1. Introduction

Understanding poverty and its dynamics by focusing on welfare levels and distribution in a certain socio-economic context doesn’t suffice and present the real picture of the underlying process behind the observed levels of deprivation. It is understood that besides many other factors that explain the welfare and poverty dynamics, risk and shocks are important causes of persistent poverty. People in developing countries face numerous uninsured risks such as human illness, sickness, death of livestock, crop pests and diseases, erratic rains or droughts, political strife , etc. ( Dercon 2002 and 2005, Hoogeveen et al.2005).

Risk affects whether people can maintain assets and endowments, how these assets are transformed into incomes and earnings are translated into broader development outcomes, such as health and nutrition. Risky events are treated as ‘exogenous’, not directly under the control of people. However, an essential part of analyzing risk and its consequence on poverty are that households use sophisticated ex-ante and ex-post strategies to manage, reduce, or cope the consequences of risk (Dercon 2005b).

To smooth income and consumption, poor people use different risk-coping strategies, markets or technologies, conservative production and employment decisions such as storage of grain; land fragmentation; borrowing and saving, depleting and accumulating nonfinancial assets, adjusting labor supply, sell assets, or send their children to work instead of school to supplement income, and employing formal and informal insurance arrangements such as informal credit and gifts among friends, relatives and neighbors, borrowing from local money lenders, rotating savings and credit associations (ROSCAs), interlinkages in agricultural contracts, and so forth (Daniel 2003, Dercon 2000, Jacoby and Skoufias 1997, Morduch 1995, Townsend 1995 and1993,Udry 1994, Deaton 1992, Paxon 1992, Deaton 1990 and Rosenzweig 1988).

In developing countries self-insurance is inadequate to protect households from the risk of fluctuating income. In the absence of formal and other inter-temporal markets as an alternative ex-post mechanism, households resort to informal risk-sharing schemes (Daniel 2003). As discussed in Carter 1997, it is rational for households to voluntarily share with their less fortunate neighbors in the hope that their neighbors will help them out sometime in the future. This kind of reciprocity sharing is denoted as “endogenously enforced” because it does not depend on any external norms or authority to function. Reciprocity schemes can be described as vertical or horizontal. Horizontal reciprocity refers to sharing rules between households that have approximately equal wealth endowments, which permit a group to enjoy benefits across individuals in the group. However, there are costs associated with horizontal reciprocity. In addition, to the extent that reciprocity works like a marginal tax on output, it would depress work incentives and potentially result in reduced mean output.

Anthropologists considered informal risk sharing to play a role in securing social status and signaling commitment to the community, however, Economists tend to scrutinize it as they do other transfers like public aid. According to Morduch 1999, these coping strategies, although effective in reducing vulnerability and current poverty, can reduce economic growth, long-term welfare, or social mobility.

Households in rural Ethiopia employ a variety of coping strategies. Skoufias and Quisumbing 2003 (using the ERHS longitudinal data of 1994, 1995 and 1997) and Niggusse 2005 (using panel data of year long intensive monitoring 5^th round rural household survey in selected villages of rural Ethiopia) identified the presence of consumption smoothing where households use all means of risk management to insulate themselves from risk and identified the different coping strategies. Applying limited commitment model to empirically test the role of credit transactions and the effect of informal networks on risk-sharing between rural households in Ethiopia, Daniel 2003 found evidence of risk-sharing arrangement through credit transactions, where enforcement problem limits the direct credit transactions in risk sharing arrangements between rural households. Although, there are voluminous works in the area of the impact of risk and shocks on growth, welfare and poverty, there are still gaps on revealing the impact of coping strategies (especially IRS strategies) on welfare, growth, or poverty dynamics in rural Ethiopia.

Recently, there are emerging views and shift concerning the implication of risk upon poverty dynamics. This study takes part in revealing the role of shocks and informal risk-sharing strategies on the dynamics of poverty. Even if, IRS strategies have their own advantage of reducing risk, under imperfect enforceability this may create adverse incentive problem.

The study uses longitudinal household data of the Ethiopian Rural Household Survey (ERHS) collected by the Department of Economics, Addis Ababa University covering 1477 rural households. However, the sample doesn’t include pastoral households or urban areas. Since some of the questions are retrospective and self-reported, there may be memory tumble and observation bias that may lead to over- or under-reporting of asset levels, consumption, etc.

In section two, we briefly review literatures regarding the concept of risk, shocks and the associated coping strategies, especially IRS in rural village economy settings to relate with poverty and its dynamics. In Section three, we presented the theoretical model of consumption insurance and specified our poverty dynamics model that is appropriate for our panel data. Section four and five discusses the data, descriptive and empirical, respectively. with some implication, we concluded in section five.

