Welfare Impact of Wheat Farmers Participation in the Value Chain in Tanzania

The paper examines the link between value chain participation and welfare changes for wheat farmers in Tanzania. Specifically, the paper analyzes the wheat value chain from production to consumption, explores participation in the value chain, and examines the net effect of farmers’ participation in the value chain. A logistic model is used to explore the factors influencing farmers’ participation in the value chain and to estimate propensity scores to match the covariates for participants and nonparticipants. Applying the nearest neighbor and caliper radius matching algorithms found that only a few farmers are vertically (~17%) and horizontally (~39%) coordinated based on participation in contracts and associations, respectively. At the vertical coordination level, characteristics are significantly different for farmers with and without contracts in terms of land size, technical efficiency, allocative efficiency, output per acre, frequency of extension visits, frequency of village meetings attendance, and off-farm income. At the horizontal coordination level, farmers who join associations differ significantly from nonmembers in terms of level of education, frequency of village meetings attendance, output per acre, technical efficiency, and allocative efficiency. Vertical coordination participants receive a profit of 126 TSh/kg more for wheat than nonparticipants, with the difference significant at the 1% level. Horizontal coordination participants receive a profit of 46 TSh/kg more for wheat than nonparticipants, with the difference significant at the 5% level. The sensitivity analysis reveals that farmers’ participation in the value chain is generally insensitive to unobserved covariates. The findings suggest that establishing more contracts and stronger associations that specifically deal with wheat production has a positive impact on farmers’ welfare.


Introduction
The rapid growth in the consumption of wheat in Tanzania and more importantly the widening of the gap between consumption and domestic wheat production have become a major concern for the Tanzanian government. The situation is troubling given the additional burden that wheat imports place on the demands for the country's scarce foreign exchange and the fact that there are areas of the country agronomically suitable for the production of the crop that are underutilized. Of equal concern is the fact that past efforts by the government to spur domestic production have not generated the intended effect. Between 2005 and 2010, the Tanzanian government implemented the East African Community Common External Tariff (35% ad valorem) and expensive import procedures at the Dar es Salaam port in Tanzania with the hope that increasing the domestic wheat prices farmers received would boost domestic production.
Despite these efforts, growth in production failed dismally to keep pace with that of consumption. The failure of many of the small-scale to medium-scale farmers to participate in the value chain hinders farmers from accessing high-value wheat markets and obtaining the returns that would enable them to increase their productivity and profitability. For example, most small-scale farmers in Tanzania sell their crops at the farm gate to intermediaries (brokers), often at a low price [1]. Lack of strong linkages between farmers and postharvest actors in the wheat value chain marginalizes farmers' welfare gain because prices received from intermediaries are much lower and rarely cover the cost of production.
There are two types of linkages within the value chain literature: horizontal and vertical. Horizontal linkage for farmers refers to their membership and participation in farm associations. Farmers' participation in horizontal coordination has shown progressive outcomes through their collective actions as documented in the literature. Acting collectively enables farmers to reduce their transaction costs for accessing inputs and transporting outputs, ease their access to market information and extension services, and improve their bargaining power with postharvest actors [2] [3] [4] [5] [6]. Despite such apparent advantages to farmers, the findings in the literature are not clear-cut regarding the value of participation in associations to farmers' welfare due to differences in production locations and agro-economics [7].
Vertical linkage refers to various associations between farmers and postharvest actors that entail formal and informal contracts that secure market outlets for farmers' output and make it easier for smallholders to overcome constraints such as inadequate or limited access to improved inputs, modern technology, and credit [8]- [17]. Masakure and Henson [18] list the benefits of contracts in reducing market uncertainty, enhancing knowledge acquisition, and increasing farmers' income.
Despite the economic importance of associations and contracts, most small-scale farmers in Tanzania do not participate in the formal value chain. Rather, they operate within a framework/system that is characterized by weak or poor coor- actors. Moreover, few farmers belong to associations that could help them gain access to guaranteed markets and collective bargaining to influence the market [19].
This article makes two major contributions to the literature. First, a review of the literature indicates that there has been no prior formal comprehensive assessment of the value chain participation of wheat farmers in Tanzania based on horizontal and vertical coordination as key indicators for farmers' value chain participation. For example, SAGCOT [20] only maps the value chain to trace the flow of inputs, goods, and services from production point to the ultimate consumer, and a USAID [21] report on staple-food value chain analysis focuses mainly on production and consumption trends, constraints, and opportunities.
Second, this study is the first to demonstrate a plausible explanation for the Tanzanian farmers' lackluster response to what appears to be a market opportunity to satisfy domestic demand for wheat and wheat products. It also offers useful policy suggestions to address the situation.
In addition, not controlling endogeneity caused by unobservables could result in biased estimates, especially when the unobservables affects participation in the value chain [22] [23] [24] [25] [26]. These unobservable factors as pointed out by Barrett et al. [12] may include individual risk aversion behavior, social capital, and trust/distrust of associations and contracts. To control for unobservables, we conducted a sensitivity analysis on outcome results, given that the propensity score matching (PSM) can only solve observed factor bias, thus producing less biased results in assessing the impact of value chain participation on wheat farmers' welfare in Tanzania.
Following the introduction, section 2 provides an overview of the concept of value chain, as well as a discussion of the theoretical underpinnings of the analysis undertaken. In particular, we discuss the rationale for using the propensity scoring technique and the additional steps needed to improve the robustness of the technique. Section 3 lays out in detail the methodological framework, including the specifications used in the analysis and the source of the data used in the analysis. The results of the investigation are presented and discussed in section 4. Section 5 provides the concluding remarks.

