Abstract
This study uses panel data from 1,142 Kenya smallholder households over four survey periods to examine the determinants of participation in land rental markets and to quantify the impact of renting land on households’ income and poverty status. Overall, the study finds that land rental markets in Kenya promote farm productivity and significantly raise the incomes of land-constrained farm households. However, these percentage increases in the incomes of renters are often not large in absolute terms, and hence participation in rental markets alone is not sufficient to meaningfully affect rural poverty rates. (JEL O12, Q15)
I. Introduction
Land is one of the most important productive assets of rural households in developing countries. How land is owned, used, and exchanged has far-reaching implications for productivity, equity, and overall economic growth. While the impacts of land tenure security on land investments, credit access, and agricultural productivity have been widely studied in the literature (Feder and Feeny 1991; Besley 1995; Alston, Libecap, and Schneider 1996; Brasselle, Gaspart, and Platteau 2002; Jacoby, Li, and Rozelle 2002; Do and Iyer 2008), there have been relatively few studies of the performance and impact of land rental markets (Pender and Fafchamps 2006).
Within the literature on land rental markets, previous studies have focused mainly on Asia. With a few notable exceptions (e.g., Holden, Otsuka, and Pace 2009), there is little evidence from Africa on the reasons why rural households would want to rent or lease land in the first place and the consequences of participating in land rental markets. Many areas of Sub-Saharan Africa are experiencing a closing of the land frontier, land constraints, declining farm sizes, reduced fallows, and increasing intensification as a result of rising population density in such areas (Woodhouse 2003; Holden, Otsuka, and Place 2009; Jayne, Chamberlin, and Muyanga 2012). In the meantime, land rental markets in Africa have recently been found to be more widespread than previously realized, especially in the more densely populated areas, though the types of exchange, contractual arrangements, and impacts on farm efficiency and equity remain largely unknown (Holden, Otsuka, and Place 2009). The rising importance of land rental markets in Sub-Saharan Africa, the largely unknown role that they play in influencing national policy goals, and the need for empirical evidence to guide African governments’ policy stance toward land rental markets are the main motivations for this study. In particular, we seek to understand whether land rental markets raise agricultural productivity, reduce the disparities in income across rural households, and contribute to poverty reduction.
To address these issues, we use panel data from 1,142 farm households in Kenya over four surveys covering a 10-year period. To our knowledge, this is the first study in Sub-Saharan Africa that uses a relatively long panel to quantify the impact of land rental market participation on households’ income and poverty status.
We find that rental markets in Kenya contribute both to efficiency and equity. Land rental markets transfer land from less efficient to more efficient producers and also improve access to land for households with relatively small farms. The results are highly consistent across a number of alternative estimation methods (ordered probit, simple probit, and tobit models, and panel fixed effects linear probability models that control for unobserved heterogeneity). The findings from the dynamic panel income models further suggest that the overall income gains to the smallest farms from renting land are quite remarkable. Renting-in land would lead to an increase in per capita total net income and per capita net crop income by 6.6% and 25.1%, respectively. However, these percentage income gains are not always large in absolute terms and, hence, appear to be insufficient to pull a substantial proportion of rural households out of poverty.
II. Land Access and Land Rental Markets in Kenya and Africa
Cross-country African estimates of the prevalence of land leasing are spotty, but a review by Holden, Otsuka, and Place (2009) suggests that land rental markets are most active in densely populated areas where land is relatively scarce and highly fragmented. Evidence from several districts of Kenya in the 1990s suggests that less than 10% of households rented-in land (Wangila 1999), but more recent evidence from 15 districts in 2004 indicates that 17.9% of households rented-in land (Yamano et al. 2009). In our survey drawn from 22 districts, the proportion of households renting-in land rose from 18% in 1997 to 20% in 2007, an expansion of roughly 1% per year. Fixed rental rates paid in cash are by far the most common form of informally arranged land rental contracts (Yamano et al. 2009). A secondary form of rental payment is partial payment in advance with the remainder being paid at harvest.
Theoretically, both sales and rental markets can potentially improve production efficiency by equilibrating land and nonland factor ratios across farms in the presence of imperfections in nonland factor markets (Deininger 2003). But in practice, the motives and outcomes of participating in sales markets and rental markets could be quite different. First, land purchases require a much greater up-front payment than renting land. The development of land rental markets is largely a response to the need for a more flexible range of compensation and cost sharing options (Hayami and Otsuka 1993). The practice of deferring full rent payment until after harvest is well suited to overcome poor households’ credit constraints that would otherwise preclude them from participating in either land sales or rental markets. Second, the purpose and time frame of land purchase and rent are usually different. Land sales markets are better suited for individuals who are unable to inherit land but otherwise planning to make a long-term commitment to farming, whereas rental markets are better suited for making minor or shortterm adjustments in factor ratios (Yamano et al. 2009). Third, buying and selling land is potentially riskier than renting and leasing land (Otsuka 2007). Farmers may not be able to repossess land after they have sold it, so the sale of land may have more prolonged effects on their livelihoods than the renting of land, which is generally more flexible and of a shorter contract duration. Past studies identified a number of worrisome aspects of land sale transactions, such as distress sales and the dispossession of land rights for the poor and women (Wilson 1972; Collier and Lal 1986). For these reasons, land sales markets are generally much less active than rental markets in Africa (Holden, Otsuka, and Place 2009).
Unlike most other countries in the region, where land sales markets have been heavily restricted, both land sales and rental market transactions are legal and active in Kenya (Holden, Otsuka, and Place 2009). The recent study by Yamano et al. (2009) concluded that both land sales and land rental markets improve equity and efficiency. Moreover, they found that the amount of land purchased by a household either had no effect on their decision to participate in land rental markets, or its effect was controlled for after taking into account the total amount of land owned. Based on the findings of Yamano et al., and given that participation in land sales and rental markets stem from different motivations, it would appear reasonable to study participation in land rental markets separately from land sales markets.
The government of Kenya’s National Land Policy (2007) takes a decidedly positive stance toward land leasing, stating that it has “the potential to provide access to land to those who are productive but own little or no land” and that government policy is to “encourage the development of land rental markets while protecting the rights of smallholders by providing better information about transactions to enhance their bargaining power (Government of Kenya 2007, paras. 162 and 163). Given the explicitly promotional position of the Kenyan government toward land rental markets and the fact that an increasing proportion of farmers are participating in land rental markets over time, it seems important to better understand the productivity and equity effects of land rental markets within the smallholder farming sector.
