中国农村对风险作出的行为反应

WPs iq- V POLICY RESEARCH WORKING PAPER 1978 Behavioral Responses Does risk perpetuate poverty in a credit-constrained to Risk in Rural China econory7 Income risk appears not to discourage Jyotsna jalan schooling but does inhibit the Mart'in Ravallion out-migration of labor. Only a small share of wealth is held in unproductive liquid for to protect against income risk. The World Bank Development Research Group Poverty and Human Resources September 1998  POLICY R;ESEARCH WORKING PAPER 1978 Summary findings Does risk perpetuate poverty in a credit-constrained households do not, it seems, need to hold unproductive economy? cautionary wealth, and the poor probably cannot afford Jalan and Ravallion study portfolio and other to do so. behavioral responses to measured risk using household The authors find no evidence that income risk panel data for rural China. discourages schooling, but risk does inhibit the out- One-quarter of wealth is held in unproductive liquid migration of labor. forms. But only a small share of this appears to be a Generally, the results provide only limited support for precaution against income risk. the idea that uninsured risks promote unproductive The authors estimate that eliminating income risk portfolio behavior in this setting. There is such an effect, would reduce the share of wealth held in liquid form by but it is small in magnitude and cannot be deemed an less than 1 percentage point. Moreover, that effect is important cause of poverty. confined largely to middle-income groups; high-income This paper - a product of Poverty and Human Resources, Development Research Group - is part of a larger effort in the group to better understand the causes of poverty. The study was funded by the Bank's Research Support Budget under the research project "Dynamics of Poverty in Rural China" (RPO 678-69). Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Patricia Sader, room MC3-632, telephone 202-473- 3902, fax 202-522-1153, Internetaddresspsader@worldbank.org. The authors maybe contacted atjjalan@worldbank.org or mravallionCaworldbank.org. September 1998. (35 pages) The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the view of the W/orld Bank, its Executive Directors, or the countries they represent. Produced by the Policy Research Disszrnination Center  Behavioral Responses to Risk in Rural China Jyotsna Jalan and Martin Ravallion' World Bank Does risk perpetuate poverty in a credit-constrained economy? We study portfolio and other behavioral responses to measured risk in household panel data for rural China. One quarter of wealth ,s held in unproductive liquidforms. But only a small share of this appears to be a precaution against income risk We estimate that eliminating income risk would only reduce the share of wealth held in liquidform by less than one percentage point. Furthermore, this effect is largely confined to middle income groups; high-income households do not, it seems, need to hold unproductive precautionary wealth, and the poor probably cannot afford to do so. We find no evidence that income risk discourages schooling. However risk inhibits the out migration of labor. Keywords: Insurance; precautionary wealth; poverty JEL classification: D91, Q12 Address for correspondence: M. Ravallion, World Bank, 1818 H Street NW, Washington DC, 20433, USA. Our thanks go to staff of China's State Statistical Bureau for their assistance. Helpful commnents were received from Shubham Chaudhuri, Paul Glewwe, and Dominique van de Walle. The financial support of the World Bank's Research Committee (under RPO 678-69) is gratefully acknowledged.  1. Introduction As a stylized fact, there is great uncertainty about incomes and health in underdeveloped rural economies, and the instruments to insure against those risks are weak or absent. At the same time, there is pervasive poverty. Are these features of poor rural economies causally connected? In particular, does risk promote forms of rational behavior which help perpetuate poverty? One way that risk might create poverty is by inducing poor credit-constrained households to hold high levels of relatively unproductive liquid wealth.2 If borrowing is not an option when there is a sudden drop in income, then liquid wealth will be needed to protect consumption. Famously, Keynes (1973, p.170) identified a "precautionary motive" as "..the desire for security as to the future cash equivalent of a certain proportion of total resources". Less famously, he also believed that India was "..a country impoverished by a preference for liquidity" which stifled the "growth of real wealth" (Keynes, 1973, p.337).3 IThe idea that (rational) portfolio behavior in the presence of uninsured risk can help perpetuate poverty appears to be longstanding. It can be found in relatively early writings on 2 Strictly, borrowing constraints are not necessary for precautionary saving. If the marginal utility of current consumption is a convex function of consumption then (by Jensen's inequality) a mean- preserving increase in uncertainty about future incomes will increase the marginal utility of future consumption; current savings will rise to preserve intertemporal equilibrium even without borrowing constraints (see, for example, Gersovitz, 1988). Such a model.does not, however, explain an effect on the composition of wealth holdings; higher risk will encourage higher saving, but it can be in any form. Nor is precautionary saving the only way in which risk can create poverty. Another is via effects on production decisions; for example, outmoded agricultural technologies may persist because they are less risky (see, for example, Morduch, 1995). 