Data for the present study came from a household survey conducted in Uganda between June and August 2014. The geographical focus was eastern Uganda, where the DT maize seed dissemination activities have been concentrated. The region constitutes 32 districts lying over an area of 39,478.8 hectares covering about 16% of the total area of Uganda. The elevation of the area ranges between 1,075 and 1,524 m above sea level. The region's population is estimated at 9,154,960 of which about 90% lives in rural area. The average household size for the region is about 4.9 vs. 4.7 for the country as a whole [Uganda Bureau of Statistics (UBOS), 2014]. The mean annual rainfall varies from 1,374 to 2,058 mm with a range between 895 and 3,001 mm. The rainfall exhibits significant annual and seasonal variation in the amount and distributional pattern (Kansiime et al., 2013). The region has three distinct agroecological zones (AEZs) viz., Lake Victoria Crescent; Southern and Eastern Lake Kyoga basin; and Mount Elgon high farmlands. These AEZs also capture variability in altitude, soil productivity, cropping system, livestock systems, and land use intensity. Agriculture is the main source of livelihood in the region. Crop production is dominant. The major crops produced in the region are finger millet, maize, rice, sweet potato, and cassava. The region accounted for about 38% of the total maize area and half of the total maize produced in Uganda in 2008/09 [Uganda Bureau of Statistics (UBOS), 2012]. Important livestock includes cattle, small ruminants (goats and sheep), pigs, and poultry (chicken, ducks, turkey). While about 57% of agricultural households face food shortages in certain periods of the year, the Eastern region was reported to have the highest percentage (30%) and the least was for the central region (17%). Households with access to credit are low both at the national (10%) and in the Eastern region (9%). From 31% of the agricultural households using improved seed, the region has the highest share (43.7%) vs. the lowest in the Western region (16.6%) [Uganda Bureau of Statistics (UBOS), 2010]. The Eastern region is subject to the vagaries of climate variability. The effect of climate variability is assumed to be complicated by poorly developed economic and social services and infrastructures, and severe poverty status in the region. Generally, the region is characterized by a combination of acute poverty, vulnerability to drought, floods, landslides, and natural resource degradation (Kansiime et al., 2013).

Unlike many cross-sectional data sets, the current data has information on the spatial heterogeneity of selected districts which we use as a proxy for the temporal dimension of rainfall variability. Multistage sampling was employed to identify sample households for the study. In the first stage, three districts were purposively selected from Lake Victoria Crescent agroecological zones (LVC AEZs) of the eastern region where drought risk is more likely and where the DT maize seeds are disseminated. The LVC AEZ is one of the three distinct AEZs in the eastern region of Uganda which includes the southern and eastern Lake Kyoga basin and Mount Elegon high farmlands (Figure 1). These AEZs are mainly differentiated by the amount of rainfall. According to 40-year (1971–2010) observational rainfall data for the region, there is significant inter-annual rainfall variation within and across AEZs. Particularly, the long-term data show that the LVC exhibits significant intra- and inter-seasonal rainfall variations (Kansiime et al., 2013). The sample districts (Iganga, Tororo, and Bulambuli) are geographically well spread across the LVC AEZ (see Figure 1). Accordingly, they are assumed to account for the spatial heterogeneity as well as the temporal rainfall variability within this AEZ. In the second stage, we used probability proportional to size sampling to select a total of 34 out of 3,119 villages from the study districts. Details of the number of households in each village were acquired from the 2012 Uganda population and housing census of enumeration areas by the Uganda Bureau of Statistics (UBOS). Finally, from each sampled village, 12 households were selected for interview using a simple random sampling technique. The interview involved 696 individuals (householders and their spouses) drawn from a total of 408 sampled households.

Both household and plot-level data were collected through face-to-face interviews using two structured questionnaires, viz., a household and an individual questionnaire. A village questionnaire was also used to interview key informants [including extension workers, local council chairpersons (an elected head of the village- the lowest administrative level), progressive farmers, and local opinion leaders] regarding village-level variables such as input/output prices, distance to markets, and subjective assessments of rainfall patterns.