III. Review of Relevant Literature
In this section we review the developing country literature on land rental markets. In particular, we examine the traditional factor equalization hypothesis about land markets in the development process, and discuss conditions under which land market performance may deviate from the outcome of the traditional view. This discussion provides the foundation for our hypotheses and estimation strategy in the sections to follow.
It is well established in the theoretical literature that land endowment has no implication for efficiency and equity if all markets are functioning perfectly and the production function has constant returns to scale (Feder 1985; Bardhan and Urdy 1999). Land-labor ratios in this case would be the same across all farms, as would yields and output per labor unit. This result would hold even in the absence of land markets as long as other factor markets were functioning perfectly. But in practice, imperfections in agricultural labor and other factor markets are common in developing areas (de Janvry, Fafchamps, and Sadoulet 1991). Principal-agent problems cause farmers to use more family labor than they otherwise would if hired labor supervision were not a problem (Eswaran and Kotwol 1985; Binswanger and Rosenzweig 1986); hired labor is normally confined to easily monitored tasks such as land preparation and weeding (Hayami and Otsuka 1993). In the absence of labor and other factor markets, land rental markets enable farmers with insufficient land relative to their labor and/or other productive assets to gain additional land to absorb excess family labor and production capacity. This, in combination with the widely accepted inverse farm size-productivity relationship1 caused by market imperfections, implies that land rental markets promote both farm efficiency and equity.
However, these happy outcomes might not obtain if credit market imperfections and transaction costs are sufficiently problematic. If credit markets function poorly, farmers’ ability to access land through land markets would be directly correlated with their wealth and land endowment (Deininger and Jin 2008). Under certain conditions, land markets could even transfer land from land-constrained to land-abundant households. Indeed, land markets were found to exacerbate land concentration in a variety of settings, including Rwanda (Andre and Platteau 1998), Burkina Faso (Zimmerman and Carter 1999), and India (Kranton and Swamy 1999). However, this outcome is less likely for land rental markets than for land sales markets, because the former involves smaller and more flexible payments, including delaying partial payment until after harvest in some cases (Jin and Deininger 2009), and using share contracts instead of fixed rental contracts in other cases (Hayami and Otsuka 1993). In fact, studies from Ethiopia (Pender and Fafchamps 2006), Ghana (Migot-Adholla et al. 1994), Rwanda (Andre and Platteau 1998), and Malawi (Holden, Kaarhus, and Lunduka 2006) all found that rental markets helped to equalize land-labor ratios and enhance equity, even in places where sales markets had the opposite effect (Andre and Platteau 1998).
Transaction costs associated with land rental contracts have also long been recognized in the literature (Alston, Datta, and Nugent 1984; Otsuka and Hayami 1988). Transaction costs arise due to the need to monitor tenants’ behavior, including the renter’s management practices and potential mismanagement (see Otsuka and Hayami 1988 for a comprehensive review). Other types of transaction costs include the costs of negotiating, searching for partners, and enforcing contracts (Carter and Yao 2002; Holden, Otsuka, and Place 2009, ch. 2). These types of transaction costs are especially high in environments where land rights are not secure (Vranken and Swinnen 2006; Macours, de Janvry, and Sadoulet 2010) and where land transfer is restricted (Kimura et al. 2011; Deininger and Jin 2005). Cross-country empirical evidence shows that transaction costs reduced rental market participation (Deininger and Jin 2005) and prevented farmers from attaining optimal operational farm size (Skoufias 1995; Teklu and Lemi 2004; Tikabo, Holden, and Bergland 2008). However, there are exceptions. For example, Pender and Fafchamps (2006) found no impacts of transaction costs on market participation and efficiency in Ethiopia.
In spite of the emerging evidence, there remain quite entrenched perceptions that land rental markets may contribute to land concentration and increased poverty, and therefore that close government regulation of land rental markets is necessary. To our knowledge, there is no empirical evidence to date using long household panel data to rigorously assess the impacts of participation in land rental markets on farm household incomes and poverty. Evidence of impacts over time through relatively long term panel data can help inform and guide these policy debates.
IV. Hypotheses
Based on the literature review in the previous section, a farmer’s decision to participate in land rental markets and the ensuing implication on efficiency and equity are influenced by several factors. First, in an environment characterized by other factor market imperfections, land rental markets enable farmers who are endowed with less land relative to their labor and/or other productive assets to cultivate more land and generate higher production and income. As farming skills are not tradable, land markets also enable farmers with higher farming ability to cultivate more land and produce greater farm output than would otherwise be produced on rented land. In this context, land rental markets contribute both to productivity and equity. However, these favorable outcomes of land rental markets could be reduced or even reversed if credit market imperfections and transaction costs associated with land rental participation are sufficiently severe.
Hence, whether participation in land rental markets improves productivity and equity is eventually an empirical issue. To guide our empirical analysis, we develop the following main hypotheses:2
Hypothesis 1. Land rental markets enhance efficiency by transferring land from less productive to more productive households regardless of the presence of transaction costs.
Hypothesis 2. Land rental markets improve equity by transferring land from bigger farms to smaller farms.
Hypothesis 3. Land rental markets allow smaller and more productive farmers to cultivate more land and generate a greater value of output net of production costs and rental payments. Hence, we hypothesize that rental markets increase tenant households’ net crop and net total income. Relatedly, we hypothesize that rental markets reduce poverty for those who rented land.
V. Estimation Strategy
We first develop the estimation strategy to identify the determinants of rental market participation. An ordered probit model is developed to account for the fact that each farmer faces three alternative rental choices (i.e., renting-in, not renting, and renting-out). Several alternative estimation approaches were employed as robustness checks. These estimation results allow us to draw inferences about the effects of land rental markets on agricultural productivity and equity. We then develop a panel dynamic model to analyze the impact of land rental market participation on farmers’ crop income, total income, and poverty status.
Determinants of Rental Market Participation
We estimate a rental market participation model using an ordered probit model following Deininger, Jin, and Nagarajan (2008) and Jin and Deininger (2009). These previous studies provide detailed discussions on why the ordered probit is an appropriate approach. Specifically, the three rental participation regimes follow
[1]where A* is the optimal operational land size, 
is the household’s land endowment, r is the market rental rate, and Tin and Tout are transaction costs associated with renting-in and renting-out land. 
is the marginal value product of cultivating an extra unit of land evaluated at the level of autarkic land endowment, and εi ~ N(0,1) represents the part of the marginal value of product that the farmer observes but is not observed by the econometrician.
