The Case of Tanzania: Effectiveness of Management Training on Rice Framing and Farmer-to-Farmer Extension

Abstract

This chapter discusses the effectiveness of agricultural training and farmer-to-farmer extension (F2FE) on rice cultivation technologies and productivity in irrigated and rain-fed rice-growing areas. For this purpose, the chapter describes the empirical results of two studies conducted in Tanzania. The first study examines management training for the Modified System of Rice Intensification (MSRI) by assessing its impacts on technology adoption and rice farming productivity in a rain-fed area. The second study focuses on the effectiveness of F2FE training in an irrigated area. For both studies, we consistently find positive impacts of training on technology adoption and productivity of participating farmers. We also observe spillover effects from participants on non-participants in irrigated areas.

This chapter draws heavily on Nakano et al. (2018a, b) with the permission from John Wiley and Sons and Elsevier. We highly appreciate the kind cooperation from TANRICE project team members, including Mr. Motonori Tomitaka, Nobuhito Sekiya, Nobuaki Oizumi of JICA, and the staff of the JICA Tanzania and JICA Ogata Research Institute.

1 Introduction

Agricultural training has been considered a promising way to diffuse new technologies (Anderson and Feder 2004; Otsuka and Larson 2013; Takahashi et al. 2020). As discussed in Chap. 1, the training is particularly important for rice cultivation, which requires the adoption of improved management practices. Several studies have already found that agricultural training effectively increases technology adoption and the productivity of rice cultivation (deGraft-Johnson et al. 2014; Kijima et al. 2012). However, given that the impact of agricultural training may depend on agro-ecological conditions (Mgendi et al. 2021), it is important to accumulate evidence in different areas to ensure the external validity of the results of previous studies.

As discussed in Chap. 2, since it would be prohibitively expensive for extension workers to train all farmers, it is important to find appropriate methods for diffusing technologies taught to a small number of farmers who then pass them on to non-trained farmers. Considering this, growing attention has recently been paid to the effectiveness of farmer-to-farmer extension (F2FE). The empirical results on the effectiveness of F2FE, however, are still inconclusive. Several studies have shown that the knowledge taught to the lead farmers is adopted and disseminated to other farmers through social learning (Emerick and Dar 2021; Fafchamp et al. 2020; Lee et al. 2019; Morgan et al. 2020; Nakano et al. 2018b Takahashi et al. 2019; Yamada et al. 2015). On the other hand, Kondylis et al. (2017) showed that direct training to lead farmers enhanced the adoption of technologies by lead farmers, while it did not affect the adoption of the surrounding farmers.

The purpose of this chapter is to discuss the effectiveness of agricultural training and F2FE on technology adoption and productivity of rice cultivation. For that purpose, we show the results of two case studies conducted by the author in irrigated and rain- fed areas in Tanzania (Nakano et al. 2018a, b). The first study is on the impact of the Modified System of Rice Intensification (MSRI), provided by a private company called Kilombero Plantation Limited (KPL) in a rain-fed area (Nakano et al. 2018a). The SRI is a set of low-input rice cultivation technologies developed in the 1980s in Madagascar. It is said to produce higher paddy yields by adopting several agronomic practices without additional external inputs.¹ In our study site, the major recommended practices include the use of modern varieties (MVs) called SARO5 and chemical fertilizer, as well as improved agronomic practices. Since these recommended practices differ from the original SRI, which prescribes no MVs or chemical fertilizers, we call this set of recommended technologies the MSRI. This study mainly examined the impact of training on participants and found that training successfully increased the adoption of technologies and paddy yield in rain-fed areas.

The second study is on the effectiveness of TANRICE training, which is a regular F2F training conducted in an irrigated area by the Japan International Cooperation Agency (JICA) and the Ministry of Agriculture Training Institute (MATI) of Tanzania (Nakano et al. 2018b).² In TANRICE training, intensively trained farmers (designated “key farmers”) were responsible for inviting five additional farmers to training sessions at a demonstration plot in the village. The invited farmers were referred to as “intermediate farmers” and were expected to train other non-trained “ordinary farmers” later. We found that there are direct positive impacts of training on key farmers, and ordinary farmers caught up with key farmers in a few years in terms of technology adoption and paddy yield because of this F2F diffusion system.