Brief Overview of the Concept of Value Chain and Value Chain Development
The value chain concept carries various definitions based on the question the researcher wants to address. Based on Kaplinsky and Morris [27] and Donovan et al. [28], a value chain can be defined as an organized system of transforming products in various forms from production to consumption.
As the product moves along the value chain, it increases its value through transformation/processing, relocation, and distribution. In agriculture, food safety and food functionality also add value to the products through product differentiation. The incremental value of the resultant products can be identified by their price differences. Value Chain Development (VCD) is geared toward analyzing the value chain and addressing key weaknesses in a manner that contributes to the development or improvement in the value chain. Therefore, VCD is a positive or desirable change in a value chain to extend or improve productive operations and generate socioeconomic benefits toward poverty reduction, income and employment generation, economic growth, environmental performance, gender equity, and other development goals [29]. The value chain concept in agriculture involves linkages of actors and their agri-food products toward adding value for consumers. The features of value chain development include mapping, coordination, governance, upgrading, meeting consumer demand, and competitiveness. Products gain value as they move along the value chain to various actors, say from input suppliers (such as seed providers) to farmers, then to intermediaries such as processors, wholesalers, retailers, and ultimately to consumers. Therefore, there must be linkages between actors to facilitate the movements of these products. These links need to be effective so that the benefits of the value chain are distributed among the chain actors. The value chain is not sustainable if only one actor receives all the benefits.
Often, farmers receive the lowest share of the consumer dollar, which is attributed to several factors, including the risk of product damage, high product transport costs to urban markets, and weak linkages with actors farther up the value chain. To deal with these challenges, value chain actors need to be organized and have external support to participate effectively in the high-value markets, including better rural infrastructures, educational institutions, and research and extension services.
Many studies have shown the impact of value chain participation in various farm aspects, including farmers' welfare. For example, the study by Birthal et al. [30] on vertical coordination in high-value commodities found that contracts reduce transaction costs and improve market efficiency to benefit smallholders. Coordinated farmers were paid better prices and enjoyed the benefit of assured procurement of their products. Valkila et al. [31] employed the value chain approach to assess whether the Fair Trade system empowers traders. They found that despite the premium prices set by Fair Trade, farmers still received the lowest price share in the value chain. Warsanga [32] employed marketing margins to assess price variations among actors within the banana value chain in Tanzania and found that farmers received the lowest price share. Unlike these studies, this paper examines the impact on farmer's welfare by comparing identical groups of participant and nonparticipant farmers in the value chain. The findings are used to explain the lacklustre response of farmers to opportunities in the wheat market in Tanzania.