The intuition underlying the three conditions corresponding to the three rental participation regimes (equation [1]) are straightforward. The rent-in regime implies that a farmer chooses to rent-in land only if his marginal product of farming additional land (evaluated at the level of own endowment) is greater than the rental payment and the associated transaction costs. The rent-out regime (the bottom line of equation [1]) means that a farmer chooses to rent-out land if the marginal product of farming additional land is smaller than the net rental income to be made by rentingout land, again after accounting for transaction costs. And finally, the middle condition means that a farmer chooses to remain autarkic if the marginal product of cultivating additional land is smaller than the net rental payment he would pay as a tenant and greater than the net rental income he would receive as a landlord. Under this circumstance, not participating in land rental markets is in the best interest of the farmer.
Let 
, where X is a vector of household and village characteristics that determine the marginal product of cultivating an extra unit of land, β is a vector of parameters to be estimated, and y is an ordered indicator for rental participation, taking on the values {0, 1, 2}, respectively, for {rent-in, autarky, and rent-out}.
One difference of our strategy compared to previous studies (Deininger, Jin, and Nagarajan 2008; Jin and Deininger 2009) is that rather than specifying T as including specific village-level land policy or regulatory variables, we specify T as village-level dummies. While this approach does not allow us to identify the effects of T explicitly, which is not our objective, the use of village dummies arguably accounts for village-level differences in transaction costs more comprehensively than specific policy variables, which may not capture other spatial differences in transaction costs.
Let the upper and lower cutoff points be Cin=r + Tιn and Cout=r - Tout, respectively. The maximum likelihood function corresponding to equation [5] can be written as
[2]Equation [2] is a well-behaved maximum likelihood function and can be easily estimated (Wooldridge 2002).
The two key variables in X are the household’s farming ability and land endowment, the coefficients of which allow us to test hypotheses (1) and (2). Following Mundlak (1961) and more recently (Jin and Deininger 2009), the household-specific farming ability variable, αi, is obtained as the estimated fixed effects parameter in the household panel crop production function. Let household i in village j in year t have the following Cobb-Douglas crop production function (in logarithm form):
[3]where Qijt, Aijt, Lijt, and Kijt are, respectively, the value of agricultural output; total cultivated area, family labor, and value of agricultural assets; Xijt includes a vector of variable inputs (such as seed, fertilizer, hired labor) and a vector of household and village-level control variables (e.g., head’s age, education, gender of the head, rainfall, distance to extension services, distance to main road, etc.) that could potentially affect the household’s crop production. αi is the time-invariant household fixed effect that serves as our measure of farming ability, and Dt is a time dummy to control for systemwide productivity changes over time.
Equation [3] is estimated using household fixed effects. The estimated betas can then be used to recover the farming ability (αi). The variable captures the household-specific residual contribution to total crop output not explained by the variables in the model. Because an objective of this study is to understand whether land markets promote improvements in land productivity, particular interest revolves around the sign of the coefficient on αi in land rental participation models.
The coefficient on αi is likely to be a lower-bound estimate for several reasons. First, αi is likely to be correlated with unobserved timevarying household factors, investments in land quality being perhaps the most important. Prior studies consistently find that land being offered for rent tends to be of lower fertility than plots being cultivated by farmers owning their own land (Benin et al. 2006). Specific to Kenya, Yamano et al. (2009) found that rented land had a significantly lower soil fertility index than inherited or purchased land. For these reasons, any potential bias in the farmer ability variable αi for renting farmers is likely to be downward. This problem is likely to be partially addressed with the use of fixed effects models compared to ordinary least squares (OLS) models to the extent that land quality changes very little from one year to another for the majority of farmers. Moreover, to control for villagewide changes in farm productivity caused by, for example, investments in village infrastructure, we include the interaction terms of village and year dummy variables in the fixed effects production function model. Second, estimates of the fixed effects for a short panel, even though unbiased, are not consistent (e.g., see Wooldridge 2002, 272-74). To assess the robustness of the other results of interest to the inclusion/ exclusion of αi, we estimate all ordered probit and tobit models both with and without the farming ability variable.
Other variables that are expected to affect a household’s rental participation decision include household member composition, household agricultural assets, the education of the household head, a dummy for female-headed households, distance to markets, village population density, weather shocks, and a dummy for topology and soil type. Finally, agroecological zone and village dummies are included.
Because only the 2007 survey collected data on the amount of land rented-in and rented-out,3 we utilize this particular survey wave to estimate tobit models on the quantity of land rented-in and rented-out. Doing so allows us to check the robustness of results from the ordered probit model. Moreover, the estimated coefficient on land endowment in the tobit model indicates whether and to what degree land rental markets allow farmers to adjust their farm size to achieve the optimal cultivated area—the same testing strategy adopted for several other studies (Skoufias 1995; Yamano et al. 2009).
Impact of Renting Land on Household Income
Finally, we specify a dynamic panel model to quantify the impact of rental participation on household income and poverty status. The dynamic model has the following desired features: (1) it accounts for the fact that annual household income is likely to be persistent over time, that is, the level of current income is likely to be affected by income in prior years; (2) it is possible for the endogenous variables such as rental participation to be instrumented by the same variable in prior years; and (3) the potential estimation bias caused by many omitted time-invariant variables can be eliminated through the first-differencing process. Specifically, the dynamic income regression equation can be defined as
[4]where Yijt is alternately the household’s per capita total income, per capita agricultural income, per capita off-farm income, or poverty status, with the subscripts i, j, and t respectively denoting household, village, and time. Yijt - 1 is lagged income or poverty status (to capture the persistent effect of past income or poverty status), Rijt is a dummy for renting-in land, Xijt is a vector of household and village characteristics that are expected to affect household i’s income or poverty status; Vj and Dt are village and time dummies; λi is the time-invariant and unobserved fixed factor.