The survey covered plot-level information where for each plot, the respondent recounted the names and details of maize varieties cultivated during the 2013/2014 production year. Other plot-level data collected included slope, soil fertility, plot size, irrigation access, plot tenure, erosion incidence, crop production estimates, and input use. Important socioeconomic and demographic variables collected were age (number of the year lived), gender, education (number of school years), family size (number of household members), access to extension service (dummy variable for the source of information), the likelihood of getting credit (dummy variable), and social capital (an index calculated based on the number of groups the respondent is a member out of a list of selected group related to agricultural activity). Moreover, maize network size was defined as the number of other farmers the respondent regularly talks to get information regarding maize farming, the number of progressive farmers in the respondent's village, the number of DT maize seeds known to the respondent, and sources of information on new maize seed were also retrieved from the respondents.²

We collected data on drought shocks based on the number of times the farmer could remember having encountered drought and drought-induced maize harvest loss in the previous 5 years prior to the date of the interview (i.e., 2010–2014). Experiencing frequent episodes of drought may influence farmers' decisions to cultivate DT or other types of maize that withstand droughts and help minimize yield losses.

Data related to rainfall patterns were collected from village key informants based on their subjective assessment of the timeliness, amount, and distribution in the major growing season immediately prior to the date of the survey. We considered farmers' perception of the timeliness (onset and cession), adequacy (whether the amount received is enough to support maize production), and distribution [based on rainfall availability at critical growth stages and extent of dry-spell (based on farmers' estimate of the time interval between any rainy day and the next rainy day) experienced] of rainfall to construct an index variable for the adequacy of rainfall or lack thereof. Perceived responses to each of these items (as “yes” or “no”) were coded as favorable and unfavorable rainfall outcomes and averaged over the number of questions. The index values range from zero to one where one stands for a favorable outcome and zero for the worst. While farmers' decision to cultivate a particular maize seed (i.e., DT or non-DT maize seed) may depend on their expectations regarding rainfall pattern, their perception of the actual rainfall³ pattern may help to explain the performance of a particular type of maize cultivated (Teklewold et al., 2013).

Also, the survey involved the administration of a risk elicitation experiment to measure farmers' risk preferences (risk-taking behavior) using Gneezy and Potters (1997) risk elicitation mechanism. The experiment provided a hypothetical situation that elicits the risk preference of farmers based on actual pay-offs made following: (1) respondents' choice from the combination of high and no-risk maize seeds and (2) probabilistic determination of weather conditions (favorable or unfavorable outcomes) which is based on the toss of a coin. Farmers' risk-taking behavior is important in explaining decisions regarding the cultivation of new maize varieties as a potential adaptation strategy. Although DT maize seeds are developed to deal with drought risk, a decision to switch to new unfamiliar varieties may involve some risk, the potential benefits notwithstanding. Farmers who are more vulnerable to extreme weather events are less likely to use improved varieties as an effective means to cope with drought; instead, they tend to delay it (Cavatassi et al., 2011; Liu, 2013). In this study, we expect that farmers who are averse to risk could decide to cultivate DT maize if their belief in drought tolerance of the variety outweighs the potential risk associated with its being new.

Given that the DT maize seeds are developed for drought tolerance and yield, farmers who have a preference for these two traits are expected to cultivate these varieties. Farmers' preferences for these traits are captured based on whether he/she mentioned them among his/her most preferred traits of maize variety. Social capital may exert behavioral influences in various ways (that is, information sharing, positive externality, resource sharing, and relaxing liquidity constraints through informal credit). The possible roles of farmers' social capital in the decision to grow DT maize are captured using the social capital index constructed based on group membership, roles assumed in each group, the nature of the group composition (homogeneous or diverse), and its functionality.

The adoption of new technology is often modeled as a utility-maximizing choice between two or more alternatives. It can be influenced by the characteristic features embodied in the technology and many other factors. Observed adoption choice of agricultural technology (for example, modern crop varieties) is hypothesized to be the end result of a process of preference comparisons by farmers.

