Using Stated Preferences and Beliefs to Identify the Impact of Risk on Poor Households ∗

Ruth Vargas Hill International Food Policy Research Institute November 2007

Abstract Whilst the importance of uncertainty in shaping economic behaviour of poor households is widely acknowledged, empirically identifying the impact of risk is difficult. By using data on risk preferences and perceptions of risk collected through hypothetical questions in combination with more traditional measures of a household’s ability to deal with risk, this article identifies the impact of risk on production decisions. It shows both that data on stated preferences and beliefs can be usefully utilised to explain household behaviour, and that risk has a significant impact on the production decisions of poor households.

1.

Introduction

There is an increasing body of empirical work highlighting the presence of poverty traps among very poor people in Sub-Saharan Africa. While some studies find little evidence of poverty begetting poverty, micro-level evidence of poverty traps has been found for a number of countries in Sub-Saharan Africa suggesting that poverty and hunger do put into play mechanisms that cause their persistence, and that for some, poverty does entrap.

∗

The author is grateful to Chris Mukiza, Robert Waggwa Nsibirwa, Aliziki Kaudha and the Uganda Bureau of Statistics without whom this work would not have been possible. Funding for this study was provided by the Economic and Social Research Council, UK and the Commodity Risk Management program of the World Bank. Many thanks to Marcel Fafchamps, Stefan Dercon, Simon Appleton, Pramila Krishnan, Sylvie Lambert, Jean Louis Arcand and participants at seminars and workshops in Oxford and Bonn for comments on an earlier draft.

1

Poverty traps have been found to be present in Madagascar, Kenya (Barrett et al. 2006), and South Africa (Adato et al. 2006), but have not been found to be present in Russia (Loshkin and Ravallion 2004), China (Jalan and Ravallion 2004), and Mexico (Antman and McKenzie 2005). Understanding when and why poverty traps exist is important in designing policies and programs that will enable some of the poorest households to improve their incomes and welfare. Several dynamics of behaviour of very poor households that give rise to poverty traps have been identified. The lack of energy and ill-health of those who are severely malnourished makes the poorest least able to undertake productive employment (Dasgupta 1997); the inability of poor households to invest in the education and assets of their children makes it more likely that children who experience poverty will continue to do so into adult life (Quisumbing 2006, Fafchamps and Quisumbing 2005); and the inability of credit constrained low-income households to invest in high-return activities in the face of risk causes low-income households to engage in low-return economic activities (Barrett and Carter 2006). It is this last dynamic that provides the motivation for this paper. This paper considers whether uninsured risk influences the amount of labour coffee-producing households in Uganda allocate to coffee production in a given season. Eswaran and Kotwal have shown that for a given degree of risk aversion, underinvestment in risky production activities will be greater for households who are less able to insure consumption from uncertain returns (Eswaran and Kotwal 1986). Given higher average returns to risky production activities, this finding implies that households less able to insure consumption are left in lower return activities. Eswaran and Kotwal’s theoretical finding is the basis for studies in which involvement in risky production activities is compared across households with different abilities to smooth consumption (Morduch 1991; Rosenzweig and Binswanger 1993; Dercon 1996). This paper considers these predictions as they apply to short-term labour allocation decisions made by poor coffee-producing households in Uganda. These models have been applied to short term production decisions of agricultural households before by Morduch (1991) for India and Dercon (1996) for Tanzania. These papers provide insight as to how households that are less able to smooth consumption are less able to produce high-risk, high-return crops in the face of uncertainty. Given the data available in these

2

studies, the impact of uncertainty in determining farmers’ crop choices is identified through wealth alone. The contribution of this paper is to use a unique data set with data collected on risk preferences and perceptions of risk to better identify the impact of risk on short-term labour allocation decisions and to determine whether these decisions are really constrained by uncertainty, or by other impacts of credit constraints on production (Barrett and Moser 2002). Data on risk preferences and perceptions of Ugandan coffee farmers was collected through hypothetical questions in addition to a standard socioeconomic survey. This data is used to identify the impact of risk on the share of labour allocated to coffee production by these farmers. There are a number of studies that present evidence that stated risk preferences correlate as expected with risk taking behaviour (Barsky et al 1997; Feinerman and Finkelshtain 1996; Knight et al 2003). However these studies do not take into account the differential impact of risk on production for poorer and wealthier farmers. To identify the role risk plays in constraining the production decisions of poor households, measures of risk preference need to be used in combination with measures of household wealth. This paper also uses a measure of perceived risk not used in other studies. The significance of variables included in the regression analysis to capture risk preference and perceptions of risk suggest that these variables are useful and that risk does impact households’ production decisions, particularly for poorer farmers. The next section considers further the empirical framework underpinning the analysis and some of the caveats in applying this model to this data and context. Section 3 describes the data; Section 4 presents the main empirical results and results of robustness tests undertaken. Section 5 concludes.

2.

The relationship between wealth, risk and labour allocation

Susceptibility to risk is a distinguishing feature of what it means to be poor. Poor households in rural areas of developing countries have little access to formal credit or insurance markets with which to fully insure their consumption. Although income shocks can be traded across time by a household through the accumulation of assets in good

3

years and the liquidation of assets for consumption in bad years, this can be costly if it results in the sale of productive assets (Udry 1995, Lim and Townsend 1998). When productive assets are sold the future income streams from these assets that are lost. Transfers from family and friends can also provide some insurance, but the size and geographical limits of these networks limits their effectiveness (Lund and Fafchamps 2000, Morduch 2004). As a result, the ability to trade income shocks across time and area varies across households, with wealthier households more able to trade income shocks as a result of their greater access to limited credit markets, larger asset stock, and ready access to the necessary social networks in times of need (Dercon 2002). Consequently, the rich are more able to choose the level of risk they are happy accepting and the fate of poor households is more strongly bound to the draws it experiences from the stochastic income distribution it faces. This inability to deal with risk that is characteristic of poor households impacts decisions that a household makes about its livelihood. Rural agricultural households are both consumers and producers and when they are unable to choose the level of risk they are happy accepting in their consumption; production decisions are made to avoid risk, often at a cost. In the standard agricultural household model (Singh, Squire et al. 1986) a household maximises a utility function, with respect to the resource constraint it faces, determined in part by the profits earned from agricultural production. The consumption and production decisions of the household can be modelled separately, or as is more common, recursively with production decisions made first to maximise profits and consumption decisions made subsequently, given the level of income from this first stage. When markets—such as the market for credit and insurance—are missing, the production and consumption activities of a household cannot be separated and a household’s characteristics and preferences are important in understanding its production decisions (Lopez 1986; Kurosaki and Fafchamps 2002). A much simplified version of the model presented in Kurosaki and Fafchamps (2002) is presented here to illustrate how separation of consumption and production decisions breaks down when uncertainty is present and a full set of contingent assets does not exist. In each period a household chooses current consumption, c t , and inputs for

4

production at cost D t . Consumption and production expenditures are allocated from current household resources which is given by the real value of assets the household holds, A t , and the crop income accruing to the household as a result of the production decisions made in the previous period, Π t . The amount from total resources that is withheld for future consumption is denoted by w t . The return to w t is certain and given by w t+1 = (1+r) w t while the return to purchased inputs for production is uncertain. The production function is given by Π t+1 =π(D t , l t ) where π is increasing in D and π′ is decreasing in D, but is subject to multiplicative risk, θ: an identical and independently distributed process with a mean of 1 and variance of σ. The value function is given by

(1) and the first order conditions are thus,

which yields,

(2) In the absence of uncertainty θ t is always equal to 1 and (1+r) = π′ which means resources are allocated to next period production until the return to assets equals the return to production. Uninsured uncertainty causes a deviation from this profit maximisation strategy by E t [V′ t+1 θ t+1 ] / E t [V′ t+1 ] and, thus, production decisions are no longer separable from consumption preferences captured in the concavity of V t (.). The concavity of the value function is in part dependent on the concavity of the household's utility function. As exemplified by the Arrow-Pratt measures of absolute and relative risk aversion, the greater the household's risk aversion the higher will be the concavity of the utility function. However, the value function of a household will in general be less concave than its utility function. If the household does not hold any assets nor has access to a market for assets, such that it cannot transfer income from one period 5

to the next, the concavity of the value function is equal to the concavity of the utility function. The more a household can disassociate consumption from income earned in one period through inter-temporal transfer of resources the flatter the value function becomes with respect to current income (Deaton 1991). The concavity of the value function increases with the household's risk aversion and falls with the household's ability to insure consumption from income fluctuations. When households can transfer resources from period to period without limit, its value function is linear, i.e. V′ t+1 is constant and can be factored out of the equations. When borrowing constraints exist, this ability depends entirely on the wealth of the household. When uncertainty is present and the value function is not linear the effect of the presence of E t [V′ t+1 θ t+1 ] / E t [V′ t+1 ] on the amount of resources allocated to risky production depends crucially on whether this ratio is greater than or less than unity. Sandmo (1971) shows that under certain conditions this ratio is less than 1 which, given π′ is decreasing in D, means D is less than that predicted by the profit maximisation strategy. This means, in the example outlined, that a household with a higher value of V′ t+1 will choose a lower D t earning a lower but safer w t . Inter alia households who are more risk averse will devote less labour to risky crops. Also for a given degree of risk aversion, underinvestment in risky production activities will be greater for households who are more constrained in their consumption smoothing activities. This is the point Eswaran and Kotwal (1990) make when they show that assuming constant risk aversion across a population, risk is less costly for richer households. This implies risk and poverty combine to create a vicious circle: resource-poor households use the assets they have less productively because of poverty, ensuring future poverty. The positive relationship between wealth and return to assets that this causes is a key factor in defining a poverty trap (Barrett and Carter 2006). However, Sandmo's results have been qualified by two papers which suggest that the impact of risk on labour allocation decisions of risk averse households is less clear. Finkelshtain and Chalfant (1991) show that if the risky good that is being produced is also a good that is consumed by the household then price risk can have a positive effect on output. Similarly Barrett (1996) shows that risk will have a different impact on production decisions for rich and poor households. Although large farmers behave as