Estimating equation [4] using OLS will yield inconsistent estimates due to the fact that Yijt - 1, Rijt, and λi are included on the righthand side of the equation. We can eliminate λi by transforming equation [4] from level form to first-difference form as
[5]We also add village and time interaction terms 
to control for the impact of localized shocks that could affect household income over time. OLS estimation of equation [5] will still yield inconsistent estimates because both ΔYijt - 1 and ΔRijt are endogenous, that is, E[ΔYijt-1Δεijt]≠ 0 and E[ΔRijtΔεijt]≠0. To obtain consistent estimates, we estimate equation [5] through an instrumental variable (IV) approach. Anderson and Hsiao (1981) propose using Yjt-2 to instrument ΔYjt-1 in equation [5] under the assumption that the error term is not serially correlated. Yijt- 2 is a valid IV because it is expected that E[ΔYijt-1Yijt-2]≠0 and E[ΔYijtYijt-2]=0 under the assumption of no serial correlation. Other right-hand-side endogenous variables can be similarly instrumented, for example, by using Rij,t- 2 to instrument ΔRijt. The exogenous right-hand-side variables in first-differenced form, ΔXijt, can be their own instruments. Additional lagged dependent variables (e.g., Yij,t-3, Yij,t-4,...) can also be used as instruments for ΔYijt- 1 (Arellano and Bond 1991). This comes at the cost that the more distant lagged dependent variables (or other endogenous variables) and the change in dependent variable (or change in other endogenous variables) may be only weakly correlated or not correlated at all. Prior studies following this approach have adopted various instrumenting strategies using lags of different lengths, as instruments depend partially on the number of available panel waves.
The relatively long time gap between two consecutive surveys (typically three years) in our case is likely to make the correlation between the endogenous variables and their more distant lagged values weak. Because of this, we use Yijt- 2 and Yijt- 3 to instrument ΔYijt-1. And we use Rijt-1 rather than Rijt-2 to instrument ΔRijt. The choice of Rijt- 1 rather than Rijt- 2 is justified because land rental decisions are normally made at the beginning of the crop season, so Rijt is more likely to be affected by Yijt- 1 (or εijt- 1) but not Yijt (or εijt), which are not yet observed at the beginning of the crop season. In other words, we can reasonably assume E[Rijtεijt-1]≠0, but E[Rijtεijt]= 0. Inthis case, Rijt- 1 qualifies as a valid IV (Bond 2002). We use the Hansen’s J-statistics to check whether the instrumental variables as a group are exogenous (Caselli, Esquivel, and Lefort 1996). Other variables on the right-hand side of equation [5] are either treated as predetermined or exogenous.
VI. Data and Descriptive Evidence
This study uses four rounds of rural farm household surveys (1997, 2000, 2004, and 2007). The variables used in the analysis are defined in Table 1. These include sociodemographic household characteristics (such as the number of household members, adult equivalents, age and educational attainment of the household head, and whether the household incurred the death of an adult over the prior three years), indicators of income and asset wealth (such as landholding size; the value of nonland assets per capita; information on crop production, livestock, and off-farm activities), and variables describing the household participation in land rental markets.
Household Characteristics by Region (Means)
Household incomes were computed as the sum of gross crop income, income from animals, and income from off-farm income, minus input costs.4 Data was available for only a partial set of cash input costs. Crop income includes production that is sold as well as retained for home consumption. Selling prices were collected from each household respondent for all crops and animal products sold. Village-specific median selling prices were then computed to value and sum production across all crops. Buying prices might represent a more appropriate opportunity cost of production for commodities that are purchased instead of sold by households, but retail price data were not available for most goods. Village-level median prices were then used to value the quantities of production. In the case of animal products, data were collected on the sale of 24 different types of animal products, and the production of eggs, fish, and dairy products.
The asset variable in this study is computed as the summed value of 14 assets that were consistently collected in each of the four panel rounds, including ploughs, harrows, ox-carts, farming implements, irrigation equipment, boreholes, wells, motorized vehicles, bicycles, and the stock of animal assets. These assets are priced at their selling price as reported by respondents. In this case, respondent-specific prices were used because of the typically great variation in the quality and value within each asset category (e.g., cows can vary greatly in their weight, health, and milk producing ability). For the purpose of summing crop income across crops, we used villagelevel median prices because in most cases households did not sell any particular commodity and hence lacked a household-specific price for valuation purposes.
Survey and Sample Design
The panel household survey was designed and implemented under the Tegemeo Agricultural Monitoring and Policy Analysis Project (TAMPA), implemented by Egerton University/Tegemeo Institute, with support from Michigan State University. The sampling frame for the panel was prepared in consultation with the Kenya National Bureau of Statistics in 1997. Twenty-four districts were purposely chosen to represent the broad range of agroecological zones (AEZs) and agricultural production systems in Kenya. Next, all nonurban divisions in the selected districts were assigned to one or more AEZs based on agronomic information from secondary data. Third, proportional to population across AEZs, divisions were selected from each AEZ. Fourth, within each division, villages and households in that order were randomly selected. A total of 1,578 households were selected in the 24 districts within eight agriculturally oriented provinces of the country. Farms over 50 acres and two pastoral districts were excluded from the sample.
The initial survey was implemented in 1997. Subsequent panel surveys were conducted in 2000, 2004, and 2007. The initial number of observations in 1997 was 1,500. Our analysis focuses on the 1,142 panel households who were interviewed in all four survey rounds and for which data were available on the complete set of variables used in the analysis. The average attrition rate between two consecutive rounds (about three years between each round) is about 5%, which is roughly comparable to attrition rates of most other household panel surveys in developing countries.5
Attrition bias is a potential problem in panel estimation. If sample attrition occurs randomly, then we do not need to worry about selection biases caused by attrition, although efficiency will be lost because of a reduced sample size. But if sample attrition occurs systematically, then attrition may create selection bias. We estimate reinterview models to assess the degree to which sample attrition is a problem and use the inverse probabilities of being reinterviewed as weights to control for attrition in the subsequent analyses. We follow Wooldridge’s (2002, 587–90) two-step estimation procedure to control for attrition. In the first step, probit models are used to estimate the probability that observation i remains in the next survey round and all subsequent survey rounds. Regressors include household and community characteristics and survey team dummies from the subsequent panel round (see Appendix Table A1). For t = 2,..., T, let 
be the fitted probability for household i to remain in year t. Then a set of probability weights 
can be constructed as the product 
In the second step, the equations of main interest can be estimated using 
as the weights for household i and year t.
Most of the household and community variables are statistically insignificant between households who were reinterviewed and those who dropped out of the sample in subsequent panel periods, suggesting that attrition bias is relatively small (Appendix Table A1). The variables that are significantly different include the interview team dummies, household land and nonland assets, and age composition of household members. For example, for the 1997 and 2000 panels, compared to survey team 1 (base group), households that were interviewed by survey team 4 in the previous period are 5.7 percentage points less likely to remain in the next round of survey. Households with more dependent members (those <14 or >60 years of age) are more likely to be reinterviewed in the next round of survey.
