The Population-Level Gains in Life Expectancy from Improved Blood - - PDF document

The Population-Level Gains in Life Expectancy from Improved Blood Pressure Control in Indonesia: A Parametric G-Formula Approach

Nikkil Sudharsanan Harvard University

Abstract Hypertension is the leading risk factor for mortality in many low- and middle-income countries, causing stroke, ischemic heart disease, and chronic kidney disease. Although the importance of controlling hypertension has been extensively recognized in the scientific community, rates of hypertension diagnosis, treatment, and control continue to remain low in many LMICs. Therefore, controlling blood pressure has the potential to be a powerful and cost-effective way of achieving large, unrealized improvements in mortality at the population level. Yet to date, the expected size of life expectancy gains from improving blood pressure control in many LMICs remains unclear. Using nationally representative longitudinal data on Indonesian adults, I combine epidemiological and demographic models to estimate the gains in adult life expectancy that would result from improving blood pressure control. I also investigate the distributional effects of blood pressure control by estimating the gains in life expectancy by quintiles of wealth. Finally, I test the sensitivity of my

• bservational estimates to unobserved confounding through simulations. I find that improving

blood pressure control would result in a four-year, gain in life expectancy at age 40 for both men and women. The size of this change is substantial: as a reference, a four year change in life expectancy at 40 corresponds to the difference between life expectancy at age 40 in Indonesia between 1960 and 2010, or equivalently, 50 years of progress in improving adult mortality. Second, I find that the benefits of blood pressure control are not concentrated within any single wealth-strata

• f the Indonesian population, but rather are equally distributed across rich and poor sub-
• populations. Based on the results of a simulation-based sensitivity analysis, I find that even under

high levels of unobserved residual confounding, the size of the gain in adult life expectancy from improving blood pressure control remains large. Overall, my results suggest that across LMICs, improving blood pressure control has the potential to result in large improvements in longevity. Keywords: blood pressure, hypertension, mortality, life expectancy, longevity, epidemiology, demography, Indonesia, low- and middle-income countries.

Introduction Hypertension is the leading risk factor for mortality in many low- and middle-income countries, causing stroke, ischemic heart disease, and chronic kidney disease. While hypertension is often thought of as a disease of high-income individuals and countries, numerous studies have documented extremely high levels of hypertension and hypertension-related mortality in lower- income, often pre-nutritional transition, populations (1-4). Over the coming decades, this burden will only continue to increase as LMICs undergo rapid population aging. Although the importance of controlling hypertension has been extensively recognized in the scientific community (4), rates of hypertension diagnosis, treatment, and control continue to remain low in many LMICs (3, 5). Since hypertension is strongly related to mortality, extremely prevalent, and poorly controlled, controlling blood pressure has the potential to be a powerful and cost- effective way of achieving large, unrealized, improvements in mortality at the population level. Yet to date, the expected size of life expectancy gains from improving blood pressure control in many LMICs remains unclear. My study aims to fill this gap in the literature by estimating the gains in life expectancy that would result from improving blood pressure control in Indonesia. Indonesia is an important context to study hypertension since it is the third most populous LMIC, is rapidly aging, and has extremely high rates of uncontrolled hypertension. In addition to estimates for the overall population, I also investigate the distributional effects of blood pressure control by estimating the gains in life expectancy by quintiles of wealth. This approach reveals whether the benefits of improving blood pressure control are concentrated among wealthier individual (as is often assumed) or equally distributed across the population. I address the challenge of estimating population-level effects by

first using nationally representative longitudinal data to estimate the individual-level relationship between blood pressure and mortality then aggregating this effect across individuals using a demographic reweighting approach (also known as a parametric g-formula). I also investigate the sensitivity of these observational estimates to unobserved confounding through a simulation-based sensitivity analysis. As LMICs, such as Indonesia, continue to age over the coming decades, estimating the population-level benefits of improved blood pressure control is essential for setting health policy priorities and informing health decision-making at national and sub-national levels. Background Blood pressure in low- and middle-income countries Although high blood pressure if often believed to be a condition of high-income, or post nutrition- transition, populations, many studies document high levels of hypertension in less developed contexts (1-5). For example, the prevalence of hypertension among adults over the age of 40 is 57.1% in Ghana, 32.3% in India, and 58.2% in Mexico (3). Similarly, hypertension and hypertension- related mortality are not clustered within richer segments of the population within LMICs: in Malawi, the prevalence of uncontrolled hypertension is greater than 40% among the rural poor and hypertension-related stroke is the leading cause of death in one of the poorest districts of rural Maharashtra (2). Indonesia in particular has very high levels of hypertension coupled with low levels of blood pressure treatment and control. Hussain et al. (2016) find that nearly 50% of individuals above the age of 40 in Indonesia have hypertension. More importantly, only 30% of these individuals are aware