2. Review of Literature

2.1. Linking Risk and Poverty

The consensus after the works of Sen 1999 is that poverty encompasses more than just low levels of income or consumption. Although, studies on poverty analysis emphasizes on the welfare levels and distribution, there are two consequences of risk on poverty; there is the impact of shock[2] and the behavioral impact. The impact of shocks and the coping strategies may destroy or reduce the physical, financial, human or social capital of the household. The behavioral impact, on the other hand, is that households faced with risk and with access to limited insurance alternatives, such as assets or safety nets, are pushed towards risk management strategies such as low risk activities and asset portfolios, at the expense of lower mean return and incomes. As in least developing countries, if credit market and insurance markets are poorly developed, exposure to risk may induce household to hold least productive asset portfolios for the purpose of buffering consumption (Dercon 2005b).

The direction of causation can also be reversed so that poverty causes exposure to risk. As discussed in Hoogeveen et al.2005, to avoid extreme income poverty households may choose to cultivate in insecure areas, land infested with landmines, areas where rebels are active, or live in an unhealthy/unsafe environment.

2.2. Risk and coping strategies

In addition to coping strategies employed by households, there are different forms of ex-ante and ex-post institutional coping strategies to manage risk and its consequences. These strategies can be categorized into three main institutional arrangements. First, market based arrangements; these have great potential and, where available, households and individuals take advantage of the financial products offered by insurance and banking institutions. Second, public arrangements; there are arrangements made by governments to deal with social risks such as unemployment, old age, work injury, disability, widowhood, and sickness. Third, informal arrangements; in a situation where there is missing market or public institutional arrangements to deal with idiosyncratic and common risks, individual households respond to risk through informal arrangements. They involve a system of mutual assistance between family networks or community members. The first two institutional arrangements are none or limited in LDCs (Hoogeveen et al.2005).

2.3. The theory of Full-risk sharing and the theory of limited commitment

Any two agents may be said to share risk if they employ state-contingent transfers to increase the expected utility of both by reducing the risk of at least one. Risk-sharing can be viewed as the cross-sectional equivalent of consumption smoothing over time. Full risk sharing is a situation in which all idiosyncratic risk is eliminated. Since risks are shared, the marginal utilities of consumption are perfectly correlated across all agents. That is, movement in average group consumption represents aggregate risk. Full risk sharing is an important feature of any Pareto efficient allocation in an Arrow-Debreu economy; provided that agents have von Neumann-Morgenstern preferences, no one is risk-seeking, and at least one agent is strictly risk averse[3] (Townsend 1994).

There are a number of empirical works that tests whether household consumption allocations replicate the Pareto-efficient full risk-pooling outcomes resulting from a complete set of competitive state-contingent markets, i.e, testing the null hypothesis of full risk-sharing. For instance, Mace 1991 and Cochrane 1991 (US data), Deaton 1992 (Cote d’Ivoire, Ghana, and Thailand), Townsend 1994 (ICRISAT data from semi-arid India), Fafchamps and Lund 1997 (Philippines), Skoufias and Quisumbing 2003 based (Bangladesh, Ethiopia, Mali, Mexico, and Russia), Daniel 2003 (Ethiopia), Niggussei 2005 (Ethiopia), etc are some of the empirical works in the area of risk sharing. In sum, the finding reveal the estimated response of consumption to income shocks is small but significant, suggesting a rejection of the null hypothesis of full risk sharing or perfect insurance.

2.4. Limited information, limited commitment and risk-sharing

It has been argued that information asymmetry among insiders is not a stern problem in rural village economies. However, the setting in these rural villages doesn’t support the assumption of full information (Udry 1994 and Kocherlakota 1996 in Daniel 2003).

Among the efforts made by different authors to explain the failure of full risk-sharing in the context of developing countries, Ligon 1998 and Ligon et al. 2000b are the most citable one, who suggests to relax the assumption of full information and then replace it by a system of private information that excludes some contracting possibilities due to moral hazard and adverse selection problems. The failure of full risk-pooling may be due to either problems of limited information, limited commitment or both (Ligon et al. 2000b). Cognizant of this problem, recent papers appeal to the theory of limited commitment to explain the observed positive relationship of individual consumption with current and lagged individual income[4].

In their successive work, Ligon et al. 2000a, examine a dynamic limited commitment model of mutual insurance by introducing intertemporal substitution possibilities, such as intertemporal production, storage, or access to external credit market. They show that under certain conditions savings enhance the use of mutual risk-sharing as a subgame perfect equilibrium, while under another condition it encourages agents to renege by tightening their sustainability constraints as it increases utility derived from autarky.

