Explaining smallholder maize marketing in southern and eastern Africa: The roles of market access, technology and household resource endowments

Food Policy 43 (2013) 248–266

Contents lists available at ScienceDirect

Food Policy journal homepage: www.elsevier.com/locate/foodpol

Explaining smallholder maize marketing in southern and eastern Africa: The roles of market access, technology and household resource endowments David Mather ⇑, Duncan Boughton1, T.S. Jayne Department of Agricultural, Food, and Resource Economics, Justin S. Morrill Hall of Agriculture, Michigan State University, 446 W. Circle Dr., Rm 207, East Lansing, MI 48824-1039, USA

a r t i c l e

i n f o

Article history: Received 30 December 2011 Received in revised form 1 September 2013 Accepted 12 September 2013

Keywords: Smallholder grain marketing behavior Food security Sub-Saharan Africa

a b s t r a c t Research on household food grain sales behavior in developing countries has tended to focus on the roles of market access and prices to explain why many rural households do not sell staple crops, though recent literature suggests that low household asset endowments may also be key constraints. We use econometric analysis of panel data from smallholders in Kenya, Mozambique, and Zambia to inform the design of public investments that will enable smallholders to increase their maize sales. Results show that investments that raise farm-level productivity and land access are an essential complement to investments that improve market access. Ó 2013 Elsevier Ltd. All rights reserved.

Introduction Given the new international food price environment and continuing rapid urbanization, African governments are anxious to increase the availability of domestically-produced marketed food staple surpluses. In most African countries, smallholders account for the majority of marketed food staples, even though only a small proportion of the rural population are net sellers. Broadening the base of smallholder maize market participation and increasing their ability to respond to price incentives therefore represents both a means to improve food security and a potential opportunity to raise smallholder incomes. But what kinds of investments are most likely to achieve this objective? During the 1990s it was a widely held view that marketfriendly policies, combined with investments in public or collective goods to increase farm productivity and reduce marketing costs, would be sufficient to overcome barriers to specialization and trade by rural smallholders. In practice, however, levels of investment in important public goods such as national agricultural research and extension systems were woefully inadequate in the 1990s and are only beginning to increase after two decades of neglect. Furthermore, some recent literature has questioned whether heterogeneity in household resource endowments might also constrain an important number of poor households from taking advantage of lower market access costs (Boughton et al., 2007; ⇑ Corresponding author. Tel.: +1 517 449 9694. E-mail addresses: [email protected] (D. Mather), [email protected] (D. Boughton), [email protected] (T.S. Jayne). 1 Tel.: +1 517 432 6659. 0306-9192/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.foodpol.2013.09.008

Barrett, 2008). In other words, increases in public good-type investments to improve market access may be a necessary but not sufficient condition to enable a significant number of households to escape poverty and become food secure. In this paper, we use econometric analysis of nationally-representative smallholder panel data sets from Kenya, Mozambique, and Zambia to inform the design of public investments that will enable smallholders to increase their marketed food staple surpluses in a financially sustainable manner. The heterogeneity across and within the smallholder sectors of these countries allows us to analyze the extent to which smallholder maize marketing patterns vary by levels of market access, household assets, use of improved inputs, and agro-ecological potential. The principal conclusion from our analysis is that investments that raise farm-level productivity and land access are an essential complement to investments that improve market access (reduce marketing costs). The paper is organized as follows. Data sources describe the data sources, and Conceptual framework discusses the conceptual framework for modeling smallholder maize marketing. Methods discuss the econometric model and estimation issues, and Results the key empirical results. The concluding section briefly discusses implications for investment programs aimed at increasing smallholder marketed maize surpluses in eastern and southern Africa. Data sources Kenya: The Tegemeo Institute of Egerton University and Michigan State University designed and implemented smallholder farm surveys in eight agro-ecological zones where crop cultivation predominates. The sampling frame for the survey was prepared in

D. Mather et al. / Food Policy 43 (2013) 248–266

consultation with the Central Bureau of Statistics. Households and divisions were selected randomly within purposively chosen districts in the eight agro-ecological zones. The nationwide survey includes 106 villages in 24 districts. Sampling details are provided in Argwings-Kodhek et al. (1998). A total of 1578 small-scale farming households were surveyed in 1997. Of these, we drop 48 households because either they were found to be mainly pastoral farmers or their landholding size exceeded 20 ha. The 1997 survey therefore constituted 1530 sedentary households farming under 20 ha. Subsequent panel waves were conducted in 2000, 2004, and 2007. The 2007 sample contains 1342 households of the original 1578 sampled, a re-interview rate of 85%. Due to problems with the 2000 survey implementation, for this paper, we use a balanced panel of households interviewed in each of the 1997, 2004 and 2007 surveys. Mozambique: In 2002, the Mozambican Ministry of Agriculture and Rural Development (MADER) in collaboration with the National Institute of Statistics (INE) conducted a national rural household income survey – the Trabalho do Inquerito Agrìcola (TIA). The sampling frame was derived from the Census of Agriculture and Livestock 2000, and used a stratified, clustered sample design that is representative of small- and medium-scale farm households at the provincial and national levels. The sample was stratified by province (10 provinces) and agro-ecological zones, and included eighty of the country’s 128 districts. A total of 4908 small and medium-sized farms were interviewed in 559 communities (clusters). A subsequent panel wave was conducted in 2005, with a re-interview rate of 82.7%, and replacement of attrited households to retain a representative sample of the population. Zambia: Data is drawn from the Central Statistical Office’s Post Harvest Survey (PHS) of 1999/2000, and the linked 2001, 2004, and 2008 Supplementary Surveys (SS) designed and conducted jointly by the government’s Central Statistical Office and Michigan State University. A 3-wave household panel data set is available for the three agricultural production seasons covered by the three Supplemental Surveys, 1999/2000 (SS 2001), 2002/2003 (SS 2004), and 2007/08 (SS 2008). The PHS is a nationally representative survey using a stratified three-stage sampling design. Census Supervisory Areas (CSA) were first selected within each district, next Standard Enumeration Areas (SEA) were sampled from each selected CSA, and in the last stage a sample of households was randomly selected from a listing of households within each sample SEA. The SEA is the most disaggregated geographic unit in the data, which typically includes 2–4 villages of several hundred households. The 2001, 2004, and 2008 surveys are based on a sample frame of about 7400 small-scale (0.1–5 ha) and medium-scale farm households, defined as those cultivating areas between 5 to 20 hectares. For our panel econometric analysis of the Zambia data, we use the 68.2% of households that were successfully interviewed in both the 2001 and 2008 survey waves. We do not use cases from the 2004 survey wave as it failed to collect information on a key variable for our analysis (total landholding). Because the 2001 Supplemental Survey covered the 1999/2000 agricultural year, we refer to this survey wave as 2000 from here on. Because this descriptive and panel work is focused on small and medium-scale farmers, we drop sample households with greater than 20 ha cultivated in Mozambique and Zambia (n = 7 and n = 74 households, respectively), and households with greater than 20 ha owned in Kenya (n = 16 households). For this paper we only use households that were interviewed in each of the panel years (attrition bias tests are discussed in Panel attrition). This enables us to take advantage of panel econometric techniques that help us to control more effectively for potential endogeneity bias that could arise if unobserved time-constant household-level factors are correlated with explanatory variables (as explained in Household-level unobserved heterogeneity)

249

Conceptual framework The conceptual framework for many of the existing empirical papers on marketed food staples is based on seminal theoretical work by de Janvry et al. (1991), who used a household model to demonstrate that costs associated with market transactions can explain why some households avoid engaging in food and cash crop markets. Their results derive from the premise that the typical rural household in a developing country faces a wedge between the sales price of a given commodity and its purchase price. This wedge may be due to a combination of factors related to marketing, production, or consumption. Market-related factors include transport costs between the farm household’s village and the relevant market, non-competitive behavior among local traders, poor access to price information, and shallow local markets. Production-related factors include lack of access to finance for key inputs and low food crop productivity, while consumption-related factors include lack of insurance mechanisms (e.g., credit) against risks of excessive variation in food market prices and/or availability. The larger the wedge between sales and purchase prices, the greater the width of the price band or wedge in which the costs of selling exceed a household’s willingness to sell, and the costs of purchasing the commodity are greater than a household’s willingness to pay. A household whose internal or shadow price for the commodity falls within this price band or wedge will thus chose to not participate in the market, as either a seller or buyer. This condition is sometimes referred to as a missing market or as a market failure. In this context, market failure is household- and not commodityspecific. The principal strand of empirical literature on smallholder participation in staple food markets has built upon these theoretical results, yet has focused primarily on the role of transaction costs in discouraging market participation (Goetz, 1992; Key et al., 2000; Renkow et al., 2004). In general, these studies find that transportation and search costs (usually proxied for by distance from the village to the nearest road or town) are negatively associated with market participation, while household ownership of transportation assets such as bicycles, pack animals, carts, and motorized vehicles (which would tend to reduce search costs) have a positive association with market participation. Based on these results, they argue that the effects of price policy are muted for a majority of rural households due to insufficient investment in institutional and physical marketing-related infrastructure. However, de Janvry’s theoretical model does not explain the missing market outcome on the basis of transaction costs alone. For example, while transaction costs define the width of the price band, the location of the household’s individual shadow price for the commodity is also influenced by its supply curve, which is determined by household asset levels (landholding, farm equipment), input choices (including technology choice), local agro-ecological potential, etc. Alene et al. (2007) recent empirical study on the role of transaction costs in impeding market participation is one of the few papers in this area which also highlight the role of non-price factors such as household landholding and technology choice in household maize marketing decisions. The location of household-specific price band is also determined by the household-specific demand curve, which is a function of not only household income but also socio-demographic factors. More recent literature has questioned whether the lack of smallholder response to the market reforms of the 1980s–90s in eastern and southern Africa (ESA) is due to heterogeneity in household resource endowments, which prevents a large number of poorer households from taking advantage of lower market access costs (Boughton et al., 2007; Barrett, 2008). The conceptual framework underlying these papers comes from the theory of asset

250

D. Mather et al. / Food Policy 43 (2013) 248–266

poverty traps (Carter and Barrett, 2006), which argues that lack of assets may preclude many smallholders from being able to produce a surplus necessary for participating in markets as sellers, and which give rise to the existence of minimum asset thresholds which must be overcome for a household to escape from poverty. Evidence presented or reviewed in Boughton et al. (2007) and Barrett (2008) suggest that increases in public good-type investments to improve market access may be a necessary but not sufficient condition to enable a significant number of households to escape poverty and become food secure. Thus, increased public good-type investments in improving the productivity of existing household assets (e.g., crop science, farmer know-how) may be an important compliment to investments to promote market development. A secondary strand of literature has demonstrated the importance of using the household model framework of Singh et al. (1986) to study the marketed surplus of semi-subsistence rural households, given that such households are both producers and consumers of food crops. Strauss’ theoretical work (1984) explicitly recognizes the importance of wealth effects on home consumption of a food crop, which may result from the impact of price changes on farm profits. This work has important implications for estimation of the responsiveness of marketed surplus to price changes, as his results demonstrate that these wealth effects tend to dampen supply response, and in some cases may be large enough to induce negative marketed surplus response. Renkow (1990) builds on Strauss’ work by considering how post-harvest stocks influence household wealth and thus the marketing decision. He demonstrates that incorporating post-harvest stocks into the wealth effect further dampens long-run supply response, and may result in negative supply response in the short-run. A more recent study demonstrates theoretically and empirically that even if households have access to storage, they may actually be more likely to sell maize in the immediate post-harvest period (when prices are relatively low) due to liquidity constraints caused by credit market failures (Stephens and Barrett, 2011). The scenario of negative marketed supply response has been observed empirically in a few cases (Bardhan 1970; de Janvry and Kumar 1981). Scandizzo and Bruce, (1980) survey of supply response elasticities for major staples in 103 developing countries also finds that supply response to higher prices is quite limited in many cases; they found that 62% of the supply elasticities were less than 0.50 and 27% were negative. Therefore, for the countries in our analysis, we might expect to see low or even negative responsiveness of marketed surplus to price changes where we observe either an asset-poor grower in a poor agro-ecological environment (i.e., low supply elasticity), or one for whom maize constitutes a large portion of his household income (i.e., high income elasticity) and who has a low substitution effect between food and other goods. When one assumes separability between household production and consumption, the first-order conditions for profit maximization give household production as function of input and output prices, the household’s productive assets and technology. However, the assumption of separability is unlikely to hold in developing countries like Mozambique, Zambia, and Kenya, with their imperfect credit and labor markets along with the risk factors caused by high weather variability and other shocks. Therefore, our study recognizes that household production and consumption decisions are likely non-separable in this context, so a household’s socio-demographic characteristics (such as the household’s dependency ratio) will affect its desired production level, which in turn affects its level of marketed surplus (Sadoulet and de Janvry, 1995). Because household production and consumption decisions in a non-separable framework also depend upon consumer prices, we deflate all prices and asset values by the consumer price index in each country.