In sum, we observed the positive training impact on technology adoption and productivity among participants in both irrigated and rain-fed areas. Furthermore, we observed spillover effects from key to intermediate and ordinary farmers in TANRICE training, which implies the effectiveness of the F2FE program in the irrigated area.

The remaining part of this chapter is organized as follows. Section 4.2 explains the first study on the training of MSRI in a rain-fed area; Sect. 4.3 discusses the effectiveness of TANRICE training in the irrigated area; Sect. 4.4 summarizes our main conclusions and offers some policy implications as well as suggestions for further research.

2 The Impact of Direct Training in Favorable Rain-Fed Areas

This section mostly relies on Nakano et al. (2018a).

2.1 Study Sites

The first study was conducted in Kilombero district, Morogoro region. While there is no irrigation infrastructure, the study site is in the Kilombero valley, where farmers enjoy plenty of rainfall and thus, can be classified as a favorable rain-fed area. A private company called Kilombero Plantation Limited (KPL) provided training on rice cultivation technologies to surrounding farmers. KPL operated a large-scale plantation of approximately 5,000 ha and a rice miller as their primary business. KPL offers extension services at the request of the Tanzanian government, which was responding to the complaints of neighboring farmers that a single large company cultivates such a huge area.

The recommended practices in our study site include (1) use of MV called SARO5, (2) chemical fertilizer use, (3) seed selection in salty water, (4) straight-row dibbling³ or transplanting, and (5) spacing of 25 cm by 25 cm or more. In addition, the use of a rotary weeder and dibbling or transplanting one to two seeds or seedlings per hole or hill was suggested. However, the adoption rates of the latter two technologies are generally low. As a result, only one key component of SRI technologies is adopted in our study site: wide spacing of 25 cm by 25 cm. Furthermore, the recommended practices are substantially different from the original SRI, requiring no MVs or chemical fertilizers. Thus, we call the recommended technologies in our study site MSRI technologies.

The SRI office, established as a section of KPL, is in charge of extension services to the local farmers. They trained 25 farmers in a village in 2010 and expanded their extension service to an additional 1,350 farmers in 2011, 2,850 farmers in 2012, and 2,250 farmers in 2013. The extension services provided by the SRI office are financially supported by the United States Agency for International Development (USAID) and operate in 10 surrounding villages. When they start the training program, officers call for a village meeting and ask those interested in training to form a group of 25 farmers. The criteria for the participants are that they must be residents of the villages, must be farmers, and must not have been trained by the SRI office before. A group of participants has to provide a quarter-acre piece of land called a demo plot. The extension officers, qualified agronomists hired by KPL and USAID, provide training on the demo plot during the cultivation season. During the training, each participant is provided with 26 kg of chemical fertilizer and 4 kg of seeds of SARO5, which are recommended amounts for a quarter acre. Each farmer is recommended to cultivate a quarter acre of his or her own land following the technology and management practices taught by the training program.

One year after receiving training, instead of free modern inputs, trainees are eligible to receive in-kind credit of chemical fertilizer and seed from NGOs associated with the SRI office. Farmers are obliged to repay part of the loan every two weeks during the cultivation season for five months, resulting in 10 installments. In addition, farmers need to sell six bags (approximately 600 kg) of paddy at the agreed price to KPL at the time of harvest so that KPL can repay its remaining balance to the lending NGOs. However, this credit service has not been popular among farmers. First, it is difficult for the farmers to repay the loan every two weeks during the cultivating season, as they generate most of their cash income at harvest time. Furthermore, there is sometimes disagreement over the selling price of the rice between farmers and KPL due to a fluctuation of the market price of paddy in the harvesting season. Only 11 households out of 25 eligible farmers received the loan from NGOs associated with KPL in 2013.

2.2 Data Collection

Data collection was carried out from February to March 2014 and covered the cultivation season from October 2012 to May 2013. To examine the impact of the training program, we selected three villages where training was held (henceforth referred to as training villages). We also covered two nearby villages where no training was held (we refer to these as non-training villages). Training villages and non-training villages were adjacent and in a similar agro-ecological condition. In each training village, we interviewed on average 37 training participants and 35 non-participants. In addition, we interviewed on average 35 farmers per village in non-training villages, generating a total sample size of 283 households.

