Vulnerability of households to health shocks: an Indonesian study

Abstract

We examine the effect of illness and injury shocks on work hours and household consumption in Indonesia. Using indices of activities of daily living to measure health shocks, we find that both labour hours and household consumption are influenced by health shocks to household heads. Further, farm households seem to be more seriously affected than non-farm households by health shocks. However, the magnitude of the health effect on household consumption is small, implying that even farm households are well protected on average by the presence of formal and informal risk-coping mechanisms.

Acknowledgements

We are grateful to Albert Park and Yasuyuki Sawada for generous comments on an earlier draft and to Kathleen Beegle for her helpful responses to our data inquiries. We benefited from constructive suggestions provided by the editor and two anonymous referees.

Notes

¹For example, it is well known that mortality is higher among poorer individuals.

²The death of a productive household member is in a sense the most extreme health shock to the household. ADLs cannot capture this potentially severe health shock, and this is a limitation of our health measure. Later, we examine briefly the welfare effect of the deaths of household heads.

³Differences in levels of ADLs across households at one point in time would largely reflect health inequality caused by differences in past health investments and in health endowments at birth. Health shocks should be construed as adverse changes in health over a relatively short time period. Our panel data set consists of three rounds of data (1993, 1997 and 2000), so changes in ADLs between surveys may be gradual rather than abrupt. Even so, change in ADLs is conceptually a better measure of health shock than level of ADLs.

⁴Labouring is also physically demanding. However, less than 4% of the sample household heads (to be used for the subsequent regressions) were labourers.

⁵The IFLS is described at <http://www.rand.org/labor/FLS/IFLS/index.html>; funding sponsors are listed at <http://www.rand.org/labor/FLS/IFLS/teamfund.html>.

⁶North Sumatra, West Sumatra, South Sumatra, Lampung, DKI Jakarta, West Java, Central Java, Yogyakarta, East Java, Bali, West Nusa Tenggara, South Kalimantan and South Sulawesi. Data from the fourth wave of the survey (undertaken in 2007–08) are now publicly available, but they are not used here.

⁷Frequently purchased non-food items include utility services; household items such as soap and detergent; transport; and recreation. Less frequently purchased non-food items include clothing; furniture; and medical, education and ceremonial items and services.

⁸By using 2000 Jakarta prices as the base (not only for Jakarta but for other regions as well), we generate real consumption figures that are comparable not only inter-temporally but also inter-regionally.

⁹One weakness of our study is that we do not take into account health shocks to productive household members other than household heads. If other working household members contribute significantly to household income, ignoring health shocks to them leads to imprecise estimates of the effect on household consumption of health shocks to household heads.

¹⁰Both urban and rural areas exist in each of the 13 sample provinces except for Jakarta, where there are no rural areas, so there are 25 regions in the sample. We created region dummies for each change of years (the change from 1993 to 1997 and the change from 1997 to 2000). Thus, there are 50 region–year dummies. One of them, (urban) Jakarta from 1997 to 2000, is the reference group, so we include 49 region–year dummies in the regressions. Because of the inclusion of region–year dummies, the constant α _jt is specific for each region–year.

¹¹Of course, the region–year dummies cannot address potential differences in omitted characteristics within region–years.

¹²The ordinary least squares (OLS) technique implicitly assumes that sample observations are randomly drawn from the population, implying that any observation should be a realisation from random sampling. In our case, however, a change in per capita household expenditure is observed twice for each household, both between 1993 and 1997 and between 1997 and 2000. These two observations for an identical household would not be two realisations from random sampling. For example, if the health of the household head deteriorated between 1993 and 1997, it is likely to have become even worse between 1997 and 2000. Because we have clustered standard errors at the household level, the resulting standard errors are unbiased, even if two observations from an identical household are not realisations from random sampling.

¹³A weakness of our study arises from the timing of the measurement of health status and labour supply. Health status was measured at the time of the survey interview, whereas labour supply was measured for the period one year before the date of the interview. Thus, our measures of labour supply are not sensitive to health shocks that occurred shortly before the date of the interview, especially in the case of the dummy for the household head's participation in work. A similar weakness applies to our measures of household consumption.

¹⁴Strictly speaking, a 1 SD decrease in intermediate ADLs is not comparable with a 1 SD decrease in basic ADLs. Thus, the magnitudes of the coefficient estimates on the ADL indices should be interpreted with caution. However, the degree of statistical correlation between the dependent variable and each ADL index should be useful in determining which kinds of daily activities move more closely together with the dependent variable (labour supply in this sub-section and household consumption in the next sub-section).

¹⁵Since the logarithm of 0 is undefined, relying on the logarithm of real per capita medical expenditure would reduce the number of sample households by about 30%.

¹⁶ Table 5 confirms the validity of our econometric identification strategy. It supports the conjecture that poor or unhealthy household heads are more likely to experience negative health shocks than those in better health or financial circumstances. In other words, household heads who later experience adverse health shocks during a given period of time are on average already less healthy or poorer than household heads who do not later experience adverse health shocks. If the level of household consumption is regressed on indicators of adverse health shocks, such as a sizeable decline in BMI or a long in-patient spell (as in Wagstaff 2007), the estimation would be biased, because households that later experience adverse health shocks tend to be relatively unhealthy and poor to begin with.

¹⁷The coefficient estimates on the control variables are similar, regardless of the categories of consumption and the measures of health.

¹⁸The estimated value of θ in model (2) is statistically significant at conventional levels only when the dependent variable is either food consumption or total non-medical consumption (no matter which health measure is used). For both types of consumption the estimated value of θ is positive, implying that farm households on average increased (reduced) food consumption and total non-medical consumption proportionately more (less) than non-farm households.

¹⁹Column 1 of table A2 suggests that after the deaths of the household heads, medical costs decline in non-farm households but increase in farm households, but neither effect is statistically significant.

²⁰Our argument is based on the degree of statistical correlation between the dependent variable (labour supply or household consumption) and each of the ADL indices.

²¹Gertler and Gruber (2002) examined the effect of ADL changes on changes in labour supply using the overall ADL index only (that is, with no separate estimation for intermediate and basic ADL changes). They used intermediate and basic ADL changes separately as the instrument for changes in income in estimating the effect of income changes on expenditure changes. Since they imputed wages to all workers based on wages in the formal labour market (no matter whether each sample household head worked in the formal labour market or in a family farm or enterprise), their income imputation, which is the product of the imputed wages and labour-supply hours, ignored the effect of health shocks on the returns to labour, as the authors acknowledge on page 57 of their paper. If intermediate ADL shocks have a larger effect on the return to labour than basic ADL shocks, then Gertler and Gruber (2002) under-estimate the importance of intermediate ADLs relative to basic ADLs in their estimation of the effect of ADL changes on changes in household expenditure through changes in household income.

²²The present study did not examine the impact of a large health shock (such as a 1 SD decrease in intermediate ADLs) on household income. Thus it is possible that a 1 SD decrease in intermediate ADLs reduces household income (and consumption) by only 3%, which would imply the absence of any risk-coping (consumption insurance) mechanism. However, given the nature of the ADL indices, it seems implausible that a 1 SD decrease in intermediate ADLs would have so little impact on household income.