Methods Double-hurdle model An econometric concern for modeling market participation is the fact that only a minority of households sell maize, thus the maize sales of non-sellers – the majority of cases – is zero. If the distribution of such a dependent variable exhibits a reasonably large number of cases lumped at zero (as in this case), this can create problems for standard OLS regression. We approach the statistical challenge posed by cases where market sales equal zero not as a missing data problem (which is typically modeled using a variant of the Heckman two-step approach, as in Goetz (1992)), but rather as a corner solution. Within the context of a study of the determinants of food staple sales by staple-growing households, the rationale for a corner solution model is that a sales value of zero is a valid economic choice to be explained, not a reflection of missing data. The standard approach to modeling a corner solution dependent variable is to use either a Tobit (Tobin, 1958) or a double-hurdle (DH) model. Our research objectives are to understand both the factors affecting the probability that a household sells maize and the factors affecting the amount sold. When the household’s maize sale participation and quantity decisions are made simultaneously, the Tobit model is appropriate for analyzing the factors affecting the joint sales decision. However, DH models such as the one proposed by Cragg (1971) offer a more flexible version of the Tobit in that they allow the household decision regarding whether to sell maize (participation) and what quantity to sell to be determined by different underlying processes. We estimate a double-hurdle model of household maize sales in two stages – the first stage being a Probit of the decision to sell maize or not, and the second a log normal of maize quantity sold (i.e. OLS on log of maize sale quantity on non-zero observations of maize sales). We then use a Vuong (1989) test of non-nested models to compare the Tobit versus the log-normal DH model and find that the DH model is clearly preferred in each case.2 Double-hurdle model variables Dependent variables The binary dependent variable for the Probit stage of the double-hurdle model in each country =1 if the household sold maize that year, or =0 otherwise. The dependent variable in the Lognormal stage is the natural log of maize quantity sold (kg) that year. In the extant literature, explanatory variables typically used to explain food market sales behavior can be divided into three major categories: agroecological conditions, market characteristics (e.g., prices and market access), and household characteristics. Agro-ecological potential and rainfall To control for spatial variation in agroecological potential in a national sample, we include dummies for either agroecological zones or provinces. Using rainfall data from each country at the district or village level, we create variables to measure the effect of drought stress on maize production during each season. In Kenya, we also include binary indicators for three of the principal soil-group categories (out of seven) created by Sheahan (2011) using village-level information on soil type, depth and percentage of sand and clay. The soil groups for which we create binary indicators include: volcanic soils, humic or highly productive soils, and Rankers soils with high sand. In Zambia, we use a binary variable which equals one if soil characteristics in the village are 2 In the case of each country, the Vuong test comparing a tobit with a lognormal DH demonstrates that the lognormal DH is preferred to the tobit (with a p-value of 0.000 on the test statistic in each case).

D. Mather et al. / Food Policy 43 (2013) 248–266

agro-ecologically suitable (in terms of rainfall and soil type) for low input fertilizer use, and equals zero otherwise (FSRP, 2012 as used in Mason, 2011a). Market access In Kenya and Zambia, we use distance to the nearest feeder road to proxy for transport cost to the nearest market, while for Mozambique, we use a variable ‘travel time to nearest city of 10,000 residents or more.’ While the distance to road measure of market access does not account for the costs of transport from the road to the relevant market itself, the majority of the transport cost to market is often from the village to the nearest good road. Household marketing assets are assumed to reduce search costs, thus we include binary variables for: ownership of a bicycle; ownership of a cart; ownership of a vehicle (motorcycle, truck). We note that it is possible that ownership of such assets might be endogenous if correlated with unobserved time-varying factors, thus significant partial effects of these variables on our dependent variables may best be interpreted as associations and not causal effects. Household receipt of market price information is recorded in the Mozambique survey and should serve to dramatically reduce information costs for rural households. To reduce the potential for endogeneity of this variable due to from simultaneity bias or reverse causation, we only use cases in which the household which received price information also owns a radio, which accounts for roughly 75% of all households which report receipt of market price information (which is broadcast via radio in many rural areas of the country). We use radio ownership as a proxy of market information receipt in Kenya, and cell phone ownership as a proxy in Zambia. That said, we note that use of a variable such as radio or cell phone ownership may potentially be endogenous due to correlation with unobserved time-varying factors, for which we cannot control. Household membership in a farmer association may also reduce the costs of obtaining market information (or market access), though this is only observed in Mozambique. An additional factor which may affect the market access of smallholders in Zambia is farmers’ expectations regarding future maize purchases in their district by the Food Reserve Agency (FRA), a government parastatal. Large volume purchasing by the FRA should tend to put upward pressure on prices, which would theoretically induce an increase in maize area planted, maize production, and marketed maize. To measure the potential influence of the activities of the FRA on household maize marketing decisions, we include log of FRA district-level maize purchase quantity in the marketing year that occurs just prior to the planting for the 2000 and 2008 supplemental surveys.

251

and marketing experience. In the interest of testing for gender disparities in maize marketing behavior, we include a binary variable which equals one if the household is headed by a single female (and zero otherwise) (i.e. single, divorced, widowed), and a separate binary variable which equals one if the head is a female with a resident spouse (and zero otherwise). Because the binary variable for single female-headed households may pick up negative effects of adult mortality (in the event that she was recently widowed), we also include a binary variable which equals one if the household suffered a death of an adult age 15–59 within the past 3 years. Thus, if the average partial effect of having a single female head is statistically significant, this should represent gender disparities in the ability to produce and market maize, free of any adverse effects due to recent adult mortality. In each country, we also include a measure of access to credit in the event that some observed sales are due in part to liquidity constraints in the post-harvest period (Stephens and Barrett, 2011). On the other hand, households with access to credit may be more capable of financing inputs such as hired labor, which could have a positive effect on maize productivity and therefore sales. Due to potential endogeneity concerns, the measure of credit access we use in each country is the percentage of village households that received credit for purchase of farm inputs that season.

Household production and consumption characteristics Household productive assets include physical assets such as: the log of total landholding size3 and landholding size squared and the log of total farm assets. To proxy for the availability of family labor for agricultural activities, we use the number of prime-age adults (those 15–59) and its square. In Kenya and Zambia, we are able to adjust the number of adults by their reported actual months of residence (over the 12-month recall period), while in Mozambique we use the number of prime-age adults who claim agriculture as a primary or secondary occupation. We include head’s years of education as a measure of human capital (in Kenya, we use maximum education among adults due to data limitations in the first panel year). Head’s age and age squared is included as a proxy for lifecycle wealth effects, though it may also measure human capital in terms of years of farming

Input use and improved technologies In each country, we include a binary variable which equals one if the household owns mechanized or animal traction (a suitable animal and equipment) – and zero otherwise – which may increase crop productivity due to more timely planting and improved soil aeration and weed control. In Kenya and Zambia, we include a binary variable which equals one if the household used a purchased hybrid maize seed variety that season, as well as the log of the quantity of chemical fertilizer applied per hectare of maize planted. We do not include such variables for Mozambique as information on purchases of improved maize seed were not available for both panel years and information on fertilizer use was only collected at the farm level. Household choices regarding input use are typically considered to be endogenous due to either omitted variable bias or the simultaneity of input decisions and realized outputs. While omitted variable bias from time-constant unobserved factors is likely controlled for with our CRE terms (see Household-level unobserved heterogeneity), we address the remaining potential for endogeneity bias of hybrid seed use and fertilizer use by including time-varying shocks such as main season rainfall and drought shocks and using an adapted Control Function (CF) approach developed by Rivers and Vuong (1988) to control for a continuous endogenous explanatory variable, and by Vella (1993) to control for an endogenous variable that is also a corner solution.4 We use village-level median fertilizer price for DAP and the median distance to fertilizer seller as the instruments for both use of hybrid seed and fertilizer use on maize (and in Kenya we also add the instrument distance to hybrid seed seller). The CF approach involves two steps. First, using the fertilizer example, we run a reduced form Tobit regression of the quantity of fertilizer used per hectare of maize planted (kg/ha) as a function of all the variables in our structural regression plus the instruments for both fertilizer use and hybrid seed use. Second, we include the residual from the Tobit regression of fertilizer use as a regressor in the structural equations of the double-hurdle model, along with the endogenous variable, household fertilizer use (kg/ha). The fertilizer variable is deemed endogenous if the partial effect of the Tobit

3 Landholding is defined as land for which the household has title or use-rights, excluding land which is rented in.

4 See Ricker-Gilbert et al. (2011) for a recent application of this adapted control function approach.

252

D. Mather et al. / Food Policy 43 (2013) 248–266

reduced form residual is significant in either of the stages of the double-hurdle. In Kenya, we run the reduced form Tobit regression of fertilizer quantity (Supplemental Table S.1) and find that the instruments distance to fertilizer seller and distance to seed seller both have a significant effect on fertilizer use (Mather et al., 2011). Because the coefficient on the residual from the reduced form model (fertilizer quantity used) is insignificant in both stages of the doublehurdle (p = 0.97; p = 0.85), we conclude that household fertilizer use on maize is exogenous, and we leave the reduced form Tobit residual out of the double-hurdle model. To test for the potential endogeneity of household use of purchased hybrid maize seed, we estimate a reduced form Linear Probability (LPM) regression of a binary variable which =1 if the household used purchased hybrid maize seed that season, =0 otherwise. We find that two of the instruments, fertilizer price and distance to fertilizer seller, have a significant effect on hybrid seed use, while distance to seed seller is nearly significant (p = 0.105) (Supplemental Table S.2). Following the CF approach, we include both the potentially endogenous variable – the binary variable for ‘1 = household used purchased hybrid maize seed’ – and the reduced form residual (from the LPM) in the double-hurdle model of maize sales. Because the coefficient on the reduced form residual is insignificant in both stages of the double-hurdle (p = 0.71; p = 0.70), we conclude that the binary variable for hybrid use is exogenous (Mather et al., 2011). There is an additional aspect of household fertilizer use in Zambia that may be endogenous to their maize sales decision, relating to government-subsidized fertilizer from programs implemented in 2000 and 2008. Because receipt of subsidized fertilizer by households was not likely random, we need to test and control for the potential endogeneity of the quantity of subsidized fertilizer received by the household in our Tobit regression of household fertilizer applied per hectare of maize. We therefore use the CF approach and run a Tobit reduced form regression of quantity of subsidized fertilizer received by the household, as a function of all the variables in the double-hurdle model, plus the village median commercial fertilizer price, and two IVs created and used by Mason (2011a) in her study of the effect of fertilizer subsidies on household demand for commercial fertilizer in Zambia. The first IV is a binary variable equal to one if the household’s constituency was won by the ruling party (the Movement for Multi-Party Democracy, MMD) during the last presidential election, and equal to zero otherwise. Presidential and parliamentary elections in Zambia take place every five years and the MMD candidate has won every presidential election since 1991. The second IV is the percentage point spread between the MMD and the lead opposition party in the household’s constituency in the last presidential election. These two variables provide reasonable instruments with which to test and control for the potential endogeneity of household quantity of subsidized fertilizer received, because each of them are expected to have a significant effect on subsidized fertilizer received (we test this below) yet would not be expected to have a significant effect on household maize sales, conditional on our other household- and village-level controls (this is a maintained assumption). Beginning with the Tobit reduced form regression of the quantity of subsidized fertilizer received by the household, we find that the first IV (a binary variable indicating that MMD won the household’s constituency in the previous election) is significant (p = 0.00) in our Tobit reduced form regression of the quantity of subsidized fertilizer received by the household (Supplemental Table S.3). Because the average partial effect (APE) of the residual from the Tobit of subsidized fertilizer quantity is significant (p = 0.04) in our Tobit reduced form regression of household fertilizer quantity applied per hectare of maize (Appendix Table S.4), we conclude that quantity of subsidized fertilizer received by the household is

endogenous to household fertilizer use, and we leave this residual in the Tobit of household fertilizer quantity (Mather et al., 2011). Next, we find that the IV distance to nearest fertilizer seller is significant (p = 0.001) in the reduced form regression of household quantity of fertilizer applied to maize (Supplemental Table S.4). Finally, we find that the residual from the reduced form of household quantity of fertilizer applied to maize is significant in the Lognormal stage (p = 0.00) of the double-hurdle, thus we leave this residual in the double-hurdle model and conclude that household fertilizer quantity applied per hectare of maize is endogenous to household maize sales (Mather et al., 2011). We also use the CF approach to test for the potential endogeneity of household use of purchased hybrid maize seed in Zambia, running a LPM on a binary variable which =1 if the household purchased hybrid maize seed that season, and 0 otherwise. Because one of the GRZ’s fertilizer subsidy programs also made hybrid seed available to farmers, we include in our LPM regression of hybrid seed use the variable quantity of subsidized fertilizer received by the household along with the residual from its Tobit reduced form (as discussed above). The IV distance to fertilizer seller is significant (p = 0.024) in the LPM reduced form regression, supporting its validity as an instrument (Supplemental Table S.5). Because the LPM reduced form residual is insignificant in both stages of the double-hurdle model of household maize sales (p = 0.68 in Probit, 0.85 in Lognormal), we conclude that household use of hybrid maize seed is exogenous to household maize sales (Mather et al., 2011). Expected farmgate maize prices Due to data limitations and/or their choice of theoretical model, all but one of the studies of grain marketing in developing countries of which we are aware use either an annual staple price (Goetz, 1992) or, more commonly, the post-harvest price (Key et al., 2000) in their sales model. Most of the papers on this topic are based on a theoretical model which assumes that the farmer uses a single staple price to make staple food production, consumption and sales decisions simultaneously; this further assumes that the farmer has perfect foresight with respect to price prediction. The sole exception is Renkow (1990), who notes that a farm household’s food staple marketing decision is likely to be a function of staple prices at two different points in time. First, the farmer’s output supply and factor demand decisions are based on the farmer’s expectation of the farmgate staple price at harvest, which itself is a function of price and other information available at the time of planting. Second, the farmer’s decisions regarding whether or not to sell the staple and how much to sell are based on observed output quantities and prices observed by the farmer in the post-harvest period. Renkow (1990) argues that a household’s marketed supply response is therefore a function of both the household’s response to the staple food price as a producer (supply response) and later as a consumer of the staple (consumption response); he notes that these two components of marketed supply response should ideally be estimated from two separate functions (that each include the staple price at a different point in time). Due to data limitations with respect to village-level retail or farm-gate maize prices observed in the year prior to each of our surveys for each country, we generate expected farmgate maize prices for each survey year using a method similar to quasi-rational expectations described by Nerlove and Fornari (1998). Using the subsample of maize sellers, we regress the household farm-gate post-harvest maize sale price (of maize sellers) in year t as a function of variables observed by the farmer at planting time such as lagged wholesale market prices of maize from the nearest regional market and household and village characteristics that might affect the maize sale price received by a given household. Household characteristics include factors such as head’s age, head’s education,