In this study, adoption is defined as the reported use of DT maize variety (that is when the farmer exactly mentions the variety by name) on a specific plot managed by the head or spouse of a household during the 2013/14 growing season. Adopting the utility maximization concept, a maize farmer (also referred to as plot manager⁴)—who has to make decisions regarding what type of maize to grow- has three possible alternatives, which are DT modern maize, non-DT modern maize, and “unimproved” or local maize seed.⁵ Each one of these categories of choice includes a distinct list of maize varieties. Plots are identified with the type of maize varieties cultivated by matching the name of the varieties reported by farmers with the list under each category. The DT maize seeds are mainly developed for drought tolerance and yield whereas the non-DT modern maize seeds are developed for traits other than drought (mostly higher yields under optimal conditions). The actual choice is assumed to be made based on farmers' utility derived from the adoption of one of these maize varieties.

Farmers' decisions to adopt a particular maize seed are assumed to be driven by maximization of the utility derived from the alternative types of maize seeds. For farmer i to choose any maize type, j, from available alternatives, m, it is required that U_ij > U_im, m ≠ j; the expected benefit, U_ij, that a farmer derives from the adoption of maize type j (that is, DT, non-DT modern, or local maize) is a latent variable determined by observed household, individual, plot, and location characteristics (X_i), and unobserved characteristics (e_ij):

Where X_i is a vector of observed exogenous variables; β_j is a vector of parameters to be estimated for each type of maize; and e_ij are unobserved characteristics. Let the choice from among the alternative maize types by a farmer is denoted by an index variable, I, such that:

Equation (2) implies that ith farmer will adopt modern maize type j to maximize his/her expected gain (η_ij) if j provides greater benefit than any other alternative type of maize m; that is, if ηij= max(Uij*-Uim*)>0,  m≠j (Bourguignon et al., 2007).

Assuming that e_ij are independently and identically Gumbel distributed, the probability that farmer i with characteristics X will choose maize type j can be specified by a multinomial logit (MNL) model (McFadden, 1973):

Since MNL is a model where regressors do not vary over choices, coefficients are estimated for any choice. MNL requires identification, thus one of the choices of m maize types (the local maize seed in our case) is treated as the base category (correspondent β_m is constrained to equal 0).

Based on the adoption literature, farmers' decisions to adopt a new agricultural technology depend, among others, on socioeconomic, demographic, and institutional factors. Hence, the choice of explanatory variables in this study is made based on a review of past studies on technology adoption and impact in developing countries (these include Feder et al., 1985; Bandiera, 2006; Deressa et al., 2009; Matuschke and Qaim, 2009; Kafle, 2010; Kassie et al., 2011; Asfaw et al., 2012; Teklewold et al., 2013; Fisher and Carr, 2015; Jain et al., 2015). This body of literature indicates that many factors influence adoption and thus affect our outcome variable. The factors are categorized as demographic and farmer characteristics (family size, gender, age, education, risk attitude, preference to drought tolerance, preference to yield, drought risk perception, experience of maize loss due to drought), social and institutional characteristics (social capital, number of progressive farmers in the village, main sources of information, credit access, distance to input market, distance to extension office), maize plot characteristics (plot size, tenure, extent of erosion, soil fertility, slope, irrigation), and geographic characteristics (captured through district dummies).

A fundamental problem when comparing adoption outcomes between individuals is that adoption is not randomly assigned- farmers endogenously select themselves into adopters or non-adopters of particular technologies. So decisions are likely to be influenced by both observed and unobserved characteristics. Unobserved factors that can influence adoption include aptitude, motivation, experience, or other factors, which are not readily observed in the present data. These “other” factors may be correlated with the adoption outcome. As adopters and non-adopters can be systematically different, estimating the effect of adoption without accounting for this implied endogeneity problem would lead to inaccurate results where outcomes are attributed to adoption when in fact significant aspects of the outcome are the result of other factors not accounted for. We apply an endogenous switching regression treatment effects approach to correct for self-selection bias (Dubin and McFadden, 1984) and provide treatment effects that take both observed and unobserved factors into account.