• f their hypertension. While around 70% of these individuals report taking medication, less than

25% achieved blood pressure control (5). Importantly, hypertension is not a recent phenomenon in

Indonesia: Witoelar et. al. (2012) show that these levels have remained virtually unchanged since 1997 (6). Blood pressure and individual-level mortality Descriptively, many studies find a continuous increasing relationship between blood pressure (especially systolic blood pressure) and mortality. For example, a meta-analysis of 61 prospective cohort studies concludes that a 20-mmHg increase in systolic blood pressure or a 10-mmHg increase in diastolic blood pressure is associated with a more than two-fold increase in stroke and ischemic heart disease mortality rates (7). Importantly this relationship continues to be large and significant across the range of both age and blood pressure. Moving beyond observed associations, the results from numerous clinical trials have established the causality of this relationship by showing that lowering blood pressure through a combination of medication and lifestyle treatments causes large reductions in cardiovascular and all-cause mortality. For example, a recent meta-analysis of 42 blood pressure clinical trials finds a near linear relationship in the level of blood pressure reduction and mortality: individuals who reduced their systolic blood pressure down to 120-124 mmHg had 27% lower all-cause mortality compared those who achieved a blood pressure of 130-134 mmHg, 41% lower than those with a blood pressure between 140-14 mmHg, and 53% lower than those with a blood pressure of 160 mmHg or more (8). A similar meta- analysis of 112 blood pressure lowering trials finds that a 10-mmHg reduction in systolic blood pressure causes a 13% reduction in all-cause mortality regardless of baseline systolic blood pressure (9).

Blood pressure and population-level mortality While these studies have established the salience of hypertension for individual survival, they do not easily inform population health priorities. This is because individual-level risk estimates do not take into account the prevalence and distribution of hypertension in a population. For example, if hypertension substantially increased the risk of mortality but was not very prevalent, then policies aimed at controlling blood pressure would not produce large gains in longevity at the population

• level. In contrast, if some other risk factor displayed a weaker individual-level relationship with

mortality but was far more prevalent, treating this risk factor could potentially produce far greater population-level gains in longevity. Currently, the most widely used measure of population burden is the disability adjusted life year (DALY). While DALYs provide an overall measure of burden, they are difficult to interpret in terms

• f counterfactual policy questions such as “how many years of life expectancy would be gained from

improving blood pressure control?” Second, they do not provide information on the social distribution of a disease within a country (e.g. are the gains in life expectancy from hypertension treatment equally distributed across socioeconomic groups or does the burden of hypertension disproportionately affect some sub-populations?). Since the goal of many health policies is to maximize equity in addition to improving population health, DALYs are less than ideal for decision making since they are difficult to use to understand the social distribution of proposed intervention effects. In contrast to the DALY, other approaches use models to simulate population-level morbidity and mortality under various levels of risk factors. For example, the IMPACT model developed by Capewell et. al. (1998) combines information on trends in population-level risk factors with

information on the relationship between risk factors and outcomes to decompose changes in coronary heart disease mortality over time into the contribution of changes in specific risk factors (10). Applications of the IMPACT model in the United States, for example, attribute 44% of the decline in coronary heart disease mortality between 1980 and 2000 to reductions in risk factors including total cholesterol, systolic blood pressure, smoking, and physical inactivity (11). These types

• f models can also be applied prospectively to generate policy counterfactuals: for example,

Capewell et al. apply the IMPACT model prospectively to the US population for the year 2000 and conclude that half of coronary heart disease deaths could be averted by 2010 if the population were to achieve the Health People 2010 goals (no smoking, lowered total cholesterol, lowered blood pressure, and lowered body mass index) (12). My approach follows the spirit of the IMPACT model by first using nationally representative data to estimate the individual-level relationship between blood pressure and mortality, setting a counterfactual distribution of blood pressure, and then aggregating these counterfactual estimates across individuals to generate population-level inferences. Data and Methods Empirical approach My main question of interest is how many years of life expectancy will be gained if all individuals in Indonesia were brought to controlled levels of blood pressure? To begin, I first estimate the individual-level effect of uncontrolled hypertension on mortality. I then use a population-level reweighting technique, known as a parametric g-formula, to aggregate the individual-level effects to the population level under the policy scenario of full control of blood pressure. I express and estimate the individual and population-level effects using a potential outcomes framework. This approach explicitly identifies the causal question of interest and which quantities are observed and which are counterfactual. The potential outcomes framework also allows me to simulate the effects

• f unobserved confounding to determine how sensitive my results are to omitted variables.