D. Mather et al. / Food Policy 43 (2013) 248–266

household marketing assets, household access to market price information, etc., and their time-averages, and village-level characteristics include factors such as distance to the nearest road or district town, the maize price in the planting quarter from the nearest wholesale market, quarterly lags of the wholesale maize prices, etc., and their time-averages. Prior to running that regression, we use a control function approach to test and control for potential sample selection bias in the distribution of maize prices that we observe. We then use the coefficients from the farmgate maize price model to predict the farmgate price expectation for each household –seller or non-seller of maize – which we then use in our double-hurdle models of household maize sales. We use the expected price by itself rather than the post-harvest sales price (by itself), because we find that the latter often results in ‘findings’ of negative marketed supply response that appear to be driven by the scenario of thin markets in areas of low food crop productivity, which is common to much of Mozambique and parts of Zambia.5 In each country, our farmgate price data consists of observations of farmgate sales prices of maize in the panel surveys. Wholesale market prices of maize include the price in the planting month from the nearest regional wholesale, as well as 11 months of lagged wholesale prices. In Mozambique, our lagged wholesale prices are quarterly average prices. In Kenya, we also include a variable distance to regional market to control for variation across villages in transport costs between the village and the regional market. Household characteristics which might influence the price received by a farmer include: age of the household head, a proxy for market experience; and education level of the head, a proxy for negotiation skill. We also use measures of the household value of storage assets6, total value of farm assets, and dummies for truck and bicycle ownership, which could proxy for improved market price offers due to improved access to market price information and/or better market access. Distance to the nearest motorable road serves as a proxy for transport costs to the relevant market. Because weather conditions may influence market prices, we include variables to measure expected rainfall levels and expected rainfall shocks. Finally, we also include the long-term average of each time-varying variable in the model, used to control for unobserved time-constant household heterogeneity using the correlated random effects (CRE) approach. Additional details on the price expectation models for each country are provided in Mather et al., (2011). Estimation issues Household-level unobserved heterogeneity If unobservable time-constant household-level characteristics such as soil quality, farm management ability, or risk preferences are correlated with observable determinants of household maize sales (assets such as total land area owned, head’s education level, etc., or measures of household input use), this can lead to biased coefficient estimates (i.e. termed omitted variable bias by Wooldridge (2002)). The household data sets used in this paper are longitudinal, which offers the analytical advantage of enabling us to control for time-constant unobservable household characteristics.

5 For example, under such conditions, following a good production season with adequate rainfall, maize prices at harvest will often be low, yet the probability of maize sale would be higher than average. By contrast, following a poor production season (due to drought), we would expect to observe relatively high farmgate maize prices, yet a lower than average probability of maize sale. Either scenario can result in a negative correlation between maize sale prices and the probability of maize sale. 6 Our survey instruments as respondents about the numbers of various farm and storage assets that they own, and their current market value. We value these assets based on household-specific reported market value of each asset, with some cleaning of extreme values. Farm assets include the value of livestock, farm tools/equipment, and household goods like radios and bicycles.

253

The fixed effect (FE) estimator is usually the most practical way to control for these unobserved time-constant household characteristics, since using FE requires no assumption regarding the correlation between observable determinants and unobservable heterogeneity. However, the FE estimator is problematic for this application as the FE Probit estimator has been shown to be inconsistent (Wooldridge, 2002). We therefore estimate each of the double-hurdle models in this paper using a pooled balanced sample and include Correlated Random Effects (Mundlak, 1978; Chamberlain, 1984), which explicitly accounts for unobserved heterogeneity and its correlation with observable variables, while yielding a fixed-effects-like interpretation. In practice, this means that we add the household’s timeaverage of each time-varying explanatory variable – as additional explanatory variables – to each stage of the double-hurdle model. In contrast to traditional random effects, the CRE estimator allows for correlation between household-level unobserved heterogeneity (ci) and the vector of explanatory variables across all time periods (Xit) by assuming that the correlation takes the form of: ci = s + aXibar + ai where Xi-bar is the time-average of Xit, with t = 1, . . ., T; s is a constant, and ai is the error term with a normal distribution, ai|Xi  Normal(0, r2a ). This assumption implies that household-level time-constant unobserved factors (ci) are correlated with these time-average CRE terms, thus enabling our explanatory variables Xit to remain uncorrelated with unobserved time-constant factors in the error term. We estimate a reduced form of the model in which s is absorbed into the intercept term and Xi-bar are added to the set of explanatory variables. Using an adjusted Wald test, we reject the hypothesis of zero correlation (n = 0) between unobserved heterogeneity and explanatory variables in the participation and sales equations in all but one case in each country,7 indicating that the CRE approach is superior to the traditional pooled or random effects estimators. To facilitate interpretation of the results, we compute average partial effects8 (APE) for each regressor, along with bootstrapped standard errors. Panel attrition We use only households which were re-interviewed in each of the subsequent panel surveys. Given that over time, some households move away from a village and others dissolve as part of a typical household life-cycle, panel household surveys typically have to contend with at least some sample attrition over time. To test for attrition bias, we follow the regression-based approach described in Wooldridge (2002, p. 585). Only in Zambia do we find evidence of attrition bias in the double-hurdle model (Supplementary Table S.1). For auxiliary or double-hurdle models which are affected by attrition bias, we use sampling weights which are adjusted for panel attrition bias using the Inverse Probability Weighting (IPW) method (Wooldridge, 2002).9 In the case of Kenya, we have chosen to use a balanced rather than an unbalanced panel, which results in us dropping 3.8% of households from the initial 1997 survey. Tests for potential attrition bias due to dropping these cases does not show evidence of such bias, nor are the significance or general magnitude of our key results changed by using a balanced rather than an unbalanced sample. Anticipating that average partial effects for some variables may differ across levels of household total landholding, we rank all 7 The test results by country are as follows for the participation and sales equations: Mozambique (p = 0.582, p = 0.018); Zambia (p = 0.000; p = 0.000); Kenya (p = 0.000, p = 0.000). 8 We compute the partial effect for each household, and then take the average partial effect across the entire sample (or subsample), and compute bootstrapped standard errors for inference (Wooldridge, 2002). 9 The attrition-correction factors for Mozambique were computed and described by Mather and Donovan (2007), those for Zambia were computed by the authors, and those for Kenya were computed by Bill Burke, as described in Burke and Jayne (2008).

254

D. Mather et al. / Food Policy 43 (2013) 248–266

Table 1 Market access and household asset indicators in Kenya, Mozambique, and Zambia. Source: (A) World Bank; (B) rural household surveys referenced in the data section of this paper. Kenya (2007)

Zambia (2008)

Mozambique (2005)

All rural householdsA Rural population (% of total population) Rural poverty (% of rural population below the poverty line) Rural population density per hectare arable land (people/ha) Agriculture, value added (% of GDP) Fertilizer use (kilograms per hectare of arable land)

79 49 5.6 20 36.4

65 77 3.4 19 50.1

66 57 3.0 27 1.6

Panel small and medium-holder rural householdsB Total gross rural HH income/AE ($US/AE) (mean/median)1 Tropical Livestock Units Share of total gross HH income from crops and livestock (%) % of Total HH agriculture production marketed (mean/median)2 Total area cultivated/Adult equivalents (hectares/AE) (mean/median) Distance from village to fertilizer seller (km) % of HHs which purchase chemical fertilizer for use on maize (%) Fertilizer applied to maize (kg/ha), among users (mean/median) % HHs which purchase hybrid or improved maize seed (%) % HHs using animal or mechanized traction % HHs renting in land % HHs with irrigation (%)3 % HHs receiving extension visit in past year (%) % HHs with access to credit (%) Observations

662/406 4.0 63 47/50 0.42/0.29 3 74 145/123 73 48 21 11 60 53 1166

179/97 3.1 70 25/17 0.38/0.27 19 32 265/246 25 37 0.6 0.6 61 12 3701

154/68 1.2 68 19/8 0.53/0.42 67 4 259/141 2 11 0.5 1 17 4 3181

Notes: (1) Total income includes the value of retained and sold crops, livestock product sales, remittances, wage earnings, and own business income; (2) computed as (total value of crop sales/total value of crop production); (3) irrigation definition by country: ownership of irrigation equipment (Kenya); ownership of water pump (Zambia); use of mechanized or gravity irrigation (Mozambique).

households in each of our country samples from largest to smallest in terms of landholding, and divide them into four quartiles. We then run the equivalent of a Chow test on the sales equation, which shows that the model coefficients are (jointly) significantly different enough by landholding quartile (and AEC zone) to warrant the use of separate regressions in each country by landholding quartile. We also group households into four agroecological (AEC) zones in each country (three in Kenya) that represent maize-production potential: Low potential, Low-Medium, Medium, and High. We find the same result when testing for differences among coefficients by AEC zone. While most of our results are from national-level regressions, we also report some results by landholding quartile or AEC zone. Results

times that of Zambia. Kenya thus demonstrates that smallholder agriculture can provide a pathway out of poverty when households have the necessary assets and access to improved agricultural technologies which enable them to take advantage of investments in market development. In all three countries, only a minority of rural households are net sellers of maize (43% in Kenya (2007); 27% in Zambia (2008); 16% in Mozambique (2005)) and maize sales are highly concentrated among those who do sell. For example, when we define households with net sales of more than 100 kg per AE, we find that 28% of smallholders are large net sellers in Kenya, in Zambia between 10% and 20% depending on the year, and in Mozambique only 3.8% (Table 2). In all three countries, smallholders with net sales greater than 100 kg/AE have maize production levels that are five times that of their counterparts who have negligible or no sales.