The endogenous switching regression (ESR) framework we used to estimate the impact of adoption while accounting for selection biases involves two stages. In the first stage, farmers' choices of maize types are modeled using the MNL selection model as specified above; and in the second stage of the estimation, the impact of adoption on the outcome variable is evaluated using ordinary least squares (OLS) which includes a selectivity correction term from the first stage.⁶

The relationship between the outcome variable and the set of exogenous variables Z is estimated for the maize type chosen. The outcome equation for each possible regime j is given as

Where Y_ij are the outcome variables of the i^th farmer in regime j, Z is as defined above, and the error terms (u) are distributed with E(u_ij|X,Z) = 0 and var (uij|X,Z)=σj2. For consistent estimation, Equation (4) will be augmented with plot-specific unobservable characteristics such as land quality that can help in controlling for unobserved heterogeneity among plots. Y_ij is observed if and only if alternative j is adopted which occurs based on the utility maximization indicated earlier. If the error terms of the selection equation (e) and that of the outcome equations (u) are correlated, unbiased estimation of Equation (4) requires the inclusion of the selection correction terms of the alternative choices as recovered from the MNL models. Bourguignon et al. (2007) show that consistent estimates of the parameter in the above equation can be obtained by estimating the following multinomial endogenous switching regression models:

Where θ_j is the covariance between e's and u's; and λ_j is the inverse Mills ratio (IMR) computed from the estimated probabilities in Equation (3) as follows: λj=∑m≠jJρj[Ṕimln(Ṕim)1-Ṕim+ln(Ṕij)]; ρ is the correlation coefficient of e's of the selection equation and u's of the outcome equations, and ω represents error terms with an expected value of zero. In the multinomial choice setting, there are J-1 selection correction terms, one for each alternative maize type. The standard errors in Equation (5) are bootstrapped to account for the heteroscedasticity arising from the generated regressor (λ_j).

For Equation (5) to be identified, it is good practice in empirical analysis to use variables that affect the choice decision but not the outcome variable (as exclusion restrictions) in addition to those automatically generated by the non-linearity of the selection regression. The specification chosen for the outcome equations in Equation (5) follows the common practice in the agricultural economics literature (see, e.g., Solis et al., 2007; Di Falco et al., 2011; Teklewold et al., 2013), allows us to use variables related to information sources, farmer and farm household's characteristics as well as including the number of DT maize varieties known to the farmer as exclusion restriction. The validity of the exclusion restriction is tested based on the significance of these variables in the selection model but not in the outcome equation.

Using the above framework, the average treatment effect on the treated (ATT) which is the average effect of adoption for adopters (DT maize growers), and on the untreated (ATU) which is the average effect of adoption for non-adopters, are both examined by comparing the expected actual and counterfactual outcomes. The computation of the counterfactual and average treatment effects using endogenous switching regression is implemented in the same manner as in Di Falco et al. (2011), Teklewold et al. (2013), and Kassie et al. (2014).

Actual outcomes for

Counterfactual outcomes for

While Equations (6) and (7) represent the actual outcome observed in the sample for adopters and non-adopters, respectively, Equations (8) and (9) are the respective counterfactual outcomes. These conditional expectations are used to compute ATT and ATU. The average adoption effect for adopters (ATT) is calculated as the difference between Equations (6) and (8). The difference between Equations (7) and (9) provides the average adoption effect for non-adopters (ATU). ATT, for example, is shown as follows:

The first term on the right-hand side (RHS) of Equation (10) represents the expected change in adopters' average outcome if adopters' characteristics had the same return (coefficient) as the characteristics of non-adopters. In the second term (on the RHS), λ_j, is the selection correction term that captures all potential effects of difference in unobserved variables.

The average treatment effect (ATE) can be calculated by taking the difference between the two actual outcomes (Equations 6, 7). In observational studies, ATE gives the treatment effect without accounting for selection bias. The computation of the conditional expectations and treatment effect are summarized in Table 1.