To begin, define, Dij as a Poisson random variable that measures the number of deaths in a population of sex j between ages i and i+1. Since Dij depends on the size of the population, it is useful to divide by the person-years of exposure (Lij) to estimate the age- and sex-specific mortality rate (in life table calculations, this rate is often represented in as mx or 1mx). Based on this, E(Dij)/Lij is the observed age- and sex-specific mortality rate in the population; since Lij is assumed to be fixed, from here on define E(Dij)/Lij as E(Mij). Now consider the counterfactual scenario where everyone in the population remains the same except that blood pressure is controlled for those with uncontrolled hypertension. Define the age- and sex-specific mortality rate that results under this counterfactual scenario as E(Mij

*). The average effect of removing uncontrolled hypertension on

mortality can now be defined as the following contrast:

!(!!"

∗ ) − !(!!")

To explicitly reveal the potential outcomes, expand E(Mij

*) by conditioning on hypertension status:

! !!"

∗

= !! !!" ℎ!" = 0 ℎ!" = 0 Pr !"# = 0 + !! !!" ℎ!" = 0 ℎ!" = 1 Pr !"# = 1 .

Here, the statement in parentheses after M represents setting hypertension to the specified value; for example, Mij(hyp=0)|hyp=0 is the mortality rate for those without hypertension if they were set to not having hypertension (this potential outcome can be observed). Conditioning on hypertension status reveals that E(Mij

*) is a weighted average of the mortality rate for those who do not have

uncontrolled hypertension, E(Mij(hyp=0)|hyp=0), and the mortality rate for those who do have uncontrolled hypertension if they were set to not having uncontrolled hypertension, E(Mij(hyp=0)|hyp=1). While the first term is observable, the second term is counterfactual, since the potential outcome E(Mij(hyp=0)|hyp=1) is not observable. The causal inference problem is to find a sufficient set of covariates X1,…,Xn such that

! !!"(ℎ!" = 0) ℎ!" = 1, !!, … , !! = !!(!!"|ℎ!" = 0, !!, … , !!).

Assuming there exist such a set of observable variables, the counterfactual outcome can be estimated by estimating the conditional mortality rate for each combination of all the X variables when hypertension = 0, multiplying by the marginal probability of that specific combination of X’s in the overall population, and then summing over all observed combinations of the X’s. This estimate is sometimes referred to as a “standardized mortality rate.” For the sake of clarity collapse all the X’s into one confounder C with c values representing all strata of the joint distribution of X1,…,Xn:

!(!!"

∗ ) = !

E(!!"|ℎ!" = 0, ! = !)Pr!(! = !)

! !!!

This is also known as the non-parametric formulation of the g-formula. Although this quantity could be estimated directly, the number of strata becomes extremely large as the size of the covariate set

• increases. An alternative is to specify a parametric model for E(Mij|hyp,C), resulting in the parametric

g-formula. Specifically, for a set of observed covariates X1, X2,, …, X3, estimate the following Poisson regression model:

ln !(!

!) = !!! + !!!"# + !!ℎ!" +

!!!! + ln!(!!)

! !!!

There are two important things to note with this model: first, age is now a covariate in the model, resulting in only models separate by sex, rather than sex and age; second, the person-years of exposure are entered in the model as an offset, resulting in a rate regression. For simplicity again, collapse all the X’s into a single composite variable C. To estimate E(Mij

*), substitute the model

predicted rates into the g-formula:

! !!"

∗

= exp !! + !!! + !!!

! !!!

Pr!(!"# = !, ! = !, ! = 1)

Note that the value of the offset, L, is set to one for all strata, resulting in an estimate of deaths per person-year of exposure. The standardized mortality rate now has the interpretation: “what would the age- and sex-specific mortality rate be if everyone in the population had their same values of C, but no one had uncontrolled hypertension.” At this point, the population level effect of hypertension on mortality could be determined by directly comparing differences in age-specific mortality rates. However, interpreting differences in rates over a wide range of ages is difficult. For this reason, sets of age-specific mortality rates are

• ften summarized as a period life expectancy, or the average number of years an individual would

live if they were exposed over the remainder of their lives to the observed age and sex-specific mortality rates. This approach moves beyond estimating the effect of hypertension on mortality rates to a more interpretable contrast of the effect of hypertension on adult life expectancy (in this case, life expectancy at age 40). After estimating life expectancies from the rates, the effect of uncontrolled hypertension on life expectancy is just the difference in life expectancy between the

• bserved and counterfactual life expectancies. Assuming an unbiased estimate, this difference has

the interpretation of the number of additional years of life expectancy that would be gained if hypertension were eliminated. Bias analysis The estimated effects will only be an unbiased estimate of the true causal effects under the strong assumption that the exposure status (having uncontrolled hypertension) is conditionally independent from mortality given C. That is, conditional on the set of observed covariates C, the mortality rate of those without hypertension is equal to the mortality rate of those with hypertension, if they were set

as having no hypertension. If there were unobserved confounders, collapsed here into a single variable U, then the estimate would only be unbiased conditional on U:

! !!"(ℎ!" = 0) ℎ!" = 1, !!, … , !!, ! = !!(!!"|ℎ!" = 0, !!, … , !!, !).