Descriptive statistics Market access As context for the discussion of the econometric results, we first note the marked differences in smallholder agriculture across our case study countries (Table 1). For example, there is a wide range of market access conditions, as conventionally measured by distance to a tarmac or feeder road or to an input dealer. A Kenyan smallholder farmer need travel just 3 km on average to purchase from a fertilizer retailer, compared to 19 km in Zambia and 67 km in Mozambique. Across and within these three countries there is also wide variation in household assets and agro-ecological potential. Among our three study countries, Kenya has both the highest level of smallholder commercialization and the lowest rural poverty rate. For example, the median Kenyan smallholder sells 50% of the value of their crop production, while the median smallholders in Zambia and Mozambique sell only 17% and 8% of the value of their crop production, respectively. Even though the share of total household income from crops and livestock in Kenya is slightly lower (at 63%) than either Zambia or Mozambique (70% and 68% respectively), median household income per adult equivalent (AE) in Kenya is almost six times that of Mozambique and four

Distance to road/town We now turn our discussion to the econometric results (Tables 3–5), focusing on results for market access, rainfall and drought stress, input use, land resource endowments, and price responsiveness. Summary statistics for all variables used in the auxiliary and double-hurdle models in each country are presented in Appendix Tables A1–A3. The conventional explanation for observing a low percentage of rural households in sub-Saharan Africa that sell food grains is that it is caused by poor market access (which results in high transaction costs for farmers – costs both of search and for transportation) and/or oligopolistic trader behavior. Consistent with this hypothesis, previous studies of food marketing behavior in rural areas of developing countries have used variables such as ‘‘distance to the nearest road (or town)’’ as a proxy for households’ market access, and have consistently found such variables to have a significant and negative relationship with the probability of selling food grains. By contrast, while our study uses similar proxies for market access, such as distance to the nearest road (Kenya/Zambia) or travel time to nearest town of 10,000

255

D. Mather et al. / Food Policy 43 (2013) 248–266 Table 2 Production and marketing characteristics of panel households by net sales category, by country. HH maize production/marketing characteristics1

HHs with large net sales Mean value by group

HHs with small net sales

HHs with negligible sales

Deficit HHs

All HHs

Kenya (2007) % of Rural HHs by group Total HH landholding (ha) % HH that apply fertilizer to maize % of HH using purchased hybrid seeds Maize production (kg/AE) Maize sales (kg/AE) Share sold (%) Maize purchases (kg/AE)

28.3 3.42 89 89 1033 662 55.2 8

11.7 1.74 84 81 270 68 27.9 11

36.6 1.68 69 65 224 6 3.2 10

23.4 1.40 59 59 191 1 0.8 66

100.0 2.11 74 73 451 198 20.4 23

Zambia (2008) % of Rural HHs by group Total HH landholding (ha) % HH that apply fertilizer to maize % of HH using purchased hybrid seeds Maize production (kg/AE) Maize sales (kg/AE) Share sold (%) Maize purchases (kg/AE)

17.7 4.49 71 56 927 555 57.5 9

8.5 3.54 44 30 269 70 36.2 11

42.8 2.36 22 16 197 4 3.4 8

31.0 2.17 21 20 175 3 1.4 86

100.0 2.78 32 25 338 107 15.2 33

Mozambique (2005) % of Rural HHs by group Total HH landholding (ha) % HH that apply fertilizer to maize % of HH using purchased hybrid seeds Maize production (kg/AE) Maize sales (kg/AE) Share sold (%) Maize purchases (kg/AE)2

3.8 3.29 8.6 4.5 483 246 57.3 10

7.0 2.84 6.4 1.5 230 50 33.6 1

51.0 2.23 5.2 1.4 94 3 5.0 3

38.2 2.04 2.6 1.8 30 2 6.1 91

100.0 2.24 4.4 1.7 94 15 9.8 37

Notes: (1) Maize production statistics computed among those who grew maize; (2) purchases/AE for Mozambique computed as = average consumption/AE (production/ AE sales/AE), using average consumption/AE by province from 2003 IAF expenditure survey data. HH = household; AE = Adult consumption equivalents; net maize sales of large net sellers = 100+ kg/AE, small net sellers = 25–99 kg/AE; negligible sales = 24 to 24 kg/AE; deficit has < 25 kg/AE net maize sales; autarkic households have zero net sales and are included with negligible seller group.

or more residents (Mozambique), we find that these market access proxies do not have a significant negative effect on smallholder maize sales in Kenya or Mozambique (Tables 3 and 5), and while market access has a significant and negative effect on maize sales in Zambia, the magnitude of these effects is negligible (Table 4). Recent rapid appraisal work on maize value chains in our three case countries offers an explanation for the discrepancy between our results regarding measures of ‘market access’ such as ‘distance to road/town’ and those from other studies. Contrary to the conventional depiction of poor to negligible market access in rural sub-Saharan Africa, recent rapid appraisal work in Zambia, Kenya, and Mozambique found that maize sellers in both remote and nonremote villages claimed that upwards of 30 traders had visited their village during the most recent post-harvest period (in 2009) (Kirimi et al., 2011). Likewise, our survey data shows the percentage of rural households which live in a village in which at least one grower sold maize on their farm or within the village to a privatesector trader (i.e., not including sales to other households) is 53% for Mozambique (2002), 67% for Zambia (2004), and 82% for Kenya (2007). This suggests that, while there still may be significant transport costs to the nearest relevant market for farmers in remote villages (which would cause traders to adjust their maize buying prices lower), even these farmers now face considerably lower search costs for price information and access to traders than they did a decade ago – a finding consistent with recent empirical study of market access measures in Kenya (Chamberlain and Jayne, 2011). Trader presence has likely increased in recent years due to increased investments in road construction in some countries (Kenya, Zambia) as well as the recent proliferation of cell phones in rural areas, which has dramatically reduced the risks involved in maize trading and thus lowered barriers to entry. Note that because expected farmgate maize sales price is included in our model, and that the location for the majority of sales

observed in these countries is the village, then transport costs for most sales should be reflected included in the farmgate sales prices which we observe. Thus, while prices may still be too low to induce some farmers to sell (because of low maize productivity and/or high transport costs to relevant markets), our results suggest that lack of access to traders is not a principal constraint to maize marketing in most areas of these three countries. The implication of this finding is not that further investment in road infrastructure is unnecessary – because better road infrastructure should lead to lower input prices for farmers, and higher output prices – but rather that road investments alone are not likely to be sufficient to elicit broad-based increases in maize market participation. Household ownership of transportation assets Only in Kenya do we find evidence that household ownership of a transportation asset improves either the probability of household maize sale or quantities sold. For example, in Kenya, we find that ownership of a vehicle improves probability of maize sale by 13%, while ownership of a cart improves quantity sold among sellers by 63% (Table 3). By contrast, in Mozambique and Zambia, ownership of a bicycle or cart does not affect the probability of household maize sale or sale quantities (Tables 4 and 5). This result is surprising considering that average market access in Kenya is considerably better than in Zambia or Mozambique. However, we would expect that if there are significant transaction costs associated with poor market access (as measured by measures such as ‘distance to road/town’) that are not already captured by our household-level maize price expectation variables, then transportation assets that theoretically reduce such transaction costs should condition the effect of market access on household maize sale behavior. In other words, if transaction costs of poor market access are present, then we would expect that households that do not own a bicycle or car would more likely face a

256

D. Mather et al. / Food Policy 43 (2013) 248–266

Table 3 Double-hurdle model of maize market sales participation and level of maize sold, Kenya, 1997–2004–2007. Probit

Independent variables

Lognormal

Dept. variable = 1 if HH sold maize, 0 otherwise APE of Xj on P(y > 0)

Probit and lognormal Dept variable = ln(kgs of maize sold)

APE (conditional) of Xj on lny, given y>0

APE (unconditional) of Xj on lny

APE

p-Value

APE

p-Value

APE

p-Value

Main season drought shocks Ln(total landholding) Ln(total assets) # Prime-age adults 1 = HH owns irrigation equipment

0.185 0.022 0.006 0.009 0.065

0.019** 0.151 0.725 0.209 0.116

0.409 0.216 0.081 0.002 0.261

0.109 0.002*** 0.238 0.959 0.237

1.013 0.273 0.090 0.034 0.528

0.004*** 0.002*** 0.201 0.405 0.151

1 = HH used hybrid maize Ln(fertilizer applied to maize) (kg/ha) Ln(village agricultural wage) % Village hhs which received credit 1 = HH owns animal or mechanized traction

0.031 0.016 0.141 0.089 0.016

0.122 0.027** 0.104 0.231 0.686

0.161 0.118 0.095 0.389 0.072

0.080* 0.000*** 0.635 0.148 0.587

0.267 0.171 0.367 0.679 0.128

0.013** 0.000*** 0.295 0.105 0.575

Distance from village to nearest extension Head’s age (years) Maximum adult education (years) Distance from village to motorable road (km) 1 = HH owns bike

0.004 0.004 0.000 0.003 0.032

0.575 0.045** 0.900 0.882 0.202

0.007 0.001 0.004 0.063 0.065

0.732 0.715 0.215 0.280 0.444

0.005 0.014 0.012 0.053 0.040

0.880 0.078* 0.346 0.557 0.722

1 = HH owns vehicle 1 = HH owns cart 1 = HH owns radio Ln(expected farmgate maize price) Dependency ratio

0.134 0.073 0.005 0.453 0.013

0.093 0.160 0.836 0.000*** 0.370

0.066 0.498 0.209 0.268 0.069

0.699 0.031** 0.043** 0.305 0.315

0.380 0.161 0.225 1.215 0.028

0.336 0.560 0.057* 0.025** 0.732

1 = HH headed by single female 1 = HH headed by female with spouse 1 = HH had prime-age death in past 3 years

0.060 0.060 0.021

0.105 0.317 0.657

0.033 0.178 0.479

0.787 0.377 0.054*

0.240 0.009 0.380

0.192 0.977 0.241

Observations Psuedo R-squared; R-squared

3506 0.212

1496 0.548

3506

Model includes dummies for agroecological zone (7) and the year 2004 and 2007. Also included are long-term averages of each of the time-varying regressors. APE = average partial effect. * An APE significant at 90% level of confidence. ** An APE significant at 95% level of confidence. *** An APE significant at 99% level of confidence.

negative effect on their maize sales due to a market access measure such as ‘distance to road’ or ‘travel time to town’. To further assess the effect of market access on household maize sales, we interact our market access measures in each country – distance to the nearest road or town (Kenya, Zambia); travel time to the nearest town of 10,000 or more residents (Mozambique) – with our measures of household ownership of transportation assets such as bicycles, carts, or vehicles. When we interact household transportation assets with distance to road/town in Zambia, we find that ownership of these assets does not change the general finding from our original model, which is that there is a negative and significant yet very small effect of distance from a tarmac road on the probability of maize sale. For example, an additional kilometer of distance reduces the probability of sale by only 0.1% (Table 4). There are no significant interaction effects in Zambia between transportation assets and distance to road/town with respect to quantities sold. In the case of Kenya, none of the transport asset and distance interaction terms are significant. By contrast, in Mozambique, we find that for households that do not own a bike (cart), a reduction in the travel time to the nearest town of 1 h (about a 10% increase at the mean of travel time) would improve the probability of maize sale by 0.3% (0.2%), while travel time does not have a significant negative effect on households that own a bike or cart. Yet, the interaction of ownership of transportation assets and travel time is not significant in the case of quantities of maize sold. The fact that Mozambique is our only case country in which not owning a bicycle is associated with a negative effect on probability of maize sale is consistent with the fact that it has the lowest road density among our three case

countries and also the lowest percentage of households living in a village in which at least one household sold maize to a trader either on-farm or in the village itself (53%). Access to market price information Another factor related to market access is access to market price information, which would be expected to improve farmers’ bargaining power with intermediaries and thus improve both participation probability and sales quantities. We find that access to market price information is associated with higher maize market participation. For example, in Mozambique, we find that household receipt of market price information is associated with an 18% increase in the probability of sale as well as a 21% increase in the conditional quantity sold (i.e. by current sellers) and a 41% increase in the unconditional quantity sold (i.e. by any given household, whether a current seller or not). In Zambia and Kenya, we use radio ownership and cell/landline phone ownership as proxies of household access to market price information. While in each country we find that neither proxy has a significant association with probability of maize sale, in Zambia, radio ownership is associated with an increase in conditional sale quantity of 15%, while cell phone ownership is associated with an increase in conditional sale quantities of 81% and unconditional sales of 76%. In Kenya, radio ownership is associated with an increase in conditional and unconditional sales quantities of 23% and 25%, respectively. While more research is warranted to document the link between radio and cell phone ownership and household access to market price information, especially considering that we are not able to control for potential endogeneity of these factors due to unobserved time-varying

257

D. Mather et al. / Food Policy 43 (2013) 248–266 Table 4 Double-hurdle model of maize market sales participation and level of maize sold, Zambia, 2000–2008. Probit

Independent variables

Lognormal

Dept variable = 1 if HH sold maize, 0 otherwise APE of Xj on P(y > 0)

Probit and lognormal Dept variable = ln(kgs of maize sold)

APE (conditional) of Xj on ln y, given y>0

APE (unconditional) of Xj on lny

APE

p-Value

APE

p-Value

APE

p-Value

Ln(seasonal rainfall) Rainfall stress 1 = SEA soils suitable for low input fertilizer Ln(total landholding) Ln(total assets)

0.134 0.004 0.016 0.061 0.002

0.012** 0.645 0.314 0.000*** 0.652

0.510 0.177 0.160 0.169 0.010

0.024** 0.000*** 0.046** 0.000*** 0.659

1.072 0.195 0.230 0.502 0.015

0.002*** 0.003*** 0.066* 0.000*** 0.490

# Prime-age adults Head’s education level (years) Head’s age (years) 1 = HH used hybrid maize Ln(fertilizer per hectare applied to maize)

0.014 0.005 0.003 0.150 0.030

0.039** 0.087* 0.000*** 0.000*** 0.000***

0.005 0.009 0.005 0.496 0.087

0.858 0.493 0.053* 0.000*** 0.000***

0.057 0.029 0.019 1.529 0.213

0.180 0.185 0.000*** 0.000*** 0.000***

Residual from reduced form fertilizer regression % Village hhs which received credit 1 = HH owns animal or mechanized traction Distance from village-district capital, 2000 (km) Distance from village to feeder road, 2000 (km)

0.001 0.061 0.039 0.000 0.007

0.762 0.297 0.124 0.779 0.132

0.064 0.137 0.023 0.000 0.027

0.009*** 0.588 0.871 0.994 0.188

0.069 0.392 0.178 0.001 0.056

0.055* 0.310 0.234 0.885 0.120

Distance from village to main/tarred road, 2000 (km) 1 = HH owns bike 1 = HH owns cart 1 = HH owns radio 1 = HH owns cell phone

0.001 0.008 0.015 0.005 0.007

0.041** 0.654 0.628 0.749 0.773

0.000 0.053 0.267 0.146 0.594

0.973 0.511 0.138 0.051* 0.000***

0.003 0.087 0.193 0.173 0.563

0.213 0.451 0.428 0.137 0.007***

Ln(expected farmgate maize price) dependency ratio 1 = HH headed by single female 1 = HH headed by female with spouse 1 = HH had prime-age death in past 3 years

0.048 0.012 0.095 0.125 0.015

0.110 0.228 0.001*** 0.002*** 0.494

0.135 0.066 0.233 0.149 0.332

0.424 0.195 0.076* 0.443 0.017**

0.067 0.117 0.600 0.571 0.413

0.786 0.078* 0.000*** 0.002*** 0.031**

ln(district FRA purchases, prior season) Observations Psuedo R-squared; R-squared

0.002 7402 0.163

0.592

0.029 2552 0.435

0.180

0.038 7402

0.219

Model includes dummies for provinces and for the year 2008. Also included are time-average terms for each of the time-varying regressors. APE = average partial effect. An APE significant at 90% level of confidence. ** An APE significant at 95% level of confidence. *** An APE significant at 99% level of confidence. *

factors, these are at least two implications from these results. First, our results suggest that funding to increase the spatial coverage and frequency of radio broadcasts of agricultural market price information could significantly increase quantities of maize sold in rural Mozambique, while investments in rural cell phone coverage could lead to increases in quantities of maize sold in Zambia.

