**3. Methodology**

## **3.1. Empirical analysis of adoption**

Adoption of a different technology (i.e., improved variety of sweetpotato) from the existing one usually entails both costs and benefits. In the case of OFSP, these costs could include conventional (such as having to purchase new planting materials from neighbors) and transaction costs related to search and transportation of planting materials to the firm.

The decision to adopt of improved technology is often modeled using a logit or probit regression model. In the case where OFSP varieties coexist with other improved local varieties, it is also expected that farmers make the decision to adopt OFSP fully knowing of the existence of the other varieties. Thus the adoption decisions relating to OFSP and improved local varieties are interdependent [15, 16]. More formally, this means that the decision to adopt variety j by farmer i is correlated to the decision to adopt variety k. The correlation in decision to adopt multiple varieties of sweetpotato, in turn, suggests the likelihood of correlation of the error terms across the different equations. This correlation renders the use of probit or logit regressions models inappropriate because the estimates coefficients will be biased [16].

To overcome the above problem, a multivariate probit regression was used to model the choice of sweetpotato varieties to plant. The multivariate probit regression technique has been used in several past studies [17–19]. It is an extension of bivariate probit model [20] and uses Monte Carlo simulation techniques to jointly estimate the multivariate probit regression equation system [21].

The implicit form along with the variables included in the estimation of the model is given by

variety = f(*intervened*, *foodsec*, *gender*, *lneduc*, *lnage*, *credit*, *lnusedfarm*,      *valley*\_*bottom lndistmkt*, *lncrpinc*1, *grpmember*, *assetindex*, *mo reyield*, *sugary*, *aezP*4*N*10, *aezP*8 ) + error term

The dependent variable in the equation above is a dummy variable equal 1 if a farmer *i* decided to plant variety *j* and 0 otherwise. That is, the dependent variable takes the value of 1 if *variety* = *j*, for *j* = Kabode, Ejumala, Jewel, or New Polista, Ukerewe is planted, and 0 otherwise.

The explanatory variables used in the estimation of the model are: *intervened* is a dummy equal 1 if a farmer participated in the Marando Bora project, 0 otherwise; *gender* is a dummy equal 1 if the respondent is a male, zero if otherwise; *lneduc* is natural log of education in years; *foodsec* is an index of food security computed as discussed above; *valley* is a dummy equal 1 if farmer has access to valley bottom, 0 if otherwise; *lndistmkt* is natural log of distance to market in walking minutes; *lnfamsize* is natural log of size of farm in acres ; *grpmember* is a dummy equal 1 if the farmer belongs to a farmer organization, 0 if otherwise; *credit* is a dummy equal 1 if a farmer received credit in 2012; *lncropinc* is natural log crop income in Tanzania Shillings; *assetindex* is wealth index computed from physical (household and farm) assets. The index was computed following [22]; *moreyield* is a dummy equal 1 if a higher root yield is important in decision to conserve vines, 0 if otherwise; *sugary* is a dummy equal to 1 if the sweetness (high sugar content) is important to the farmer, 0 if otherwise; *aezP5N10* is a dummy equal to 1 if a farmer is in the more wet area, 0 if otherwise; *aezP8* is a dummy equal to 1 if a farmer is in moderately wet area, 0 if otherwise; *eazP4W3* is a dummy equal 1 if a farmer is in drier area, 0 if otherwise.

In this study, other variables that should be included in the model, based on a priori expectations, but are highly correlated with participation in the Marando Bora project were dropped. This is because their inclusion along with intervention (i.e., participation in Marando Bora) could result in "double counting" (For example, see [23] for similar treatment but in an unrelated study). The variables dropped are voucher, dummy equal to 1 if a farmer received a voucher that allowed him/her to get sweetpotato vines at discounted rate, 0 otherwise and know\_DVM is a dummy equal 1 if a farmer knew where to find a DVM, and 0 otherwise. In addition, the Wald joint-exclusion restriction test is used to assess and drop variables that do not explain much of the variability in the decision to adopt OFSP.