This implies that when U is not conditioned on:

! !!"(ℎ!" = 0) ℎ!" = 1, !!, … , !!, ≠ !!(!!"|ℎ!" = 0, !!, … , !!, ).

By making a few assumptions about the structure of U, I can assess the robustness of my estimated effect to different levels of unobserved bias. First, I assume that the relationship between U and mortality does not vary based on the levels of the other covariates and that the difference in the prevalence of U between those with and without hypertension is the same in all strata of C. Then, the bias B (expressed as the difference between the true and estimated effect) is:

! = (Pr !!" = 1 ℎ!" = 1, ! − Pr!(!!" = 0|ℎ!" = 0, !)) ∗ (!(!!" ℎ!", !, ! = 1 − !(!!"|ℎ!", !, ! = 0))

Therefore, the degree of bias is a product of the difference in the prevalence of the unobserved confounder between those with and without hypertension (the relationship between the confounder and hypertension) and the difference in the probability of mortality for those with and without the unobserved variable (the relationship between the confounder and mortality). For a proof of this formula see: (13). By setting different values of the two sensitivity parameters, adding the bias to the estimated age- and sex-specific mortality rates, and then estimating counterfactual life expectancies with the bias adjusted rates, I can see how my estimate of the effect of hypertension on life expectancy varies under different levels of unmeasured confounding. Data Data are from the 2007 and 2014/2015 waves of the Indonesian Family Life Survey (IFLS). The IFLS is a longitudinal survey of individuals and households from 14 of Indonesia’s 34 provinces (the IFLS is representative of 83% of Indonesia’s population). The IFLS contains detailed information

• n health and socioeconomic conditions as well as a host of measured biomarker and

anthropometric data. This analysis is limited to target respondents above the age of 40 (other members in the household were sometimes also measured but since they were not explicitly followed-up in the 2014/2015 wave, I cannot construct reliable estimates of mortality for these individuals) with non-missing information on all the covariates for a total sample size of 10,047 individuals (66,986 person-age observations). Primary outcome The primary outcome is mortality—measured based on household reported information on the age and date of death of individuals surveyed in the 2007 wave. The IFLS has excellent tracking follow up for target respondents: the mortality status of only 3 target individuals surveyed in the 2007 wave was unknown in the 2014/2015 data. Supplementary Table 1 presents the age and death distribution

• f the sample.

Exposures The primary exposure in this study is uncontrolled hypertension (hypertension based solely on blood pressure levels). For each individual, three separate measures of blood pressure were taken by a trained assessor using an Omron HEM-7203 device. Based on the average of the three blood pressure measurements, individuals were classified as having uncontrolled hypertension (from here

• n hypertension) if they had a systolic blood pressure greater than or equal to 140 mmHg or a

diastolic blood pressure greater than or equal to 90 mmHg. As a secondary exposure, I also classified a subset of the non-hypertensive individuals as prehypertensive if they had a systolic blood pressure greater than or equal to 130 mmHg and less than 140 mmHg, or, a diastolic blood pressure between 80 mmHg and 90 mmHg.

Other covariates My empirical approach requires identifying and adjusting for a set of covariates that plausibly make the assignment of hypertension conditionally random. Based on prior studies on potential causes of hypertension and mortality (14-16), I identify the following set of sufficient covariates that encompass both proximate and distal potential causes of hypertension and mortality: urban or rural residence (based on census classification); province of residence (dummy variables for each province); religion (grouped into Islam, Hindu, Protestant, and other); marital status (grouped into never married, current married, and formerly married); self reported occupation type (grouped into retail, manufacturing, agriculture, service, housewife, retired, and not working); self reported completed schooling (grouped into no schooling, some primary school, primary school or more); wealth quintiles (constructed using principle components analysis on indicators for asset ownership); body mass index (measured continuously as height in meters divided by weight in kilograms squared), and the average number of days a week an individual engages in moderate or vigorous physical activity. Diet may also be an important confounder however diet information is extremely limited in the IFLS. However, assuming that diet is mostly determined by social and environmental influences, conditioning on the socioeconomic and geographic covariates included would control some part of the effect of diet. Results Table 1 presents descriptive characteristics of the sample for the baseline 2007 wave. The mean age