Weather-related shocks Given that nearly all maize production in these countries is rainfed, and that maize is a principal staple food crop in most areas, we would expect that agroecological potential and weather-related shocks would play an important role in the probability of maize sale and quantities sold. In each country, we find that indicators of rainfall or drought stress have significant and large effects on the probability of selling and/or amounts sold. For example, in Kenya, we find that a 20% increase in the percentage of 20-day periods during the main season with less than 40 mm rain leads a significant reduction of 8.8% in probability of maize sale and a 12.6% reduction in unconditional quantity sold (Table 3).10 In Zambia, a seasonal rainfall has a significant though 10 Because the range of the variable (% of village households who report yield shock) is from 0 to 1, a one-unit increase in this variable represents the entire range of the variable. Thus, a standard way to interpret the marginal change in a fractional variable is to multiply the standard deviation of the variable by the standard deviation of the variable by the partial effect, or by a value such as 0.20, which we use here.

relatively small effect on probability of maize sale, as a 1% increase in rainfall leads to a 0.5% increase quantity sold by current sellers (conditional quantity) and a 1.0% increase in quantities sold by the average household (sellers and non-sellers) (Table 4). In Mozambique, a 20% increase in the percentage of village households reporting maize yield loss results in a 6.5% decrease in a household’s probability of selling maize and a 10% reduction in unconditional quantity sold (Table 5). These results highlight the sensitivity of marketed maize surplus to weather shocks, and thus the potential value of: (a) investment in development and dissemination of drought-tolerant maize varieties; and (b) widespread promotion of smallholder access to low-cost methods of irrigation and/or conservation farming techniques to reduce the impact of drought – in contrast to the recent emphasis of heavy investment in formal perimeter irrigation schemes, which tend to benefit only a small proportion of the smallholder population.

Use of improved seed and fertilizer We note again that for both use of improved seed and fertilizer quantity, we are using the CRE approach to control for potential omitted variable bias due to time-constant unobservables, and have used the Control Function approach test and control for potential endogeneity due to time-varying unobservables. That said, our findings of significant partial effects on these explanatory variables on our dependent variables in Zambia and Kenya suggest but perhaps do not prove a causal effect of these inputs on household

258

D. Mather et al. / Food Policy 43 (2013) 248–266

Table 5 Double-hurdle model of maize market sales participation and level of maize sold, Mozambique, 2002–2005. Probit

Independent variables

Lognormal

Dept variable = 1 if HH sold maize, 0 otherwise APE of Xj on P(y > 0)

Probit and lognormal Dept variable = ln(kgs of maize sold)

APE (conditional) of Xj on lny, given y>0

APE (unconditional) of Xj on lny

APE

p-Value

APE

p-Value

APE

p-Value

# of Days of drought (district-level) % Village hhs which report maize yield loss Ln(total landholding) Ln(total assets) # Prime-age adults working in ag

0.000 0.066 0.039 0.010 0.007

0.809 0.203 0.018** 0.040** 0.544

0.006 0.394 0.314 0.025 0.078

0.200 0.020** 0.000*** 0.319 0.269

0.005 0.709 0.470 0.072 0.044

0.393 0.023** 0.000*** 0.058* 0.571

1 = HH owns animal traction % Village hhs received extension visit Head’s age Head’s education level Ln(distance to fertilizer seller)

0.108 0.054 0.000 0.001 0.012

0.008*** 0.362 0.890 0.871 0.044**

0.353 0.343 0.007 0.036 0.000

0.440 0.201 0.520 0.145 0.995

0.290 0.603 0.000 0.041 0.059

0.405 0.066* 0.988 0.253 0.272

Travel time to nearest 10k town (hours) 1 = HH owns bike 1 = HH owns cart 1 = HH received market price info 1 = HH belongs to farm association

0.002 0.023 0.130 0.035 0.001

0.254 0.266 0.198 0.101 0.982

0.002 0.083 0.101 0.211 0.018

0.711 0.308 0.844 0.028** 0.927

0.010 0.201 0.522 0.412 0.015

0.240 0.202 0.431 0.008*** 0.956

Ln(expected farmgate maize price) Dependency ratio 1 = HH Headed by single female 1 = HH headed by female with spouse 1 = HH had prime-age death in past 3 years

0.005 0.007 0.005 0.021 0.027

0.974 0.611 0.919 0.640 0.312

1.179 0.072 0.333 0.190 0.269

0.080* 0.298 0.098* 0.416 0.081*

1.205 0.106 0.358 0.275 0.368

0.235 0.214 0.195 0.288 0.035**

Observations Psuedo R-squared; R-squared

6352 0.111

1468 0.300

6352

Model includes dummies for agroecological zone (7) and for the year 2005. Also included are time-average terms for each of the time-varying regressors. APE = average partial effect. * An APE significant at 90% level of confidence. ** An APE significant at 95% level of confidence. *** An APE significant at 99% level of confidence.

maize sales. In Zambia, we find that household use of purchased hybrid maize seed increases probability of maize sale by 15% and conditional (unconditional) quantities sold by 64% (361%) (Table 4).11 In Kenya, household use of purchased hybrid maize seed increases conditional quantity sold by 17% and unconditional quantity sold by 30% (Table 3). We also find that hybrid seed is scale-neutral in both countries with respect to farm size, as in Zambia, it has a significant positive effect on probability of sale and/or quantities sold for households both lower and higher landholding quartiles (Table 6). While effects of hybrid maize are not significant for any specific landholding quartile in Kenya, the magnitudes of the effects demonstrate that hybrid seed use appears to be scale-neutral. This is consistent with the proportion of households using hybrids, ranging from 62% among those in the lowest landholding quartile to 72% among those in the highest. These results are not surprising given that hybrid maize seed is a highly divisible technology, which involves minimal fixed costs (i.e., transport costs to and from a seed dealer), and are thus accessible by a wide range of smallholders. While maize hybrids are typically developed for use with fertilizer and adequate water, the results from the rainfall variables in each country suggest that increased investment in the development and dissemination of drought-resistant maize varieties could improve maize production and marketed surpluses in lower potential zones.

11 Results for conditional and unconditional effects on the log of sales quantities report the actual change in the natural log of the dependent variable, not the percentage change in the first column (APE); note that the actual percentage change in the dependent variable needs to be adjusted since the logarithmic transformation approximates small changes well (those under 20%) but larger changes less well (Wooldridge 2002). The necessary adjustment is as follows: % change in y = [exp(B) 1] and is made for the APEs in the column denoted ‘Adjusted APE’.

In both Zambia and Kenya, we also find that marginal increases in fertilizer applied to maize have relatively small effects on the probability of sale and quantities sold by current sellers, though larger effects on unconditional quantities sold, as a 10% increase in fertilizer use increases unconditional quantity sold by any given household by 1.8% in Kenya and by 2.3% in Zambia (Table 7). In both Zambia and Kenya, we find that marginal increases in fertilizer applied to maize have relatively small effects on the probability of sale and quantities sold by current sellers, though larger effects on unconditional quantities sold, as a 10% increase in fertilizer use increases quantity sold by any given household by 1.8% in Kenya and 2.3% in Zambia. While these APEs at first appear relatively small in magnitude when compared with the rather large effects of hybrid use, the APEs of fertilizer use indicate considerably larger effects if we consider the difference between predicted unconditional maize quantity sold under observed fertilizer use levels and the predicted quantity sold assuming no fertilizer use (i.e. subtracting from the predicted value the APE of fertilizer quantity applied to maize times the household’s observed fertilizer quantity applied to maize). For example, holding other factors constant, Kenyan (Zambian) households using fertilizer at observed levels are 5% (5%) more likely to sell maize than if they were to use no fertilizer, and would sell 32% (16%) more maize – conditional among sellers – than if they were to not use any fertilizer, and would sell 34% (17%) more maize, unconditional on selling. If we restrict this comparison to households currently using fertilizer, Kenyan (Zambian) households using fertilizer at observed levels are 8% (18%) more likely to sell maize than if they were to use no fertilizer, and would sell 40% (37%) more maize – conditional among sellers – and 49% (61%) more maize, unconditional on selling. Like hybrid maize seed, we also find that fertilizer use is scale-neutral with respect to farm size as it has a significant

259

D. Mather et al. / Food Policy 43 (2013) 248–266 Table 6 APE of hybrid seed use on probability of maize sale and log quantity of maize sold (conditional), by landholding quartile, Zambia and Kenya. Kenya

Zambia

APE on probability of maize sale Landholding quartile 1-Low 2 3 4-High National

APE on probability of maize sale APE 0.043 0.046 0.041 0.019 0.031

p-Value

Landholding quartile

APE

p-Value

0.212 0.366 0.417 0.724 0.122

1-Low 2 3 4-High National

0.148 0.164 0.153 0.148 0.150

0.015 0.001*** 0.001*** 0.003*** 0.000***

p-Value

Landholding quartile

Adjusted APE

p-Value

0.465 0.174 0.264 0.560 0.080*

1-low 2 3 4-high National

1.203 0.280 0.819 0.512 0.643

0.093* 0.260 0.003*** 0.004*** 0.000***

APE on log quantity of maize sold (conditional) Landholding quartile 1-low 2 3 4-high National

Adjusted APE 0.263 0.335 0.325 0.127 0.175

APE on log quantity of maize sold (conditional)

APE = Average Partial Effect; APEs and SEs computed from subgroup regressions; Adj. APE = APE adjusted for logarithmic transformation of the dependent variable. ⁄⁄ An APE significant at 95% level of confidence, * An APE significant at 90% level of confidence. *** An APE significant at 99% level of confidence

Table 7 APE of log of fertilizer applied to maize on probability of maize sale and log quantity of maize sold (conditional and unconditional), by landholding quartile, Zambia and Kenya. Kenya

Zambia

APE on probability of maize sale

APE on probability of maize sale

Landholding quartile

APE

p-Value

Landholding quartile

APE

p-Value

1-Low 2 3 4-High National

0.038 0.008 0.003 0.018 0.016

0.017** 0.589 0.852 0.156 0.027**

1-Low 2 3 4-High National

0.022 0.034 0.041 0.021 0.030

0.002*** 0.000*** 0.000*** 0.009*** 0.000***

APE on log quantity of maize sold (conditional)

APE on log quantity of maize sold (conditional)

Landholding quartile

Adjusted APE

p-Value

Landholding quartile

Adjusted APE

p-Value

1-Low 2 3 4-High National

0.174 0.184 0.166 0.066 0.125

0.132 0.021** 0.003*** 0.130 0.000***

1-Low 2 3 4-High National

0.100 0.085 0.023 0.141 0.091

0.111 0.037** 0.578 0.000*** 0.000***

APE on log quantity of maize sold (unconditional)

APE on log quantity of maize sold (unconditional)

Landholding quartile

Adjusted APE

p-Value

Landholding quartile

Adjusted APE

p-Value

1-Low 2 3 4-High National

0.412 0.222 0.175 0.122 0.187

0.013** 0.024** 0.025** 0.061* 0.000***

1-Low 2 3 4-High National

0.276 0.271 0.203 0.215 0.237

0.004*** 0.000*** 0.000*** 0.000*** 0.000***

APE = Average Partial Effect; APEs and SEs computed from subgroup regressions; Adj. APE = APE adjusted for logarithmic transformation of the dependent variable. * An APE significant at 90% level of confidence. ** An APE significant at 95% level of confidence. *** An APE significant at 99% level of confidence.