6

Sandmo predicts, reducing the use of costly inputs whenever revenues are stochastic, smaller farmers who are net-buyers of food will utilize extraordinary amounts of labour beyond their shadow valuation of labour on account of the food insecurity created by food price risk. The case being considered in this paper is the allocation of labour to coffee production vis-à-vis staple crop production. There is a risk associated with food prices rising in time of need, and indeed this risk can be assumed into the risk of coffee production in that part of the risk of coffee production is that it forces a household to become a net-buyer of food-crops. Whilst, in this case, we might expect to find lower levels of labour allocation to coffee as a result of risk, these papers highlight that the response of agricultural households to price risk across multiple crops, particularly when they are net buyers of some of the crops being considered, is complex. This discussion suggests that the choice of how much labour to expend on coffee production will be a function of risk preferences (denoted by P), wealth (W t ) and expectations about r t+1 (denoted by E t (r t+1 )) such that λt = f (Wt , P, Et (rt +1 ), Z t )

(3)

where λ t is the share of labour allocated to risky crop production (0 ≤ λ t ≤ 1) and Z t is a vector of control variables. This can be applied to data by running a reduced form regression of measures of wealth, risk preference and risk perceptions on the share of labour allocated to risky crop production. Studies including wealth in reduced form regressions have born out theoretical predictions. Morduch considers crop choices for the Indian ICRISAT households and finds that households less able to smooth consumption allocate a smaller share of land to high yield, high risk varieties of rice and castor (Morduch 1991). In Tanzania, Dercon shows that households with fewer livestock holdings allocate a higher proportion of land to low risk, low return sweet potato production (Dercon 1996). Studies that included measures of risk preference found risk-averse Israeli farmers less likely to be involved in risky production activities (Feinerman and Finkelshtain 1996) and risk averse farmers in Ethiopia were less likely to adopt new technology (Knight et al 2003) Studies including measures of individual risk perceptions are not known however, measures of price or yield volatility have been included in studies on North American farmers using aggregate

7

crop data and have shown a negative relationship between crop acreage and a measure of the volatility of crop price or yield (e.g. Chavas and Holt 1990, Ozanne 1998). This study is unique in using all three measures, and in allowing measures of risk preferences to vary across households with different wealth levels. Whilst it is the role of constraints that is ultimately of concern and of policy interest, the reliance on measures of wealth to identify the impact of risk in many contexts is problematic as it is not fully possible to deal with the endogeneity issues entailed in identifying the causal relationship between a measure of wealth and production decisions. An unobserved preference for risk will affect not only current production choices, but also past production choices and thus the asset wealth of a household (given the higher return expected from high risk activities), causing a household's ability to deal with risk to be endogenous to production choices. Including a measure of innate risk preference allows this endogeneity problem be solved. 2.1.

Perennial crops

Thus far we have considered a general crop choice scenario, but in the case considered in this paper the risky crop is a perennial crop—coffee. In Uganda coffee trees become productive three years after they are planted and stay productive for up to seventy or eighty years, although a tree's yield potential falls after thirty to forty years (UCTF 2002 and Magambo 2000). At the beginning of each season a household already has a predilection to produce a certain amount of coffee based on the share of its land planted to coffee trees (the result of previous periods’ investment and production decisions). The analysis thus needs to take into account the fact that the return to labour allocated to perennial crop production depends on the stock of trees held by the household in any given period. In other words the empirical application we need to consider is really: λt = f (Wt , P, Et (rt +1 ), K t , Z t )

(4)

Where K t is the share of land planted to coffee. There is much anecdotal evidence that suggests the choice of K t is quite separate from the choice of λ t . For many Ugandan coffee farmers the coffee trees they farm are very old— the majority of trees are aged about 40 years and some trees are still being farmed at 70 years—and anecdotally households do not cut down trees when the price

8

falls, but keep them in case the price rises. In these cases little labour is applied to the trees, and sometimes the trees are not harvested. Over the past few years there has been little new planting observed and changes in stocks of coffee trees have been driven more by wilt disease shocks than investment. These observations indicate that perhaps production capacity is set much earlier for these households, at a very different time. Planting coffee trees is an investment decision with benefits reaped over many years. Many papers have modelled planting perennial crops using standard models of investment (for example: Hartley et al 1987; and Trivedi 1992), but the uncertainty around the returns to investment and the irreversible nature of the investment (there is no second hand market for coffee trees—the tree has no value when it is no longer in the ground—and the opportunity cost of land lying with no output for three years until the trees bear fruit cannot be recouped) suggest that models of irreversible investment under uncertainty characterise this decision better (Dixit and Pindyck 1994; Abel and Eberly 1994). There is also some theoretical and empirical analysis that suggests that, indeed, the decision to plant or uproot coffee trees can be characterised by models of irreversible investments under uncertainty (Malchow-Moller 2001; Hill 2003). Whilst it is likely that different decision processes are at work in determining the share of land allocated to coffee production, K t still remains a choice variable. It is thus most likely erroneous to consider K t as completely exogenous to the household’s labour allocation choice. As such K t needs to be instrumented and there are two instruments that can be used: a lagged value of the share of land allocated to coffee production, and the year in which the plot of land in which coffee is planted was acquired by the household. The year of acquisition is used as households who had this land before coffee liberalization changed the coffee price distribution may have been more or less likely to plant coffee on their land than households who acquired land after liberalization. 2.2.

Using data on stated preferences and perceptions

Stated preferences on loss aversion, trust and time preferences collected through hypothetical questions have been used to explain real world behaviour in Fehr and Gotte (2002), Karlan (2005) and Ashraf et al (2006) respectively. As noted above a number of studies present evidence that stated risk preferences correlate as expected with risk taking

9

behaviour. Additionally Nyarko and Schotter (2002) have shown that stated beliefs can be a good predictor of actual behaviour. Kahneman and Tversky (1979) advocated the use of hypothetical choices between real world scenarios as a more reliable way to understand preference to risk than "contrived gambles for small stakes" on two conditions: "the assumption that people often know how they would behave in actual situations of choice, and on the further assumption that the subjects have no special reason to disguise their true preferences". However empirical work has questioned whether these assumptions can validly be made. In their literature review Camerer and Hogarth (1999) found three studies in which there was no significant difference between choices of lotteries when monetary incentives were introduced; five similar studies in which respondents were more risk averse with payoffs, and two in which respondents were actually more risk seeking when monetary payoffs were introduced. However, they concluded that if there was an effect to offering incentives it was to increase the level of risk aversion reported. Holt and Laury (2002) show that the difference between an individual's response to questions with and without real payoffs increases with the size of payoffs resulting in large differences for lotteries that approach real life choices. This work suggests that using hypothetical payoffs is problematic when choices between lotteries are used to estimate the size of risk aversion parameters or to undertake detailed studies on the nature of risk preferences over a distribution. However if the bias is uniform across individuals using responses to a given set of lotteries as an explanatory variable in regression analysis is not problematic (Bertrand and Mullainathan 2001). When we examine studies which have offered both real and hypothetical payoffs to farmers in developing countries, it appears that the main impact of offering real incentives is to reduce measurement error rather than any bias in the responses. Binswanger found measures of risk aversion calculated from hypothetical choices were more dispersed than measures of risk aversion calculated from the experimental game (Binswanger 1980, p.398). However Binswanger finds no overall bias from posturing in the hypothetical responses: respondents believed they would act more or less aversely to risk than they actually did in the real game. 1 The usefulness of stated preferences as an explanatory variable depends on the nature of this measurement error: its size and

10

correlation with other variables (observed or not) that impact the dependent variable. Measurement error will result in attenuation of the coefficient towards zero and makes it more difficult to reject the null hypothesis of no significant effect. Both of these should be born in mind when the empirical results are presented. An additional cause for concern is the extent to which data collected on risk preferences, whether from questions with real or hypothetical payoffs, capture a household's innate preference for risk or some combination of innate preference and ability to deal with risk. It is unclear whether data from questions on risk preferences provide a measure of the concavity of a household's utility function—as is usually assumed—or the concavity of a household's value function. As a first cut at analysis we assume that the data collected is a measure of the concavity of a household's utility function (interacting data collected on risk preferences with household wealth in the regression analysis), we then test the robustness of our results to making this assumption.