• f the sample was 57.3 years. The sample was split evenly between men and women (52.5% female)

and urban/rural (50.3% urban). The vast majority of individuals were married at baseline (80.2%) with Islam as the primary religion (88.6%). Less than half of individuals had more than primary schooling (48.1%) with agriculture and retail as the most common occupations (31.0% agriculture

and 18.2% retail). The sample was on average normal weight (mean BMI of 23.0) and fairly active (on average individuals engaged in moderate of physical activity for 4.3 days in a week). Figure 1 graphs the continuum of hypertension in Indonesia. Levels of hypertension are extremely high in Indonesia (48.6% of adults ages 40 and above are hypertensive). Despite these levels, only 37.5% of hypertensive individuals have been diagnosed. While 71.3% of diagnosed individuals reported taking hypertensive treatment, only 22.7% of those under treatment achieved blood pressure control. Across the full continuum only 6% of hypertensive individuals in Indonesia have controlled blood pressure. While overall prevalences of hypertension provide a measure of burden, the potential gains in life expectancy depend strongly on the age-patterns of blood pressure. Figure 2 graphs the age and sex- specific prevalence of prehypertension and hypertension. For both men and women, hypertension rises very steeply with age: for men, the prevalence of hypertension starts around 30% at age 40 and reaches slightly more than 60% by age 75. This rise is more pronounced with women: by age 75, nearly 80% of women over have hypertension in Indonesia. In contrast, the prevalence of prehypertension decreases slightly over age as more individuals become hypertensive: for men, prehypertension moves from around 35% at age 40 to 20% by ages 80+; for women, prehypertension decreases more steeply to around 10% by age 80+). Table 2 presents the rate regression estimates of the relationship between prehypertension, hypertension, and mortality across three models separately by sex. Model 1 only adjusts for age, model 2 adds a wide array of health and sociodemographic confounders, and model 3 includes province fixed effects. For both men and women, I find a strong relationship between hypertension

and mortality that robust to the inclusion of a large set of potential confounders. Assuming for now that the observed estimates are unbiased, in the fully adjusted model hypertension increases the rate

• f mortality by 60% for men and 50% for women (RR: 1.6, 95% CI: 1.3, 2.1 for men; RR: 1.5, 95%

CI: 1.2 – 2.0). In contrast to hypertension, I find no evidence that being prehypertensive increases the rate of mortality relative to normotensive individuals. While the regression estimates capture the individual-level relationship between hypertension and mortality, they do not readily reveal the importance of hypertension at the population level. In Table 3, I present the population-level implications of the hypertension-mortality relationship by aggregating the estimated effect across individuals for two policy scenarios: the base-case or

• bserved scenario is based on the actual distribution of hypertension in the population while the

counterfactual scenario moves all hypertensive individuals to the normotensive range. Controlling blood pressure results in large improvements in mortality at the population level: for both men and women, life expectancy at age 40 would increase by around 4 years. For men, controlling blood pressure would move life expectancy at age 40 from 34.2 (95% CI: 33.9, 34.6) to 38.2 (95%CI: 36.9, 39.5)—a difference of 4.0 years (95% CI: 2.4, 5.5; p < 0.001). For women, controlling blood pressure would move life expectancy at age 40 from 37.3 (95% CI: 36.1, 38.4) to 41.5 (95%CI: 38.0, 344.9)—a difference of 4.2 years (95% CI: 1.6, 6.7; p < 0.001). The causal interpretation given to my estimates is predicated on the assumption of no unobserved variables (no residual confounding). If my observed result was driven by the influence of an unobserved confounder, then the estimates above may overstate the benefits of blood pressure

• control. In Figure 3, I assess the sensitivity of my estimate of life expectancy years gained to

departures from my identifying assumption. For the sake of interpretability, the role of unobserved

variables is collapsed into a single binary confounder. The effect of this confounder on my estimate depends on two factors: how strongly this unobserved variable is related to mortality and how much more prevalent it is (net of all observed control variables) among the hypertensive than the

• normotensive. The sensitivity results reveal that the size of life expectancy gains is very robust to
• mitted variables for both men and women. For example, if an unobserved confounder was 40%

more likely among hypertensive individuals (net of all controls) and increased the risk of mortality by 15%, the years of life gained would still be above 3 years for both men and women. Indeed even under extreme levels of unobserved confounding, where the unobserved variable is 80% more likely among hypertensive individuals and increases mortality by 20%, eliminating hypertension would still result in around 2 years of life expectancy gained at the population level. Distributional effects of blood pressure control To determine whether the gains in blood pressure control are equally distributed across the population or concentrated among richer individuals, I first estimate the age-standardized prevalence

• f hypertension across wealth quintiles (Table 4). For both men and women, I do not find evidence

that the burden of hypertension is clustered in any specific wealth segment of the population. Levels

• f hypertension are high and similar across the spectrum of wealth (between 37-42% for men and

between 45-50% for women). Given the similarly in the levels of hypertension across wealth quartiles, I find that the potential gains in life expectancy are evenly distributed across wealth groups in Indonesia (Figure 4). For men and women across all five quintiles, fully controlling blood pressure results in around a 4-year gain in life expectancy at age 40. Similar to the overall results, these estimates are quite robust to unobserved confounding (results available upon request).

