The household‐level economic burden of heart disease in India

To estimate healthcare use and financial burden associated with heart disease among Indian households.


Introduction
In India, the burden of cardiovascular disease (CVD) is rising. Ischaemic heart disease was the 4th largest cause of years of life lost due to premature mortality in 2010, up from its 8th ranking in 1990 and related conditions such as stroke are becoming more common as a cause of death (Institute of Health Metrics and Evaluation 2012). This rising health burden of CVD reflects multiple risk factors, including low dietary intake of fruits and vegetables, high levels of tobacco and alcohol consumption, low levels of physical activity, high rates of abdominal obesity, untreated hypertension and ageing (Popkin et al. 2001;Reddy & Katan 2004;Prabhakaran et al. 2005;Goyal & Yusuf 2006). Moreover, morbidity and mortal-ity associated with CVD are occurring at younger ages in India relative to other developed countries (World Health Organization 2010).
The growing health burden of CVD potentially translates into increased economic burden at both the national and household levels in India. Estimates from recent studies indicate that CVD-related aggregate medical care costs ranged from approximately US$3 billion in 2004 (Mahal et al. 2010) to US$7.5 billion in 2010 (Bloom et al. 2011). Leeder et al. (2004) also project major income declines due to CVD-related losses in productive labour and Gupta et al. (2006) estimated that CVD costs could amount to as much as 20% of the domestic product in the Indian state of Kerala.
Few studies have examined the household-level impacts of CVD in low-and middle-income countries. In India, there are two previous studies. Rao et al. (2011) assessed out-of-pocket spending by households, the frequency of hospitalization and ambulatory care visits among patients Michael Engelgau: Formerly World Bank Washington, D.C. The copyright line for this article was changed on 22 April 2016 after original online publication.
reporting heart disease, hypertension or diabetes. Huffman et al. (2011) examined 1657 hospitalized patients in four countries including Argentina, China, India and Tanzania (500 patients were from India) and estimated the impact of CVD (defined in their study as myocardial infarction, unstable angina, stroke, acute heart failure or peripheral vascular intervention) on catastrophic spending, 'distress' financing (borrowing or sale of assets) and workforce participation of household members. Although both studies used multivariate analyses to identify potential correlates of catastrophic spending and distress financing, both likely overestimated the potential economic burden of CVD on households at the population level. In the first study, an appropriate comparison group, which is the amount that similar households without CVD-affected members spent on health services, was not considered. In the second study, the focus on hospitalized cases possibly resulted in higher than average estimates of household economic burden and again, an appropriate comparison group was missing.
Our study uses data from a large cross-sectional household survey and considers comparison households from the same survey. Our outcome measures include healthcare utilization and out-of-pocket expenses for inpatient and outpatient care, financing patterns to cover these expenses and employment among household members. We also assessed outcomes by socio-economic status. Collectively, our results provide a comprehensive assessment of outcomes among households with a member reporting heart disease relative to households that did not contain a member reporting heart disease.

Methods
We used propensity score matching (PSM) to match case households containing a member reporting heart disease (CVD-affected households) to control households that did not contain a member reporting heart disease (Rosenbaum & Rubin 1983). This definition did not include stroke or acute heart failure, information for which was not available in our survey data. We estimated propensity scores to predict the probability of a household containing at least one member reporting heart disease (CVDaffected households) based on household socio-economic and demographic characteristics. Each CVD-affected household was matched to one control household (not CVD-affected) with the closest propensity score. As a quality check, we conducted three additional assessments: first, for each covariate used in the regression model that generated the propensity scores, we compared the means between the CVD-affected households and matched control households using a t-test; second, we re-estimated our results after excluding households that experienced a death in the previous year, and after excluding the 1% of households with the highest out-of-pocket spending on illness; and third, we re-estimated our results using the stratification matching method, which essentially matches subgroups of cases and controls instead of on a one-to-one basis (Dehejia & Wahba 1999).