Theoretical Framework
A frequent problem encountered by social scientists in the quest to determine  [33] proposed using propensity score matching (PSM) to estimate the probability (propensity score) of participating in the value chain. The expected value of ATT is the difference between the expected value of "with" and "without" treatment. That is, the ATT parameter is the actual gain from participation in the program and can be compared with its cost to determine whether the program should proceed, assuming it has a positive impact [23]. Accordingly, this paper focuses on ATT by fixing counterfactual information from the untreated group using PSM [33]. The major assumption of the treatment effect for evaluation studies is that the treatment satisfies the exogeneity condition, referred to in the literature by several names. The Unconfoundedness assumption states that

Propensity Score Matching
1 D = if treatment was received, and 0 D = if no treatment was received. The expression above implies that with a given set of covariates K unaffected by treatment, the potential outcomes R 0 , R 1 would be independent of treatment assignment. Further, it implies that all covariates that might affect the treatment and the outcome simultaneously must be observed to reduce any biasness that could alter inference. The overlap assumption states that ( ) implying that participants and nonparticipants with the same K values both have a positive probability of being treated [23]. Assumptions 1 and 2 are strongly ignorable [33], where ATE and ATT can be defined for all values of K. In this paper, the logit model is chosen due to the presence of binary dependent variables [24].

Matching Algorithms
After the propensity scores have been obtained, the second necessary step is to choose matching algorithms. The most common matching algorithm techniques are Nearest Neighbor Matching (NN), Caliper and radius, stratification and interval, and Kernel and Local linear [24]. All these types techniques can be done with or without replacement. "With replacement" means an individual from the control group can be matched more than once, while "without replacement" means the individuals from the control group are matched once only with individuals in the treated group. Both "with" and "without" involve a tradeoff between bias and efficiency (variance). "With replacement" is useful for dispersed propensity score distribution between the control and treated groups. The choice will depend on the nature and availability of the data. It would make sense to use "with" when there are more observations for the treated group than for the control group, and to use "without" in the opposite situation. Of the various matching techniques, NN is the one most frequently used, often in combination with others. NN matches individuals with the closest propensity score from the control group to those in the treated group. Caliper and radius matching resolve the problem of NN when the closest neighbor is far away. Caliper imposes a common support condition, whereby observations that are out of radius are dropped [34]. One advantage of the caliper technique is that it uses all the individuals within the caliper range. When there are suitable matches within the range, extra-individuals can be used; otherwise fewer individuals are used. Thus, caliper shares the attractive feature that avoids the risk of unsuitable matches [24]. Once the matching algorithms and their combinations have been chosen, the next stage of the process is to check for overlap between participants and nonparticipants.

Overlap
Overlap (common support condition) ensures that only comparable observations are used in the matching algorithm before proceeding with the analysis  [26]. Several techniques to accomplish this can be found in the literature, including visual distribution of propensity scores before and after matching, minima and maxima comparison, and trimming [25]. Determining overlap involves identifying and retaining those individuals inside the region of a suitable match and discarding those outside the region. In other words, once the minimum and maximum propensity scores from both groups have been determined, individuals below the minimum or above the maximum of the control (untreated) group are discarded [23]. The next step requires revisiting the covariates and assessing the quality of matching. The process of doing so is discussed in the next subsection.