3 It is not at all clear from Keynes (1973) what his own views on the causes of poverty in India were based on. In his biography of Keynes, Skidelsky (1983, p. 176) writes that, "although he was to write and advise extensively on Indian affairs, the furthest east he ever got was Egypt; the only Indians he ever met were at Cambridge or London; the only books he ever read on India were specialized tomes on finance". 2  finance and development (such as Patrick, 1966) as well as recent discussions (World Bank, 1998), and emerges in more formal terms in the Bencivenga and Smith (1991) model of endogenous growth with multiple assets. The idea has also been seen to strengthen the case for public efforts to promote better institutions for financial intermediation in poor rural economies. The plausibility of the claim that precautionary portfolio behavior can cause poverty is not self-evident, however. One can readily agree that there are benefits to poor people from self- insurance; they are likely to be more credit constrained, ancl (possibly) more averse to risk and more exposed to it. But there are costs too, and arguably the poor will not be able to afford to hold a large share of their wealth in unproductive forms. Poverty is surely a strong inducement to assuring that one's own resources are not idle. Adequate insurance may still be possible by holding only moderately liquid but still relatively productive forms of wealth. This paper aims to contribute to knowledge about behavioral responses to risk in poor rural economies and the role this might play in perpetuating poverty. Formal economic models of saving behavior have only recently begun to systematically incorporate uncertainty.4 A new body of microeconometric work using data from underdeveloped agrarian economies has looked for effects of rainfall variability and other income risks on tlle flow of consumption or savings (including asset transactions), or on growth rates of consumption.5 However, as Keynes argued, 4 Contributions include Zeldes (1989), Caballero (1990), Kimball (1990), Carroll (1992, 1997) and Deaton (1991, 1992). 5 For surveys of past research on risk and savings in developing countries see Gersovitz (1988), Alderman and Paxson (1992) and Besley (1995). Empirical studies of the effect of income risk or variability on savings behavior (including asset transactions) in poor rural economies include Paxson (1992), Rosenzweig and Binswanger (1993), Rosenzweig and Wolpin (1993), Alderman (1996), Dercon (1998) and Fafchamps, Udry and Czukas (1998). Carroll and Sarnawick (1997) review recent research on 3  and has since been formalized in theories of demand for money balances when the.future interest rate is uncertain (following Tobin, 1958), it is the stock of their liquid wealth which protects people from risk, not the flow into wealth as such. So instead of looking for effects of risk on the flows of savings or consumption, this paper tests directly for portfolio effects of risk. In particular, we test whether those households facing higher idiosyncratic risk to their incomes tend to hold a higher share of their wealth in unproductive liquid form than one would expect given 1heir permanent income and other characteristics. We also test for other potentially costly responses to risk. The setting for our empirical work is rural areas of southwest and southern China.6 Our data cover a period (1985-90) after reforms began which entailed abandoning the old commune system of agricultural production in favor of allowing individual farmers the freedom to make their own production choices. Previous research found considerable vulnerability to idiosyncratic risk in this setting (Jalan and Ravallion, 1998a). So an aggregate (village- or area- wide) measure of risk, such as rainfall, is unlikely to provide a good estimate of individual incom,e risk. Here we use instead the household-specific income process over time to identify income risk, following Carroll and Samwick (1997). We extend this method by controlling for any persistence in the errors of the income process. We find that the share of wealth held in liquid form has a severely skewed and kurtotic distribution, and that this non-normality persists after controlling for a broad set of household precautionary saving in developed countries. 6 For a comprehensive overview of what is currently known about household savings behavior in China see Kraay (1997). 