We asked farmers to list all of their farming plots during the interviews. Among those listed, we selected two paddy plots for plot-level analysis. In our study sites, trainees differentiated between plots where they adopted MSRI technologies (called MSRI plots) and plots where they did not adopt these technologies (called non-MSRI plots). Presumably, because the technologies are newly introduced, farmers cultivate at most one plot using MSRI technologies. Thus, for sample farmers who have attended MSRI training, we automatically selected the MSRI plot and selected at most one more plot randomly where rice is grown using traditional cultivation methods.⁴ For sample farmers who have not attended MSRI training, we randomly selected up to two plots where rice is grown. Since farmers do not necessarily adopt all the MSRI technologies even on MSRI plots, we investigate the adoption rate of each component of MSRI technologies.

The number of sample plots for the cultivation season of 2013 was 406 cultivated by 283 households. After dropping households and plots with missing values in crucial variables, the total sample size became 398 plots of 281 households. Note that a significant number of farmers cultivate only one plot either by MSRI or traditional methods. We also collected recall data on paddy yield and the adoption of critical technologies from 2010 to 2013 to construct a panel data set before and after the training. Our panel data sample size was 398 plots for four years, generating a total sample size of 1351 plots. Note that our sample is unbalanced because some farmers do not grow rice in some years.⁵

Out of 110 training participants in our sample, no farmers were trained before 2011, 25 farmers were trained in the main season of 2012, and 85 farmers were trained in 2013. This implies that 25 farmers trained in 2012 and 85 trainees in 2013 received the free inputs for a quarter acre from KPL in 2012 and 2013, respectively, while 25 trainees in 2012 were eligible for the KPL credit program in 2013. Since trainees in 2012 and 2013 received different support (i.e., credit and free inputs) from KPL in 2013, we name the trainees from 2012 early trainees and those from 2013 late trainees. In the following analyses, we differentiate the training impacts on each group. Note that the impact of the MSRI training in 2013 for early trainees partially includes the impact of the credit service. Regardless of this, it is difficult to distinguish the effects of training and credit statistically.

2.3 Descriptive Analyses

Table 4.1 compares the adoption of modern inputs and improved practices—including straight-row dibbling or transplanting, wide spacing, and seed selection in salty water—separately by early trainees, late trainees, non-trainees in training villages, and farmers in non-training villages in 2013. The most important observation is that trainees, regardless of their training year, achieved an average yield as high as 4.7 tons/ha on their MSRI plots in 2012 and 2013. This yield is remarkably high compared to the average yield of 2.9 tons/ha on the trainees’ non-MSRI plots and 2.6 tons/ha for the non-trainees in the training village.

Table 4.1 Yield and technology adoption in the sampled rice plots in 2013 by Modified System of Rice Intensification (MSRI) training participation

Importantly, we do not observe a significant difference between yields before training (2010–2011) on the MSRI plots (2.6 tons per ha) and the non-MSRI plots of trainees (2.7 tons per ha). This suggests that farmers do not necessarily select plots of good quality to apply MSRI technologies. The high yield of trainees on MSRI plots may be attributed to the high adoption rate of new technologies on these plots. Of the five MSRI technologies,⁶ SRI trainees adopt 3.7 of them on average on their MSRI plots but only 0.3 on their non-MSRI plots. However, there is some variation in the adoption rate of each technology: 90.9% for MV, 78.2% for straight-row dibbling, 56.4% for wide spacing, and 71.8% for seed selection in salty water. Trainees apply much more chemical fertilizer (52.4 kg/ha) on their MSRI plots than on their non-MSRI plots (6.1 kg/ha). Note also that there is no significant difference in the performance between early and late trainees in 2012 and 2013 (Table 4.1, columns b-c). This suggests that the high yield and high rate of technology adoption are not due to the free inputs distributed only for the 2013 trainees in our survey year. In fact, 2012 trainees achieve as high a yield and technology adoption rate as the 2013 trainees without receiving the free inputs.