3.

Coffee production in Uganda: the context and data

This section describes the context and data used for the analysis. For coffee-producing households in rural Uganda, Robusta coffee production is a relatively risky yet high return activity. Robusta coffee production is characterised by a very low technological level, a low use of purchased inputs, and limited use of modern farming methods such as irrigation. The costs of production, amounting mainly to labour and land costs, are estimated to be around $0.10 per kilo. 2 The average price for a kilo of unmilled coffee post coffee market liberalisation in 1991 is $0.30 (in 2001 prices), which suggests an average return of $200 per hectare. The return to matooke production, a banana-like staple produced by the majority of households in Uganda, is calculated as $150 per hectare using information in Bibagambah (1996). However, although the average price of coffee since coffee market liberalisation is $0.30, there is a large degree of variation in this price as Figure 1 shows, much more so than for other crops. During the period in which data was collected the median price recorded for a kilo of unmilled coffee was $0.16 which implies a per hectare return of about $60.

11

Although coffee production was always somewhat risky (both in the return received and the fact that producing coffee subjected farmers to the food crop market), there has been an increase in price variation as a result of market liberalisation in 1991. The coefficient of variation of the farm-gate price increased from 0.38 pre 1991 to 0.62 after, while the coefficient of variation of the prices of matooke and sweet potatoes were 0.25 and 0.28 respectively during this time. Additionally, yield risk has increased in the last decade as a result of the spread of coffee wilt disease. A survey carried out by the Ugandan government in 2003 revealed coffee wilt had affected almost half of the area planted to Robusta coffee (Munyambonera 2004). Once a tree is infected with coffee wilt disease it withers and dies, so the threat of infection poses a considerable source of uncertainty in yields. On the basis of price risk and in the presence of substantial yield risk there is a case for assuming the return to coffee is more volatile than that of other crops. This makes it suitable for consideration as a risky and high return crop in comparison to other crops grown by Ugandan farmers. Data on production decisions, risk preferences, perceptions of risk and wealth was collected for 300 Robusta coffee farmers at the beginning of 2003. 3 Data was collected in four districts in Uganda--Mukono, Luwero, Masaka and Bushenyi—that together produce about half of Uganda’s Robusta coffee. The sample of coffee producers was drawn randomly from a sampling frame consisting of coffee farmers identified in the 1999/2000 National Household Survey. Many of the questions on household characteristics and production decisions were kept the same, allowing a small panel to be constructed. As the period between the baseline and the follow up survey was relatively short, there was little attrition resulting from death or migration. Most households were still in existence within the village and it was relatively easy to trace them. Summary statistics of the data used in this analysis are presented in Table 1. The majority of coffee grown in Uganda is grown by small holders, and this was true for this sample also, with more than two-thirds of the households owning land less than or equal to five acres. Heads of coffee-producing households tend to be older than the national average, at 50 years. The mean level of education of household heads is 5 years. Most

12

households have had a problem with coffee wilt disease in the last few years and on average a quarter of trees have been lost. 3.1.

Variables used in analysis

The share of labour going to coffee production was calculated using questions that asked households to report the time allocated to crop production in total (in days) by household members and paid labourers; and the time allocated to coffee crop production (in days) by household members and paid labourers. Although the data analysis focuses on explaining the share of labour allocated to coffee production, regressions using the share of revenue coming from coffee as dependent variables are also conducted. The share of revenue from coffee production was calculated by the share of crop income (from produce sold at market and consumed at home) coming from coffee. To the extent revenue from coffee in a given period is dependent on labour share, a similar pattern should be observed, although the dependence of revenue share on realised price and yield might weaken the hypothesised relationship. A measure of wealth is used to proxy for a household's ability to smooth income shocks. If credit and insurance markets are incomplete, as we expect in rural Uganda, a household's ability to deal with fluctuations in income will depend on its wealth - both as a means to insure itself and to facilitate access to the limited credit market that does exist. A measure of liquid and land wealth per household member is used; and the log of it is taken to minimise the impact of outliers on the results. Land wealth is included as studies have shown land markets to be active in rural Uganda (Deininger and Okidi 2001; Baland et al 2001). 4 The number of people that can be asked for financial help in a time of need is also included as a regressor to capture access to informal borrowing not picked up by this per capita measure of wealth. Intercropping is commonplace among the households interviewed making it hard to estimate the area of land allocated to coffee. However, data on the total area of land cultivated and the total number of coffee trees owned was collected. Using an estimate of an average of 900 trees planted per acre, a measure of the share of cultivated land planted to coffee could be calculated. The share of land planted to coffee could be calculated for both 1999 and 2002 using information on land collected in the 1999 survey and

13

information on changes to the number of trees in the last three years from the 2002 survey. Various controls are included in all regressions. Although technology is assumed constant across households, if coffee production is more or less labour intensive than other crop production and markets are not perfect for either labour or land, technological considerations will be of importance in determining portfolio choice. To allow for this a land to labour ratio is included in the regression (measured by a ratio of total cultivatable land owned to available household labour 5 ). Regional dummies are used to account for agro-climatic differences in regions and for considerable geographic variability in the crops grown with coffee (Pender et al 2004). Distance from the nearest market is also included to control for differences in crop prices resulting from differential market access. Traditional cash crop production in some countries is favoured by men rather than women and some research suggests gender may also be correlated with risk preference (Wik and Holden 1998). As such it is important to account for gender of the household head in the analysis. Age has been shown to be correlated with risk taking behaviour (Barsky et al 1997) and in Uganda there may be a higher perceived return to coffee production among older household heads for whom the traditional status of coffee is still important. Age of household head is also included in all regressions. 3.2.

A measure of risk preference

To provide a measure of risk aversion for the analysis farmers were asked to choose one of five lotteries offered to them. Their choice of lottery was then used to classify farmers as more or less risk averse. The choice of lotteries method was used as it was simple for the respondent to understand, and the use of a constant probability of 0.5 has the advantage of being easy for the farmers to conceptualise as it could be explained as an equal chance of either option being realised. The lotteries offered increased in mean and spread such that each successive lottery represented an increase in risk. The lotteries offered were given contextual specification to encourage the farmer, as much as possible, to reveal his preference for crop income risk relevant to the production choices being considered. The degree to which the choice could be made realistic was limited in that no commodity has only two outcomes for choice or yield, but the

14

contextualisation helped the farmer conceptualize the risk. The lotteries offered to the farmer increased in mean and spread in the same manner for each contextual specification. Preference data was collected for coffee yield and price risk and also for matooke yield risk. Table 2 provides examples of the options using the examples of coffee prices and yields. The current farm-gate price for twenty kilos of kiboko at the time of the survey was 6,000 shillings ($0.16 per kilo) and 100kg / 0.25 acre was an average yield for these farmers. Visual aids were used depicting the different options offered to the farmer, to make it easier to understand the options offered and easier for a choice to be made. Table 3 indicates the mean and variance of each lottery, and the risk preference parameters that would be associated with each choice under the assumption of a specific expected utility functional form. 6 The option chosen by the median coffee farmer was option 3 for the lottery specified in terms of coffee price risk and option 2 for the lottery specified in terms of coffee and matooke yield risk. This suggests an average risk aversion parameter similar to that in other studies coming from the coffee price risk question, but higher than other studies for the response to the yield questions. The option chosen by the median farmer in the Binswanger (1980, 1981) study is comparable to option 3. An analysis of the covariates of measured risk aversion suggests the measure covaries as the literature would suggest with wealth and past shocks to the household, although the significance of its covariance with wealth is weak. 7 The data on risk preferences enters the analysis both as the calculated parameter and as a categorical variable. When it is entered as a categorical variable, a test of joint significance across all categories is performed. This measure of preference for risk is assumed to capture a household’s innate preference, and as a result it is interacted with household wealth (W) in the data analysis given wealth will mediate the effect of an innate preference for risk on labour allocations decisions. The assumption that the measure is a measure of a household’s innate preference is later relaxed. The questions remained hypothetical given survey budget considerations, and as a result there is some concern that the data collected suffer from a high degree of measurement error. This issue is discussed in more detail when the empirical results are presented and some robustness checks performed.

15

3.3.