Because very few variables are significantly different between households remaining in the sample and those dropping out over time, we expect the bias caused by attrition to be small. Consistent with our expectation, the results based on the two-step attrition correction estimation procedure are very close to the uncorrected estimates, and there are no changes in either the signs or significance levels of the main variables of interest. Nevertheless, the regression estimates reported in this study are corrected for possible attrition bias.
Household Characteristics and Rental Participation
Household characteristics and rental participation across regions are reported in Table 1. We divide the total sample into four main zones according to their agroecological conditions and agricultural productivity potential: Eastern and Western Lowlands, Western Transitional and Western Highlands, High Potential Maize, and Central Highlands (see Table 1).6 The lowlands zone is of relatively low agricultural potential, while the latter two zones are of relatively high potential. On average, household size and adult equivalents have both declined considerably (by almost one member) during the past 10 years, with household size in the Central Highlands Zone (4.2 adult equivalents) being much smaller than in the other zones (6.0 to 6.5). About 24% of households were headed by women, but this varied from 19% in the High Potential Maize Zone to 31% in the Eastern and Western Lowlands Zone. Roughly 60% of household heads completed primary school, while a quarter of heads completed secondary school. Educational attainment was higher in the high potential zones than in the lower potential zones. About 12% of households lost at least one adult member between 2004 and 2007; 5% of households suffered the death of their heads during the same period. Headcount poverty rates over the full sample declined from 57% in 1997 to 40% in 2007. This compares to poverty estimates of 46% in 2005 by the World Bank (World Bank 2005).
Interzonal differences in mean land endowment, assets, incomes, and poverty status are quite remarkable. The mean per capita land endowment is 0.81 acres, but this varies from 0.64 acres in Western Transitional and Western Highlands to 1.0 acre in the Lowlands. Mean per capita household incomes (productive assets) in the Central Highlands are more than double (triple) that in Western Transitional and Western Highlands. Poverty rates across regions are inversely related to mean incomes.
About 20% of households rented-in land and 12% of households rented-out land in 2007. Survey data frequently show that the proportion of households renting-in land is greater than that renting-out (Deininger and Jin 2008; Yamano et al. 2009), primarily because some households renting-out land are absentee landholders living outside the sample area. Moreover, an individual landlord may lease land to multiple renters. Rental market activity also varies considerably across regions, with the relatively densely populated Western Transitional and Western Highlands and the High Potential Maize Zones being much more active than the other zones both in terms of participation and size of land area (relative to land endowment) being transferred.
Evidence on Determinants of Rental Participation
Table 2 reports household characteristics, land and productive assets, and the composition of income for four rental participation groups: those who rented-out land, those who remained autarkic, those who rented-in land, and those who rented-in only during 2007 but not in any of the previous periods. First, we note that the initial per capita land endowment for households renting-in land is 0.56 acres, which is only half of the initial landholding size for those renting-out land (1.07 acres). Households renting-in land also tend to have more household members and adult equivalents than those renting-out land. Second, the derived farming ability coefficient is higher for the rent-in group than for the group not participating in the rental market or the group renting-out land, although the differences are not statistically significant. Female-headed households accounted for only 13% of the households renting-in land; they accounted for 26% of the households not participating in land rental markets. The possible transfer of land from female-headed households to maleheaded households would be consistent with the fact that female-headed households have higher land-labor ratios than male-headed households (1.49 vs. 1.25 acres per person).
Household Characteristics by Rental Participation Status, 2007
To control for the fact that some households renting-in land in 2007 may have already benefited from renting-in land during the previous periods, we report the characteristics in Column 4 of Table 2 for those who rented-in land in 2007 but not in any prior year. This group, which comprises 43% of those renting-in land in 2007, has the lowest level of both per capita income and per capita assets compared to the other groups, but these differences are not statistically significant. Gini coefficients of landholdings across the full sample drop from 0.55 to 0.53 after including rented land. When computed on the basis of per capita landholdings, the Gini coefficient declines from 0.60 to 0.57. Finally, neither the level of nor the change in the share of households below the national poverty line is significantly different across rental participation status, though a relatively higher share of households who rented-in land for the first time in 2007 ascended out of poverty compared to autarkic households ( -0.08 vs. -0.046). While these descriptive results provide some bivariate support for our hypotheses, they do not control for endogeneity or the effects of other factors affecting incomes and land rental decisions. The next section examines whether these bivariate relationships are borne out by econometric evidence.
VII. Econometric Results
In this section we report estimation results for three sets of regressions, namely, the Cobb-Douglas production function that is used to derive the farming ability variable, the ordered probit model on determinants of rental market participation, and the dynamic income model to quantify the impact of renting land on household income and poverty status.
The results are largely consistent with our expectations. The coefficients of the production function for the main inputs are all statistically significant at 10% or better, have the expected signs, and are of plausible magnitude. Results from the ordered probit model are consistent with our hypothesis that land rental markets enhance productivity and transfer land from bigger farms to smaller farms. Alternative probit and fixed effects panel linear probability models produce highly consistent findings. Lastly, the results from the income regression further support our expectation that participation in land rental markets by small and poor land holders is associated with significant income gains. However, poverty regression results indicate that these income gains are alone not sufficient to reduce poverty. Each of these findings is discussed in detail below.
Production Function
Following the earlier discussion in the section on estimation strategy, fixed effects panel models (Column 3) are used to recover the farming ability coefficient. For purposes of comparison, the results from pooled OLS estimation with village dummies (Column 1) and those from random effects estimation (Column 2) are also reported in Table 3. The crop production function R2 ranges from 0.67 for the pooled OLS estimation to 0.32 for the household fixed effects model. The results from the random effects model are extremely close to the pooled OLS regression, as expected. The coefficients for all the main factors of production have the expected sign and are statistically significant at the 10% level or higher. Doubling total cropped area leads to a 50% to 58% increase in total output, a finding that is consistent with the familiar inverse farm size-productivity relationship found by other studies. Compared to land, returns to adult equivalent labor endowment are relatively low, as is expected given the relatively densely populated conditions that characterize most of Kenya’s rural farming areas. A doubling of adult equivalent labor endowment leads to an increase in output of only 5.3% to 9.0%. A doubling of expenditure on seed leads to a 12% to 17% increase in total crop production. The productivity of a femaleheaded household is 7% to 8% lower than a male-headed household based on pooled OLS and random effects models, other factors constant. By contrast, the household fixed effects model indicates no significant gender difference, most likely because the headship status of a given household does not vary much over time. Finally, it is interesting to note that rainfall has an important and highly significant effect on production in all models. A change from the 25th to the 75th percentile of annual rainfall over the 1990-2008 period (roughly a 36% increase) is associated with a 13.5% to 15.3% increase in total production, indicating the sensitivity of rain-fed agriculture to weather variations.