Lastly, the yield of the non-MSRI plots of trainees (2.9 tons/ha) is slightly higher than that of non-trainees (2.6 tons/ha). However, the adoption rate of technologies on the non-MSRI plots of trainees is not significantly higher than those of non-trainees in the training villages (columns d-g), except for slightly higher chemical fertilizer use and seed selection in salty water. Furthermore, non-trainees in the training village do not show higher yield or technology adoption rates than farmers in the non-training villages (columns g-h). These observations suggest that spillover effects from the MSRI plots to the non-MSRI plots of trainees and from trainees to non-trainees are limited, at least during our observation periods. Whether MSRI technologies further diffuse to non-MSRI plots of the trainees and all of the plots of non-trainees over a longer period is an important remaining issue.

2.4 Estimation Methods

To investigate the impact of MSRI training on the adoption of rice cultivation technologies and paddy yield, we estimate the difference-in-differences (DID) model by using recall panel data (Imbens and Wooldridge 2009). The dependent variables are paddy yield (tons/ha) and technology adoption using a dummy variable that takes 1 if a farmer adopts MVs, dibbling or transplanting in rows, a recommended spacing of 25 cm by 25 cm, or seed selection in salty water, and chemical fertilizer use (kg/ha). The base model is:

$$ \begin{aligned} {\text{Y}}_{{\text{ijt}}} & \, = \,\tau \left( {{\text{MSRI}}_{{\text{ij}}} *{\text{trainee}}_{\text{i}} *{\text{post-training}}_{{\text{it}}} } \right)\, + \,\delta \left( {{\text{trainee}}_{\text{i}} *{\text{post-training}}_{{\text{it}}} } \right) \\ & \quad + \,\eta {\text{trainee}}_{\text{i}} \, + \,\lambda {\text{MSRI}}_{{\text{ij}}} + \theta {\text{year}}_{\text{t}}\, + \,{\text{p}}_{{\text{ij}}} \, + \,{\text{u}}_{{\text{ijt}}} , \\ \end{aligned} $$

(4.1)

where Y_ijt is the outcome variable of individual i’s plot j at time t; MSRI_ij is the time-invariant dummy variable that takes 1 if the plot is an MSRI plot in any single year; trainee_i is a dummy variable that takes 1 if the cultivator of the plot is either an early or late trainee; post-training_it is a year_t 2012 dummy for early trainees and year 2013 dummy for early and late trainees; year is a year dummy; p_ij is a plot- and household-specific time-invariant characteristics, and u_ijt is an error term.

We estimate the DID model controlling for plot-level fixed effects. By doing so, we attempt to control for plot- and household-specific, time-invariant characteristics (p_ij) that may affect a farmer’s endogenous selection of the MSRI plot and innate household characteristics. Note that in the basic DID model, the terms of the time-invariant training status dummy for trainee_iand MSRI_ij are included. In our case, however, these terms are absorbed by p_ij, as we control for the plot fixed effects.

The most important independent variable in the model is the interaction term of the MSRI plot, trainee, and post-training dummies. This interaction term is intended to measure changes in the outcome values induced by the MSRI training. Note that the MSRI plot dummy is a time-invariant variable that takes 1 if the plot is an MSRI plot in any single year. We include post-training dummies (i.e., the year 2012 dummy for early trainees and the year 2013 dummy for both early and late trainees) since trainees have attended the training by the indicated years. As discussed earlier, farmers receive free input during the training period and are eligible for the credit program in subsequent seasons. Thus, to examine the differential impact of the training in 2012 and 2013, we constructed three interaction terms representing, respectively, (1) the interaction of the MSRI plot, the early trainees, and the year 2012 dummies; (2) the interaction of the MSRI plot, early trainee, and the year 2013 dummies; and (3) the interaction of the MSRI plot, the late trainee, and the year 2013 dummies. Since only three households discontinued their adoption of MSRI technologies after their initial adoption, we consider the coefficient τ to estimate the impact of the training on the MSRI plots relative to non-MSRI plots of the trainees.

We also include interaction terms for the trainee and the post-training dummies. The coefficient δ was expected to capture the impact of the training on productivity and technology adoption in the non-MSRI plots of trainees due to their labor reallocation from non-MSRI plots to MSRI plots or the positive spillover effect of the training on the non-MSRI plots of the trainees. Again, to estimate the impact of training on the trainees in 2012 and 2013 separately, we include the interaction term of the 2012 trainee and year 2012 dummies, that of the 2012 trainee and year 2013 dummies, and that of the 2013 trainee and year 2013 dummies. We also include year dummies to capture the effects of general trends, including non-trainees in both training and non-training villages.