A measure of risk perception

Farmer perceptions of relative returns and their variability are also important in determining input allocation decisions. To capture heterogeneity in households' beliefs about the distribution of the future coffee price, data was collected on household's expectations about the future price. Respondents gave an idea of their expected distribution of the price of coffee in six months (July 2003) and in three years through the following exercise. The respondents were given twenty beans and a handout marked with three squares of different price categories (less than $0.10 (200 shillings), between $0.10 and $0.20 (between 200 and 400 shillings), and more than $0.20 (400 shillings)). They were asked to place beans on the squares in accordance with what they thought was the chance of that outcome. If the respondent thought one option was very likely they were instructed to put many beans on the corresponding square, if the respondent thought the option was unlikely they were instructed to place few beans there. So for example if the respondent was sure the price was going to rise in the next six months and felt sure that she would receive at least $0.20 per kilo for her coffee in July she would place all her beans on the square marked more than $0.20. If the respondent could not predict at all what the price would do, she would place the beans evenly among the squares. Once the exercise had been explained to the respondents, they found it easy to place the beans. Estimates of the mean and variance of the farmer's beliefs about the future price distribution were calculated from the response to these questions. 8 The average expected price in 6 months was (standard deviations in brackets) $0.22 ($0.06) and the average expected standard deviation of the price was $0.08 ($0.06). An analysis of covariates of this measure suggest the estimates of future return and variance vary with past experience of the coffee price and access to price information. A number of studies (Kahneman and Tversky 1979; Shahabuddin et al 1986) suggest that downside risk is particularly important in affecting behaviour under uncertainty. As a measure of perceived risk to enter into the regression analysis we calculate the perceived probability that the price falls below $0.10 per kilo. As $0.10 is approximately the average cost of producing one kilo of coffee, this approximates the perceived probability of a negative return to coffee production.

16

Coffee wilt disease has significantly affected expected coffee yields in the last few years. Areas that have recently experienced disease are more likely to be infected again, and certainly households in those areas are more likely to perceive a future risk, so we might expect past experience with the disease to impact expectations about its future prevalence and future expected yields. The number of trees cut down or abandoned as a result of disease in the last three years is included as a measure of exposure to coffee wilt disease under the assumption that the higher a household's exposure to disease in the past, the lower expected future yields will be. However only when the share of land allocated to coffee is included in the regression can the impact of coffee wilt disease on beliefs be separated from the impact of coffee wilt disease on tree stocks.

4.

Empirical results

This section presents results from the empirical analysis. Because the sample contained only coffee farmers, the results are only representative of coffee farmers in Uganda. Also because of this few corner observations (where the share of labour allocated to coffee production equals zero) were observed, and a Tobit specification is not needed. 9 In all regressions standard errors are controlled for clustering at the village level. Clustering is observed in the data at the village level as households from the same village are likely to be more similar each other than households from different villages. Also, enumerators varied across villages causing cluster specific measurement error to occur. 4.1.

Main results

Before turning to multivariate regression analysis non-parametric results examining the relationship between the portfolio decision and two of the variables of interest - wealth and risk preference - are presented. First, the log of per capita liquid wealth was regressed non-parametrically on the share of labour allocated to coffee production and the share of crop revenue coming from coffee production. Results for the two relationships are shown in Figure 2. Both show an increasing relationship between wealth and portfolio allocation

17

to coffee. For the share of labour allocated to coffee the relationship is constant across all levels of wealth, for revenue share some non-linearity seems to be present. Second, Table 4 shows how the mean share of labour and revenue allocated to coffee production varies across different risk preference categories. Results are presented for two of the three risk preference measures coming from questions on coffee price and coffee yield respectively. For both labour and revenue share it appears that both risk averse and risk loving households are less likely to hold a high share of coffee in their portfolio. This might be a result of not controlling for other explanatory factors, but it may also be an indication that whilst coffee production is risky, it is not the most risky production activity available to farmers. This is explored further below. Turning now to multivariate analysis, results from cross-sectional reduced form regressions with labour share as the dependent variable are shown in Table 5. P is interacted with W in the analysis, as the sensitivity of production decisions to risk preferences will be higher for households that are less able to insure their consumption. In the set of regressions presented in Table 5, risk preferences from the clonal coffee yield lotteries are simply allowed to vary between the top 25% and bottom 75% of the wealth distribution. In columns 1 and 2 risk preferences are entered as dummies reflecting the household’s lottery choice, and in columns 3 and 4 risk preferences are entered as the calculated risk aversion parameter, P. Columns 1 and 3 present results from a linear regression. Columns 2 and 4 present results from an instrumental variable regression in which the share of land planted to coffee is instrumented. Dummies reflecting the household's choice of lottery (columns 1 and 2) are positive and jointly significant for the bottom 75% of the wealth distribution, but are insignificant in explaining labour share for the top 25%. This is as the model predicts: the curvature of the value function diverges more from the curvature of the utility function at higher wealth levels. This is also true when risk preferences are entered as the calculated coefficient of risk aversion. The parameter is negative and significant in the bottom 75% of the distribution but are insignificant for the top 25% of households. Although dummies reflecting the household's choice of lottery (columns 1 and 2) are positive and jointly significant, the coefficients do not increase linearly as the model would suggest - a similar pattern to that observed in Table 4 is present in that both risk

18

neutral and risk averse farmers devote a lower share of labour to coffee production. This could reflect either that risk averse and risk loving households are less likely to hold coffee in their portfolio, or that those who are less likely to hold coffee in their portfolio are more likely to be outliers on decisions framed in terms of coffee. The latter explanation is examined in subsection 4.3 below and the results suggest that it is not the framing of the lotteries offered that explains this pattern. Examining the other crops grown by the least risk averse respondents, we see that they are more likely to grow pineapples, tomatoes and other garden vegetables. It may be that these are more risky. Additionally Finkleshtain and Chalfant (1991) and Barrett (1996) both highlight that the Sandmo (1971) result may not hold for agricultural households when the household is a net buyer of some of the crops it produces. Whilst for the large part the Sandmo result does seem to hold for these households (households with more concave value functions are less likely to produce risky coffee), the non-linear relationship may be explained by the fact that for agricultural households that are net buyers of food crops with price risk, the relationship is somewhat more complex. Wealth is positive and significant across all regressions. This is consistent with the hypothesis that as a household’s wealth and ability to deal with risk increases the share of income it is willing to derive from coffee production also increases. It may also indicate that richer households are better able to produce coffee for other reasons. The measure of access to informal borrowing (number of people that can be asked for financial help if a serious problem arises) is not significant in any of the regressions. The data on expectations about the future variability of the price allow us to look specifically at the effect of a household's beliefs about the riskiness of coffee income on its portfolio decisions. A significant relationship between the perceived probability of the price falling below $0.10 and allocation of labour to coffee is observed. This is a consistent result across different model specifications. Additionally a household's experience of coffee wilt disease in the last three years negatively affects its allocation of labour to coffee production. Given the household's reduced capacity to produce coffee as a result of wilt is controlled for by the inclusion of the share of land allocated to coffee this may be because it captures household beliefs about future yield risk.

19

Instrumenting K t with the lagged value of K t and the year of plot acquisition causes K t to be insignificant (columns 2 and 4), but does not otherwise affect the results on the variables of interest. The first stage regressions for K t are shown in the appendix. The lagged value of K t is positive and significant as expected. The year land was acquired has the expected sign but is insignificant. The joint F-test on the dignificance of the instruments is F(2, 256)=19.74***, and the predictive power of the instruments (Rsquared due to the instrument) is 0.11, suggesting the instruments are quite good. Although instrumenting causes K t to be insignificant, the Hausman test shows we cannot reject the null of no significant difference in coefficients (χ2(1)=0.00) This indicates there is not a problem of endogeneity and suggests the decision of how much land to plant to coffee is a different one to that of how much labour to allocate to coffee production in a given year. Regressions results presented in the following tables do not instrument for K t , however in all cases instrumenting for K t does not affect the results on the variables of interest. The results suggest that risk does impact the labour allocation decisions of agricultural households, and that the impact of risk is larger for poorer households. In the following three subsections the robustness of these results is tested by altering the way in which wealth and risk preference were interacted, the measure of risk preferences used in the analysis, and running regressions on the share of revenue from coffee production. 4.2.