Discussion The populations of many LMICs are projected to age dramatically over the coming decades. In the backdrop of population aging, identifying which health conditions will provide the largest gains in well-being at the population level is important for setting health priorities and informing decision making at the global, national, and subnational levels. Hypertension is already a leading risk factor for mortality in many LMICs and will only increase in prevalence as the share of older individuals

• grows. Although many studies have documented high-levels of uncontrolled hypertension in several

LMICs, little research has sought to determine the population-level mortality impacts of improving blood pressure control in these contexts. Using rich longitudinal data on Indonesian adults, I combine epidemiological and demographic modeling approaches to estimate the gains in adult life expectancy that would result from fully controlling blood pressure. I find that improving blood pressure control would result in a large, four-year, gain in life expectancy at age 40 for both men and

• women. The size of this change is substantial: as a reference, a four year change in life expectancy at

40 corresponds to the difference between life expectancy at age 40 in Indonesia between 1960 and 2010, or equivalently, 50 years of progress in improving adult mortality. Beyond overall population health, the goal of many health policies is to improve health equity by addressing conditions that are salient among vulnerable populations. I find that the high prevalence

• f hypertension is not concentrated within any single wealth-strata of the Indonesian population but

almost equally distributed across rich and poor sub-populations. Given this distribution of hypertension, the four-year gain in adult life expectancy from improving blood pressure control is found for men and women across the entire wealth distribution. This finding stands in contrast to many popular understandings of hypertension as a disease of the well off, or a disease of post- nutrition transition populations. However, other researchers have also documented high rates of

hypertension among poorer population, arguing that the popular conception of hypertension as a disease of the rich is not consistent with epidemiological evidence and may lead to suboptimal health policy or resource allocations (17, 18). The key limitation to this study is the use of observational data to estimate the relationship between uncontrolled hypertension and mortality. For my results to have a causal interpretation, I make the identification assumption that the assignment of hypertension is random conditional on a set of

• bserved characteristics. If there were unobserved variables that resulted in residual confounding,

my four-year estimate would be a biased estimate of the true effect of blood pressure control. However, based on the results of a simulation-based sensitivity analysis, I find that even under high levels of unobserved residual confounding, the size of the gain in adult life expectancy from improving blood pressure control remains large. While the results of these simulations do not solve the identification problem, they suggest that bias from unobserved variables is unlikely to affect the policy conclusions drawn from my results. Within the IFLS, individuals only have blood pressure measurement from one point in time prior to their mortality follow-up. This misses the changes in blood pressure that occur prior to follow-up wave, introducing measurement error into the estimate

• f blood pressure. However, this type of measurement error would result in a downward bias in the

estimate, making my conclusions conservative. While the use of nationally representative data alleviates many issues of selection bias, the final analytic sample is smaller than the overall IFLS sample due to missing data for some individuals. This may result in some degree of selection. Indeed, I find that the life expectancy of the analytic subsample is higher than that over the overall sample, suggesting the analytic sample may be positively selected on health.

Despite these limitations, this study has a number of important strengths. The IFLS is one of the

• nly sources of large nationally representative data with measured blood pressure and reliable

mortality follow-up in an LMIC. Empirically, this study is one of the first to estimate the population- level mortality impact of improving blood pressure control using a clearly specified causal contrast that is motivated by policy-relevant counterfactual scenarios. Second, this study moves beyond showing differences in mortality rates or probabilities by using demographic life table techniques to express mortality differences in terms of an easily interpretable format (life expectancy at age 40). This approach frames the importance of blood pressure control in a common and widely used metric of population health. Similarly, I estimate the gains for the overall population and across wealth quintiles to investigate whether the benefits of hypertension are disproportionately higher for wealthy individuals. Finally, I explicitly confront the potential biases in my observational estimates by showing that the conclusions drawn here are robust to a substantial level of unmeasured confounding. The results of this study promote hypertension prevention and control as a promising strategy for improving mortality at the population level; however, further research is needed to realize this

• potential. First, research is needed to establish the cost-effectiveness of various hypertension

prevention and treatment strategies to identify which policy options provide the highest rates of

• return. Next, implementational research is needed to identify the best ways to introduce and scale up

hypertension interventions at the population level. Finally behavioral research is needed to promote health seeking behavior, preventative health behaviors, and treatment compliance among individuals. Within Indonesia, improving blood pressure control has the potential to greatly reduce mortality at the population level. While the results presented here are for Indonesia, high-levels of uncontrolled

hypertension are not unique to Indonesia. My results suggest that across LMICs, improving blood pressure control can result in massive, unrealized, gains in longevity. In contrast to many other chronic health conditions, interventions to control blood pressure are also comparatively straight forward and treatments relatively affordable. Therefore, improving blood pressure control has the potential to also be a cost-effective and achievable way of improving longevity in Indonesia and

• ther LMICs.