2.5 Estimation Results

Table 4.2 presents DID estimation results for the training impact on paddy yield and technology adoption with plot fixed effects. All three interaction terms of the MSRI plot, trainee, and post-training dummies have positive and significant coefficients in all the regressions except for the chemical fertilizer use of the 2012 trainees in 2012. This suggests significant effects of MSRI training on paddy yield and improved technology adoption. Compared to the non-MSRI plots of trainees, adoption rates are higher on the MSRI plots to the order of 50 to 90 percentage point, even after we control for plot fixed effects. Trainees also increased their chemical fertilizer application by 20–27 kg/ha on their MSRI plots in 2013.

Table 4.2 Difference-in-differences estimates of impact of Modified System of Rice Intensification (MSRI) training on yield and technology adoption in 2010–2013 (Plot fixed effect—Unbalanced panel)

As a result of the high adoption rate of the new technologies, trainees’ paddy yield increases by 1.3–1.4 tons/ha on their MSRI plots. This result strongly indicates that rice cultivation training has a significant yield-enhancing effect even under rain-fed conditions. Note that trainees in 2012 received free inputs in 2012 and trainees in 2013 received the same in 2013, while trainees in 2012 were eligible for the credit program instead of receiving free inputs in 2013. However, for paddy yield, the difference between the estimated coefficient of the interaction term for the MSRI plot, the year 2012 trainee, and year 2012 dummies (indicated as a) and that for the MSRI plot, the year 2012 trainee, and year 2013 dummies (indicated as b) is not statistically significant, as the F-test statistics shown in Table 4.2 indicate. This result implies that the training was effective even after KPL stopped providing free inputs to the trainees.

The interaction terms of the trainee and the post-training dummies, which capture the impact of the training on productivity and technology adoption on trainees’ non-MSRI plots compared to non-trainee plots, have negative coefficients for some technologies such as recommended spacing and positive coefficients for chemical fertilizer use. This suggests that the adoption rates of labor-intensive technologies such as wide spacing may be lower due to the increased labor requirement on MSRI plots, while there is some positive spillover effect on the adoption of chemical fertilizer. However, the coefficients are insignificant for paddy yield, suggesting no significant training impact on the productivity of non-MSRI plots of the trainees relative to non-trainees’ plots.

3 The Impact of Farmer-to-Farmer Training in Irrigated Areas

This section largely depends on Nakano et al. (2018b).

3.1 Study Site and Data

Having established the positive impact of agricultural training on rice productivity, this section moves on to the question of whether F2FE can complement direct training by extension workers. It specifically relies on the case of agricultural training on rice production technologies conducted by JICA in the Ilonga irrigation scheme in the Kilosa district, Morogoro region of Tanzania before and during the main crop season from November 2008 to May 2009 (hereafter, this particular crop season will be referred to as the 2009 crop season). The irrigation scheme is located nearly 15 km from the nearest town of Kilosa. The program, called the TANRICE training, covered several technologies: the use of MVs of rice, the application of chemical fertilizer, improved bund construction, plot leveling, and transplanting in rows. Improved bund construction entails piling soil solidly around the plots, while plot leveling involves flattening the ground for better storage and uniform water distribution on paddy fields. Transplanting seedlings in rows allows rice growers to control plant density precisely and remove weeds easily.

Intensive training was offered to 20 farmers, called key farmers, at the nearby training institute (MATI Ilonga) over 12 days in November 2008, prior to the 2009 crop season. Subsequently, during the 2009 main crop season, each key farmer was requested to invite five intermediate farmers to training sessions held at a demonstration plot within the irrigation scheme. The key farmers and MATI jointly provided three-day training sessions to the intermediate farmers at three different stages of farming—nursery preparation, transplanting, and harvesting. Following these “in-field training” sessions, key and intermediate farmers were expected to disseminate technologies to the remaining farmers (i.e., the ordinary farmers). One day of in-field training was open to all the farmers in the scheme, including the ordinary farmers. The key farmers were competent and leading farmers selected by MATI based on such criteria as age, literacy, gender balance, residence within the irrigation scheme, and the practice of rice farming and were confirmed at an all-villagers meeting. The intermediate farmers were selected personally by the key farmers with no formal involvement by MATI. Thus, the selection of the key and intermediate farmers was rather purposive. Neither the key nor the intermediate farmers were paid for attending the training.