positive effect on probability of sale and/or quantities sold for households in both low and high landholding quartiles (Table 7). That both hybrid maize seed and fertilizer are scale-neutral is not surprising as they are highly divisible inputs with low fixed costs. However, our three case countries present very different cases with respect to the current status of private sector fertilizer and seed markets. For example, in Mozambique, less than 4% of households use fertilizer on maize and fewer than 2% use improved maize varieties. At the other end of the spectrum is Kenya, where 71% of rural households use fertilizer on maize and 70% use hybrids. Zambia represents an intermediate case as fertilizer and hybrid seed is used by a sizeable minority of growers (37% and

41%, respectively). In Mozambique, moving beyond the current situation of a near absence of fertilizer and hybrid seed use is a large challenge which will require policymakers to address constraints to private sector development of seed and fertilizer markets, and effectively linking agro-dealer network development with improved extension services. Household land endowments We have three principal findings with respect to the relationship between landholding and maize market participation. First, we find that marginal increases in landholding have significant effects on the probability of maize sale in Mozambique and Zambia,

260

D. Mather et al. / Food Policy 43 (2013) 248–266

though the effects are relatively small. For example, controlling separately for a household’s long-term average landholding size, a 10% increase in landholding would improve probability of sale by 2% in Mozambique, by 1.9% in Zambia and by 1.17% in the medium potential zones of Kenya (Table 8). Second, marginal increases in landholding also have relatively small (though significant) effects on sale quantities by existing sellers (conditional effect), though considerably larger effects on sales of all growers (unconditional effect) in Mozambique and Zambia. For example, a 10% increase in landholding increases unconditional quantity sold by 6% in Mozambique and by 6.5% in Zambia. Third, to explore the question of whether there might be threshold effects across landholdings among smallholders in these countries, we also consider the APE of landholding on maize market

participation in each country by landholding quartile. Only in Mozambique does the APE of landholding on probability of maize sale or quantity sold appear to increase in land endowment (Table 9). For example, in Mozambique, the APE of landholding on probability of sale for the top two landholding quartiles are significantly larger than those for the lower two quartiles, though there is a significant difference in APEs of landholding by quartile on unconditional quantities sold (Table 9). The insignificant and small marginal effects found at the lowest initial landholding quartiles in Mozambique likely reflects the need for a minimum threshold of landholding to meet a critical proportion of the household’s maize consumption requirements, combined with the almost complete absence of fertilizer and improved maize seed (less than 4% of Mozambican households

Table 8 APE of landholding on probability of maize sale and sale quantities, Mozambique, Zambia, and Kenya. Average partial effect of 10% increase in landholding on Probability of maize sale

Quantity sold, conditional on selling (existing sellers)

Quantity sold, unconditional (sellers or non-sellers)

2.4 1.8 3.6

3.1 6.5 6.0

% Change Kenya Zambia Mozambique

1.1a 1.9 2.0

Notes: (a) effect significant only in the medium potential zones.

Table 9 APE of landholding on probability of maize sale and maize quantity sold (unconditional), by landholding quartile, Mozambique, Zambia, Kenya. Landholding quartile

Mozambique APE

APE on probability of maize sale 1-Low 2 3 4-High National

0.016 0.032 0.117 0.092 0.039

Zambia APE

p-Value

0.712 0.312 0.001*** 0.014** 0.018**

0.093 0.031 0.073 0.026 0.061

0.001*** 0.247 0.000*** 0.114 0.000***

0.053 0.079 0.030 0.005 0.022

p-Value

Adj. APE

p-Value

Adj. APE

p-Value

0.862 0.021** 0.011** 0.001*** 0.000***

1.028 0.951 0.628 0.257 0.652

0.006*** 0.001*** 0.000*** 0.002*** 0.000***

0.455 1.326 0.164 0.252 0.313

0.089* 0.005*** 0.328 0.117 0.002***

Mozambique Adj. APE APE on maize quantity sold (unconditional) 1-Low 0.088 2 1.071 3 1.322 4-High 1.160 National 0.601

Kenya

p-Value

Zambia

APE

p-Value 0.050** 0.042** 0.360 0.862 0.151

Kenya

APE = Average Partial Effect; APEs and SEs computed from subgroup regressions; Adj. APE = APE adjusted for logarithmic transformation of the dependent variable. * An APE significant at 90% level of confidence. ** An APE significant at 95% level of confidence. *** An APE significant at 99% level of confidence.

Table 10 APE of log of farmgate maize price on probability of maize sale, by agroecological zone, Mozambique, Zambia, Kenya. Agroecological zone

1-Low 2 Low-medium 3 Medium 4-High National

Kenya

Zambia

APE

p-Value

0.654

0.006***

0.641 0.630 0.453

0.003*** 0.033*** 0.000***

APE = Average Partial Effect; APEs and SEs computed from subgroup regression. An APE significant at 95% level of confidence. An APE significant at 90% level of confidence. *** An APE significant at 99% level of confidence.  *

APE 0.147 0.107 0.345 0.148 0.048

Mozambique p-Value 0.773 0.100* 0.456 0.178 0.110

APE 0.421 0.059 0.822 0.705 0.005

p-Value 0.091* 0.927 0.311 0.309 0.974

Notes: (1) All figures computed among the sample of households which planted maize in the given year; (2) dependency ratio defined as: (# of children age 0–14 + # adults age 60+ + chronically ill adults)/(health adults age 15–59).

93.5 85.6 85.7 54.2 41.5 47.6 65.5 32.7 1.8 Kenya, 2007 Male Single female Female with spouse

46.5 46.6 57.1

863 657 696

2.22 2.02 1.91

0.50 0.62 0.63

0.21 0.28 0.25

371.1 393.0 614.4

666 562 574

3.0 8.3 2.1

0.52 0.46 0.52

78.4 67.5 57.1

76.8 64.6 81.0

24.1 34.4 33.3

65.8 35.3 53.3 65.4 30.3 45.3 36.9 27.4 26.1 1.00 0.71 1.01 11.5 21.3 20.6 186 160 240 337.6 307.0 425.0 0.22 0.25 0.22 0.66 0.70 0.70 3.09 1.91 3.15 675 259 339 75.4 21.6 3.0 Zambia, 2008 Male Single female Female with spouse

39.9 30.4 28.8

4.8 1.9 5.7

29.9 19.5 22.3

6.9 7.1 10.4

62.6 26.5 52.1 41.3 6.3 29.3 30.0 44.8 36.8

Mean

1.2 1.3 1.5 4.0 11.3 6.7 168 119 101 97.9 78.7 85.3 0.19 0.26 0.19 0.57 0.69 0.56 2.38 1.63 2.14 67 17 57 25.7 14.9 23.5 74.2 15.8 10.0 Mozambique, 2005 Male Single female Female with spouse

Mean

2.0 0.4 1.1

HH owns bicycle % Maize production (kg/AE) Mean Maize area planted (ha/ AE) Mean Total landTotal landholding (ha) holding (ha/ AE) Mean Mean Maize quantity sold (kg/AE) Mean % of HH HHs1 sold maize % % Type of household head

Table 11 Characteristics of maize-producing households by type of household head in Mozambique, Zambia, and Kenya.

Total income per AE ($/AE)

Remittances as share of total income %

Dependency ratio2

HH used fertilizer on maize %

HH used hybrid maize seed %

HH in low potential zone %

HH owns radio %

D. Mather et al. / Food Policy 43 (2013) 248–266

261

use fertilizer on maize and less than 2% use improved maize varieties). Together, these two factors imply that Mozambican households must rely on extensification to increase their maize production and thus probability of maize sale. By contrast, the lack of threshold effects in Zambia and Kenya is likely due to the fact that fertilizer and hybrid use are much more common in these countries than in Mozambique, even for households in the lower landholding quartiles. Farm gate maize prices In Mozambique, while we do not find significant positive effects of higher expected maize prices on probability of sale at the national or zonal levels (Table 10), we do find that current sellers respond to higher expected maize prices by selling more maize (Table 5). However, we also find that households in the lowest potential region react to higher expected maize prices by reducing their probability of selling maize (Table 10). The negative price response in this zone may be due to the combination of a poor agroecological environment (i.e. low supply elasticity), the fact that maize constitutes a large portion of household income (i.e. high income elasticity) and the possibility that stock effects could turn the price response negative in this zone, especially if household preferences are especially strong to store food rather than rely on the market and a low substitution effect between food and other goods. In Zambia, we also find that there is considerably spatial heterogeneity in the responsiveness of household maize sales to changes in expected farmgate maize prices. For example, households in the Low-Medium and High potential zones behave as we would expect more commercially-oriented farmers to behave, as there is some evidence that they increase their probability of maize sale and/or quantities sold in response to higher expected maize prices (Table 10). By contrast, many households in the Low and Medium-3 potential regions appear to have little to no positive household responsiveness of maize market participation to higher expected maize prices In contrast to Mozambique and Zambia, the majority of smallholders in Kenya respond to higher expected maize prices by increasing the probability that they sell some maize (Table 10). The reason for this is likely related to the widespread use of fertilizer and hybrid maize by both smaller and larger smallholders in Kenya (as well as enjoying a larger percentage of arable land in areas of medium and higher-potential, relative to say, Mozambique), which enable smallholders to increase their maize yields in response to higher expected output prices. The heterogeneity found in the responsiveness of household maize sales to changes in expected farmgate maize prices in Zambia and Mozambique indicates that while improved infrastructure may elicit a positive sales response in some regions, non-price factors are vital for increasing maize production and sales in other regions. For example, in the lower potential zones of Mozambique, until productive assets such as landholding and animal traction are increased, and returns to existing assets are improved via adoption of inorganic fertilizer and improved maize varieties, it is questionable whether improved prices alone (through better infrastructure) will elicit a positive supply response from maize producers who currently do not sell maize (i.e. 80% of maize growers). Gender of household head Given the potentially large negative welfare effects of adult mortality on rural households in our case countries, we also look to see if households headed by a single female are less likely to sell maize. Our findings with respect to gender of the household head

262

D. Mather et al. / Food Policy 43 (2013) 248–266

are mixed. For example, in Kenya, we find that female-headed households are no less likely to sell maize than male-headed households or to sell lower quantities. By contrast, in Zambia, we find that households headed by either a single female or female with spouse are 9.5% (12.5%) less likely to sell maize relative to male-headed households, and have considerably lower unconditional quantity sold. In Mozambique, we find that households headed by a single female are just as likely to sell maize as maleheaded households, and sell about 33% less (among sellers). Because our regressions separately control for factors such as recent adult mortality shocks, total landholding, farm equipment value, market-related assets, and input use, this suggests that lower maize sales by households headed by a single female in Zambia and Mozambique are related to unobserved factors. To gain more insight into these gender-specific results, we consider bivariate statistics of household maize sales and other characteristics by gender of the household head (Table 11). There appear to be various reasons why households headed by a single female are less likely to sell maize (in Zambia) or to sell lower quantities of maize (Zambia and Mozambique). First, while these household cultivate a similar area to maize per AE relative to male-headed households (on average), they have lower maize production per AE on average, which may be related to lower input use, higher probability of being in the low potential zone (in Mozambique), or unobserved factors such as poorer quality soils, fields with shorter fallows (Goldstein and Udry, 2008), and/or less knowledge regarding improved crop management practices in maize production. Second, households headed by a single female have lower total household income per AE on average, which may mean that they have a lower tolerance for maize price risk in the lean season, especially considering that most households in Zambia and Mozambique are net maize buyers. Related to this is the fact that these households have a higher share of total income from remittances (on average) – the timing and magnitude of which might be quite unpredictable, which might increase their risk aversion related to selling own maize production when they expect to need to buy some maize in the lean season. Conclusions Previous research on household food grain sales behavior in developing countries has tended to focus on the role of market access and price-related factors to explain why many rural households do not sell staple crops such as maize. However, a key concern raised in recent literature is that low household asset endowments may constrain the ability of many smallholders to take advantage of public goods that reduce the cost of market access. In this paper, we use econometric analysis of nationallyrepresentative smallholder panel data sets from Kenya, Mozambique, and Zambia to examine the question of how to achieve increases in marketed surplus of maize, the most widely marketed cereal food staple of eastern and southern Africa. Our analysis leads to six main findings. First, while a combination of market liberalization and improved road infrastructure would likely to improve input and output prices facing smallholders, we present evidence that suggests that many smallholders in these countries already enjoy reasonable market access. For example, our econometric analysis of household maize sales finds that typical market access proxies (distance to physical infrastructure or towns) are not significant or of negligible magnitude in our case countries. These findings are also consistent with recent rapid appraisal work in each country which found that trader presence even in ‘remote’ villages has greatly improved compared with the situation a decade ago, perhaps a result of increased investments in road construction as well