References 1. Heiniger S, Viswanathan B, Gedeon J, Paccaud F, Bovet P. Trends in prevalence, awareness, treatment and control of high blood pressure in the Seychelles between 1989 and 2013. Journal of

• Hypertension. 2017;35(7):1465-73.

2. Kalkonde YV, Deshmukh MD, Sahane V, Puthran J, Kakarmath S, Agavane V, et al. Stroke is the leading cause of death in rural Gadchiroli, India. Stroke. 2015;46(7):1764-8. 3. Lloyd-Sherlock P, Beard J, Minicuci N, Ebrahim S, Chatterji S. Hypertension among older adults in low-and middle-income countries: prevalence, awareness and control. International journal

• f epidemiology. 2014;43(1):116-28.

4. Cooper RS, Rotimi CN, Kaufman JS, Muna WF, Mensah GA. Hypertension treatment and control in sub-Saharan Africa: the epidemiological basis for policy. BMJ: British Medical Journal. 1998;316(7131):614. 5. Hussain MA, Al Mamun A, Reid C, Huxley RR. Prevalence, Awareness, Treatment and Control of Hypertension in Indonesian Adults Aged≥ 40 Years: Findings from the Indonesia Family Life Survey (IFLS). PloS one. 2016;11(8):e0160922. 6. Council NR, Committee on Population. Aging in Asia: Findings from new and emerging data initiatives: National Academies Press; 2012. 7. Collaboration PS. Age-specific relevance of usual blood pressure to vascular mortality: a meta-analysis of individual data for one million adults in 61 prospective studies. The Lancet. 2002;360(9349):1903-13. 8. Bundy JD, Li C, Stuchlik P, Bu X, Kelly TN, Mills KT, et al. Systolic blood pressure reduction and risk of cardiovascular disease and mortality: a systematic review and network meta-

• analysis. JAMA cardiology. 2017.

9. Ettehad D, Emdin CA, Kiran A, Anderson SG, Callender T, Emberson J, et al. Blood pressure lowering for prevention of cardiovascular disease and death: a systematic review and meta-

• analysis. The Lancet. 2016;387(10022):957-67.

10. Capewell S, Morrison C, McMurray J. Contribution of modern cardiovascular treatment and risk factor changes to the decline in coronary heart disease mortality in Scotland between 1975 and

• 1994. Heart. 1999;81(4):380-6.

11. Ford ES, Ajani UA, Croft JB, Critchley JA, Labarthe DR, Kottke TE, et al. Explaining the decrease in US deaths from coronary disease, 1980–2000. New England Journal of Medicine. 2007;356(23):2388-98. 12. Capewell S, Ford ES, Croft JB, Critchley JA, Greenlund KJ, Labarthe DR. Cardiovascular risk factor trends and potential for reducing coronary heart disease mortality in the United States of

• America. Bulletin of the World Health Organization. 2010;88(2):120-30.

13. VanderWeele TJ, Arah OA. Bias formulas for sensitivity analysis of unmeasured confounding for general outcomes, treatments, and confounders. Epidemiology (Cambridge, Mass). 2011;22(1):42-52. 14. Cutler D, Deaton A, Lleras-Muney A. The determinants of mortality. The Journal of Economic Perspectives. 2006;20(3):97-120 %@ 0895-3309. 15. Kaufman JS, Owoaje EE, James SA, Rotimi CN, Cooper RS. Determinants of hypertension in West Africa: contribution of anthropometric and dietary factors to urban-rural and socioeconomic gradients. American journal of epidemiology. 1996;143(12):1203-18 %@ 476-6256. 16. Van Rooyen JM, Kruger HS, Huisman HW, Wissing MP, Margetts BM, Venter CS, et al. An epidemiological study of hypertension and its determinants in a population in transition: the THUSA

• study. Journal of human hypertension. 2000;14(12):779 %@ 0950-9240.

17. Lloyd-Sherlock P, Minicuci N, Corso B, Beard J, Chatterji S, Ebrahim S. Diseases of the Rich? The Social Patterning of Hypertension in Six Low-and Middle-Income Countries. The European Journal of Development Research. 2016:1-16 %@ 0957-8811. 18. Lloyd-Sherlock P, Ebrahim S, Grosskurth H. Is hypertension the new HIV epidemic? : Oxford University Press; 2014.

Table 1: Descriptive characteristics of the sample in the baseline wave. Adults ages 40+, Indonesian Family Life Survey, 2007, N = 10,047.