Interacting risk preferences and wealth

In Table 5 risk preferences are allowed to vary very crudely between rich and poor households, exogenously determining that those households in the top quarter of the distribution are rich. Regressions were also repeated using a dummy taking the value 1 if the household was in the top half of the distribution and were very similar. However, ideally risk preference would be allowed to vary with wealth more fully. To allow for this, household wealth is interacted with the calculated coefficient of risk aversion by diving P by W (causing household P to be smaller for households with higher levels of wealth). Results for this are presented in Table 6. In column 1 the coefficient of risk aversion is divided by W. As an increase in wealth might have a greater impact in reducing the concavity of the value function at lower levels of wealth, the risk

20

preference parameter is divided by lnW in column 2. In both cases the interacted term has a negative significant coefficient as expected. The second regression has a higher R2 suggesting some non-linearity. To further explore this non-linearity this relationship is measured nonparametrically using a Robinson partial kernel regression. The relationship we want to estimate is the following:

λi = xi β + Pi ϕ (ln W ) + ε i

(5)

which differs from the usual partial linear regression in that P i appears interacted with φ(.). To revert back to the usual partial linear regression model we divide through by P i which gives

λi Pi

=

xi β ε + ϕ (ln W ) + i Pi Pi

(6)

allowing φ(.) to be estimated in the usual way. However, as Equation (6) shows the error term in this specification is clearly heteroscedastic. As with parametric estimation heteroscedasticity will not bias the estimation of φ(.), but it will be inefficient and make inference problematic (Dette and Munke 1998). However, for the purpose of gaining a better understanding of the shape of the relationship between P i and φ(.) this procedure will suffice. Figure 3 shows the estimated relationship between λ i / P i and φ(.). As can be seen it is quite linear indicating the specification used in column 2 of Table 6 is adequate for this data. 4.3.

Testing data on risk preferences

Contextualising the risk preference questions in terms of familiar situations of coffee price and yield risk, allowed the farmer to give a response to a situation he or she knows well, hopefully encouraging a meaningful response. However, as some studies have encountered, it may be the case that farmers who are less dependent on coffee income are less concerned about coffee revenue risk, and as a result will respond in a less risk averse

21

fashion that their true preferences when questions are framed in this way (Moscardi and de Janvry 1977; Feinerman and Finkelshtain 1996). This would result in a negative reverse causality observed between the share of coffee in production and the measure of risk aversion coming from a response to a question on coffee price or coffee yield risk. It is thus important to repeat the regressions using the measure of risk from the hybrid matooke yield risk question. Table 7 presents results using responses to lotteries of coffee price risk and matooke yield risk. Comparing results in Tables 5, 6 and 7 shows that results are very similar across measures of risk preferences, suggesting the way the question was framed (as coffee or matooke income risk) did not impact the response. As the earlier discussion highlighted, given the questions were hypothetical it is likely that risk preferences are measured with some degree of error, attenuating the coefficient estimates. The significance of the data on risk preferences indicates measurement error is not so large as to render the measure useless. However if individual's error in reporting their risk preferences varies with observed or unobserved determinants of crop choice it will bias all coefficient estimates. To control for any question specific measurement error an instrumental variables specification was used in which one measure of risk aversion is instrumented with the other two using the estimated risk aversion parameters. This does not change the results significantly and so the results are not reported. This suggests question specific measurement error in the risk aversion measures is not important, but this method does not allow us to control for any systematic measurement error that might be present. In the presence of systematic measurement error the correct interpretation of the results is that households that report themselves as risk averse are less likely to undertake risky production activities. It is questionable as to whether households abstract from their current economic circumstances, and ability to deal with risk when answering questions about the level of risk they are happy to accept. As a result it is not clear as to the extent to which questions on risk preferences, hypothetical or real, capture a measure of innate risk aversion or a combination of this and a household's ability to deal with risk, i.e. whether a measure of the curvature of a household's utility function or the curvature of a household's value function is captured. So far the analysis has assumed that questions on risk capture a measure of the concavity of a household's utility function. The conformity of the results

22

with the predictions from theory suggests that P may indeed reflect innate preferences. However, if this is not the case and instead these questions measure the concavity of a household’s value function, then P should be entered directly in the regression, rather than interacted with wealth as it is currently. To test the robustness of the results to relaxing this assumption, P is entered directly in the regression and the results are presented in Table 8. In both specifications the share of labour allocated to coffee falls with reported risk aversion. The results are thus consistent with both reported risk aversion reflecting innate preferences and reported risk aversion reflecting the concavity of a household's value function. The remainder of the results on the other variables of interest remain unchanged. 4.4.

Using revenue share as a dependent variable

To test the robustness of the results that have been presented for labour share, results using the share of coffee in a household’s crop revenue as a dependent variable were also estimated. We may expect these results to conform less well to the model as the construction of the share of revenue variable includes realised coffee prices around which there was uncertainty. The inclusion of realised prices dilutes the relationship between production decisions and the variables of interest, but the regressions should provide, to some degree, a robustness check of the validity of the labour share results. These results are presented in Table 9. For these regressions the share of land devoted to crop production is omitted. Perhaps as expected, the model performs less well with revenue share as the dependent variable; however for many of the variables of interest the results remain unchanged. Wealth is still positive and significant, and poor risk-averse households are still less likely to derive their income from coffee production than richer and less riskaverse households. The relationship between risk preferences and revenue share holds when risk aversion parameters are entered in a number of different ways, showing the revenue share results to also be robust to the checks performed in previous sub-sections. However the perceived probability of a negative return to coffee production is not significant in explaining the share of revenue to come from coffee production.

23

5.

Conclusion

This article has gone someway into looking at how risk affects production of coffee for households in Uganda and to indicate the merit of simultaneously considering the effect of a household's ability to deal with risk, risk preferences and perceptions of risk to identify the impact of risk on production decisions. By using data on households' risk preferences and perception of risk in combination with more traditional measures of a households' ability to deal with risk this article identifies the impact of risk on production in a way previous empirical studies have not. Risk averse households were found to be less likely to allocate labour to coffee production, but the effect of risk preference on labour allocation was minimal for households in the top quarter of the wealth distribution. This suggests risk preferences are only important to the extent households cannot insure against income fluctuations. The perceived probability of a negative return was also found to have a strong negative impact on labour allocation. This was not the case for the variance of the coffee price which suggests that perceived downside risk is more important in explaining such decisions. Whilst the analysis has shown the utility of collecting information on risk preferences and perceptions in identifying the impact of risk on production behaviour of poor households, it also highlights the burden of risk on small-holder farmers. Using the results in Table 5 we see that if a risk averse farmer (with median values of all other characteristics) were to move from the 10th to the 50th wealth percentile he would increase his share of his household's labour allocated to coffee by 34 days. This study has only considered the impact of risk on short-run labour allocation decisions. The results on the exogeneity of coffee tree stocks suggest that long term production decisions cannot be analysed within the same model. However, given the dependence of many small-holder farmers on risky perennial crop production this issue merits further research.

24

References Abel, A. B. and J. C. Eberly (1994). A unified model of investment under uncertainty. American Economic Review 84 (5), 1369-1384. Adato, M., M. Carter, and J. May. (2006). Exploring poverty traps and social exclusion in South Africa using qualitative and quantitative data. Journal of Development Studies 42(2): 226-247. Antman, F., and D. McKenzie. (2005). “Poverty traps and nonlinear income dynamics with measurement error and individual heterogeneity.” World Bank Policy Research Working Paper 3764. Washington, D.C.: World Bank. Ashraf, N., D. S. Karlan and W. Yin (2006). Tying Odysseus to the mast: Evidence from a commitment savings product in the Philippines. Quarterly Journal of Economics 121 (2), 635 --672. Baland, J.-M., F. Gaspart, F. Place, and J.-P. Platteau (2001). The distributive impact of land markets in central Uganda. CRED, Department of Economics, University of Namur, Belgium. Barrett, C. B. (1996). On price risk and the inverse farm size-productivity relationship. Journal of Development Economics 51, 193-215. Barrett, C., and M. Carter. (2006). The Economics of Poverty Traps and Persistent Poverty: An Asset-Based Approach. Journal of Development Studies 42(2): 178-199 Barrett, C., P. Marenya, J. Mcpeak, B. Minten, F. Murithi, W. Oluoch-Kosura, F. Place, J.C. Randrianarisoa, J. Rasambainarivo, and J. Wangila. (2006). Welfare dynamics in rural Kenya and Madagascar. Journal of Development Studies 42(2): 248-277. Barrett, C. and C. Moser (2002). The system of rice intensification in practice: Explaining low farmer adoption and high disadoption in Madagascar. Cornell Department of Applied Economics and Management Working Paper. Barsky, R. B., F. T. Juster, M. S. Kimball, and M. D. Shapiro (1997). Preference parameters and behavioral heterogeneity: An experimental approach in the health and retirement study. Quarterly Journal of Economics 112 (2), 537--579. Bertrand, M. and S. Mullainathan (2001). Do people mean what they say? Implications for subjective survey data. MIT Economics Working Paper No. 01-04. Bibagambah, J. R. (1996). Marketing of Smallholder Crops in Uganda. Kampala: Fountain Publishers.