Estimation of Cobb-Douglas Production Function
Rental Market Participation
As explained earlier, the ordered probit model is appropriate for jointly estimating a farmer’s decision to be in one of the three rental market participation regimes. To compare the sensitivity of our findings to alternative estimation procedures, we also report the results from probit and linear probability models based on the panel data.7 The main results, however, are highly consistent across different modeling strategies. The estimated results from the ordered probit model are reported in Table 4. To help interpret coefficients from the ordered probit model, Table 5 reports the change in probability with respect to the change of a few key variables that are shown to be important.
Determinants of Participation in Land Rental Market, Pooled Data (Ordered Probit Model)
Change in Probability of Falling into One of the Three Rental Regimes with Respect to Changes in Each of the Key Variables
The ordered probit results strongly support our Hypothesis 1. The positive and statistically significant coefficient on farming ability suggests that relatively productive farmers are more likely to rent-in land and are less likely to rent-out land. The probability that a household at the top 10% productivity ranking will rent-in land is 3.9 percentage points higher than a household of average productivity. On the other hand, a household of average farming productivity is 2.5% more likely to rent-out land or 1.4% more likely to remain in autarky than a household at the top 10% productivity ranking.
The second main finding is that the positive coefficient on adult equivalents and negative coefficient on acres owned are both statistically significant at 1% across all three model specifications. The findings are consistent with our Hypothesis 2, indicating that land rental markets transfer land from land-abundant and labor-constrained households to labor-abundant and land-constrained households. According to the marginal probabilities (Rows 2 and 3 of Table 5), doubling the size of the farm would lead to a 10.7 percentage point decline in the likelihood of renting land. By contrast, doubling the amount of adultequivalent labor in the household is associated with a 9.0 percentage point increase in the probability of renting land. A doubling of landholding size is also associated with a 7.1 and 3.6 percentage point rise in the probability of renting-out land or remaining autarkic.
Our results also show that after controlling for other effects, land rental markets transfer land to households with more agricultural assets and male heads. Doubling productive assets would increase an average household’s tendency to rent-in land by 5.2 percentage points. Female-headed households are 6.1 percentage points less likely to rent-in land than male-headed households. While statistically insignificant, the negative coefficient on the dummy variable for household heads having completed primary school across all the specifications at least weakly indicates that households with more education are more likely to rent-out land and take on nonfarm jobs, which is consistent with other findings in the literature.
To check the validity of our assumption of variable transaction cost and no credit constraints in our conceptual model, we include squared terms for landholding size and lagged household income in the equations shown in Column 3 of Table 4. The insignificant coefficient on lagged household income suggests that the decision to rent land is not strongly related to household income prior to renting. The positive coefficient on the squared term and negative coefficient on the linear term of land endowment is inconsistent with the fixed transaction cost argument, as it suggests that relatively small farms are more likely to rentin land (with a turning point of 5 acres). Our assumption of no credit constraints is further supported by the insignificant coefficient on the credit dummy in all regressions. The results on all other variables nonetheless are highly consistent with the other specifications.
The results from the tobit model (Table 6) are highly consistent with those from the ordered probit model. The positive (negative) and significant coefficient on labor and agricultural assets in combination with the negative (positive) coefficient on land endowment in the rent-in (rent-out) model suggest that rental markets contribute to the equalization of land-other factor ratios and improve efficiency. The productivity effect is further bolstered by the positive and significant coefficient on farming ability in the rent-in model and negative (though insignificant) coefficient on the rent-out model. As before, we estimate the models with and without the ability variable; both models produce highly consistent results. The magnitude of the coefficient on land endowment (-0.14 in the rent-in equation and 0.25 to 0.28 in the rent-out equation) indicates a very partial adjustment of landlabor ratios through rental market participation.8 The finding that the smallest farms retain relatively high labor-to-land ratios even after participation in land rental markets suggests the presence of transaction costs of participating in land rental markets.
Determinants of Area Rented-in and Rented-out (Tobit Model Results)
Finally, results from separate panel probit and panel linear probability models of renting-in land are reported in Appendix Table A2.9 Because farmer-specific ability is constant over time, this variable cannot be in the fixed effects model, but the results on the remaining variables, especially on labor and land endowment, female headship, and lagged income, are again similar to those of the pooled probit model. They reconfirm each of the main findings from the ordered probit model, namely, that rental markets transfer land from farmers with relatively low agricultural ability and agricultural assets, less labor, and more land to farmers with greater ability, more labor, and less land.
Impact of Rental Markets on Household Income
Table 7 reports the results of the dynamic income model (equation [5]) for per capita total income, per capita agricultural income, and per capita off-farm income. We estimate equation [5] in three ways. We first present the OLS results for each of the three measures of income (Columns 1, 4, and 7), then the instrumental variable generalized method of moments (IVGMM) results, treating as endogenous only the lagged dependent variable and the land rental dummy (Columns 2, 5, and 8). Finally, using IVGMM again, we treat the other variables as predetermined and instrumented by their own values at t-1 and t-2 (Columns 3, 6, and 9).
Effect of Renting Land on Household Per Capita Total Income and Net Crop Income (Dynamic Panel Regression)
Before discussing results, we note that the Hanson J overidentification test does not reject the null hypothesis of all instrumental variables being uncorrelated with the error term. The tests for no serial correlation suggest that original error terms (in levels) are serially uncorrelated in most of the equations, indicating that the moment conditions are valid (Loayza, Schmidt-Hebbel, and Servén 2000).