as the recent proliferation of cell phones in rural areas. Thus, while there still may be significant transport costs to the nearest relevant market for farmers in ‘remote’ villages (which would cause traders to adjust their maize buying prices lower), even these farmers now face considerably lower search costs for price information and access to traders than they did a decade ago. Second, in Mozambique, we find that household receipt of market price information results in large increases in the probability of maize sale and sale quantities. In Kenya and Zambia, we use radio and cell phone ownership as a proxy for household access to market price information; we also find significant positive effects of such assets on quantities sold. These findings suggest that funding to increase the spatial coverage and frequency of radio broadcasts in these countries could potentially lead to large increases in both quantities of maize sold as well as the numbers of households selling maize. Third, in each country, we find that village or district-level measures of either rainfall or drought stress have significant and large effects on smallholders’ probability of selling maize and/or amounts sold. These results highlight the sensitivity of marketed maize surplus to weather shocks, and thus the potential value of investment in climate change adaptation measures. Such investments include the development and dissemination of drought-tolerant maize varieties, as well as widespread promotion of smallholder access to low-cost methods of irrigation and/or conservation farming techniques. Fourth, our results from Kenya and Zambia show that use of divisible improved technologies such as hybrid seed and chemical fertilizer can significantly increase the number of households selling maize as well as quantities sold. In addition, the positive and relatively large effects of these inputs on smallholder maize sales are also significant among farmers of various landholding sizes and from various agro-ecological zones. While the question of how best to increase smallholder access to such inputs is currently the focus of much debate, it is clear that improvements in access to input markets and extension to enable smallholders to deploy profitable technology packages are at least as important as access to output markets, especially in countries like Mozambique and Zambia where the majority of farmers have negligible amounts of surplus food staples to sell. That said, Tittonell and Giller (2013) argue that declining soil fertility in many areas of sub-Saharan Africa as well as lack of proper agronomic management (planting dates, spacing, cultivars, early weeding, etc.) may make it difficult for smallholders to reap the full benefits from the expected productivity effects of fertilizer use. The implication is that improving smallholder access to improved seed and fertilizer alone may not have the expected productivity payoffs without concurrent investments in extension and more region-specific development and targeting of recommendations regarding optimal fertilizer type and dosage. Fifth, we find that marginal increases in landholding in Mozambique and Zambia have significant and relatively large effects on the quantities of maize sold by both current sellers and all households (whether currently selling maize or not). Given that Zambia and Mozambique both contain large tracts of uncultivated land, there are clear opportunities in these countries to address the extremely low levels of landholding among the bottom half of the land distribution, though this will require investment in public goods, such as investments to eradicate disease constraints to animal traction use in Mozambique, and infrastructure investments in unsettled areas to promote migration in Zambia. In the short run, expanding access to improved seed and fertilizer is a powerful way to overcome smallholder land constraints, while expanded access to animal traction and/or re-settlement in more land abundant areas can further increase labor productivity and incomes in the medium to longer term.

263

D. Mather et al. / Food Policy 43 (2013) 248–266 Table A1 Summary statistics of variables in auxiliary & double-hurdle models, Kenya, 1997–2004–2007. 1996/97

2003/04

2006/07

Obs.

Mean

SE

Mean

SE

Mean

SE

Dependent variables Maize sale price (Ksh/kg) Ln(maize sale price) 1 = HH sold maize Quantity of maize sold (kg) Ln(quantity of maize sold)

1658 1658 3506 3506 3506

11.481 2.407 0.321 481.636 2.016

0.169 0.014 0.014 55.329 0.089

13.186 2.554 0.473 737.824 3.054

0.139 0.010 0.015 53.408 0.099

12.332 2.491 0.486 805.838 3.160

0.119 0.010 0.015 63.318 0.099

Independent variables 6-Year moving average seasonal rainfall 6-Year moving average seasonal drought shock Main season drought shocks Total landholding Ln(total landholding)

3506 3506 3506 3506 3506

558.977 0.318 0.234 2.155 0.162

5.840 0.006 0.007 0.077 0.034

573.765 0.284 0.237 2.303 0.347

4.314 0.006 0.007 0.087 0.029

512.744 0.338 0.294 2.152 0.287

5.422 0.007 0.006 0.078 0.028

Ln(total assets) # Prime-age adults 1 = HH owns irrigation equipment 1 = HH used hybrid maize Ln(fertilizer applied to maize) (kg/ha)

3506 3506 3506 3506 3506

10.878 3.456 0.126 0.683 2.267

0.064 0.054 0.010 0.014 0.057

10.268 3.107 0.110 0.634 3.225

0.065 0.051 0.009 0.014 0.066

10.590 3.113 0.112 0.723 3.428

0.045 0.055 0.009 0.013 0.064

Ln(village agricultural wage) % Village hhs which received credit 1 = HH owns animal or mechanized traction Distance from village to nearest extension Head’s age (years)

3506 3506 3506 3506 3506

4.876 0.401 0.118 5.036 50.321

0.008 0.009 0.009 0.078 0.388

4.587 0.346 0.081 4.731 56.520

0.010 0.010 0.008 0.082 0.387

4.400 0.526 0.110 4.398 58.814

0.009 0.008 0.009 0.082 0.384

Maximum adult education (years) Distance from village to motorable road (km) 1 = HH owns bike 1 = HH owns vehicle 1 = HH owns cart

3506 3506 3506 3506 3506

3.999 1.091 0.413 0.032 0.034

0.042 0.034 0.014 0.005 0.005

12.185 1.017 0.469 0.049 0.045

0.162 0.025 0.015 0.006 0.006

11.510 0.501 0.499 0.051 0.039

0.114 0.014 0.015 0.006 0.006

1 = HH owns radio Ln(expected farmgate maize price) Dependency ratio 1 = HH headed by single female 1 = HH headed by female with spouse

3506 3506 3506 3506 3506

0.758 5.282 0.837 0.121 0.003

0.013 0.007 0.024 0.010 0.001

0.886 3.463 0.566 0.190 0.014

0.009 0.004 0.018 0.011 0.003

0.907 2.500 0.499 0.328 0.018

0.008 0.003 0.016 0.014 0.004

1 = HH had prime-age death in past 3 years Ln(district median farmgate fertilizer price) Village median distance to fertilizer seller Village median distance to hybrid seed seller 1 = volcanic soils in village

3506 3506 3506 3506 3506

0.000 4.095 6.933 4.488 0.262

0.000 0.004 0.224 0.123 0.013

0.057 3.611 3.129 2.861 0.262

0.007 0.001 0.081 0.057 0.013

0.047 3.558 2.864 2.971 0.261

0.006 0.001 0.056 0.056 0.013

1 = humic soils in village 1 = Rankers soils with high sand in village 1 = sale quarter is Jan-Mar 1 = sale quarter is April–June 1 = sale quarter is July–September

3506 3506 3506 3506 3506

0.109 0.218 0.032 0.021 0.084

0.009 0.012 0.005 0.004 0.008

0.109 0.218 0.222 0.105 0.061

0.009 0.012 0.012 0.009 0.007

0.110 0.215 0.172 0.093 0.065

0.009 0.012 0.011 0.009 0.007

1 = sale quarter is October–December distance to regional wholesale market (km) 1 = HH buyer type: NCPB 1 = HH buyer type: processor/miller 1 = HH buyer type: other

3506 3506 3506 3506 3506

0.862 75.420 0.008 0.004 0.002

0.010 1.371 0.003 0.002 0.001

0.612 75.270 0.009 0.009 0.002

0.014 1.375 0.003 0.003 0.001

0.669 74.968 0.009 0.004 0.000

0.014 1.373 0.003 0.002 0.000

1 = HH buyer type: other household Ln(value of storage assets) Village-level effective NCPB purchase price at planting Ln(NCPB district-level purchases, last year) Regional wholesale price in planting month

3506 3506 3506 3506 3506

0.355 3.030 4.337 3.513 2.011

0.014 0.119 0.074 0.152 0.006

0.221 2.869 7.969 5.208 2.559

0.012 0.119 0.067 0.132 0.004

0.290 2.772 10.848 5.848 2.584

0.013 0.122 0.111 0.152 0.004

Regional Regional Regional Regional Regional

wholesale wholesale wholesale wholesale wholesale

price price price price price

in in in in in

planting planting planting planting planting

month, month, month, month, month,

t t t t t

1 (months) 2 3 4 5

3506 3506 3506 3506 3506

1.961 1.947 1.934 1.957 1.948

0.006 0.006 0.007 0.005 0.005

2.504 2.517 2.522 2.462 2.431

0.004 0.005 0.005 0.005 0.005

2.654 2.607 2.644 2.648 2.632

0.003 0.004 0.004 0.004 0.004

Regional Regional Regional Regional Regional

wholesale wholesale wholesale wholesale wholesale

price price price price price

in in in in in

planting planting planting planting planting

month, month, month, month, month,

t t t t t

6 7 8 9 10

3506 3506 3506 3506 3506

1.974 2.054 2.110 2.120 2.089

0.003 0.002 0.003 0.003 0.002

2.302 2.212 2.293 2.327 2.383

0.003 0.003 0.004 0.005 0.005

2.646 2.620 2.772 2.786 2.752

0.003 0.004 0.002 0.003 0.003

Regional wholesale price in planting month, t

11

3506

2.070

0.003

2.244

0.008

2.689

0.002

Sixth, while we find that the responsiveness of smallholder maize sales to changes in expected farmgate maize prices is significant and positive in most areas of Kenya, in higher potential

zones in Zambia, and among current sellers in Mozambique, we also find insignificant or negative household responsiveness to maize prices in lower potential zones of Zambia and Mozambique.

264

D. Mather et al. / Food Policy 43 (2013) 248–266

Table A2 Summary statistics of variables in auxiliary and double-hurdle models, Zambia, 2000–2008. 1999/00

2007/08

Obs.

Mean

SE

Mean

SE

Dependent variables Maize sale price (Kw/kg) Ln(maize sale price) Quantity of subsidized fertilizer received by HH (kg) 1 = HH sold maize Quantity of maize sold (kg) Ln(quantity of maize sold)

2586 2586 7402 7402 7402 7402

257.8 5.478 22.7 0.279 262.1 1.647

3.5 0.011 3.1 0.008 16.6 0.048

671.8 6.484 43.3 0.348 534.0 2.167

4.693 0.006 2.635 0.009 35.258 0.055

Independent variables Ln(6-year moving average seasonal rainfall) 6-Year moving average seasonal drought stress Ln(seasonal rainfall) Rainfall stress 1 = SEA soils suitable for low input fertilizer

7402 7402 7402 7402 7402

8.514 13.010 6.723 1.207 0.554

0.004 0.118 0.003 0.019 0.009

8.622 11.299 7.004 1.228 0.554

0.004 0.115 0.005 0.026 0.009

Total landholding (ha) Ln(total landholding) Ln(total assets) # Prime-age adults Head’s education level (years)

7402 7402 7402 7402 7402

2.812 0.621 9.497 2.60 5.15

0.048 0.017 0.104 0.025 0.065

2.778 0.539 9.894 2.89 5.08

0.080 0.017 0.109 0.030 0.065

Head’s age (years) 1 = HH used hybrid maize Fertilizer per hectare applied to maize Ln(fertilizer per hectare applied to maize) % Village hhs which received credit

7402 7402 7402 7402 7402

45.72 0.118 54.02 1.138 0.141

0.283 0.006 2.54 0.039 0.004

51.32 0.254 84.92 1.721 0.129

0.272 0.008 2.80 0.045 0.004

1 = HH owns animal or mechanized traction Distance from village-district capital, 2000 (km) Distance from village to feeder road, 2000 (km) Distance from village to main/tarred road, 2000 (km) 1 = HH owns bike

7402 7402 7402 7402 7402

0.123 34.9 3.4 25.6 0.433

0.006 0.416 0.059 0.670 0.009

0.160 34.9 3.4 25.6 0.568

0.006 0.416 0.059 0.670 0.009

1 = HH owns cart 1 = HH owns radio 1 = HH owns cell phone Ln(expected farmgate maize price) Dependency ratio

7402 7402 7402 7402 7402

0.064 0.343 0.000 15.882 1.097

0.004 0.009 0.000 0.011 0.015

0.089 0.580 0.206 6.490 0.892

0.005 0.009 0.007 0.002 0.014

1 = HH headed by single female 1 = HH headed by female with spouse 1 = HH had prime-age death in past 3 years Ln(district FRA purchases, prior season) Ln(district median farmgate maize sale price)

7402 7402 7402 7402 7402

0.181 0.044 0.105 0.000 6.507

0.007 0.004 0.005 0.000 0.004

0.225 0.028 0.099 7.347 6.519

0.008 0.003 0.005 0.057 0.002

1 = HH’s constituency won by MMD in last pres. election % Point gap in last election between MMD & runner-up Ln(district median farmgate fertilizer price) Distance from village to nearest fertilizer seller 1 = district reported no fertilizer sales

7402 7402 7402 7402 7402

0.929 50.302 7.669 20.159 0.153

0.004 0.361 0.002 0.612 0.006

0.626 40.529 7.552 18.610 0.022

0.009 0.439 0.005 0.467 0.002

District median farmgate maize sales price, last Regional wholesale price in planting month Regional wholesale price in planting month, t Regional wholesale price in planting month, t Regional wholesale price in planting month, t

season 1 (months) 2 3

7402 7402 7402 7402 7402

243.2 164.8 165.2 169.8 198.1

0.622 0.920 0.496 0.428 0.818

567.5 554.9 481.7 498.6 442.2

1.865 2.240 1.738 1.939 1.412

Regional Regional Regional Regional Regional

4 5 6 7 8

7402 7402 7402 7402 7402

236.6 267.0 336.0 399.7 379.2

1.163 0.837 0.883 1.114 1.023

447.4 462.4 713.1 937.3 1034.2

1.900 2.097 2.146 3.912 2.302

9 10 11

7402 7402 7402

401.5 417.4 328.1

1.096 1.149 0.692

972.9 971.4 931.7

2.130 2.382 2.433

wholesale wholesale wholesale wholesale wholesale

price price price price price

in in in in in

planting planting planting planting planting

month, month, month, month, month,

t t t t t

Regional wholesale price in planting month, t Regional wholesale price in planting month, t Regional wholesale price in planting month, t

The heterogeneity of maize price responsiveness in Zambia and Mozambique indicates that while improved infrastructure may elicit a positive sales response in some regions, policymakers aiming to increase marketed maize surplus from smallholders need to also consider non-price factors like the distribution and level of key production assets such as land, as well as factors which affect the return to those productive assets, such as technology use and agro-ecological potential (which affects the technology needs for a given region). In the case of Mozambique, until productive assets

such as landholding and animal traction use are increased, and returns to existing assets are improved via adoption of technologies such as fertilizer and improved seed, it is questionable whether improved prices alone (through improvements in infrastructure) will elicit a positive supply response from maize producers who currently do not sell maize (i.e., 80% of maize growers). Seventh, we find that households headed by a single female are less likely to sell maize (in Zambia) and/or sell lower quantities of maize (Zambia and Mozambique), and that this appears to be due

265

D. Mather et al. / Food Policy 43 (2013) 248–266 Table A3 Summary statistics of variables in auxiliary and double-hurdle models, Mozambique, 2002–2005. Obs.