2.5. IRSS in Ethiopia

Driven by religious, culture or based on reciprocity, historical evidences from Ethiopia disclose the existence of risk sharing networks both in rural and urban areas. They have traditions of gathering to defend their village from aliens, participate in community development activities like erosion preventions, gather for funerals, wedding ceremonies, religious festivals, in sickness, etc. Institutions and activities of the informal sector in Ethiopia include rotating savings and credit association, for example – Eqqub, mutual aid association, such as Iddir, and local moneylenders, agricultural interlinkages such as sharecropping and calling work party/labor sharing, Debo/Wonfel. Eqqub and Iddir are usually formed among persons united in family and friendship, by place of work, by living in the same localities, etc (Daniel 2003).

3. The Model

3.1. Theoretical model of consumption insurance

In this part we discus the theoretical and econometric model of consumption insurance. The theoretical underpinning of consumption insurance and full risk sharing is based on the theory of full insurance introduced by Arrow (1964) in uncertain economies and later developed by others (Morduch 1995, Towensend 1994, Besley 1995, Deaton 1992, Paxon 1992, Cochrane 1991, Rosenzweig 1988, etc). Under fully functioning market, households will not be vulnerable to income shocks, where all risk should be diversified away so that shocks should have no impact on consumption level. The consumption insurance tests are based on the proposition that with full-insurance, consumption growth should be independent of idiosyncratic variables[5]. In other words, discounted marginal utility growth of consumption should be the same for all households under full risk-sharing.

Based on the theoretical model employed by Cochrane 1991, Townsend 1994 and Skoufias and Quisumbing 2003, let us consider an endowment economy with N households indexed by maximizes the sum of life time utility subjected to community resource constraints. Assume that expected utility of households is time separable and inter-temporally additive for period , defined over instantaneous utility . Households face a finite set of possible states of the world,, each of which occurs with probability . Furthermore, let and , i.e., instantaneous utility is concave[6]. This implies that households will have an incentive to smooth consumption.

For household at time and state with state-contingent consumption the problem facing a social planner is to solve

[1]

Subject to the resource constraint

For all

Where, is household i's Pareto weight which are assumed to be constant over time and, is household i's rate of time preference, is the probability that state occurs, is the i^th household utility function, is household i's consumption at date t and state, and is a preference shocks. Our Lagrange equation, therefore, will be

[2]

If we let denote the Lagrangian multiplier associated with the resource constraint for period t and state. By taking the first derivative of with respect to consumption and satisfying the condition in all periods and states for every agent, for the social planner, it follows that

[3]

For any period t, any pair of agents (i,j) and any state , so that , i.e, the marginal utility across agents will be the same and we have full risk-sharing among households .

Given specific parameterization of the utility function, such as an isoelastic utility function;

[4]

Where, is a multiplicative shock factor and risk aversion coefficient, which is assumed to be constant over time. Substituting [4] into [2] and taking the first order condition for the maximization problem, dividing the FOC for an individual household at two points in time, we obtain;

[5]

Equation [5] is the condition that marginal utility growth is equated across households for the hypothesis of full risk-sharing to hold. Taking the log of this equation and adding the error term will give the following equation

[6]

This is a simple consumption function, which is expect to be consistent with any efficient allocation. Where, and are related to the aggregate supply of the consumption good in period t and t+1, respectively, which are the only determinants of consumption depending on the random state[7]. The terms , and represent household preference shifts and is measurement error. For the theory of full risk-sharing to be true, with the assumption of homoskedasticity of the measurement error and preference shifts, which are uncorrelated across households, the coefficient of additional regressor that is cross-sectionally independent of the preference shifts and measurement error will be zero[8].

3.1. Econometric strategies of testing consumption insurance in the literature

The most commonly applied version of equation [6] in the empirical literature using panel data (in Ravallion and Chaudhuri 1997, Jacoby and Skoufias 1998, Skoufias and Quisumbing 2003 and Nigussie 2005) is of the form,

[7]

where denotes the change in log consumption or the growth rate in total consumption per capita of household i in period t ( between round t and round t-1 in our case); is the growth rate of household i income; is a vector of household or household head’s characteristics; denotes a set of binary variables identifying each community separately by survey round[9]; , , and , are parameters to be estimated; and is a household-specific error term capturing changes in the unobserved components of household preferences.

Based on the underlying theory of risk-sharing, the coefficient provides an estimate of the extent to which idiosyncratic income changes play a role in explaining the household-specific consumption growth rate. As noted in Skoufias and Quisumbing 2003, the set of discrete terms, , identifying communities by survey round, serves two interrelated functions. First, the term controls for the role of aggregate (covariate) shocks common to all households within any given community and survey round. Second, given that consumption and income are in logarithms, they also account for potential difference in the round-to-round inflation rate across communities. They also noted that including community/round interaction dummies is equivalent to deviate all variables for their respective community/round mean.