25

Binswanger, H. (1980). Attitudes toward risk: Experimental measurement in rural India. American Journal of Agricultural Economics 62 (3), 395--407. Camerer, C. F. and R. M. Hogarth (1999). The effects of financial incentives in experiments: A review and capital-labor-production framework. Journal of Risk and Uncertainty 19, 7--42. Chavas, J.-P. and M. T. Holt (1990). Acreage decisions under risk: The case of corn and soybeans. American Journal of Agricultural Economics 72, 529-538. Dasgupta, P. (1997). Nutritional status, the capacity for work, and poverty traps. Journal of Econometrics 77(1): 5-37. Deaton, A. (1991). Saving and liquidity constraints. Econometrica 59 (5), 1221-1248. Deininger, K. and P. Mpuga (2002). Land markets in Uganda: Incidence, impact and evolution over time. World Bank Policy Research Working Paper, Washington DC. Deininger, K. and J. Okidi (2001). Rural households: Incomes, productivity and non-farm enterprise. In P. Collier and R. Reinikka (Eds.), Uganda's Recovery: The Role of Farms, Firms and Government. World Bank. Dercon, S. (1996). Risk, crop choice and savings: Evidence from Tanzania. Economic Development and Cultural Change 44 (3), 385--514. Dercon, S. (2002). Income risk, coping strategies and safety nets. The World Bank Research Observer 17 (2), 141 .166. Dette, H. and A. Munke (1998). Testing heteroscedasticity in nonparametric regression. Journal of the Royal Statistical Society Series B 60 (4), 751--764. Dixit, A. K. and R. S. Pindyck (1994). Investment Under Uncertainty. Princeton, NJ: Princeton University Press. Eswaran, M. and A. Kotwal (1986). Access to capital and agrarian production organization. Economic Journal 96 (382), 482--492. Eswaran, M. and A. Kotwal (1990). Implications of credit constraints for risk behaviour in less developed economies. Oxford Economic Papers 42 (2), 473-482. Fafchamps, M., and A. Quisumbing. 2005. Marriage, Bequest and Assortative matching in Rural Ethiopia. Economic Development and Cultural Change 53(2): 347380. Fehr, E. and L. Goette (2002). Do workers work more when wages are high? evidence from a randomized field experiment. IEW Working Paper 144.

26

Feinerman, E. and I. Finkelshtain (1996). Introducing socioeconomic characteristics into production analysis under risk. Agricultural Economics 13 (2), 149-161. Finkelshtain, I., and J.A. Chalfant. “Marketed Surplus Under Risk: Do Peasants Agree With Sandmo?” American Journal of Agricultural Economics 73(August 1991):557-67. Hartley, M. J., M. Nerlove, and K. R. Peters (1987). An analysis of rubber supply in Sri-Lanka. American Journal of Agricultural Economics 69 (4), 755-62. Henstridge, N. M. (1997). The reconstruction of a macroeconomic dataset for Uganda. Centre for the Study of African Economies Working Paper Series No. 98-3. Hill, R. V. (2007). The determinants of investments in coffee trees in Uganda. Governance, Coordination, and Distribution along Commodity Value Chains Workshop Proceedings volume. FAO, Rome. Holt, C. A. and S. K. Laury (2002). Risk aversion and incentive effects. American Economic Review 92 (5), 1644--1745. Jalan, J., and Ravallion, M. (2002). 2004. Household Income Dynamics in Rural China. In Insurance against Poverty, S. Dercon, ed. Oxford: Oxford University Press. Ozanne, A. (1998). Uncertainty, duality and perversity: An empirical test of the Schultz- Baron hypothesis. Applied Economics 30 (4), 521-530. Kahneman, D. and A. Tversky (1979). Prospect theory: An analysis of decision under risk. Econometrica 47 (2), 263--291. Karlan, D. (2005). Using experimental economics to measure social capital and predict financial decisions. American Economic Review 95 (5), 1688 --1699. Kurosaki, T. and M. Fafchamps (2002). Insurance market efficiency and crop choices in Pakistan. Journal of Development Economics 67 (2), 419-53. Lim, Y. and R. M. Townsend (1998). General equilibrium models of financial systems: Theory and measurement in village economies. Review of Economic Dynamics 1 (1), 59.118. Lokshin, M. and M. Ravallion. (2004). Household Income Dynamics in Two Transition Economies. Studies in Nonlinear Dynamics and Econometrics 8(3). Lopez, R. E. (1986). Structural models of the farm household that allow for interdependent utility and profit maximization decisions. In I. Singh, L. Squire, and J.

27

Strauss (Eds.), Agricultural Household Models: Extensions, Application and Policy, pp. 306-326. Baltimore: John Hopkins Press. Lund, S. and M. Fafchamps (2000). Risk-sharing networks in the rural Philippines. Journal of Development Economics 71 (2), 261.287. Knight, J., S.Weir, and T.Woldehanna (2003). The role of education in facilitating risk-taking and innovation in agriculture. Journal of Development Studies 39 (6), 1--22. Magambo, M. J. S. (2000). Clonal robusta coffee farming. Uganda National Farmers Association: Farmers Guide Series, Kampala, Uganda. Malchow-Møller, N. (2002). Dynamic Behaviour and Agricultural Households in Nicaragua. Ph. D. thesis, University of Aarhus, Denmark. Morduch, J. (1991). Risk and Welfare in Developing Countries. Ph. D. thesis, Harvard University. Morduch, J. (2004). Consumption smoothing across space: Testing theories of risk-sharing in the ICRISAT study region of south India. In S. Dercon (Ed.), Insurance Against Poverty, Chapter 2. OUP: Oxford. Moscardi, E. and A. de Janvry (1977). Attitudes towards risk among peasants: An econometric approach. American Journal of Agricultural Economics 59 (4), 710--716. Munyambonera, E. (2004). Socio-economic impact of the coffee wilt disease and implications to coffee production and policy implications for the future. In The Coffee Yearbook 2003 - 2004. Uganda Coffee Trade Federation, Kampala, Uganda. Nyarko, Y. and A. Schotter (2002). An experimental study of belief learning using elicited beliefs. Econometrica 70, 971 --1005. Pender, J., P. Jagger, E. Nkonya, and D. Sserunkuuma (2004). Development pathways and land management in Uganda. World Development 32 (5), 767--792. Quisumbing, A. (2006). Investments, bequests, and public policy: Intergenerational transfers and the escape from poverty. Washington, D.C.: International Food Policy Research Institute. Rosenzweig, M. and H. Binswanger (1993). Wealth, weather risk and the composition and profitability of agricultural investments. Economic Journal 103, 56--78. Sandmo, A. (1971). On the theory of the competitive firm under price uncertainty. American Economic Review 61 (1), 65--73.

28

Scandizzo, P. L. and J. L. Dillon (1979). Peasant agriculture and risk preferences in northeast Brazil: A statistical sampling approach. In J. A. Roumasset, J.-M. Boussard, and I. Singh (Eds.), Risk, Uncertainty and Agricultural Development, pp. 133--160. New York, New York: Agricultural Development Centre. Singh, I., L. Squire, and J. Strauss (Eds.), Agricultural Household Models: Extensions, Application and Policy, Baltimore: John Hopkins Press. Shahabuddin, Q., S. Mestelman, and D. Feeny (1986). Peasant behaviour towards risk and socio-economics and structural characteristics of farm households in Bangladesh. Oxford Economic Papers 38, 122--30. Trivedi, P. K. (1992). A case study of cocoa replanting and new planting in Bahia, Brazil, 1966-85. Journal of Development Economics 39 (2), 279-299. Udry, C. (1995). Risk and sharing in northern Nigeria. American Economic Review 85 (5), 1287-1300. Uganda Coffee Trade Federation (2002). The Coffee Yearbook 2001 - 2002. Kampala, Uganda. Wik, M. and S. Holden (1998). Experimental studies of peasants' attitudes toward risk in northern Zambia. Discussion Paper D -14, Department of Economics and Social Sciences, Agricultural University of Norway.