2001/02

2004/05

Mean

SE

Mean

SE

1462 1462 6352 6352 6352

2.487 0.773 0.276 61.926 1.284

0.065 0.020 0.010 4.416 0.046

3.586 1.104 0.238 59.673 1.134

0.108 0.027 0.009 4.651 0.045

Independent variables # of Days of drought (district-level) % Village hhs which report maize yield loss Total landholding Ln(total landholding) Ln(total assets)

6352 6352 6352 6352 6352

32.737 0.628 2.099 0.475 6.168

0.738 0.006 0.038 0.016 0.052

48.088 0.820 2.238 0.572 6.137

0.625 0.004 0.032 0.014 0.055

# Prime-age adults working in ag 1 = HH owns animal traction % Village hhs received extension visit Head’s age Head’s education level

6352 6352 6352 6352 6352

2.407 0.026 0.150 42.917 2.221

0.026 0.003 0.003 0.313 0.049

2.496 0.035 0.172 45.414 1.985

0.029 0.003 0.004 0.312 0.054

ln(distance to fertilizer seller) Travel time to nearest 10k town (hours) 1 = HH owns bike 1 = HH owns cart 1 = HH received market price info

6352 6352 6352 6352 6352

3.565 7.602 0.271 0.012 0.229

0.029 0.131 0.009 0.002 0.009

3.566 7.601 0.346 0.016 0.266

0.029 0.131 0.010 0.002 0.009

1 = HH belongs to farm association Ln(expected farmgate maize price) dependency ratio 1 = HH headed by single female 1 = HH headed by female with spouse

6352 6352 6352 6352 6352

0.043 0.802 0.525 0.154 0.074

0.004 0.006 0.012 0.007 0.006

0.079 1.179 1.269 0.158 0.100

0.006 0.007 0.021 0.008 0.006

1 = HH had prime-age death in past 3 years Ln(distance to main district town) Ln(distance to public transport) Average regional wholesale price in October–December (planting) Average regional wholesale price in July–September (t 1 quarter)

6352 6352 6352 6352 6352

0.036 2.513 2.921 3.978 2.532

0.004 0.022 0.024 0.012 0.008

0.059 2.513 2.921 3.432 3.137

0.005 0.022 0.024 0.009 0.008

Average regional wholesale price April–June (t 2 quarter) Average regional wholesale price January–March (t 3 quarter)

6352 6352

1.514 1.729

0.008 0.007

3.071 4.416

0.012 0.013

Dependent variables Maize sale price (’000 meticais/kg) Ln(maize sale price) 1 = HH sold maize Quantity of maize sold (kg) Ln(quantity of maize sold)

to the fact they have somewhat lower maize production per AE. As our regressions control for factors such as productive assets, household labor, input use, and recent adult death, this suggests that the lower maize productivity and lower probability of maize sale among households headed by a single female may be due to unobserved factors such as poorer quality soils, fields with shorter fallows, less knowledge regarding improved crop management practices in maize production, and higher risk aversion due to their having lower total income per AE (on average) and a higher reliance on remittances. A key implication of the foregoing for Comprehensive African Agricultural Development Program (CAADP) investment strategies at country level is that there needs to be very effective spatial coordination between investments under the different ‘pillars’ of land area expansion, market access, and technology development to ensure that farmers have sufficient access to land and technology with which they can take advantage of investments in improved market access. While the private sector has a vital role to play in developing seed and fertilizer markets, there are strong public good aspects to both the development of technology packages which are adapted to varying agro-ecological conditions – especially for farmers in zones with poorer agroecological potential – as well as extension services to farmers, which address smallholder constraints related to both crop and livestock production and marketing. Yet, increased funding for such public goods depends upon governments successfully managing the challenge posed by political economy factors which have recently led many of them to funnel increased spending in the agricultural sector into subsidizing private goods (fertilizer) and grain parastatal activities

– which provide economic and thus political benefits in the shortterm – rather than investment in public goods such as agricultural research and development, extension and improved road infrastructure, whose benefits are only realized in the longer-term. Role of the funding source This research was supported by United States Agency for International Development (USAID) Bureau for Africa and the Bureau for Food Security through the Food Security III Cooperative Agreement. Acknowledgements This report would not be possible without the data collection efforts of colleagues at Tegemeo Institute (Kenya), Ministry of Agriculture and Rural Development (Mozambique) and the Ministry of Agriculture and Cooperatives and the Central Statistical Office (CSO Zambia). The authors acknowledge the invaluable contributions that Professor Jeffrey Wooldridge of the Department of Economics at MSU made to the econometrics methods employed in this paper. The authors also wish to acknowledge the time and information provided by the rural families who have participated in household surveys in each country, without which this research would not be possible, and Patricia Johannes for her editing and formatting assistance. Appendix A See Tables A1–A3.

266

D. Mather et al. / Food Policy 43 (2013) 248–266

Appendix B. Supplementary material Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.foodpol.2013.09. 008. References Alene, A.D., Manyong, V.M., Omanya, G., Mignouna, H.D., Bokanga, M., Odhiambo, G., 2007. Smallholder market participation under transactions costs: maize supply and fertilizer demand in Kenya. Food Policy 33, 318–328. Argwings-Kodhek, G., Jayne, T., Nyambane, G., Awuor, T., Yamano, T., 1998. How Can Micro-level Household Survey Data Make a Difference for Agricultural Policy Making? Nairobi: Egerton University/Tegemeo Institute of Agricultural Policy and Development. . Bardhan, K., 1970. Price and output response of marketed surplus of foodgrains: a cross-sectional study of some north Indian villages. American Journal of Agricultural Economics 52, 51–61. Barrett, C.B., 2008. Smallholder market participation: concepts and evidence from eastern and southern Africa. Food Policy 33 (4), 299–317. Boughton, D., Mather, D., Barrett, C.B., Benfica, R., Abdula, D., Tschirley, D., Cunguara, B., 2007. Market participation by rural households in a low-income country: an asset-based approach applied to Mozambique. Faith and Economics 50, 64–101. Burke, W.J., Jayne, T.S., 2008. Spatial Disadvantages or Spatial Poverty Traps: Household Evidence from Rural Kenya. International Development Working Paper No. 93. Michigan State University, East Lansing. Carter, M.R., Barrett, C.B., 2006. The economics of poverty traps and persistent poverty: an assets-based approach. Journal of Development Studies 42 (2), 178– 199. Chamberlain, G., 1984. Panel data. In: Grilliches, Z., Intriligator, M.D. (Eds.), Handbook of Econometrics, vol. 2. North Holland, Amsterdam, pp. 1247–1318. Chamberlain, J., Jayne, T.S., 2011. Unpacking the Meaning of Market Access. Staff Paper. Michigan State University, East Lansing. Cragg, J.G., 1971. Some statistical models for limited dependent variables with application to the demand for durable goods. Econometrica 39 (5), 829–844. de Janvry, A., Kumar, P., 1981. The transmission of cost inflation in agriculture with subsistence production: a case study in northern India. Indian Journal of Agricultural Economics 36, 1–14. de Janvry, A., Fafchamps, M., Sadoulet, E., 1991. Peasant household behaviour with missing markets: some paradoxes explained. The Economic Journal 101 (409), 1400–1417. Food Security Research Project (FSRP), 2012. Maize Suitability Study. Unpublished Document by Michigan State University’s FSRP Team in Lusaka Involving Collaboration with the Central Statistics Office (CSO) of Lusaka, and the Soils and Crops Research Branch (SCRB). Goetz, S.J., 1992. A selectivity model of household food marketing behaviour in SubSaharan Africa. American Journal of Agricultural Economics 74 (2), 444–452. Goldstein, M., Udry, C., 2008. The profits of power: land rights and agricultural investment in Ghana. Journal of Political Economy 116 (6), 981–1022. Key, N., Sadoulet, E., de Janvry, A., 2000. Transactions costs and agricultural household supply response. American Journal of Agricultural Economics 82, 245–259. Kirimi, L., Sitko, N., Jayne, T.S., Karin, F., Muyanga, Sheahan, M., Flock, J., Bor, G., 2011. A Farmgate-to-Consumer Value Chain Analysis of Kenya’s Maize

Marketing system. MSU International Development Working Paper No. 111. Department of Agricultural Economics, Michigan State University, East Lansing, Michigan. Mason, N., 2011a. Fertilizer Subsidies and Displacement of Farmers’ Commercial Fertilizer Purchases – The Case of Zambia, Revisited. Unpublished Ph.D. Dissertation Essay. Department of Agricultural, Food and Resource Economics, Michigan State University, East Lansing, MI. Mather, D., Donovan, C., 2007. The Impacts of Prime-age Adult Mortality on Rural Household Income, Assets, and Poverty in Mozambique. Directorate of Economics Research Paper 65E, Maputo, Mozambique. Mather, D., Boughton, D., Jayne, T.S., 2011. Smallholder heterogeneity and maize market participation in southern and eastern Africa: implications for investment strategies to increase marketed food staple supply. In: MSU International Development Working Paper No. 113. Department of Agricultural Economics, Michigan State University, East Lansing, Michigan. Mundlak, Y., 1978. On the pooling of time series and cross section data. Econometrica 46, 69–85. Nerlove, M., Fornari, I., 1998. Quasi-rational expectations, an alternative to fully rational expectations: an application to U.S. beef cattle supply. Journal of Econometrics 83 (1/2), 129–161. Renkow, M., 1990. Household inventories and marketed surplus in semi subsistence agriculture. American Journal of Agricultural Economics 72, 664–675. Renkow, M., Hallstrom, D.G., Karanja, D., 2004. Rural infrastructure, transactions costs and market participation in Kenya. Journal of Development Economics 85 (5), 1140–1146. Ricker-Gilbert, J., Jayne, T.S., Chirwa, E., 2011. Subsidies and crowding out: a doublehurdle model of fertilizer demand in Malawi. American Journal of Agricultural Economics 93 (1), 26–42. Rivers, D., Vuong, Q.H., 1988. Limited information estimators and exogeneity tests for simultaneous probit models. Journal of Econometrics 39, 347–366. Sadoulet, E., de Janvry, A., 1995. Quantitative Development Policy Analysis. The Johns Hopkins University Press. Scandizzo, P., Bruce, C., 1980. Methodologies for Measuring Agricultural Price Intervention Effects. World Bank Staff Working Paper No. 394. The World Bank, Washington, DC. Sheahan, M., 2011. MS Thesis. Analysis of Fertilizer Profitability and Use in Kenya. Department of Agricultural, Food and Resource Economics, Michigan State University, East Lansing, MI. Singh, I., Squire, L., Strauss, J., 1986. Agricultural Household Models. Johns Hopkins University Press, Baltimore. Stephens, E., Barrett, C., 2011. Incomplete credit markets and commodity marketing behaviour. Journal of Agricultural Economics 62 (1), 1–24. Strauss, J., 1984. Marketed surpluses of agricultural households in Sierra Leone. American Journal of Agricultural Economics 66, 321–331. Tittonell, P., Giller, K., 2013. When yield gaps are poverty traps: the paradigm of ecological intensification in African smallholder agriculture. Field Crops Research 143, 76–90. Tobin, J., 1958. Estimation of relationships for limited dependent variables. Econometrica 26, 24–36. Vella, F., 1993. A simple estimator for simultaneous models with censored endogenous regressors. International Economic Review 34 (2), 441–457. Vuong, Q.H., 1989. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica 57, 307–333. Wooldridge, J.M., 2002. Econometric Analysis of Cross Section and Panel Data. MIT Press.