4  characteristics. With such heavy tails in the distribution, standard estimation methods found in the literature will not be robust. To assure that our analysis of responses to risk is robust to the non-normality, we apply recent advances in quantile regression methods. We also allow for two other sources of risk that we expect to be important in this setting, namely the variability of foodgrain yields on the household's own farm, and a measure of medical risk. Transport and transaction costs in this setting could mean that risks to own-farm food output matter independently of their implications for overall income risk. Health risk could well entail large increases in medical spending which would influence savings. There has been work on the effects of such health risks on precautionary saving for developed countries, though no behavioral responses have been identified.7 There is a further issue of how "liquid wealth" should be defined in a poor rural economy. One might focus solely on money balances (cash in hand). However, a foodgrain stock can also be a good hedge, as has been recognized in the literature.8 So we define liquid wealth as grain stock plus cash in hand. In defining total wealth we exclude land, which is mainly allocated administratively in rural China; since the market is very thin, valuation is impossible. Non-liquid wealth includes bank deposits, farm capital, livestock, housiing and consumer durables. Besides holding liquid wealth, we examine two other ways in which households might protect themselves from risk in this setting, namely by not sending their children to school and 7 We refer to the work of Hubbard, Skinner and Zeldes (1994), who find no effect of health risk on savings using U.S. data. s See, for example, Patrick (1966). There is recent supportive evidence on the precautionary role of foodgrain stocks from Park's (1995) surveys in poor areas of northwest China. 5  by temporary out migration of family labor (which,though restricted in the past, is becoming more common in China). One could make theoretical arguments either way about how risk affects these variables. Consider labor export. Greater income uncertainty might encourage out migration as a risk diversification strategy (as argued in, for example, Rosenzweig, 1988). Alternatively, when rural labor markets are thin, risk can discourage migration, due to concerns about possible labor shortage on the farm. This is plausible in rural China, given that farm labor markets are also thin or non-existent. A labor surplus on average is not then sufficient for out migration - the family will also take account of the variability in demand for family labor. Consider schooling instead. It has been argued that income risk discourages investment in human capital., though there is little evidence.9 Again the effect could go either way. On the one hand, keepinig kids in school may expose the family to higher risk of family labor shortage, but (on the other hand) better educated children could be expected to directly reduce future income risk. The next section outlines our test for risk effects on liquid wealth holding. Section 3 describes our data, while section 4 presents our results. Conclusions are in section 5. 2. Modeling behavioral responses to risk It will help motivate our empirical analysis to begin by considering a simple theoretical model of a farm-household's choice between holding liquid wealth and investing in a risky production activity. 9 The only study we know of is Jacoby and Skoufias (1997), who find seasonal effects on schooling of income risk in semi-arid areas of India. 6  2.1 A model ofprecautionary wealth Consider a household facing a two-period decision on how to allocate its initial wealth W between current consumption, investing an amount K in a risky production activity, and holding an amount Mof an unproductive but secure liquid asset. Utillity at any date is a strictly increasing and concave function of consumption at that date, and goes to minus infinity as consumption goes to zero. Utility in the first period is U(W-K-M). Output in the second period is F(K, ,) where F is increasing and non-convex in K but also (lepends on the realization of a random variable t. The value of F(K, 4) exceeds K for at least some values of . F(K, 4) is also assumed to fall to zero (a total crop failure, for example) for some values of 4 irrespective of the value of K. These conditions assure that there will be positive investment in the risky activity, but that at least some liquid wealth will also be held as insuramce (for otherwise there is a positive probability of zero consumption, which gives infinite disutility). The choice of K and M maximizes expected utility: U(W -K -Mf) +E EU[F(K, 4) -,fl l Since our assumptions imply interior solutions for K and M thiese must satisfy: U'(W -K -M) =E, ,U'[ F(K,+) tM] (2) E EqU'[F(K,4) MIF K(K, 4) The choice of M and K will depend on W and the properties of the distribution