Testing the Matching Quality
Testing the matching quality involves checking all the covariates to determine if the balancing property is achieved from relevant variables of both the control and treated group. Specifically, the intent is to determine whether any systematic differences between the groups remain after the matching is completed (i.e., after conditioning on the propensity scores). The matching quality checks if ( ) where K are the covariates that are independent to treatment (D) after conditioning to their probability of participation: If there is still a dependency on K covariates, then it can be concluded that either the model is misspecified or lacks good matches between the groups [35].
That is, there should be no more new significant information about the treatment decision. In applying the test, various methods have been suggested, including standardized bias, t-test, joint significance and pseudo R 2 , and the stratification test [36] [37] [38]. This paper uses the t-test method for participants and nonparticipants for the reasons described below.
The t-test, which is used to test the means of covariates before and after matching, is the preferred test because it gives statistically significant results. After the matching quality has been checked and tested, the impact of participation is measured using the matched sample. The parameter value is the ATT (the average treatment effect on the treated). Because the PSM only reduces the observable bias, there is the need to conduct a sensitivity test for endogeneity or unobservable bias.

Sensitivity Analysis
The last step for this analysis is to check the sensitivity of confounders on our results. The treatment effect estimation is based on two major assumptions: unconfoundedness and overlap. The unconfoundedness assumption is a strong assumption that can lead to bias estimates if there are confounders that affect both participation and the outcome simultaneously [39]. This is because the estimators from matching will not be robust to the hidden bias. While the magnitude  Rosenbaum's technique of sensitivity analysis depends on the sensitivity parameter, γ, which determines the degree of departure from treatment or participation. Thus, two individuals, i and j, with the same covariates k differ in their odds of participation in the program by at most a factor, γ. In experimental studies, the randomization ensures that the γ value is always 1 to control for bias [42].
In the odds criteria, values of γ are normally generated and tested in the model to see whether the findings will change. The odds ratio in sensitivity analysis is used because it shows how great the differences in π would need to be to change our estimated results. If i π is the probability of participation for individual i, then the odds that individual i participates in the program is . The odds ratio is bounded by gamma (γ) such that The expression implies that there would be a hidden bias if two individuals with the same covariate values k have a different probability of participating in the program. That is, we would have hidden bias if for individuals i and j. The basic process for a sensitivity analysis has two steps. First, is the selection of values for γ. Second

Method
Participation of farmers in the value chain is associated with linkages among themselves (horizontal coordination) and between postharvest actors (vertical coordination). An experimental approach (treatment and control) is used to determine the extent to which farmers benefit from value chain participation and to identify a causal relationship between participation and an outcome or set of outcomes. PSM is used to evaluate the impact of participation in contracts (C) and associations (A) on wheat farmers' welfare. Based on Heckman et al. [33] and Dehejia and Wahba [34]), the propensity scores were estimated using logistic probability regression, the algorithms for matching were selected, common support conditions were checked for variables that influenced both vertical and horizontal coordination participation, the ATT was estimated, and a sensitivity analysis was conducted to check for any confounder effect.
The treatment groups for this study are the participants in contracts and associations, while the control groups are the nonparticipants in contracts and associations. The outcome (R) is the wheat net profit per kilogram (kg). The impact of participation in contracts and associations on household wheat profit (R) is estimated by taking the average difference for R across both the treatment and control groups after controlling for differences in participation due to observable variables (k). First, we use the logit model to estimate the probability of farmers' participation, assuming that the error term is logistically distributed [43]. The logit model for C and A are specified as where C represents contracts (dummy), and A represents membership in an association (dummy). C takes on a value of 1 if the farmer had a contract during the wheat sales, and zero otherwise. Likewise, A takes on a value of 1 if the farmer belonged to a wheat association, and a value of zero if not. Following the probability estimation, the nearest neighbor (NN) and caliper algorithms with varied radii are used, respectively, to match the control and treatment groups based on propensity scores. The matched sample is used to determine the average treatment effect on treated (ATT) group for net profit R. Explicitly, the treatment effect for individual i is written as where R i is the outcome of an i th individual with treatment, and R 0 is the outcome of the same i th individual without treatment. However, because R 0 is not observed the counterfactual profit is used. In this case, the expected treatment effect of participation or average treatment effect on treated (ATT) is the difference between the actual profit and the profit if the farmer did not participate in are the counterfactual net profits for contract participation and membership in a farmer's association, respectively.
A proxy is needed for the counterfactual. A ready candidate for the proxy is to use an outcome observed from the untreated group (or a subset of the group).
Comparing the average difference in the outcome of the treated vs the proxy as counterfactual, the estimated ATT is are the net profit proxies for contract and membership participation, respectively, as obtained from the matched control group. The difference between the true ATT and the estimated ATT is the estimation bias due to some farmers being selected (or self-selected) for the treated group and others for the untreated group such that proxies have to be used for counterfactual outcomes. This bias is referred to as "selection bias" in the econometric literature and is given by where i v and i u are biases given by unobserved pre-existing differences between the groups. Thus, the true parameter of ATT is only identified if the counterfactual net profit is similar to proxy net profit without considering unobserved biases. That is After the ATT is obtained, we can now further check for the unobserved effect.
Let the probability of participation in value chain for individual i be i π and for the matched individual j be j π . Assuming each individual i is exactly matched by individual j, their treatments odds are given by ( ) The odds ratio for the paired matched individuals is given by (20) where γ is the treatment odd ratio in Rosenbaum's sensitivity analysis and represents the probability ratio of participants to the matched nonparticipants of the value chain. From Assume the F function has a logistic distribution. Then odds ratio equation The individuals still differ in their odds of participation by a factor γ and their unobserved covariate w. If there are no differences in unobserved variables i j w w = or if unobserved variables have no influence on the probability of participating 0 γ = , the odds ratio is 1, implying the absence of hidden or unobserved selection bias. Following Rosenbaum [39] and Aakvik [40], the bounds for the odds ratio in equation 23 above is given by where i π represents individual i participating in the value chain, while j π represents individual j not participating in the value chain despite the similarity in the covariate value with individual i. Similarly, γ shows the difference in the odds of treatment and unobservable covariates between two individuals of the same covariate values.
In this case, e γ is the measure of the degree of departure from participation that is free of hidden bias. The package rbound in the r-program is used such that γ is the log odds of the differential assignment to treatment due to unob-

Data
Data were collected through a field survey in northern highland area of Tanzania where 90% of the total cultivated wheat is produced. Arusha and Kilimanjaro, which are two relatively homogenous regions in agricultural land use, production practices, and ecological condition, were chosen. Two districts from Arusha

Value Chain Structure
The wheat value chain in the study area consists of four main value chains: wheat input, wheat grain, wheat flour, and wheat product. We focus mainly on farmers' participation in the wheat grain value chain, which consists of producers, brokers, wholesalers, and retailers. Although landholdings in the study area range from 0.5 acres to more than 50 acres, the bulk of farmers represented in the survey are small-scale farmers with land averaging about 5 acres. Figure 1 depicts the wheat grain value chain. As can be seen in Figure 1,

Value Chain Coordination
Coordination along the chain is achieved by means of contracts and associations  As a result, many farmers have exchanged/reduced wheat cultivation for barley.

Mean Characteristics of Participants and Nonparticipants of Value Chain
The participants and nonparticipants of vertical coordination differ significantly  (Table 3). For the rest of the characteristics, the differences between participant and nonparticipant farmers were not significant.
This implies that relatively suitable matches would be available for vertical coordination participants and nonparticipants to analyze the impact of vertical coordination participation on farmer's welfare as measured by wheat profit per kg. Table 4 shows the characteristics of participants and nonparticipants of associations. Many variables are insignificant, which implies that the mean differences between participants and nonparticipants in associations are not associated with farm and farmer characteristics.