We used data from a nationally representative crosssectional household survey of morbidity and healthcare utilization undertaken by the National Sample Survey Organization (NSSO) in India in 2004. The survey sample included approximately 74 000 households and 383 000 individuals of all ages (NSSO 2006). Information on socio-economic and demographic characteristics, insurance status, healthcare utilization, out-of-pocket medical spending for self-reported health conditions, non-medical spending, sources of healthcare financing and employment status for all household members were collected from a single key adult respondent in each household.

Definition of CVD-affected household
A household was defined as CVD affected if any of its members was reported as: (i) currently living with heart disease; and (ii) hospitalized due to heart disease in the year preceding the survey, whether or not the affected member was currently alive. The survey did not collect self-reported information on strokes and heart failure.
Variables used to construct propensity scores CVD-affected households and control households were matched on several socio-economic, demographic and locational indicators: (i) educational status of household head (primary, secondary, college degree and above), (ii) type of house (solid (pucca), mud or similar (kuccha) or other), (iii) source of drinking water (piped or other), (iv) type of sanitation (covered drains, latrine, septic tank), (v) major source of livelihood (e.g. head of household was self-employed or paid wages), (vi) family demographics (share of children (0-15 years), share of young adults (15-59 years), share of elderly (age > 60 years)), the proportion of household members consisting of females of any age and household size, (vii) caste status (whether 'scheduled caste or tribe', 'other backward caste' or 'other'), religion (Hindu, Muslim or other), (viii) health insurance status (public or private insurance coverage for a member of the household) and (ix) rural/urban residence status. Indicator variables for 71 subareas of residence were also included to account for locational characteristics that might influence healthcare use, spending and employment.

Outcome variables
We constructed outcome variables at the household level (and not at the level of the individual patient), because healthcare use, spending and financing as well as employment reflect both individual and joint decisionmaking at the household level (Gruber & Madrian 2002).
Healthcare utilization outcomes were as follows: (a) hospital stays: total for all household members in the preceding year divided by household size, (b) public hospital stays: total for all household members in the preceding year divided by household size, (c) hospital days: total for all household members in the preceding year divided by household size and (d) outpatient visits: total for all household members in the preceding 15 days divided by household size.
Indicators of healthcare spending were as follows: (a) out-of-pocket (OOP) spending on outpatient care per member in the preceding 15 days, (b) OOP spending per member for hospital care in the preceding year, (c) OOP drug spending per member in the preceding 15 days, (d) OOP transportation spending per member for healthcare in the preceding 15 days and (e) the share of OOP health spending in total household spending in the preceding 15 days.
Outcome indicators for financial stress were as follows: (a) expenses on items other than healthcare, per household member (Gertler & Gruber 2002) in the preceding 15 days; (b) household borrowing and sale of assets to finance out-of-pocket spending on outpatient and inpatient health care; (c) adult employment rates (the number of members aged 15 years and over who are currently working, divided by all household members aged 15 years and older) and elderly (60 years and over) employment rates; (d) healthcare use by members not reporting heart disease in CVD-affected households. CVD might also cause less serious conditions to be neglected in the household. Hence, we compared health service use for 'non-major conditions' (specifically excluding cancer, heart disease, diabetes and injuries) of CVD-affected households and controls.
We examined outcomes by subgroups of socio-economic status, specifically (a) scheduled castes and tribes (SC/ST), two groups that are considered especially deprived in India, versus non-SC/ST households, and (b) households where the education of the headan indicator of economic statuswas below the median for the full sample of households in the survey vs. households where the education of the head of household was above the median. The economic burden of CVD across socioeconomic groups was assessed using methods described in Appendix A1. In brief, to measure group-specific burden, propensity scores were re-estimated but after excluding the indicator for the socio-economic groups of interest (e.g. SC/ST status) from the list of variables used to generate propensity scores. The subset of matched and control households (using a matching algorithm such as nearest neighbour) was retained, and all other observations dropped. Finally, ordinary least squares methods were used to estimate linear relationships between economic outcomes, an indicator for whether a household has a member with CVD (treatment), an indicator for socio-economic status (e.g. whether SC/ST or not) and the product of the indicators for CVD status and socioeconomic status.

Results
Our matched cases and control households were similar without any statistically significant differences in the variables used for generating propensity scores (Table 1).
CVD-affected households experienced an extra 10.3 hospital stays per 100 members annually (P < 0.01), and an extra 11.2 outpatient visits per 100 members in the 15 days preceding the survey (P < 0.01) compared with matched control households (Table 2). CVD-affected households also reported 1.45 extra days per member spent in hospitals (2.33 vs. 0.88, P < 0.01) relative to matched controls. Per person outpatient visits of members without heart disease (in CVD-affected households) were lower by almost two visits for every 100 members compared with control households in the 15 days preceding the survey (P = 0.01). Per person outpatient visits for 'non-major' health conditions in the 15 days preceding the survey were also lower in CVD-affected households, by 3.6 visits per 100 members (P < 0.01) compared with matched controls.
Out-of-pocket expenses were significantly higher in CVD-affected households than controls by INT$231.75 (P < 0.01) per household member for inpatient care in the preceding year and by INT$5.16 (P < 0.01) per member for outpatient care in the preceding 15 days ( Table 3). The results for outpatient spending were driven mainly by differences in OOP drug expenses between CVD-affected households and controls (INT$3.58, P < 0.01). OOP transportation expenses for healthcare accounted for 6.7% of the difference in OOP outpatient spending between CVD-affected households and controls (INT$0.34, P < 0.01). As a share of household spending, OOP spending on healthcare by CVD-affected households significantly exceeded that of controls (27.2% vs. 10.7%, P < 0.01) in the preceding 15 days.
Per person spending on non-medical items was lower in CVD-affected households than matched controls in the preceding 15 days (INT$23.90 vs. INT$28.51, P < 0.01). CVD-affected households relied to a significantly greater extent on borrowing and/or asset sales for financing OOP inpatient spending in the preceding year (32.7% vs. 12.8%, P < 0.01) and OOP outpatient spending in the 15-day preceding (5.1% vs. 2.0%, P < 0.01), relative to controls.
Our subgroup analysis results are described in Tables 4 and 5. The columns labelled (1) and (2) describe the effect of CVD on healthcare use and economic outcomes Estimates are based on calculations by authors using household-level data from National Sample Survey data for 2004. The data presented refer to all households, whether or not a death was experienced in the household. The 71 residential indicators that we used in our analysis have been consolidated into large regional blocks in the table to conserve space for presentationthe actual propensity score calculations for estimating the economic burden used 71 individual dummies for each region. The comparison of means used for the t-test for balance checking reported in column 4 was based on post-matching (using nearest neighbourhood method) data; P-values are reported in parentheses besides the t-statistic in column 5 for a two-tailed test. Columns 2-3 report 95% confidence intervals in parentheses next to the means.
among SC/ST and non-SC/ST households separately, and the column labelled (1)À(2) provides a test of the difference in the effect of CVD on these two groups. Similarly, the columns labelled (3) and (4) report our estimates of the effect of CVD on healthcare use and economic outcomes among households by education of the head of household. Analogous to column (1)À(2), the data in column (3)À(4) can be used to assess the difference in CVD effects between the two education subgroups. The data in Table 4 suggest that the impact of CVD on hospital admissions and hospital days is not statistically different between low SES and high SES households. However, low SES households rely more on public hospitals than high SES households, with a statistically significant difference in public hospital admissions for SC/ST relative to non-SC/ST households (8% vs. 4%, P < 0.01).
High SES households (non-SC/ST or education of head of household above the median) generally incurred higher OOP spending on inpatient and outpatient care due to CVD relative to low SES households, but statistically significant differences were observed only for comparisons by education of head of household and in only two cases: OOP inpatient expenditures and OOP transportation expenditures. But low SES households relied more on borrowing and/or selling assets to finance their care than high SES households. CVD also increased the share of OOP spending in total household spending across all subgroups, but subgroup differences were statistically significant only for the groups classified by education of head of household (education below median 22.0% vs. education above median 15.8%, P = 0.02). All subgroups saw a reduction in non-medical spending on account of CVD, with the exception of the SC/ST group that experienced a statistically insignificant decline. CVD was also associated with larger declines in employment rates among members of low SES households than high SES households, particularly in the 15-59 year age group.

Discussion
CVD-affected households experienced greater financial hardship due to illness than similar control households, Means of the outcome variables are reported in columns 2 and 3 for cardiovascular disease (CVD)-affected and matched controls, respectively, with 95% confidence intervals in parentheses. Column 4 reports the difference in means and the associated P-values (the probability that the outcomes for matched cardiovascular disease-affected and control households differ in a two-tailed t-test) in parentheses below coefficient estimates. 'Non-major' health conditions refer to conditions excluding heart disease, injuries, cancer and diabetes. One-year reference refers to the 1 year immediately preceding the survey; 15-day reference refers to the 15 days immediately preceding the survey. and they also reported more inpatient stays and outpatient visits. Household non-medical consumption can be expected to be lower when a member has heart disease, unless the household is able to effectively insure against associated financial risks, including any treatment expenses and earnings losses (Gertler & Gruber 2002;Islam & Maitra 2012). Based on the lower levels of nonmedical consumption among CVD-affected households relative to controls reported in Table 3 . Prima facie, this suggests households were able to rely on some form of insurance or coping mechanism to partially protect their non-medical spending against treatment costs. One coping device is the increased burden on unaffected members via their lower use of health services as well as curtailment of non-major health service use, similar to findings from other studies (Lilly et al. 2010). Another coping mechanism appears to be the lower employment among household members, especially among, females and older individuals. While some of this decline may be due to the ill person not working, it may also reflect INT$ = International Dollars. We used a conversion factor of 1INT$ = 14.52 Indian Rupees for the year 2004, as published by the World Bank. Columns 2 and 3 present mean outcomes for cardiovascular disease (CVD)-affected and matched control households, respectively, with 95% confidence intervals in parentheses; P-values are reported in parentheses below the estimates in column (4) for a 2-tailed test. 'Non-major' health conditions refer to conditions other than heart disease, injuries, cancer and diabetes. One-year reference refers to the 1 year immediately preceding the survey; 15-day reference refers to the 15-days immediately preceding the survey. Columns labelled (1) and (2) are estimates of the impact of CVD on healthcare use by SC/ST and non-SC/ST households, respectively. Column (1)À(2) provides the difference between these two estimates and the associated P-value can be used to assess whether the two impact estimates are statistically different. Analogously, the columns labelled (3) and (4) are estimates of the impact of CVD on healthcare use by households where the education of the head of the household is below the median and households where the education of the head of the household is above the median, respectively. Column (3)À(4) provides the difference between these two estimates and the associated P-value can be used to assess whether the two impact estimates are statistically different. The precise methodology for obtaining these estimates is described in Appendix A1. Confidence intervals are reported in square brackets. Data have been rounded to the nearest two decimal places. 'Non-major' health conditions refer to conditions excluding heart disease, injuries, cancer and diabetes. One-year reference refers to the 1 year immediately preceding the survey; 15-day reference refers to the 15-days immediately preceding the survey.  (1) and (2) are estimates of the economic burden of CVD on SC/ST and non-SC/ST households, respectively. Column (1)À(2) provides the difference between these two estimates and the associated P-value can be used to assess whether the two impact estimates are statistically different. Analogously, the columns labelled (3) and (4) are estimates of the economic burden of CVD on households where the education of the head of the household is below the median and households where the education of the head of the household is above the median, respectively. Column (3)À(4) provides the difference between these two estimates and the associated P-value can be used to assess whether the two impact estimates are statistically different. The precise methodology for obtaining these estimates is described in Appendix A1. Confidence intervals are reported in square brackets. Data have been rounded to one decimal place. 'Non-major' health conditions refer to conditions other than heart disease, injuries, cancer and diabetes. One-year reference refers to the 1 year immediately preceding the survey; 15-day reference refers to the 15-days immediately preceding the survey.
greater household care-giving responsibilities borne by other members. Employment declines potentially translate into large household earnings losses given that more than 90% of India's workforce is engaged in the informal sector with no health or employment insurance (Harris 2008;Ciani 2011;Passey et al. 2012). The households in our sample also coped with increased requirements for health spending associated with heart disease by borrowing or selling assets. Doing so can have long-term implications for household economic well-being if borrowing costs are high, or if income earning assets are sold.
Our study also sheds light on the potential economic burden on CVD-affected households in groups of different socio-economic status and suggests that low SES households are less able to protect themselves against the associated financial risk. This is reflected in their greater reliance on borrowing and sales of assets to finance health spending, and their considerably larger decline in adult workforce participation (and potential loss of earnings). These findings of a greater burden of OOP spending on low SES households are consistent with recent Indian literature on the impoverishing impact of illness on households (Garg & Karan 2009;Shahrawat & Rao 2012).
Our results underline the importance, also highlighted in other studies, of protecting Indian households against the financial burden from non-communicable conditions such as CVD. The heavy reliance on OOP spending in financing healthcare in India is not surprising given limited government financing (only about 1.2% of GDP) and other insurance mechanisms that continue to be inadequate. At most, only 1% of India's population is covered by private health insurance. While a number of publicly funded health insurance schemes have emerged, in many cases, their coverage is not comprehensive and non-poor households are ineligible to participate. Thus, the Rashtriya Swasthya Bima Yojana (RSBY), a publicly financed third-party payment scheme, which currently covers more than one hundred million people in India and is expected to further expand its coverage to 280 million people, is restricted to the poor, and its maximum coverage is INT $2066 for a family of five. More generous publicly financed schemes that cover a broader segment of the population do exist, but their geographic coverage is limited to a few southern Indian states (Fan et al. 2012). These coverage gaps highlight the importance of establishing risk pooling mechanisms that extend beyond the poor, when considering conditions that are expensive to treat. From a longer term sustainability perspective, mechanisms to lower risk factors such as tobacco smoking, hypertension, lack of physical activity and obesity may be warranted.
A key strength of our study is our use of a nationally representative household survey with information on healthcare use, OOP spending on healthcare, healthcare financing and information on individual-level employment. Matching CVD-affected households to control households on socio-economic, demographic and locational characteristics also addresses some of the confounding arising from non-random assignment of heart disease. Our study findings also rely on multiple checks for robustness, and our main conclusions hold up across different matching methods, as well as matching after excluding households with a death from any cause, and matching after excluding the top 1% households with the most out-of-pocket spending on treatment.
There are obvious limitations to our study. Our identification of CVD-affected households relies on self-reports of heart disease, which may lead to inaccurate estimation, although we would expect that acute cases, at least, are well reported. Household survey data tend to underestimate deaths, and estimates based on the survey we used suggest that in 2004, 211 000 deaths occurred from heart disease in India, compared with an estimated 2.3 million deaths due to CVD from the 2008 Global Burden of Disease study. Undercounting deaths could lead to the estimated household economic burden of heart disease becoming biased downwards if healthcare use and expenditures are concentrated in the time immediately preceding death, and a disproportionate number of CVD deaths are excluded or end up in the controls. On the other hand, our estimates of the economic burden could be biased upwards given that all households that had a member hospitalized due to heart disease in the year preceding the survey were automatically defined as a CVD-affected household. Another source of upward bias in our estimates is that individuals using health services and incurring OOP spending may be those most aware of their health status, either because their condition is more serious than average or because they are naturally pre-disposed to health-seeking behaviour. We partially addressed these issues by undertaking additional analyses limited to households that did not experience deaths, irrespective of cause, and further by including hospitalization as an additional indicator of matching in the propensity score equation, but some of the biases are likely to persist.
Another limitation is that matching methods cannot account for unobservable factors that drive household risks of heart disease. Our 2004 data do not include information on tobacco and alcohol consumption, dietary history or obesity in the household. Nor do we have information on the occupational history of household members that could affect heart disease risks (Goyal & Yusuf 2006).

Conclusion
Our study is one of only a few that estimate CVD-related economic hardship imposed at the household level in LMICs. Our finding provides a much better understanding of the economic burden of CVD at the household level than previous studies because we took into account some of the major potentially confounding effects. However, CVD is only one among a number of non-communicable conditions, such as cancer and diabetes that can be expected to impose a major financial burden on affected households. The key policy implication from our findings is the need to protect Indian households from the financial risks associated with major non-communicable conditions.