3.2 Data

Three rounds of the annual survey were implemented in 2010, 2011, and 2012. In the first survey, we interviewed 208 randomly selected farmers from the farmer roster in the irrigation scheme. We asked the respondents to identify the most important rice plot for their livelihood, hereafter referred to as the farmer’s “sample plot.” Farmers were asked in detail about rice cultivation on their sample plot, including their use of labor, capital, and other inputs in 2010. Similar information on the sample plot was also collected for the 2011 and 2012 crop seasons. During the first survey in 2010, we collected recall data on rice cultivation on the sample plot for the 2008 and 2009 main crop seasons, before and during the TANRICE training, respectively.

In the main analysis, we dropped households that took erroneous values in important variables and those that did not grow rice on the sample plot. The attrited households from the survey interviewed in the first round but not found in the second and third rounds were also omitted. This resulted in 171 observations for 2008, 182 for 2009, 202 for 2010, 168 for 2011, and 167 for 2012. We estimated an attrition probit model⁷ and confirmed that the attrition had occurred randomly concerning the observed set of variables. This result implies that analysis using the available observations (i.e., balanced and unbalanced panel data) will not suffer serious attrition bias. Thus, we use unbalanced panel data with more information due to the larger sample size.

3.3 Descriptive Analyses

Table 4.3 presents the changes in the average paddy yields and the technology adoption by the key, intermediate, and ordinary farmers from 2008 to 2012. The t-tests and χ²-tests comparing the key and intermediate farmers to the ordinary farmers are also shown. Note that the TANRICE training was conducted immediately before and during the 2009 main crop season. The table shows that even before the training (i.e., in 2008), the key farmers attained a slightly higher yield than the ordinary farmers, presumably due to the higher adoption rates of improved technologies and some innate abilities for rice production. Due to their increased technology adoption rates, the key farmers’ yield increased immediately after the training, from 3.1 tons/ha in the pre-training year 2008 to 4.4 tons/ha in 2009. They continued to achieve higher yields than the ordinary farmers, reaching 5.3 tons/ha in 2011 and 4.7 tons/ha in 2012. The key farmers’ adoption rates for modern varieties, improved bund construction, transplanting in rows, and chemical fertilizer use also rapidly increased in 2009 and remained significantly higher than the ordinary farmers until 2012, contributing to a high yield each year.

Table 4.3 Changes in paddy yield and technology adoption by training status for TANRICE training in irrigated area (key and intermediary farmers)

In contrast, the change in yield from the 2008 base year for the intermediate farmers is not as rapid as that of the key farmers. Soon after receiving the training during the 2009 season, the intermediate farmers’ technology adoption rates, including MVs, improved bund, and transplanting in rows, began increasing, eventually boosting the yield to a significantly higher level than the ordinary farmers in 2011. These results indicate that, although the effect of the training—both in terms of magnitude and speed—is more significant for the key farmers than for the intermediate farmers, the intermediate farmers also caught up with the key farmers in the years following the training.

It is remarkable to observe that the paddy yield of the ordinary farmers also rose from 2.6 tons/ha in 2008 to 3.7 tons/ha in 2012, even though the change was neither rapid nor drastic compared with the key and intermediate farmers. This increment can be attributed to the increased use of chemical fertilizers and improved agronomic practices. The belated yet significant technological changes seen in the behaviors of the ordinary farmers indicate that technologies taught in the TANRICE training spilled over from the key and intermediate farmers to the ordinary farmers over the years. The yield gap between the key and ordinary farmers ranged from 1.7 to 2.3 tons/ha between 2009 and 2011, while it diminished to one ton per hectare in 2012. These results suggest that the key farmers’ performance improved rapidly after the training. In contrast, the ordinary and intermediate farmers’ performance improved by learning from the key farmers, resulting in a smaller gap in yield and technology adoption in later years.

