Data
We used data from the NFHS-4, a nationally representative survey that collected health and sociodemographic information of reproductive age women from 640 districts in all the 29 states and seven union territories (total 36) in India using a stratified two-stage sampling framework [14]. Participation in the NFHS-4 was voluntary and consent was obtained prior interview; the survey protocols were reviewed and approved by the Institutional Review Boards of International Institute for Population Sciences and ICF and further reviewed by the US Centers for Disease Control and Prevention [14]. We used anonymized publicly available data for analysis. Our analytical sample contains hypertensive women aged 20 to 34 years, who were married at the time of the survey. Among 267,306 married women aged 20 to 34 years in the NFHS-4, a total of 22,140 were categorized as hypertensive, which constitutes our sample (Fig. 1). The methods were carried out in accordance with the “US Department of Health and Human Services regulations for the protection of human subjects” and relevant national guidelines.
Measures
The NFHS-4 reports respondents’ average systolic blood pressure (SBP) and diastolic blood pressure (DBP) measures. Blood pressure was measured three times during a single visit with at least 5 min interval between each reading. Respondents were also asked if they were taking any antihypertensive medication to lower their blood pressure. A respondent was categorized as hypertensive if average SBP ≥ 140 mmHg or the average DBP ≥ 90 mmHg or the respondent reported taking antihypertensive medication at the time of the survey. An individual was determined to have untreated hypertension if the average blood pressure measure exceeded the normal threshold, and the individual was not taking antihypertensive medication at the time of the survey.
The NFHS-4 also reports respondents’ age at first marriage. Women who were married before the age of 18 years were identified as child brides. Age at first marriage information was only available for those who were currently married, and was not available for those who were widowed, divorced, or separated at the time of the survey.
Statistical analysis
We estimated univariate and multivariable logistic regressions to obtain odds ratios (ORs) and adjusted ORs (AORs) in favor of not receiving hypertension treatment. Our dependent variable is a binary variable indicating if the respondent received hypertension treatment or not. The key explanatory variable is another binary variable indicating whether the respondent was married as child (i.e., before age 18 years) or as adult (i.e., at or after age 18 years).
In multivariable logistic model, we controlled for various sociodemographic correlates including age in 3-years interval: 20 to 22 (reference category), 23 to 25, 26 to 28, 29 to 31, and 32 to 34; education attainment: no education (reference category), primary, secondary, and higher; relationship to household head: head (reference category), wife, daughter, daughter-in-law, and other; parity or number of children born: 0 (reference category), 1 to 2, 3 to 4, and 5 + ; current pregnancy status; current breastfeeding (lactation) status; household size: 3 or less (reference category), 4 to 5, 6 to 8, and 9 + ; household wealth index quintiles: poorest (reference category), poorer, middle, richer, and richest; religion: Hindu (reference category), Muslim, Christian, Sikh, Buddhist, and other; caste: not socially or economically backward class (reference category), scheduled caste, scheduled tribe, other backward class; and residence: rural (reference category) and urban. To account for state level differences in health policy and health care access, we also controlled for state fixed effects.
We first estimated the crude ORs and AORs in favor of having untreated hypertension for each of these sociodemographic characteristics in subgroups of women who were married as adults and who were married as children. We then performed Chow test to examine whether the crude ORs and AORs for respective sociodemographic characteristics differ between the two groups. Estimates were obtained using complex survey weights and the level of significance was set to 0.05.
Next, to assess the relationship between child marriage and untreated hypertension, we estimated a univariate specification (model 1) only including the child marriage indicator and a constant. Subsequently we estimated four multivariable specifications as follows: model 2 includes individual level correlates (i.e., age, educational attainment, relationship to household head, parity, current pregnancy status, and current lactation status); model 3 includes household level correlates (i.e., household size, household wealth index quintiles, religion, caste, and residence); model 4 includes both individual and household level correlates; and model 5 includes state fixed effects in addition to individual and household level correlates.
Next, we offered two robustness checks of our analyses. First, instead of the binary child marriage indicator, we used the length of marriage as the key explanatory variable. Since the length of marriage, especially in the context of child marriage, varies broadly across age groups, we standardized the length of marriage using the following formula:
\(SML_{i,a\,} = \,\frac{{ML_{i,a} - \overline{{ML_{a} }} }}{{STDV_{a} }}\), where, SMLi,a is the standardized length of marriage for individual i of age group a, MLi,a is the actual length of marriage of individual i of age group a, \(\overline{{ML_{a} }}\) is the average length of marriage in age group a, and STDVa is the standard deviation of length of marriage in age group a. We estimated models 1 to 5 to assess how one standard deviation increase in length of marriage is associated with the likelihood of having untreated hypertension in our sample.
Second, exploiting the hierarchical nature of NFHS-4 data, we performed a multilevel analysis to account for potential bias in standard errors emanating from clustering of data. Following Jain et al. [15], we estimated a multilevel logistic regression model where individual (level 1) is nested within community (level 2)—defined by Census Enumeration Blocks in urban areas and villages in rural areas, district (level 3), and state (level 4). We thus fitted a four-level random intercept model. Since we are primarily interested in examining the relationship between child marriage and untreated hypertension, we did not examine whether or how community level variables impact the individual level outcome (i.e., untreated hypertension) nor we explored the extent of relative contribution of different levels.