• No. or mean

% or SD Age 53.7 10.8 Female 5275 52.5 Urban Residence 5053 50.3 Province North Sumatra 583 5.8 West Sumatra 488 4.9 South Sumatra 423 4.2 Lampung 406 4.0 Jakarta 641 6.4 West Java 1456 14.5 Central Java 1428 14.2 Yogyakarta 696 6.9 East Java 1645 16.4 Banten 246 2.4 Bali 533 5.3 West Nusa Tenggara 628 6.3 South Kalimantan 427 4.3 South Sulawesi 447 4.4 Current marital status Never married 149 1.5 Was married 1840 18.3 Currently married 8058 80.2 Completed schooling No schooling 1980 19.7 Some schooling 3231 32.2 Primary or more 4836 48.1 Religion Islam 8906 88.6 Hindu 491 4.9 Protestant 404 4.0 Other 246 2.4 Primary job Retail 1828 18.2

Housework only 1677 16.7 Retired 645 6.4 Agriculture 3117 31.0 Manufacturing 672 6.7 Service 1411 14.0 Not working 175 1.7 Other 522 5.2 Body mass index 23.0 4.2 Number of days of moderate or vigorous physical activity 4.3 2.9 Notes: Body mass index is calculated as measured weight in kilograms over height in meters squared; primary school or more is classified as seven or more years of schooling; analytical models additionally adjust for wealth quintiles but that is not shown since the index was constructed to have 20% of individuals in each quintile.

Figure 1: Continuum of hypertension care in Indonesia. Adults ages 40+, Indonesian Family Life Survey, 2007, N = 10,047.

All hypertensive 48.6% of individuals Diagnosed 37.5% of hypertensive individuals Treated 71.3% of diagnosed individuals 26.8% of all hypertensive individuals Controlled BP 22.7% of treated individuals 6.0% of all hypertensive individuals

Figure 2: Age-specific prevalence of prehypertension and hypertension in Indonesia. Adults ages 40+, Indonesian Family Life Survey, 2007, N = 10,047. Error bars represent 95% confidence intervals.

Table 2: Estimated individual-level relationships between prehypertension, hypertension, and mortality. Adults ages 40+, Indonesian Family Life Survey, 2007-2014/15. Men Women Model 1 Model 2 Model 3 Model 1 Model 2 Model 3 Prehypertensive 0.9 0.8 0.8 1.1 1.1 1.1 (0.7 - 1.2) (0.6 - 1.1) (0.6 - 1.1) (0.8 - 1.5) (0.8 - 1.5) (0.8 - 1.5) Hypertensive 1.8 1.6 1.6 1.7 1.5 1.5 (1.4 - 2.3) (1.3 - 2.1) (1.3 - 2.1) (1.3 - 2.2) (1.2 - 2.0) (1.2 - 2.0) PY Observations 31,669 31,641 31,641 35,404 35,341 35,341 Prehypertension p-value 0.365 0.242 0.251 0.692 0.706 0.703 Hypertension p-value <0.001 <0.001 <0.001 <0.001 0.003 0.003 Coefficients are presented as hazard ratios with 95% confidence intervals in parentheses. All analyses were weighted. Model 1 covariates: age (continuous) Model 2 covariates: Model 1 + job (cat) + religion (cat) + schooling (cat) + wealth (cat) + urban (cat) + marital status (cat) + bmi (continuous) + days physical activity per week (continuous) + taking blood pressure medication (dichotomous) Model 3 covariates: Model 2 + province fixed effects

Table 3: Estimated gain in life expectancy at age 40 after controlling blood pressure among hypertensive individuals. Adults ages 40+, Indonesian Family Life Survey, 2007-2014/15. Observed Counterfactual Difference Difference p-value Men 34.2 38.2 4.0 <0.001 (33.9 - 34.6) (36.9 - 39.5) (2.4 - 5.5) Women 37.3 41.5 4.2 <0.001 (36.1 - 38.4) (38.0 - 44.9) (1.6 - 6.7) Notes: Values represent life expectancy at age 40 with 95% confidence intervals in parentheses. Life expectancies were estimated using period life tables. Confidence intervals and p-values were estimated with a bootstrap procedure with 200 replications.

Figure 3: Sensitivity of life expectancy gains to unobserved variables. Adults ages 40+, Indonesian Family Life Survey, 2007-2014/2015.

Table 4: Age-standardized prevalence of uncontrolled hypertension across wealth quintiles. Adults ages 40+, Indonesian Family Life Survey, 2007, N = 10,047. Men Women Wealth Quintiles (1 - Lowest, 5- Highest) 1 0.41 0.5 [0.38,0.45] [0.47,0.53] 2 0.37 0.46 [0.33,0.40] [0.43,0.49] 3 0.42 0.45 [0.39,0.46] [0.42,0.48] 4 0.4 0.49 [0.37,0.43] [0.46,0.52] 5 0.42 0.46 [0.39,0.46] [0.42,0.49] Notes: 95% CI in parentheses. Estimates were standardized using the overall population distribution for adults ages 40+ as the standard.

Figure 4: Estimated gain in life expectancy at age 40 after controlling blood pressure among hypertensive individuals by wealth quintiles. Adults ages 40+, Indonesian Family Life Survey, 2007- 2014/15. Error bars represent 95% confidence intervals.