About 77% of the sampled respondents grew modern maize varieties of which 24% used DT maize. The average size of all maize plots was 0.40 hectares; the corresponding figure for plots allocated to all modern and only to DT maize seed was 0.41 and 0.45 hectares, respectively. DT growers tend to cultivate maize on relatively larger plots. The share of maize plots planted to DT maize was about 19%. Comparatively, non-DT modern maize seeds were widely cultivated covering 60% of the maize plots of the sample households. These maize seeds have been part of the seed systems in Uganda since the 1960s (Balirwa, 1992) and their spread might imply that over time the DT maize would also follow similar diffusion trends. At the time of this study, the uptake level of DT maize suggests that it is still at the early diffusion stages (Figure 2).

Table 2 presents descriptive statistics of variables that are used in the empirical models (selection and outcome equation) applied to make casual attribution of adoption of DT maize.

Table 3 presents the estimation results of the first stage (the MNL selection model) of the endogenous switching regression. From the three types of maize used as the categorical dependent variable, local maize is used as the base category. The tests of goodness-of-fit reported at the bottom of the table show the model fits the data reasonably well. The Wald χ² test statistics (672.26) rejects the hypothesis that the regression coefficients of the explanatory variables are jointly equal to zero (p = 0.00). Accordingly, the result showed that the gender of plot managers, size of the maize plot, preference for the drought tolerance trait, number of DT varieties known to the plot manager, source of information on new maize seed and location dummies, as expected, influenced the probability of adopting DT maize varieties. Male-headed plot managers have a higher adoption probability than their female counterparts.

Also, plot managers who receive information about new maize seed varieties from research centers and other farmers, are more likely to cultivate DT maize. Research can provide reliable technical information, but it may not be accessible to many farmers mainly due to limited presence in terms of geographical spread. As expected proximity to the nearest extension service increases the chances of adoption.

Farmers who have a preference for DT traits are expected to cultivate DT maize. The results indicate that compared to the base category, preference for drought tolerance is an important driver for the adoption of DT varieties. This suggests that those who grow DT seeds are well-informed about this unique trait. Moreover, these DT varieties maintain (to a large degree) yield parity with extant hybrids. It stands to reason then that preference for DT trait would be the main differentiation for many farmers in evaluating DT varieties compared to non-DT ones.

Finally, the positive correlation of some location dummies with adoption could be a reflection of agro-ecological factors (like rainfall) as well as other underlying unobserved spatial differences. The reference district (Iganga) received relatively better rainfall (in amount and distribution) over the main growing season (March–June; see Table A1). Comparatively, low rainfall signal at planting time (March) might have prompted a higher likelihood of growing DT maize seeds in Bulambuli and Tororo than in Iganga.

The results presented in this section are based on the conditional and unconditional average effects of DT maize adoption on expected maize yield. The ATT and ATU are computed using the predicted value of yield following the schema presented in Table 1.

The coefficient estimates from the second stage regression (OLS) are presented in Table A2. It is to be noted that the coefficients on selection correction terms are almost not significant suggesting that our results are unlikely to be driven by selection bias. The regression estimates without the correction terms remain similar but in the interests of brevity, we have not reported the results.

The figures reported in Table 4 are expected values of yield. A simple pairwise comparison of means (ATE) indicates that cultivation of DT maize, on average, gives higher maize yield to adopters than non-adopters who grew either non-DT or local maize. Such comparisons, however, could be misleading as it does not control for observed and unobserved factors that may influence the outcome variable.

We used Equation (10) to estimate the true average adoption effect and compare the expected yield of farmers cultivating DT maize with their counterfactual outcome- if the same farmers had instead cultivated local or non-DT maize seed. The results indicated that both adopters and non-adopters would benefit from adoption. The magnitude of the adoption effect, however, differs when the comparison is between DT vs. non-DT and DT vs. local maize growers. The ATT showed that farmers who actually cultivated DT maize get 30% more yield than what they would have obtained had they instead adopted non-DT maize seed. The corresponding yield effects of adopting DT instead of local maize were four times high. Similarly, ATU revealed that farmers who grew non-DT modern and local maize received 32% and 54% more yield, respectively, if they instead had adopted DT maize. That is, non-adopters would have realized higher productivity if they decide to switch to DT maize.