#### **3.2. Computation of household food insecurity index**

Food insecure households engage in several activities/actions in efforts to obtain food. Food insecurity status of the household can thus be deduced from a set of actions that it engages in to get food in situations of food shortage. Typically, the more the number of such actions undertaken by a household, the more food-insecure that household is likely to be. In this study, we use these "set of actions" undertaken by the households in the sample, and the Rasch model, to measure the food insecurity status of the household. Under the Rasch model, a farmer's behavior can be expressed as a matrix X containing the response xij of i=1,…,n farmer to j=1,…,m food insecurity statements. The actions typically correspond to a set of binary statements that capture what the farmer did in response to food insecurity situation in his/her household. An index of food insecurity was therefore computed for each of the 732 households by subjecting the collected statements to the Rasch analysis in RUMM2030. The more the actions a household undertook in response to food scarcity, the higher the food insecurity index. Households with a higher index are more food-insecure than the rest of the households in the sample. Moreover, it means that most of the statements they engaged in were extreme/severe.

#### **3.3. Data**

agroecological zones namely zone P4, zone P5, zone P8, zone W3, and zone N10. The sensitization on the use of quality planting materials focused on the sweetpotato agronomy, pest and disease diagnosis, and protection and methods for conserving vines for future planting.

Adoption of a different technology (i.e., improved variety of sweetpotato) from the existing one usually entails both costs and benefits. In the case of OFSP, these costs could include conventional (such as having to purchase new planting materials from neighbors) and transac-

The decision to adopt of improved technology is often modeled using a logit or probit regression model. In the case where OFSP varieties coexist with other improved local varieties, it is also expected that farmers make the decision to adopt OFSP fully knowing of the existence of the other varieties. Thus the adoption decisions relating to OFSP and improved local varieties are interdependent [15, 16]. More formally, this means that the decision to adopt variety j by farmer i is correlated to the decision to adopt variety k. The correlation in decision to adopt multiple varieties of sweetpotato, in turn, suggests the likelihood of correlation of the error terms across the different equations. This correlation renders the use of probit or logit regres-

To overcome the above problem, a multivariate probit regression was used to model the choice of sweetpotato varieties to plant. The multivariate probit regression technique has been used in several past studies [17–19]. It is an extension of bivariate probit model [20] and uses Monte Carlo simulation techniques to jointly estimate the multivariate probit regression

The implicit form along with the variables included in the estimation of the model is given by

The dependent variable in the equation above is a dummy variable equal 1 if a farmer *i* decided to plant variety *j* and 0 otherwise. That is, the dependent variable takes the value of 1 if *variety* = *j*,

The explanatory variables used in the estimation of the model are: *intervened* is a dummy equal 1 if a farmer participated in the Marando Bora project, 0 otherwise; *gender* is a dummy equal 1 if the respondent is a male, zero if otherwise; *lneduc* is natural log of education in years; *foodsec* is an index of food security computed as discussed above; *valley* is a dummy equal 1 if farmer has access to valley bottom, 0 if otherwise; *lndistmkt* is natural log of distance to market in walking minutes; *lnfamsize* is natural log of size of farm in acres ; *grpmember* is a dummy equal 1 if the farmer belongs to a farmer organization, 0 if otherwise; *credit* is a

*reyield*,

tion costs related to search and transportation of planting materials to the firm.

sions models inappropriate because the estimates coefficients will be biased [16].

     *valley*\_*bottom lndistmkt*, *lncrpinc*1, *grpmember*, *assetindex*, *mo*

for *j* = Kabode, Ejumala, Jewel, or New Polista, Ukerewe is planted, and 0 otherwise.

variety = f(*intervened*, *foodsec*, *gender*, *lneduc*, *lnage*, *credit*, *lnusedfarm*,

*sugary*, *aezP*4*N*10, *aezP*8 ) + error term

**3. Methodology**

24 International Development

equation system [21].

**3.1. Empirical analysis of adoption**

This study uses the data collected from 732 households in January and February 2013 as part of the Marando Bora endline household survey. The map of the study areas covered is presented in **Figure 1**. A multi-stage sampling technique was used. First, four regions (Mara, Mwanza, Shinyanga, and Kagera) were purposively selected. Within each region, farmers were categorized into those that participated in the decentralized vine multiplier scheme, the mass distribution scheme and those who did not participate in any of the two schemes. For the purposes of this study, the first two groups of farmers comprise the "intervened" group (i.e., participants) and the last group constitutes the nonintervention group (i.e., nonparticipants). The list of farmers in each category (i.e., intervened and nonintervention) was then compiled at the village levels in each of the project wards and districts. A random sample of farmers was selected from each category of farmers for personal interviews. In total 481 project participants and 251 nonparticipants were interviewed. The sample contained 221 and 511 male and female farmers, respectively. The high number of female farmers reflects the fact that women mostly grow sweetpotato and that the project also targeted female household members. Data collected included farmer and farm characteristics, asset endowments, institutional characteristics, and varietal traits.
