC/CO exposures and gestational SBP/DBP. The first model, which we call the long-term model because it estimates the effect of exposure over the entire gestational period, mimics the ITT model described above. We will use linear regression to model the difference between the first and the final BP measurement (i.e., change score) during pregnancy in relation to average HAP exposure during pregnancy, controlling for gestational age (measured via ultrasound) at the final BP measurement, and other covariates. In this model, the average HAP personal exposure level during pregnancy will be calculated as 1) a simple average of all available measurements for controls and 2) the weighted average of baseline exposure level and the average of post-baseline measurements for the intervention group, with the weight for the baseline measurement being the gestational age before randomization, and the weight after baseline exposure measurement being the duration of gestation during the intervention. The model for the long-term exposure-response analysis is: is the change score (difference between the first and final gestational BP level) for participant , 0 is the population intercept, 1 is the exposure coefficient of interest, is the average PM2.5/BC,/CO exposure over gestation (log transformed, as these fit better than untransformed), are time-independent covariates, and is the model residual, assumed to be normally distributed. The second exposure-response model will be a short-term model, which estimates the effect of exposure just before the BP measurement. This will be a mixed-effects analysis of repeated measures, where we will regress the three measurements of BP on the three measurements of exposure (exposure and BP were measured at the same visit, across all three visits). We will include a random intercept for each individual, time-varying gestational age, and gestational age squared at each BP measurement, and other time invariant covariates. The short-term exposure-response model is: where is the BP level for participant at observation ; 0 is the population intercept; 1 is the exposure coefficient of interest; is either PM2.5, BC, or CO for participant at observation ; are timedependent covariates; are time-independent covariates; is the individual random intercept; and is the model residual, both of which are assumed to be normally distributed. In an additional analysis, we also included an interaction term between gestational age and HAP exposure to determine whether the effect of HAP exposure increased over time. Covariate selection for exposure-response will be based on conceptual directed acyclic graphs (DAGs), the associated minimal set to eliminate confounding, and previous studies. We will consider the following covariates. Those with an asterisk (*) are considered a priori potential confounders will be included in all models. Other variable below will be retained in the model only if their inclusion alters the exposure-response coefficient by 10% of more.

Morning/Afternoon * Categorical
Time of the BP measurement  Morning (before 12:00 PM)  Afternoon (at or after 12:00 PM) Household food insecurity score*

Categorical
Categories (corresponding score): High, medium, low ** Gestational age at BP measurement will be considered via 1) a linear and quadratic term, or 2) a spline term, depending on which fits best, as judged by AIC. Gestational BP is known to follow a U-shaped pattern during pregnancy which should not be modeled by a linear term.
For both long-term and short-term exposure-response models, we first ran separate models for each IRC (see supplemental tables), and then combined estimates using a default random-effects combined measure, except when heterogeneity across the four IRCs was so minimal that a random effects analysis was not possible (i.e., when the Q statistic assessing heterogeneity was less than the degrees of freedom (df = 3 , in which case we calculated a fixed effects combined measure using the inverse variance method (DerSimonian R, Laird N. Metaanalysis in clinical trials. Controlled Clinical Trials. 1986;7(3):177-188).
Additional Secondary Analysis. The following sensitivity analyses will be conducted, controlling for IRC and using all IRC-covariate interactions:  Include those missing baseline BP or any post-baseline BP using the same analytic approach described above (adds 5-6% to sample size)  Baseline SBP/DBP and the average of P1 and P2 visit SBP/DBP, in relation to baseline PM2.5/BC/CO the average of P1 and P2 visit PM2.5/BC/CO.  The effect of the average of baseline, P1, and P2 visit PM2.5/BC/CO, in relation to SBP/DBP at P2, controlling for SBP/DBP at baseline.
Missing Data. For missing outcome, a complete-case analysis will be carried out by excluding participants without a baseline BP measurement or without any post randomization BP measurements. Missing confounder information will be addressed with the use of a missing categorical variable for each covariate (i.e., the missing by indication approach). Effect modification. We will carry out separate analyses by IRC, mother's age, mother's baseline BMI, and mother's baseline gestational age, the latter three variables dividing into 2 strata by their median.

Analysis Replication Plan
All components of the ITT and E-R analyses, including all secondary and sensitivity analyses, will be conducted independently by Kyle Steenland and Wenlu Le using SAS and R, respectively.
Specific analysis results to be compared include: 1. Summary statistics (e.g., mean, standard deviation, frequencies, percentages, proportion missing) in the baseline characteristic table, exposure, and outcome summary table. 2. Intention-to-treat analyses and additional secondary analyses according to models specified in Section 4. Exposure-response analyses and additional secondary analyses according to models specified                     Table S22. Unadjusted long-term (a) and short-term (b) exposure-response analyses between PM2.5/BC/CO exposure and SBP/DBP  Figure S2. Boxplots of personal exposure to PM2.5, BC and CO by intervention groups and visit (BL: baseline, P1: follow-up 1, and P2: follow-up 2) in each IRC. Dark red dashed lines in the PM2.5 and CO panels indicate the 2021 WHO recommended interim target 1 (IT-1) for annual PM2.5 (35 µg/m 3 ), and 24-hour CO (6.006 ppm = 7 mg/m 3 , at 20 °C and 1013 hPa, 1 mg/m 3 = 0.858 ppm). All plots represent 97% of the exposure data.