3.4 Estimation Methods

To evaluate the effects of TANRICE training on the adoption of rice cultivation technologies and paddy yield, we estimate fixed effect difference-in-differences (FE-DID) models with multiple periods and multiple treatment groups (Imbens and Wooldridge 2009; Meyer 1995) using our five-year panel data. In FE-DID, we use the panel structure of our data set to control for unobservable time-invariant household-specific characteristics that may influence training participation and the trends in the outcomes. Namely, the following econometric model is estimated:

$$ {{\text{Y}}}_{{{\text{it}}}} \, = \,\alpha \, + \,\beta {{\text{T}}}_{{\text{t}}} \, + \,\gamma {{\text{T}}}_{{\text{t}}} {{\text{S}}}_{{\text{i}}} + {{\text{C}}}_{{\text{i}}} \, + \,{{\text{u}}}_{{{\text{it}}}} $$

(4.2)

The dependent variables are paddy yield (tons per hectare) and the following set of technology adoption variables: a dummy variable for MV adoption, the amount of chemical fertilizer use (kg per hectare), and dummy variables for the adoption of improved bund construction, plot leveling, and transplanting in rows, respectively, in the sample plot; ${T}_{t}$ is a vector of four-year dummies in year $t$, with the base year of 2008; ${S}_{i}$ is a vector of two training status dummies (i.e., key farmer and intermediate farmer dummies, with their base group being ordinary farmers); ${T}_{t}{S}_{i}$ is a vector of all pairwise interactions between ${T}_{t}$ and ${S}_{i}$; ${C}_{i}$ is the time-invariant household-specific effect for household $i$ and ${u}_{it}$ is the stochastic error term. Since we sample one plot for each household, we use subscript i for the outcome variable while it is measured at the plot level. Note that in basic DID models, the terms of time-invariant training status dummies ${S}_{i}$ are included. In our case, however, this term is absorbed by ${C}_{i}$, as we control for the household fixed effects.

Years 2009 to 2012 are all post-treatment years, while 2008 is pre-treatment. Thus, coefficients γ associated with the interaction between the year dummies and training status dummies are the DID estimates of interest to capture the gap in the training effects between the trained (key and intermediate) farmers and the ordinary farmers. The strength of this model is that the term ${C}_{i}$ absorbs the unobservable time-invariant household characteristics, which are likely to affect training participation. In other words, a potential selection bias is largely addressed. The year-specific effects represented by $\beta$ capture the changes in the outcome variables for the ordinary farmers. These year dummies are assumed to capture the indirect effects of training on ordinary farmers through knowledge spillover from the trained farmers and other year-specific characteristics such as weather.

It is important to note that, in our case, $\gamma$ the vector of DID estimators should not be interpreted as the “pure” training impact, i.e., the difference in growth between the factual and counterfactual situations for the key and intermediate farmers with and without the training intervention, respectively. Instead, $\upgamma$ is designed to capture the differences between the effect on the key and intermediate farmers and the effect on the ordinary farmers, since $\beta$ captures the changes in the performance of the ordinary farmers, which incorporates the spillover from the key and intermediate farmers. Thus, as the ordinary farmers catch up with the key and intermediate farmers, $\gamma$ is expected to become smaller. In Nakano et al. (2018b), we also estimate a similar model by using the Propensity Score Matching-DID method. The results are largely the same for both cases.

3.5 Results

Table 4.4 presents the results of the FE-DID estimation on paddy yield and technology adoption. The year fixed effects are positive and significant in 2009–2012 for the use of chemical fertilizer and the adoption of plot leveling and transplanting in rows; in 2010–12, for improved bund construction; and in 2011–2012, for paddy yield. The adoption of MVs also increased in 2012. This indicates a steady increase in the technology adoption and paddy yield for the ordinary farmers, suggesting positive spillover effects of training over time.⁸

Table 4.4 Estimation results of difference-in-differences with household fixed effects models for paddy yield (tons/ha) and technology adoption for TANRICE training in irrigated areas. (2008–12)

For paddy yield, the DID estimates are significant for the key farmers in 2009 and 2010, indicating that training for the key farmers took immediate effect. Our results suggest that the training impact on key farmers’ yield is larger by 1.2–1.7 tons/ha than on ordinary farmers’ yield in 2009 and 2010. This rapid increase in paddy yield can be predominantly attributed to the fast technology adoption by the key farmers. The increase in chemical fertilizer use by key farmers relative to ordinary farmers is 41.8 kg in 2009, 56.3 kg in 2010, and 78.0 kg in 2011. The increase in the adoption rate of transplanting in rows is also steadily higher for key farmers from 2009 to 2012.