29

30

Kernel Regress ion with 95% Confidence Interval 1.08171

.094503 3.47369

lnW

7.6522

Figure 3: Nonparametric estimation of lnW on λ t / P

31

Table 2: Lotteries offered to the farmers

Option 1 Option 2 Option 3 Option 4 Option 5

Coffee Price A coffee price of:

Coffee Yield Clonal coffee trees that give:

6, 000 shillings per 20 kilos of kiboko (unprocessed coffee) every year 5,400 shillings with probability 0.5 and 9,600 shillings with probability 0.5 4,800 shillings with probability 0.5 and 12,000 shillings with probability 0.5 2,400 shillings with probability 0.5 and 18,000 shillings with probability 0.5 0 shillings with probability 0.5 and 24,000 shillings with probability 0.5

100kg / 0.25 acre every year

Matooke Yield Hybrid matooke plants that give: 100kg / 0.25 acre every year

90kg / 0.25 acre with probability 0.5 and 160kg / 0.25 acre with probability 0.5 80kg / 0.25 acre with probability 0.5 and 200kg / 0.25 acre with probability 0.5 40kg / 0.25 acre with probability 0.5 and 300kg / 0.25 acre with probability 0.5 0kg / 0.25 acre with probability 0.5 and 400kg / 0.25 acre with probability 0.5

90kg / 0.25 acre with probability 0.5 and 160kg / 0.25 acre with probability 0.5 80kg / 0.25 acre with probability 0.5 and 200kg / 0.25 acre with probability 0.5 40kg / 0.25 acre with probability 0.5 and 300kg / 0.25 acre with probability 0.5 0kg / 0.25 acre with probability 0.5 and 400kg / 0.25 acre with probability 0.5

Table 3: Risk parameters of lotteries offered

Option 1 Option 2 Option 3 Option 4 Option 5

Mean payoff

Standard Deviation of payoff

(For coffee price lotteries) 6,000 7,500 8,400 10,200 12,000

(for coffee price lotteries) 0 2100 3600 7800 12000

Range of P compatible with this choice (for all lotteries) 7.33 - ∞ 1.86 – 7.33 0.63 – 1.86 0.275 – 0.63 0 – 0.275

32

Table 4: Labour and revenue share by risk preference category Labour share Mean Stnd. dev. From question on coffee prices Option 1 0.272 -0.202 Option 2 0.312 -0.237 Option 3 0.371 -0.24 Option 4 0.343 -0.236 Option 5 0.296 -0.184 From question on coffee yields Option 1 0.264 -0.2 Option 2 0.323 -0.231 Option 3 0.4 -0.224 Option 4 0.385 -0.227 Option 5 0.297 -0.233

Revenue share Mean Stnd. dev.

No. of observations

0.182 0.231 0.27 0.256 0.255

-0.157 -0.213 -0.216 -0.23 -0.21

67 49 54 79 45

0.183 0.234 0.318 0.221 0.278

-0.164 -0.2 -0.25 -0.166 -0.243

107 51 55 35 46

33

Table 5: Main results with labour share as the dependent variable (1)

(2)

(3)

(4)

ln (wealth)

0.04***

(0.02)

0.04***

(0.02)

0.04***

(0.01)

0.04***

(0.01)

People to help

0.002

(0.01)

0.001

(0.01)

0.003

(0.01)

0.001

(0.01)

Trees lost to wilt

-0.08**

(0.04)

-0.09*

(0.05)

-0.08**

(0.04)

-0.09**

(0.05)

Land to labour ratio

0.10

(0.64)

0.07

(0.59)

-0.12

(0.58)

-0.12

(0.57)

Gender of head

0.03

(0.03)

0.03

(0.03)

0.03

(0.03)

0.04

(0.03)

Age of head

0.001

(0.0008)

0.0004

(0.0008)

0.0001

(0.0008)

0.00001

(0.0008)

Distance to Kampala

0.00003

(0.00005)

0.0001**

(0.0001)

0.00002

(0.00005)

0.0001*

(0.0001)

Distance to market

-0.01

(0.005)

-0.01

(0.005)

-0.01

(0.005)

-0.01

(0.005)

Kt

0.18***

(0.04)

0.16

(0.14)

0.20***

(0.05)

0.22

(0.16)

Prob. of negative return

-0.19**

(0.07)

-0.17**

(0.08)

-0.23***

(0.08)

-0.20**

(0.09

Option 2 (1st to 75th)

0.10**

(0.04)

0.08*

(0.04)

Option 3 (1st to 75th)

0.16***

(0.04)

0.15***

(0.04)

Option 4 (1st to 75th)

0.15***

(0.05)

0.14**

(0.06)

Option 5 (1st to 75th)

0.03

(0.04)

0.03

(0.04)

Dummy for top quartile

0.08

(0.06)

0.10

(0.06)

Option 2 (76th to 100th)

-0.05

(0.07)

-0.06

(0.07)

Option 3 (76th to 100th)

-0.01

(0.07)

-0.02

(0.07)

Option 4 (76th to 100th)

-0.03

(0.08)

-0.06

(0.08)

Option 5 (76th to 100th)

-0.08

(0.08)

-0.09

(0.08)

P*(Dummy for bottom three quartiles)

-0.01***

(0.005)

-0.01**

(0.005)

P*(Dummy for top quartile)

-0.01

0.01

-0.01

0.01

0.29***

(0.09)

0.27***

(0.10)

Intercept Number of observations F-test R-squared

0.19*

(0.09)

0.17*

(0.10)

293

275

293

275

F(22, 60) = 7.86***

F(22, 57) = 6.15***

F(15, 67) = 7.88***

F(15, 64) = 6.68***

0.3248

0.3281

0.2817

0.2868

Test of joint significance of risk preference dummies: Poorest three quartiles

F(4, 78) = 4.62***

F(4, 75) = 4.32***

Richest quartile

F(4, 78) = 0.35

F(4, 75) = 0.43

Table note: standard errors, corrected for clustering at the village level, are in brackets. *** denotes significant at 0.01, ** denotes significant at 0.05 and * denotes significant at 0.10. Regional dummies are included but not shown.

34

Table 6: Alternative interactions of P and W, with labour share as the dependent variable (1)

(2)

ln (wealth)

0.03*

(0.01)

0.03**

(0.01)

People to help

0.002

(0.01)

0.001

(0.01)

Trees lost to wilt

-0.07*

(0.04)

-0.08**

(0.04)

Land to labour ratio

0.19

(0.68)

0.004

(0.62)

Gender of head

0.03

(0.03)

0.03

(0.03)

Age of head

0.0001

(0.001)

0.0001

(0.001)

Distance to Kampala

0.00001

(0.00005)

0.00003

(0.00005)

Distance to market

-0.01

(0.005)

-0.01

(0.004)

Kt

0.19***

(0.05)

0.19***

(0.04)

Prob. of negative return

-0.18**

(0.08)

-0.21**

(0.08)

-0.06**

(0.02)

0.34***

(0.10)

P / ln (wealth) P / wealth

-0.52*

(0.26)

Intercept

0.32***

(0.10)

No. of observations F-test R-squared

293

293

F(14, 68) = 8.91***

F(14, 68) = 9.26***

0.2713

0.2812

Table note: standard errors, corrected for clustering at the village level, are in brackets. *** denotes significant at 0.01, ** denotes significant at 0.05 and * denotes significant at 0.10. Regional dummies are included but not shown.

35

Table 7: Alternative measures of P Coffee price preferences

Matooke yield preferences

ln (wealth)

0.04

(0.01***)

0.04

(0.02***)

People to help

0.00

(0.01)

0.00

(0.01)

Trees lost to wilt

-0.08

(0.04**)

-0.08

(0.04**)

Land to labour ratio

-0.03

(0.60)

-0.10

(0.71)

Gender of head

0.03

(0.03)

0.02

(0.03)

Age of head

0.00

(0.00)

0.00

(0.00)

Distance to Kampala

0.00

(0.00)

0.00

(0.00)

Distance to market

-0.01

(0.00)

-0.01

(0.00)

Kt

0.19

(0.05***)

0.18

(0.04***)

Prob. of negative return

-0.22

(0.09**)

-0.23

(0.09**)

Option 2 (1st to 75th)

0.10

(0.06*)

0.10

(0.04**)

Option 3 (1st to 75th)

0.13

(0.05***)

0.12

(0.04***)

Option 4 (1st to 75th)

0.10

(0.04**)

0.12

(0.05**)

Option 5 (1st to 75th)

0.08

(0.04*)

0.03

(0.05)

Dummy for top quartile

0.04

(0.07)

0.08

(0.07)

Option 2 (76th to 100th)

0.01

(0.08)

-0.02

(0.07)

Option 3 (76th to 100th)

0.03

(0.07)

-0.04

(0.07)

Option 4 (76th to 100th)

0.08

(0.08)

0.02

(0.10)

Option 5 (76th to 100th)

0.01

(0.08)

-0.04

(0.09)

Intercept

0.16

(0.10)

0.20

(0.10)

Number of observations F-test R-squared

293

293

F( 22, 60) = 5.56***

F( 22, 60) = 5.30***

0.2947

0.3039

Test of joint significance of risk preference dummies: Poorest three quartiles

F(4, 78) = 2.35*

F(4, 78) = 3.92***

Richest quartile

F(4, 78) = 0.40

F(4, 78) = 0.14

Table note: standard errors, corrected for clustering at the village level, are in brackets. *** denotes significant at 0.01, ** denotes significant at 0.05 and * denotes significant at 0.10. Regional dummies are included but not shown.