of 4. On differentiating (2) with respect to W and exploiting the second-order conditions it is readily 7  verified that M will be a strictly increasing function of W if and only if: E U'(F +M)FKK + U"(F+M)(FK- I)F K] < 0 (3) A sufficient condition is that FK> 1 for all ,. This must hold for some R since F(K)>K, but it may not hold at all values. Nonetheless, the inequality in (3) is not a strong assumption, and it implies that it will be the poorest (in terms of W) who hold the lowest amount of liquid wealth at any given level of risk. So this model must make one immediately skeptical of any claim that precautionary liquidity preference is largely confined to the poor. To give a tractable example with an explicit solution for this model, suppose that there are two possible outcomes in the second period: either the investment fails to produce anything or it succeeds, with a rate of return r>O. Suppose also that individuals hold logarithmic utility functions. Then M and K maximize In(W-K-M)+p. lnM+(l-p).ln[(l+r)K+M where p is the (positive) probability of failure. It is readily verified that the solution for M is p(l + 1 /r) W/2 which is strictly increasing in both Wand p, and decreasing in r. Notice that not only do the poorest (in terms of W) hold the lowest amount of liquid wealth at any given level of risk and rate of return, but their demand for this formn of wealth is least responsive to risk (since aM/ap is increasing in W). Total wealth carried over is WJ2 and the share of it held in liquid form (MI(M+K)) is p(l+llr)l4. For example, with a 25% rate of return and a 20% chance of failure, one quarter of wealth will be held in liquid form. This model could be extended in any number of ways. For example, one could easily introduce transaction costs which are decreasing in M, implying both a "transactions motive" and 8  "precautionary motive" for liquidity. Heterogeneity can be readily introduced by allowing for a vector of household characteristics that influence either the utility function or the production function. At high levels of initial wealth one might also conjecture that the above model will become less relevant, since more efficient means of insurance will probably become available. To sketch an extended version of the above model which incorporates an alternative insurance instrument, let us assume that crop insurance is offered to any farmer who is willing to pay some positive minimum premium in the first period, sufficient to cover a fixed administrative cost and the insurer's expected payout in the second period. Beyond some critical initial wealth (sufficient to afford the crop insurance), this option will start to be the preferred method of insurance because its payouts are state contingent. Thus one can expect that demand for liquid wealth as insurance will initially rise with wealth, but then fall after some point. 2.2 Method of testingfor precautionary responses to risk To implement an empirical test for precautionary behavior we must find a measure of the income risk facing the household. Here we follow Carroll amd Samwick (1997) in basing that measure on the estimated innovation errors from an income process of the following form: InYi, = a+X +eI,(4) where Y. is the income of household i in time t, and Xi is a vector of exogenous variables. The error structure is assumed to be: li + Vit (5) 9  where 11, is a random individual component with mean zero and variance a In the standard error component model, the errors are only correlated over time through the individc, A specific effect Ti. The v. 's are assumed to be i.i.d. random variables. However, for a variable like income it is quite possible that an unobserved shock in the current period will affect t]he behavioral relationship in at least the next period if not more. The persistence in the errors of the income process over time implies that simply using the variance of the estimated v,'s as the income uncertainty measure will understate the total income risk. In order to estimate precautionary savings, we need the variance of an i.i.d process. We assume that the random variable v., is an AR(1) process: ,it P Vit,-I it (6) where p (with I p I < 1 ) is the serial correlation coefficient and (of, is a random i.i.d. error with mean zero and variance . Ignoring the serial correlation will still give consistent estimates of the regression coefficients, but the standard errors will be biased which will bias our estimate for income uncertainty. The explanatory variables are assumed to be orthogonal to il, and xi,, i.e., E(Xi,' l) = E(Xi, I,)d = E(1 'co),) = 0 (7) We first test for p = 0 using the Bhargava, Franzini and Narendranathan (1982) generalized Durbin-Watson statistic (d p)."' Provided we reject the null that p = 0, we transform the usual 10 The test statistic is: N T N r dp= S E (u -u 2/ E f=1 t=2 1t=1 10  AR(l) model into a serially uncorrelated regression with independent observations using the Prais-Winsten transformation. Thus the transformed regression disturbances are: