Consumption Insurance and Vulnerability to Poverty: A Synthesis of the Evidence from Bangladesh, Ethiopia, Mali, Mexico and Russia

Abstract

This paper synthesises the results of five IFPRI studies using household panel data from Bangladesh, Ethiopia, Mali, Mexico and Russia, which examine the extent to which households are able to insure their consumption from specific economic shocks and fluctuations in their real income. The extent of consumption insurance is defined by the degree to which the growth rate of household consumption covaries with the growth rate of household income. All the case studies show that food consumption is better insured than non-food consumption from idiosyncratic shocks. Adjustments in non-food consumption appear to act as a mechanism for partially insuring ex-post food consumption from the effects of income changes. Food consumption is also more likely to be covered by informal insurance arrangements at the community level than non-food consumption. Households use a portfolio of risk-coping strategies, but may not be equally able to use them. Poorer households may be less able to use mechanisms that rely on initial wealth as collateral. In this regard, public transfer programmes may have a more redistributive effect.

Cet article synthétise les résultats de 5 études de l'IFRI qui utilisent des panels de données concernant le Bangladesh, le Mali, l'Ethiopie, le Mexique et la Russie, et s'interrogent sur la capacité des ménages á assurer leur consommation par rapport à des chocs économiques spécifiques et des fluctuations de leur revenu réel. L'importance de l'assurance-consommation est définie par le degré de la covariance entre le taux de croissance de la consommation des ménages et celui du revenu des ménages. Toutes les études de cas montrent que la consommation alimentaire est mieux couverte que la consommation non-alimentaire par rapport aux chocs idiosyncratiques. Les ajustements de la consommation non-alimentaires semblent constituer un mécanisme ex-post d'assurance partielle des effets des changements de revenus sur la consommation alimentaire. Cette derniére est mieux à même d'ê tre couverte par des arrangements informels au niveau communautaire que la consommation non-alimentaire. Les ménages utilisent un portefeuille de stratégies de couverture de risque, mais ne sont pas tous également capables d'en tirer parti. Les plus pauvres sont peut-être moins en mesure de faire usage de mécanismes fondés sur des garanties constituées par la richesse initiale du ménage. Dans cette perspective, les programmes publics de transfert pourraient avoir un effet plus redistributif.

Notes

Emmanuel Skoufias is at the World Bank, Washington, DC and Agnes R. Quisumbing is at the International Food Policy Research Institute, Washington, DC. Funding for the research was provided by the World Bank as part of work being undertaken by the International Food Policy Research Institute on consumption smoothing and vulnerability. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily reflect the view of the World Bank or IFPRI. The authors would like to thank Stefan Dercon, Lant Pritchett, Paul Glewwe, and Emil Tesliuc for helpful comments and suggestions. All errors and omissions are the authors’.

 1. According to the terminology of Seigel and Alwang [1999], the preceding actions represent a combination of ex-ante risk mitigating and ex-post coping actions both aimed at smoothing consumption. Households may adopt ex-ante risk-reducing management strategies such as diversifying the mix of income generating activities from their given asset base [Morduch, 1994, 1995].

 2. The Russia case study has five repeated observations per household while the Bangladesh and Ethiopia studies have four.

 3. Cochrane [1991], Mace [1991] and Townsend [1994, 1995] provide more detailed expositions of the functional forms for preferences.

 4. Other terms used to characterise income changes are ‘permanent’ versus ‘transitory’. These are related to the terms ‘anticipated’ and ‘unanticipated’ but which set of terms is used depends on whether a study adopts a microeconomic model of expectation formation (such as the rational expectations hypothesis) or a statistically oriented approach to decomposing a time series in income growth. For a paper that attempts to delineate among the predictions of various models of intertemporal consumption, see Jacoby and Skoufias [1998].

 5. Note that including the community/round interaction dummies is equivalent to deviating all variables from their respective community/round mean. For more detailed discussion of this equivalence see Deaton [1997].

 6. When prices and wages are available, one may also want to include these as explanatory variables (first-differenced) in regression equation (6b) [e.g., see Dercon and Krishnan, 2000].

 7. Note that if consumption and leisure are non-separable and labour leisure choices are endogenous, the rejection of the hypothesis that {\rm \beta = 0} does not necessarily imply the absence of risk sharing among households [Cochrane, 1991].

 8. Schechter [2001] and Ligon [2001] explore the same idea for Bulgaria and India, respectively.

 9. The higher the number of time observations per household the lower the variance of the estimated coefficient β.

10. Along similar lines, it is also possible for a wealthy household to be quite vulnerable to risk and yet not vulnerable to poverty.

11. To a large extent our emphasis on consumption insurance instead of vulnerability to poverty originates from the belief that, for any meaningful progress in measuring the latter, one must be willing to adopt a specific model for the intertemporal allocation of consumption and credit constraints faced by households.

12. Interestingly none of the vulnerability to poverty measures proposed to date seems to consider the few known facts about the variance of consumption over time. Deaton and Paxson [1994], for example, demonstrate that within any given cohort the variance of consumption increases over time and this variance may differ across cohorts. This implies that at any given point in time any attempt to characterise the variance of consumption changes of households must take into account the age distribution of the population since different households are likely to be at different points in their lifecycle.

13. This approach is taken in all of the IFPRI papers surveyed here. Details are discussed in the next section of the paper.

14. This point is also noted by Deaton [1990].

15. Deaton [1997] first noted that the coefficient of idiosyncratic income changes in specification (7) will be (mechanically) identical to the coefficient of idiosyncratic income changes in specification (5), where the community/survey round interaction dummies are used instead of the change in village mean income.

16. This draws heavily from Dercon [2002].

17. Empirically, one can distinguish between positive and negative shocks, although in the present paper we impose the same coefficient on income changes.

18. The project description at www.cpc.unc.edu/rlms provides complete information about the RLMS survey and its sampling procedure.

19. In principle, insurance arrangements are easier to organise and implement in small or closely-knit communities than in larger groups, where the moral hazard, incentive and information difficulties are more severe.

20. The only difference is the illness shock, which typically has a two-week recall period. Food consumption uses both a one-week and one-month recall period, while non-food consumption uses a longer reference period (one month for frequently purchased items, and since the last survey round for more infrequently purchased items).

21. The shock variables in the Ethiopia study are as defined by Dercon and Krishnan [2000], where a value of 1 indicates the best outcome. Thus, these shocks should be interpreted as positive shocks, and positive coefficients imply that consumption increased as a result of positive shocks.

22. The relatively higher coefficients of these shocks for non-food than for food consumption might also be explained in terms of underlying household preferences. Ceteris paribus, in so far as the incidence of these shocks represents a decrease in household income then the quantity demanded for luxury goods (non-food) will decrease more than for necessities (such as food that has an income elasticity less than 1).

23. For Mexico, see table 2 panels a and c in Skoufias [2002b]. Results for Bangladesh and Ethiopia are available from the authors.

24. We thank Stefan Dercon for pointing this out.

25. Of course, shocks arising from changes in household structure may not be completely captured by changes in household income, nor compensated by the per capita adjustment factor, which controls only for changes in household size, not household structure. Note, however, that owing to the short time interval between rounds, household structure does not change substantially from round to round. In work not reported here, we disaggregated household size into different demographic categories but the overall results were not appreciably different.

26. As already pointed out in Note 22, the relative differences in the size of the estimated income coefficients for food and non-food may also be attributed to preferences. Food is typically a necessity with a lower (<1) income elasticity while non-food is a luxury good with a higher (>1) income elasticity.

27. Measurement error in the income variable biases coefficients toward zero while imputation errors in food consumption may bias the income coefficients upward [Deaton, 1997].

28. In all five country studies, the shock variables used as identifying instruments in the first stage regressions, were significant and negatively correlated with the growth rate of income. Other instruments included changes in income from sources that were not likely to be correlated with crop production. Tests on the excluded instruments rejected the null hypothesis that they were equal to zero.

29. It necessary to acknowledge that the low degrees of freedom associated with each household-specific regression result in very high standard errors for the estimated β or consumption insurance measure of each household. Also, as noted earlier, there remain potential complications due to measurement errors in the income variable. In the absence of a better alternative it was determined that it was worthwhile to explore this approach in spite of the limitations just noted. In order to minimise the potential influence of extreme outliers, values of household-specific β's less than the 1 percentile and greater than the 99th percentile of the distribution of β's across all households were excluded from the later stages of the analysis.

30. Equations (8) and (9) were also estimated with an intercept term. This did not result in any remarkable changes in the estimates reported in Tables 6 and 7.

31. It should be noted that a similar approach was adopted in the Bangladesh study and it yielded no significant correlation between vulnerability and the probability of ‘ever being poor’ and ‘being always poor’.

32. All shock dummy variables were included simultaneously in the probit regression. Estimation using random effects (at the village level) probit did not lead to any substantive change in the results obtained using simple probit. The case study also included separate estimates for PROGRESA (treatment) villages.

33. The analysis in Skoufias [2002b] also suggests that there does not appear to be any significant differences in how households in PROGRESA villages respond to these shocks. The only notable difference is that households in PROGRESA villages seem to respond differently than households in control villages when there is shock leading to the loss of animals. Relative to households in control villages, they are less likely to respond by selling animals or borrowing, or working more, and more likely to receive help from relatives. Also, the loss of other household items or the loss of a home is more likely to result in receiving help from the government. There are also indications that the presence of the PROGRESA programme induces households to use adjustments in their labour supply less frequently than households in control villages.

34. We are grateful to Stefan Dercon for most of the points raised in this paragraph.

Additional information

Notes on contributors

Agnes R. Quisumbing

Emmanuel Skoufias is at the World Bank, Washington, DC and Agnes R. Quisumbing is at the International Food Policy Research Institute, Washington, DC. Funding for the research was provided by the World Bank as part of work being undertaken by the International Food Policy Research Institute on consumption smoothing and vulnerability. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily reflect the view of the World Bank or IFPRI. The authors would like to thank Stefan Dercon, Lant Pritchett, Paul Glewwe, and Emil Tesliuc for helpful comments and suggestions. All errors and omissions are the authors’.