We examined the treatment effect both by location (districts) and rainfall status, the latter was determined using the rainfall index (as described in the data section). Table 5 presents the ATT and ATU of DT adoption against non-DT and local maize by the district. In all the districts, the average yield of DT outperformed that of non-DT and local maize. As would be expected, the magnitude of the ATT was relatively higher against local varieties than non-DT maize varieties. The ATT against local maize indicates a yield advantage ranging from 2.6 times at Tororo to six times at Iganga. The corresponding yield advantage for ATU ranged from 0.5 to 1.5 times more at Tororo and Bulambuli, respectively. The ATT against non-DT growers shows a yield advantage of 52% at Iganga, 18% at Tororo, and 17% at Bulambuli. The ATU (for DT vs. non-DT) had a relatively better yield advantage than the ATT in all the districts suggesting that the adoption of DT would have been even more beneficial to those who grew non-DT (had they grown DT) than those who already adopted DT.

According to the rainfall records from proximate weather stations (Table A1), the amount of rainfall received by each of the study districts during the 2014 main season covered by the survey (that is, March–June 2014) was well below the long-term average (that is, 659 mm) reported by Kansiime et al. (2013). In relative terms, the Iganga district received a higher amount with better distribution during the main season of 2014. Theoretically, the treatment effect should have been pronounced more in the Bulambuli district where the amount and distribution of rainfall received during the surveyed season was relatively less favorable. But this might be an indication that district-level rainfall data adopted from the meteorological stations are less representative of the rainfall distribution under more granular scale conditions at the micro level, given the low density of weather stations in Uganda.

To mitigate this, we tracked rainfall conditions/patterns at a lower level using the rainfall index constructed for each village. The index is then used to categorize the villages into those which had potentially poor rainfall status (with a rainfall index value equal to or < 0.5) and those which had potentially good rainfall status (with an index value greater than 0.5) in the 2014 main season.

Table 6 shows the treatment effect by rainfall status of sample villages. In all cases, be it under poor or good rainfall conditions, DT adoption makes the productivity of adopters (ATT) and non-adopters (ATU) better off. Leaving the substantial yield advantage of DT maize adoption over local maize under the two states of rainfall aside, the treatment effect over non-DT reveals an interesting result. The average treatment effect (ATT) over non-DT maize under poor rainfall conditions was about 417 kg ha⁻¹, that is, cultivation of DT maize offers a yield advantage of 44% over what would have been obtained if the plot was instead planted to non-DT maize. The corresponding ATT under good rainfall conditions has an additional yield advantage of about 21%. This supports the claim that DT maize outperforms other commercial maize (non-DT) more in seasons when the rainfall condition is less favorable.

Despite its inability to account for confounding effects, we further ran ANOVA of the linear prediction of DT yield as a partial metric to compare its performance difference across the three districts. The result suggests a significant (p = 0.00) productivity difference in mean DT maize yield across the three districts. Pairwise yield comparisons between these districts further revealed statistically meaningful differences (Table 7). Bulambuli district- where the highest share (40%) of plots was planted to DT maize varieties- had the highest mean DT yield, and Tororo district had the least.

To help in visualizing the results described above, we summarized these spatial variations in the impact of DT in Figure 3 based on the ATT comparing DT with local maize varieties (Figure 3A) and DT with non-DT maize varieties (Figure 3B). Although the mean ATT is positive and significant across all districts, there are pockets where the impact as measured by ATT was below the sample average. This can be seen in the Tororo district where both the magnitude of the impact is relatively lower (see Tables 5, 6), and the DT coefficient on the district dummy is comparatively smaller (see Table A2). Further examination of the particularities of the Tororo district might be worth looking at to bring out and learn about the underlying causes.