A more striking finding is the absence of significant yield effects captured by the interaction terms between the key farmer dummy and the 2011 and 2012 dummies. This suggests that the key farmers’ “yield premium” disappeared by 2011 and 2012. Given that the performance of the ordinary farmers was steadily improving from 2010 to 2012, this suggests that the ordinary farmers caught up with the key farmers presumably because of knowledge spillover from the key and intermediate farmers to ordinary farmers.

We do not observe significant coefficients for the interaction terms of intermediate and year dummies for paddy yield. On the other hand, these coefficients are positive and significant for chemical fertilizer use in 2010 and 2011 and transplanting in rows in 2009 and 2010. The increase in chemical fertilizer use by intermediate farmers is larger than that of ordinary farmers by 22.4 kg in 2009 and 28.5 kg in 2010. The training impact on the adoption rate of transplanting in rows is also larger for intermediate farmers, by 21% in 2009 and 34% in 2010 (vs. ordinary farmers). These results imply that intermediate farmers adopted new technologies more rapidly than ordinary farmers, although their productivity increase is no faster than those of ordinary farmers. Furthermore, consistent with the previous observation, we find that ordinary farmers gradually catch up with intermediate farmers in terms of the technology adoption, and significant differences between them tend to disappear in 2012 for chemical fertilizer use and in 2011 and thereafter for transplanting in rows. Although results are not shown, we obtained consistent results even when we controlled for the participation in other programs, including fertilizer subsidy and credit program, suggesting the robustness of our results.

4 Conclusion

This chapter shows two case studies on the impact of management training on the participants and that of F2FE on non-participants in Tanzania. The first notable finding is that in both cases, trained farmers achieved as high as 4.7 to 5.3 tons/ha, which is high even compared to the Asian standard. This suggests the high potential of the African rice Green Revolution, or we can even say that Green Revolution has taken place in limited areas. In the first study, we examined the impact of MSRI training provided by a private company in a rain-fed area of Kilombero district in Tanzania. We found that the training effectively enhances the technology adoption and increases the paddy yield of trainees by 1.4–2.5 tons/ha. As a result, the farmers who apply recommended MSRI technologies achieve as high a yield as 4.7 tons/ha on average. Given that the average yield of non-trainees in the study sites is 2.6 tons/ha, this is a remarkably high yield.

In the second study, we investigated the impact of F2FE in an irrigated area. We found that participants of the training increased the adoption of recommended technologies, and their paddy yields increased significantly. In addition, the performance of non-participants also improves later, and these farmers catch up with the participants. This suggests that F2FE effectively induced spillover effects from participants to non-participants in the irrigated area. Nakano et al. (2018b) further revealed that, by using spatial econometric techniques, the social relationships between key and ordinary and between intermediate and ordinary farmers play a significant role in the adoption of technologies by ordinary farmers.

In sum, we observed the positive impact of training on technology adoption and productivity of training participants in both cases. Our results are consistent with previous studies that found a positive training impact on rice productivity in both irrigated and rain-fed areas in other African countries (deGraft-Johnson et al. 2014; Kijima et al. 2012; Takahashi et al. 2019). We also found that F2FE successfully improves the performance of non-participants, possibly due to spillover effects from participants of the training, especially in irrigated areas, which is also consistent with Takahashi et al. (2019; and Chap. 3 in this volume), who found positive spillover effects of rice cultivation training in irrigated and rain-fed areas in Côte d’Ivoire.

By contrast, we did not find strong evidence of spillover effects in rain-fed areas. The limitation of the study in the rain-fed area is that the survey was conducted soon after the training. Since it may take some time for the spillover effects to be observed, it is still not conclusive that spillovers can occur in rain-fed areas. Our study in an irrigated area also observed that the spillover from key farmers to ordinary farmers gradually took place. Further investigation is needed on the differential impact of agricultural training and F2F extensions in different agro-ecological conditions in the mid- and long-term.