As discussed earlier, the coefficients from OLS-estimated results in Columns 1, 4, and 7 are biased. If our model specification is correct, we would expect the coefficient of the lagged dependent variable from our IVGMM regressions to be smaller than estimates from a biased OLS model (Loayza, Schmidt-Hebbel, and Servén 2000). Our results confirm this expectation, as the OLS coefficient on the lagged total income of 0.18 (0.19 for crop income and off-farm income) is much greater than the corresponding figures (0.07 and 0.09) based on IVGMM regressions. The positive and significant coefficient of the lagged dependent variables in all models indicates that there is strong persistence in total, crop, and off-farm income among rural Kenyan households, as is the case in much of Sub-Saharan Africa. Second, the coefficients on household endowments and demographic characteristics in the IVGMM estimations are also consistent with our expectations. Doubling total productive assets would lead to an 8 percentage point increase in per capita income, and a 6 to 7 percentage point increase in crop incomes, respectively. Similarly, a household moving from land autarky to owning 1.73 acre of land (the average amount of land being rented by those who rented-in) is associated with a 6% to 11% increase in total income per capita and a 9% to 15% increase in crop income per capita. It is interesting to note that while the household head’s education is not significantly associated with crop income, education is positively and significantly associated with total income, suggesting higher returns to education in nonagricultural jobs. Third, the highly significant and positive coefficient of our main variable of interest, the household decision to rent land, suggests that access to additional land through rental markets would significantly improve rural households’ welfare. Renting-in land would raise household per capita total income and crop income by 16% to 18% and 32% to 33%, respectively. Finally, off-farm income (Columns 7-9) is not affected by agricultural assets, household land endowments, and the household’s decision to rent land. While these results are not surprising, they suggest that having access to additional land to rent provides an opportunity to raise total net income, as the increased resources devoted to crop production appear not to compete against other nonfarm activities.
Based on the estimated coefficients, a simple simulation is conducted to determine how the decision to lease land affects net household income. We first identified all households renting land in 2007 and stratified them into five landholding size quintiles, as presented in Table 8. Column 1 reports mean land endowment for different landholding quintiles. Column 2 presents the mean gross crop revenue generated from rented fields based on household-specific quantities of land rented-in 2007. Column 3 shows the net crop revenue from renting land after deducting the costs of fertilizer, seed, land preparation, hired labor, family labor valued at hired wage rates, and the land rental rate. Comparison of the rental payment in relation to net crop revenue in Column 4 indicates that the renting households receive the lion’s share of the net revenue produced on rented fields. Tenant households generated 2.19 shillings of net crop revenue on rented land for every 1 shilling paid to the owner of the land. There is an inverse relationship between farm size of the tenant and the net crop revenue generated by tenants per shilling of rental payment. The 20% of smallholders with the smallest farms were able to produce 2.77 shillings in net revenue from rented land per shilling paid to the landlord, while the largest 20% of farms produced only 1.62 shillings in net revenue per shilling paid for rented land.
Simulation of Impacts of Renting Land on Tenant Households’ Net Crop Income and Total Income by Farm Size Category, Based on 2007 Data
Columns 5 and 6 show the percentage change in net crop income and total household income resulting from the decision to rent land. Across the full sample, renting the mean amount of land contributed an average of 25.1% to renting households’ crop income after deducting all production costs, and contributed 6.6% to total household net income.
However, among the 20% of households with the smallest farms, their decision to rent land raised their net crop income by 41.6% and raised their total household income by 11.4%. This substantial increase in net crop income for the smallest farms is because the amount of land they own (0.70 acres on average) is a major limitation on their farm income. Land-constrained households’ ability to rent land can double or triple the amount of land they cultivate and hence greatly improve their net crop income, at least in relative terms. This is especially the case since the smallest farmers appear to generate roughly 2.7 times the net revenue per unit of land rented than the rental payment for that land. The impact on total household income, while positive, is less dramatic.
Impact of Rental Markets on Poverty
Even though renting land appears to be profitable for Kenyan smallholders, it is not clear whether the additional income generated from the existence of land rental markets is sufficient to affect rural poverty rates. To address this issue, we run a series of dynamic poverty models as reported in Table 9. The dependent variable is the poverty status dummy, which equals one for households with per capita income above the national poverty line and zero otherwise.10 As in the dynamic income model, the poverty model is estimated in several ways to examine robustness. First, we present the IVGMM results, treating as endogenous only the lagged dependent variable and the dummy for renting land (Columns 1 and 3). Then, using IVGMM again, we treat the other variables as predetermined and instrumented by their own values at t - 1 and t- 2 (Columns 2 and 4). We also run two models (Columns 3 and 4) that interact the lagged dependent variable with the dummy for renting-in land, which allows us to assess whether the impact of renting land on the household’s poverty status depends on whether the household was poor to begin with.
Effect of Rental Markets on Poverty Reduction (Dynamic Panel Regression)
As in the dynamic income models, the test results for overidentification reported at the bottom of the table again suggest the validity of the IV strategy. The regression results of the poverty model are in general consistent with those from the dynamic income model. For example, the positive and Significant coefficient of the lagged dependent variable in all four columns tends to suggest Significant persistence effects of a rural household’s poverty status. As expected, the ownership of agricultural assets, land endowment, and livestock all contribute positively to the improvement of poverty status over time. Female headship, on the other hand, is negatively associated with the reduction of poverty.
The coefficient on the land rental dummy has the expected positive sign in all four models, but it is far from statistical significance at any meaningful level when the endogeneity of the right-hand-side variables is controlled for. The interaction terms for prior poverty status and land rental (Columns 3 and 4) is negative as expected, but not significant, and the coefficient and statistical significance of the land rental dummy increases only slightly as compared to Columns 1 and 2. These findings may appear inconsistent with the positive and significant impacts of land rental on household crop and total income; however, both results are quite plausible. The 6.6% mean increases in the net incomes of renters are often not large in absolute terms given the low levels of income for households below the poverty line. Hence participation in rental markets alone may not be sufficient to meaningfully affect rural poverty rates. In this regard, our Hypothesis 3, that rental markets can improve household incomes and reduce poverty rates, is only partially correct.
VIII. Conclusions
There is considerable controversy but a paucity of empirical evidence concerning the impact of land rental markets on the distribution of income within rural communities and on agricultural productivity in Africa. This paper examines the characteristics of smallholder farm households renting-in and renting-out land in Kenya, and the impacts of renting land on household incomes. Analysis is based on panel data on 1,142 farm households in 22 districts covering four waves over a 10-year period. To our knowledge, this is the first study that quantifies the impact of land rental market participation on farm incomes and poverty status using a relatively long panel. The paper also contributes to an understanding of the processes by which land rental markets affect resource reallocations and agricultural productivity within smallholder farm sectors.