36

Table 8: Allowing P to reflect the concavity of the value function with labour share as the dependent variable (1)

(2)

ln (wealth)

0.04***

(0.01)

0.04***

(0.01)

People to help

0.004

(0.01)

0.001

(0.01)

Trees lost to wilt

-0.08**

(0.04)

-0.08**

(0.04)

Land to labour ratio

-0.15

(0.61)

-0.10

(0.62)

Gender of head

0.03

(0.03)

0.03

(0.03)

Age of head

0.0005

(0.001)

0.0001

(0.001)

Distance to Kampala

0.0001

(0.00005)

0.00003

(0.00005)

Distance to market

-0.01

(0.004)

-0.01

(0.005)

Kt

0.19***

(0.04)

0.19***

(0.04)

Prob. of negative return

-0.18**

(0.07)

-0.21**

(0.08)

Risk option 2

0.07**

(0.03)

Risk option 3

0.12***

(0.04)

Risk option 4

0.11**

(0.04)

Risk option 5

0.01

(0.03) -0.01**

(0.004)

0.30***

(0.09)

P Intercept Number of obs F-test R-squared Test of joint significance of risk preference dummies

0.22***

(0.08) 293

293

F(17, 65) = 9.24***

F(14, 68) = 9.52***

0.3055

0.2783

F(4, 78) = 4.04***

Table note: standard errors, corrected for clustering at the village level, are in brackets. *** denotes significant at 0.01, ** denotes significant at 0.05 and * denotes significant at 0.10. Regional dummies are included but not shown.

37

Table 9: Main results with revenue share as the dependent variable (1)

(2)

(3)

(using clonal coffee risk preference data)

(4) (using hybrid matooke risk preference data)

ln (wealth)

0.04

(0.02**)

0.02

(0.01*)

0.03

(0.01***)

0.02

(0.01*)

People to help

0.004

(0.01)

0.002

(0.01)

0.002

(0.01)

0.004

(0.01)

Trees lost to wilt

-0.11

(0.04***)

-0.10

(0.04***)

-0.11

(0.04***)

-0.10

(0.04***)

Land to labour ratio

-0.57

(0.53)

-0.50

(0.53)

-0.61

(0.53)

-0.47

(0.54)

Gender of head

-0.04

(0.02)

-0.04

(0.02*)

-0.04

(0.02*)

-0.04

(0.02*)

Age of head

0.0005

(0.0008)

0.001

(0.001)

0.001

(0.001)

0.001

(0.001)

Distance to Kampala

-0.0002

(0.00004***)

-0.0002

(0.00003***)

-0.0001

(0.00003***)

-0.0002

(0.00003***)

Distance to market

-0.01

(0.003***)

-0.01

(0.003***)

-0.01

(0.003***)

-0.01

(0.003***)

Prob. of negative return

0.03

(0.07)

0.01

(0.07)

0.004

(0.07)

-0.01

(0.07)

Option 2 (1st to 75th)

0.05

(0.03)

Option 3 (1st to 75th)

0.14

(0.03***)

Option 4 (1st to 75th)

0.01

(0.03)

Option 5 (1st to 75th)

0.09

(0.04**)

Dummy for top quartile

0.02

(0.06)

Option 2 (76th to 100th)

-0.05

(0.07)

Option 3 (76th to 100th)

0.09

(0.08)

Option 4 (76th to 100th)

0.03

(0.08)

Option 5 (76th to 100th)

-0.01

(0.07) -0.05

(0.02***)

-0.05

(0.02**)

0.19

(0.09**)

P / lnW P Intercept

0.04

Number of observations F-test R-squared

(0.09)

0.20

(0.08**)

-0.01

(0.003***)

0.17

(0.08**)

299

299

299

299

F( 21, 62) = 4.84***

F(13, 70) = 6.34***

F(13, 70) = 6.28***

F(13, 70) = 5.47***

0.177

0.1389

0.1422

0.1301

Test of joint significance of risk preference dummies: Poorest three quartiles

F(4, 79) = 5.88***

Richest quartile

F(4, 79) = 0.77 Table note: standard errors, corrected for clustering at the village level, are in brackets. *** denotes significant at 0.01, ** denotes significant at 0.05 and * denotes significant at 0.10. Regional dummies are included but not shown.

38

Appendix Tables Instrumenting regression for

Instrumenting regression for table

table 5, column 2

5, column 4

ln (wealth)

0.01

(0.02)

0.02

(0.02)

People to help

-0.01

(0.01)

-0.01

(0.01)

Trees lost to wilt

-0.16

(0.05***)

-0.12

(0.05**)

Land to labour ratio

-6.49

(1.21***)

-6.51

(1.22***)

Gender of head

-0.04

(0.03)

-0.05

(0.03)

Age of head

-0.001

(0.001)

-0.001

(0.001)

Distance to Kampala

-0.0001

(0.0001)

-0.0001

(0.0001)

Distance to market

-0.002

(0.005)

-0.001

(0.005)

Prob. of negative return

-0.06

(0.09)

-0.09

(0.09)

Dummy for top quartile

0.06

(0.07)

0.03

(0.06)

Option 2 (76th to 100th)

-0.02

(0.10)

Option 3 (76th to 100th)

0.05

(0.08)

Option 4 (76th to 100th)

-0.05

(0.09)

Option 5 (76th to 100th)

0.02

(0.08)

Option 2 (1st to 75th)

0.16

(0.04***)

Option 3 (1st to 75th)

0.04

(0.05)

Option 4 (1st to 75th)

0.02

(0.06)

Option 5 (1st to 75th)

0.09

(0.05*)

P*(Dummy for top quartiles)

-0.0007

(0.02)

P*(Dummy for bottom three quartiles)

0.01

(0.01*)

K t-3

0.14

(0.02***)

0.14

(0.02***)

Year land in which coffee is grown was

-0.00134

(0.001)

-0.001

(0.001)

2.81

(2.39)

2.47

(2.36)

acquired Intercept No. of observations F-test R-squared

275

275

F(23, 251) = 4.29***

F(17, 257) = 4.97***

0.2161

0.1975

39

1

Binswanger's most critical statements on the use of hypothetical questions came from an analysis of a different type of hypothetical questioning (a more complex questioning first used in Brazil by Scandizzo and Dillon (1979)). 2 The Uganda Coffee Development Authority provides some estimates of the costs of growing coffee, which allows us to calculate the return to coffee production for an average farmer. The data used comes from sensitivity studies conducted by UCDA throughout 2001 in which farmers in Masaka, Bushenyi and Kiboga regions of Uganda were interviewed. The UCDA estimates the cost of man labour to be 1,500 Ugandan shilling per day and the cost of inputs (unspecified) to be 30,000 shillings. Estimates suggest 100 days of man labour are needed to produce 1,000 kilos of kiboko (unmilled coffee) from one hectare of land. The cost of land was not factored into the UCDA analyses, but Deininger and Mpuga (2002) state the rental price per hectare to range between 3,500 and 11,000 shillings for the regions of Uganda sampled in the UCDA survey. The calculations here use the UCDA estimates and estimates of the rental price of land per hectare from Deininger and Mpuga (2002). 3 The data were collected by the Uganda Bureau of Statistics in collaboration with the Centre for the Study of African Economies at Oxford University. 4 Regressions were also run using a measure of wealth that did not include land wealth but this did not alter the qualitative results (although wealth became less significant). 5 Taken as 312 days multiplied by the number of household members older than 14 and able to work. 6 To allow comparability with other empirical studies (Binswanger 1980) a constant partial risk aversion utility function of the form U(Y)=(1-P)Y1-P was used to calculate a range of relative risk aversion compatible with each choice. To compute a unique value of P for each alternative, the geometric mean of the two endpoints was used (because as the interval length decreases the alternatives get more risky) except for the most risky alternative which has an endpoint of 0 (assuming no farmer was risk loving) and so the arithmetic mean was used. For the no risk option, the value of the lowest endpoint was used as the unique value of P. 7 Much empirical evidence suggests risk aversion is negatively correlated with wealth (e.g. Scandizzo and Dillon 1979; Binswanger 1981; Barsky et al 1997; Wik and Holden 1998). However, some qualification is needed: Binswanger found the relationship between wealth and measured risk aversion to be negative but weak; Scandizzo and Dillon found wealth only to have a negative relationship with risk aversion when subsistence was at risk; and Barsky et al found the relationship to be non-linear. Risk preferences have also been found to correlate with an individual’s past “luck” with a given outcome (Binswanger 1981; Wik and Holden 1998). 8 For each farmer the beans were split into seven 100 shilling intervals from 100 to 800. A common lower and upper limit was placed on the data. The class mark for each of these 100 shilling classes was taken as the midpoint of the class. The mean was calculated as

7

∑ f i xi where

xi is the given class mark for class i

i =1

7 and f i is the probability the price would fall into this class. The variance was given as ∑ f i ( xi − x) 2 .

9

i =1

Seven zero observations were recorded. Coffee farmers were defined as those who had coffee trees. Some farmers had coffee trees but had elected not to allocate labour to its production in the year prior to the survey.

40