Factors Influencing Farmers' Participation in Value Chain
As mentioned earlier, a logit model was used to determine the factors that influence the farmer's participation in contracts and associations to generate the fitted values (estimated coefficients) that were used to create the propensity scores. Table 5    Horizontal coordination is associated with farmers' formation of groups for collective actions in agricultural activities. Table 6

Covariate Balancing Property
Overlapping of propensity scores between participants and nonparticipants is one of two basic assumptions in PSM. The estimates that lie between 0 and 1 are used to determine the common support region and to check whether this assumption has been met. Results for both vertical and horizontal coordination are provided in Table 7 and Table 8, respectively, to show the covariate balances of observables. Visual proofs of histogram distribution (Figures 2-10) are also provided to show the balances of the matched treated and control samples.
Checking the overlap assumption before further analysis is necessary to ensure that reliable estimates are presented. Table 7 shows that the matching for the control and treated groups is properly overlapped for the selected variables. That is, there is no significant difference between the means of the control and treated groups after matching. As indicated in Table 7, all of the "after matching" mean value differences between the treated and control groups are insignificant. distributions before and after matching for treated (contract) and control (noncontract) groups reveal that after matching, the shapes for the treated and control groups are similar and that there are no significant differences between the two groups, thus suggesting that we can further use the matched sample group to examine the effect of farmers' participation in vertical coordination on wheat profit per kg. Table 8 shows that the matching for the control and the treated groups under horizontal coordination were properly overlapped. That is, there is no significant difference between the means of the control and treated groups after matching.

Impact of Vertical and Horizontal Coordination on Wheat Farmers' Net Profit
The vertical coordination effect was measured using both the nearest neighbor and caliper radius matching algorithms. As seen in Figure 6, the nearest neighbor visual distribution did not result in the best matches, so its profit effect of 130 TSh/kg in Table 9 will still be biased. Caliper radius matching is a flexible form of an algorithm that checks matching at various radii. A caliper radius of 0.01 showed a profit effect of 136 TSh/kg and was significant at the 1% level.
Despite the visual diagram ( Figure 2) showing similar distribution between the matched treated and control groups, few treated farmers were used for the analysis. Possible increases in the samples were checked further by increasing the

Conclusions
The objective of this study was to assess the impact of value chain participation on wheat farmers' welfare in Tanzania. The postulation was that the slow response to wheat production could be due to the failure of wheat growers in Tanzania to formally participate in the value chain. Nonparticipation breaks the information flow about this market opportunity and restricts the potential contribution of this crop to the welfare of farmers. In exploring that broad objective,  welfare revealed that participants received 126 TSh/kg more for wheat than nonparticipants and the difference was significant at the 1% level. Horizontal coordination participants received 56 TSh/kg more for wheat more than nonparticipants and the difference was significant at the 5% level. The sensitivity analysis revealed that our statistical inference on farmers' participation in the value chain is generally insensitive to unobserved covariates. However, we cannot ignore the fact that horizontal coordination is somewhat more sensitive to hidden bias than is vertical coordination. In this connection, it should be noted that one limitation of the study lies in the fact that the propensity score methodology can only ensure balance in measured, not unmeasured, confounders.
The overall concern of this investigation revolves around the failure of Tanzanian farmers to respond to what appears to be a market opportunity to satisfy domestic demand for wheat and wheat product. Our study suggests that even in the face of relatively weak contracts, farmers who participated in contracts (vertically or horizontally) received added benefits compared to nonparticipants.
However, the number of farmers who formally participate in contracts is relatively small, implying that the level of support for the hypothesis posited that this could be one of the major factors responsible for the lackluster response to wheat market opportunity. An implication of our findings is that policy makers and other beneficiaries should take steps to encourage and nurture contracts through upfront investments to wheat farmers in order to facilitate production through binding contracts. Further, more emphasis is needed on offering farmers more extension services, better agricultural-related meetings, greater land size for wheat production, more off-farm work opportunities, and higher levels of technical efficiency. There should also be more emphasis on improving the efficiency of horizontal coordination to improve farmers' welfare.