3.3. Econometric model of poverty dynamics in rural Ethiopia

Once we identify the poor and the nonpoor, the next step is to analyze the dynamics of poverty over the past 10 years, 1994 to 2004[10]. There are a number of empirical and theoretical models for the analysis of poverty dynamics. The most commonly applied are binary response models. Although, the econometric literature on nonlinear panel data models is growing, there are computational problems and indistinctness to clearly establish the issue of identification and estimation of nonlinear models that allow for both individual-specific effects and state dependence[11]. As discussed in Chay and Hyslop 1998, one of the fundamental issues in estimating dynamic binary response models is the issue of unobserved initial conditions of the dynamic process, i.e., initial conditions bias. Although, the practicability and performance of the models is not still well grounded, there are several methods that account for initial condition bias both in the linear dynamic regression models[12] and nonlinear models that allow for both individual effects and state dependence[13] using different simulation techniques.

For our purpose of modeling poverty dynamics in rural Ethiopia, we apply models that account for the initial conditions and state dependence. This is of interest because results in Islam and Shimeles 2005 reveals the importance of state dependence in explaining poverty dynamics in rural Ethiopia. The underlying dynamic binary response panel data model has the following specification:

, ; [8a]

[8b]

[8c]

Where, is the underlying response variable determining the latent process, is the poverty status of household i during round t as measured by the consumption expenditure of the household and takes the value of 1 if the household is poor in the relevant period, is an indicator function which is equal to one if the enclosed statement is true and zero otherwise, is a vector of exogenous determinants of poverty status, , represents all household-specific, time invariant observed and unobserved factors, is the transitory error which is assumed to be i.i.d. over time with a distribution function , and finally, and are the parameters to be estimated. In this case represents structural state dependence[14] in poverty and is the source of spurious state dependence attributable to permanent unobserved heterogeneity in the household such as intelligence, ability, motivation, attitude, etc.

Using Hausman Specification Test, which basically tests whether the vector () of random effects coefficients (efficient estimates) of the time varying variables,, are systematically different from the corresponding fixed effects coefficients (consistent estimates under Ho), the result obtained form this test using linear probability[15] fixed effects and random effects specifications for poverty dynamics model shows that we reject the null hypothesis at 10% level of significance. Therefore, the two specifications are systematically different indicating that there is endogeneity in some of the regressors, where one of the random effects model assumption () breaks down.

Hence, the random effects model is not consistent and efficient. Because fixed effects approaches do not require parametric assumptions about the conditional distribution of the individual conditions, a lot of information is being absorbed in order to “non-parametrically” condition out the unobservable individual heterogeneity. As a result, fixed effects estimators may be sub-optimal since they ‘throw away’ comparisons between individuals that may be informative about the truth. If time-constant variables are of interest like the time-varying variables, the robustness of fixed effects estimator to correlation between the unobserved effects,, and the is useless. For these reasons, we use random effects nonlinear model. But to increase the robustness of our random effects model we applied two-step procedure, as specified below, which can control for the correlation between some of the regressors and the unobserved individual effects (Wooldridge 2003).

State dependence

As noted above, empirical works show that there is state dependence in the dynamics of rural poverty in Ethiopia. For this we introduce poverty status of the household in the pervious period, as a regressor that will allow us to test for the presence of genuine state dependence. According Heckman 1981a and1981b state dependence can be spurious or genuine arising form three sources: unobserved household characteristics; the effect of time varying shocks that are not specific to the household; and behavioral and preference shifts associated with poverty spell in the past. Genuine state dependence is due to past poverty status that results behavioral and preference shift. To test for the presence of genuine state dependence, one has to control for the other spurious sources of state dependence, i.e. controlling for observable and unobservable individual characteristics.

Unobserved Heterogeneity

We base our model in line with the work of Arulampalam et al. 1997 in their application to unemployment persistence in Britain. Assume the unobservable household specific heterogeneity, , is time-invariant, from the error term in equation [8], we have;

[9]

Were denotes the household-specific unobservable effect and. Assume and is random variable. In most cases, the household specific unobserved heterogeneity are correlated with the time-varying characteristics. In this case the maximum likelihood estimates of will be inconsistent, which pick up some of the effects of the unobservable, (Arulampalam et al. 2000). To overcome this problem, we relax the assumption that and are independent. The relationship between the unobserved household specific heterogeneity and the observed household characteristics can be modeled as

[10]

Where is a vector of means of the time-varying covariates for household i over time, is independent of