The analysis highlights five main findings: First, rental markets contribute to agricultural productivity within the smallholder farming sector by transferring land from less efficient to more efficient households. In this way, land rental markets in Kenya appear to support national agricultural productivity objectives.
Second, we find little evidence to support the widespread concern that land markets may lead to land consolidation among the relatively rich and large landholders. In fact, land rental markets appear to promote the transfer of land from larger to smaller farms. Rentingin land is inversely proportional to farm size, while renting-out land is directly proportional to farm size. The Gini coefficient of landholding size per capita declines from 0.60 to 0.57 after accounting for the reallocation of land in the rental market.
Third, renters are able to garner for themselves roughly twice as much net revenue from crop production as the amount of rent payment they make to the landowner. The ratio of net revenue to the tenant/rental payment is inversely proportional to the farm size of the renter, being over 2.7/1 among smallholders with the smallest farms (prior to renting) and declining to 1.6/1 among the 20% of renters with the largest farms.
The fourth and potentially most important finding of the study is that participation in land rental markets is associated with a statistically significant contribution to farmers’ crop and overall incomes. After the rental payment and other production costs are accounted for, leasing in land was found to increase households’ crop and total income by an average of 25.1% and 6.65%, respectively, compared to not renting. The percentage improvement in net crop and net total household incomes are highest for households with the smallest farms (41.6% and 11.4%) compared to 15.9% and 3.8% for households in the highest landholding size quintile. These findings would seem to endorse and encourage the government of Kenya’s current efforts to promote the development of land rental markets. However, land rental markets alone are not sufficient to lift most of the rural poor out of poverty, at least under current rental market conditions. Findings from dynamic poverty models indicate positive but statistically insignificant effects on poverty reduction.
Finally, land rental markets in Kenya do not fully equilibrate labor-land ratios apparently due to transaction costs and/or other market imperfections. Labor-land ratios for renting households remain markedly higher than they are for nonrenting households. Mechanisms to strengthen land tenure, up-date/renew land registration, reduce land conflicts, and remove local restrictions on land rental may all improve the functioning of land rental markets and, therefore, farm households’ livelihood. Improving the functioning of rental markets in these dimensions is also likely to reduce the potential negative effects of land rental on long-term land management practices.11 Future research to identify the constraints on the functioning of land rental markets and strategies for leveraging the potential benefits of land rental markets for the rural poor could contribute meaningfully to the design of future poverty reduction strategies.
Acknowledgments
We thank two anonymous reviewers, Daniel Bromley, Edward Taylor, Gershon Feder, Michael Carter, Klaus Deininger, and seminar participants at the 2011 AAEA Annual Meeting in Pittsburg for valuable comments and suggestions.
APPENDIX
Probit Estimation of Marginal Probability of Remaining in the Next Round of Panel Survey
Determinants of Renting-In Land (Panel Evidence)
Footnotes
The authors are, respectively, assistant professor and professor, Department of Agricultural, Food, and Resource Economics, Michigan State University, Lansing.
↵1 The inverse farm-size and productivity relationship is well established in the literature, and labor market imperfections have been the leading explanation for such phenomena (Bardhan 1973; Berry and Cline 1979; Carter 1984; Gavian and Fafchamps 1996).
↵2 Deininger, Jin, and Nagarajan (2008) developed similar hypotheses (especially Hypotheses 1 and 2) from a conceptual household model.
↵3 Unfortunately, decisions to rent and/or lease land in previous survey years was collected as a dummy variable.
↵4 Nonmarketed goods such as forest products and fuel grown on-farm are not included in household income, though they may be important.
↵5 For example, the average attrition rate for a widely studied LSMS survey in Vietnam is approximately 11% between initial survey in 1992-1993 and the follow-up survey in 1997-1998 (Do and Iyer 2008). For the well-known KwaZulu-Natal Income Dynamics Study in South Africa, only 84% of the 1,393 households in the original 1993 sample were successfully reinterviewed in 1998 (Maluccio 2000).
↵6 Eastern and Western Lowlands include the districts of Kitui, Mwingi, Makueni, Machakos, Siaya, and Kisumu. Western Transitional and Western Highlands include Bungoma, Vihiga, Kisii, and Kakamega districts. The High Potential Maize Zone includes Trans-Nzoia, Eldoret, Nakuru, Bomet, Lugari, and Narok districts, while the Central Highlands is composed of Meru, Muranga, and Nyeri districts.
↵7 One limitation with the ordered probit model is that only 2007 data are used because the data on rent-out were not collected in early periods. To check whether the results from 2007 are consistent with other periods, we also estimate a panel probit and a panel linear probability model on farmers’ decision in leasing in land. We use the fixed effects estimator to estimate the panel linear probability model to account for unobserved heterogeneity.
↵8 If markets allow for full adjustment, the coefficient would have been - 1 for rent-in and 1 for rent-out, respectively (Skoufias 1995). Similar results were obtained by Yamano et al. (2009).
↵9 We are not able to do the same for those who rented-out land because data on renting-out are not available for the early rounds of survey.
↵10 We also tried the same set of regressions with the poverty status being defined on the basis of the international poverty line (one dollar per person per day). The results are very similar to those based on the domestic poverty line.
↵11 Due to the short nature of land lease contracts, the incentive of tenants to make long-term investments is likely to be limited (Holden, Otsuka, and Place 2009). However, the empirical evidence is not always clear. For example, Place and Hazell (1993) found that tenants do not make long-term natural resource investments. Nkonya et al. (2009) found that tenants in Uganda used short-term investment (such as slash and burn) rather than medium or long-term investment (manure). As a result, short-term production gains are at the cost of long-term soil depletion. On the other hand, Benin and Pender (2009) found that there is no significant difference in the probability of making a longterm investment (e.g., stone terraces) between land that was rented-in and owner-operated in Tigray, Ethiopia. They argued that rental markets are functioning better in Tigray than in other regions because farmers in the former feel sufficiently confident about their long-term ability to renew their lease contracts.
References
In this issue
Jump to section
Related Articles
Cited By...
- Adaptation to Natural Disasters through the Agricultural Land Rental Market: Evidence from Bangladesh
- Soil Investments on Rented versus Owned Plots: Evidence from a Matched Tenant-Landlord Sample in Malawi
- Transaction Costs, Land Rental Markets, and Their Impact on Youth Access to Agriculture in Tanzania