Paper short abstract:

Cost of dying prompts an imperative understanding of health shock. Nationally representative panel data from the Consumer Pyramid Household Survey spanning 2014–2023, we examined how the death of a household member affects household expenditure and socio-economic status. Data from 0.6 million individuals, highlights the spending behavior, supported by the hypothesis of Grossman Model (1972).

Paper long abstract:

High cost of dying in India: Understanding causal relationship using panel data of 0.6 million individuals

Background

The high cost of death underscores the impact of health shocks on household poverty in developing countries like India. Allocation of household funds to healthcare is a key factor in household impoverishment, reflecting India's healthcare sector characterized by low government spending, high out-of-pocket expenses, and limited financial protection. India's health spending, at 1.5% of GDP, is one of the lowest worldwide (RBI 2020-21).

Methodology

This paper aims to analyse the relationship between individual deaths and household expenditure using a sample of 0.6 million individuals. Pooled OLS regression, also known as a population-averaged model, was employed. This approach averages out individual effects and provides consistent and asymptotically normal estimates. The basic model equation is as follows:

HE_"it" =β_"0" +〖"β*" 〗_(1 ) Death_it+ β_"n " 〖SES_Cov〗_"it " +ε_(it )

Where, 〖HE〗_it captures the household expenditure at household level, where 〖Death〗_it is the main predictor variable, representing the death of an individual in the household. β_1 is the slope coefficient associated with the Death_it, which quantifies the expected change in the household health expenditure (〖HE〗_it) associated with the variable Death_it in the household. β_n represents a matrix of coefficients for ‘n’ socio-economic covariates 〖(SES_Cov〗_it) used in the regression model as control variables.

We utilized a combined approach of pooled ordinary least squares (OLS) regression and fixed effects (FE) panel models to focus on within-household variations over time. This method helps address potential biases arising from unobserved household-specific factors, particularly relevant when analysing health expenditure in households with higher death rates. These models are particularly suitable for the dataset used in this study, as suggested by Brüderl and colleagues (2015), because they eliminate between-household variation and focus solely on within-household variation. The basic model is outlined as follows:

HE_"it" =β_"i" +〖"β*" 〗_(1 ) Death_it+ β_"n " 〖SES_Cov〗_"it " +δ_t+ u_it+ε_(it )

where β captures how the outcome, in our case the health expenditure (〖HE〗_it), changes after a household change in the main independent variable, in our case death status (Death_it). The restrictive assumption of this model is that the unobserved covariates are time-constant (δ_t). In the model, 〖HE〗_it is the health expenditure at the household level and β_i represents individual fixed effects, capturing unobservable characteristics specific to each household that do not change over time. Therefore, this model accounts for individual-specific characteristics β_i and time-varying, but household-invariant factors u_it, providing a more nuanced understanding of the relationship.

The basic equation of the probit model used above is;

P (HE_"it" =1)=φ (β_"0" +〖"β*" 〗_(1 ) Death_it+ β_"n " 〖SES_Cov〗_"it " +ε_(it )

Where, P (HE_"it" =1) is the probability that health expenditure HE_"it" at household (i) and time (t) is equal to 1. In the Probit model, we are modeling the probability of a binary outcome (in this case, health spending being 1 means “spending” on health care while the other case would be “no spending”). Additionally, {φ (.)} is the cumulative distribution function of the standard normal distribution evaluated at the linear combination of the model's parameters and the error term. Thus, the above Probit Model provides insights into the relationship between the death of an individual within a household and the probability of incurring health expenditure, using the cumulative distribution function of the standard normal distribution to model this relationship.

The basic equation of the Tobit Model used above is;

〖HE〗_it*=β_"i" +〖"β*" 〗_(1 ) Death_it+ β_"n " 〖SES_Cov〗_"it " +u_(it )+ε_(it )

Where the condition,

〖HE〗_it= 〖HE〗_it* if 〖HE〗_it*>0

and,

〖HE〗_it=0 if 〖HE〗_it*<0

〖HE〗_it* is the latent variable representing the household expenditure before censoring. It's a linear combination of the model parameters, including individual fixed effects β_(i ), the effect of death status 〖"β*" 〗_(1 ) Death_it, and other factors u_(it ). The model is conditioned by 〖HE〗_it, which is the observed household expenditure, which is subject to censoring if 〖HE〗_it*>0 then the observed expenditure is the same as the latent variable 〖(HE〗_it=〖HE〗_it*). However, if 〖HE〗_it*<0, then 〖HE〗_it=0, implying that the household did not incur any health expenditure.

Therefore, the Tobit model is valuable when dealing with data that includes a substantial number of observations with zero values (censoring). By estimating the latent variable 〖HE〗_it*, the model accounts for both the observed expenditure and the censored cases where expenditure is zero. Also, β_(1 )indicates the change in the latent variable associated with a unit change in the death status variable, providing insights into how death status influences household expenditure in situations where censoring may occur.

Descriptive Statistics

Table 1 presents descriptive statistics for a sample of 600,000 individuals across representative households from the 2018–2023 panel. It provides insights into member status (alive or deceased), health expenditure, and various household demographic and socioeconomic characteristics. The mean health expenditure amounts to Rs. 534.67 ($6.44), with the highest recorded expenditure reaching Rs. 572,050 ($6,874.04). The average age of the study subjects is 34 years, and the death rate among panel individuals stands at 1.0% of the sample (n=3,338,623). It is noteworthy that 97% of individuals self-identify as 'healthy,' while only 36% possess health insurance coverage. Hospitalization records indicate that 0.1% (n=2,268,434) of individuals have reported admissions. In terms of gender distribution among study participants, 46% are female and 54% are male. On average, approximately 43% of individuals have achieved a secondary level of education. Marital status reveals that approximately 59% of individuals are 'ever-married,' surpassing the proportion of singles at around 33%.

The descriptive analysis highlights a non-linear distribution in household sizes, with health expenditure proportions varying across different household compositions. For instance, single-member households allocate 29.5% of expenditure to health, while households with 2 members allocate only 0.05%. Additionally, the introduction of the Marital Status variable after 2018 affects the sample size for this variable.

Implications of the study

Studying household expenditure in India, considering rational and emotional factors, is vital due to extended family structures and high emotional engagement. Large-scale panel studies are uncommon in developing countries, making our research involving 6.5 million individuals significant. Our goal is to identify causation on a broad scale, aiding policymakers in lowering out-of-pocket healthcare costs to achieve the Sustainable Development Goals by 2030.