Impact of agricultural technology adoption on wheat productivity: Evidence from North Shewa Zone, Amhara Region, Ethiopia

Abstract

Wheat is one of the most important cereal crops cultivated in wide range of agro-ecologies in Ethiopia. But, its productivity has remained low. Hence, this study intends to examine the impact of agricultural technology adoption on wheat productivity in north Shewa zone of the Amhara region, Ethiopia. The analysis is based on household level data covering 693 households collected in 2020. Multinomial logit model (MNL) and multinomial endogenous switching regression (MESR) models are used for analysis. The results reveal that agricultural technology adoptions are affected by the education level of the household head, off-farm employment, tropical livestock, access to credit, household saving, extension visit, and distance from the market. In addition, the study shows that the adoption of fertilizer and/or improved seed increases wheat productivity significantly. Furthermore, the adoption of a combined fertilizer and improved seed provides higher productivity than the adoption of single technologies. Therefore, this study recommends that government and other stakeholders should have to work in collaboration with rural farmers to increase rural technology generation; dissemination and adoption interventions to improve wheat productivity in the study area.

Keywords:

• Wheat
• technology adoption
• multinomial logit
• multinomial endogenous switching regression
• productivity

Public interest statement

Wheat is one of largest cereal crops produced for consumption and marketing purpose in Ethiopia. But its productivity is very low. Currently, Ethiopian government is applying different measures to improve the productivity including the expansion of the adoption of new or improved agricultural technologies. So, this study tried to examine the impact of agricultural technology adoption on wheat productivity in the north Shewa zone, Amhara Region, Ethiopia. The results revealed that the adoption of agricultural technology significantly increases wheat productivity.

1. Introduction

Majority of the poor in developing countries, particularly Sub-Sahara African countries, heavily depends on agriculture for survival; as a result, agriculture is considered as a key sector for stimulating economic growth, overcoming poverty, and enhancing food security. Agricultural productivity can be increased through the use of new technologies, and its increase will reduce poverty by increasing farmers’ income, reducing food prices, and thereby enhancing increments in consumption (Diagne et al., 2009). Correspondingly, agriculture is the backbone of Ethiopian economy. It accounts for about 34.1% of the GDP, employs 79% of the population, and accounts for 79% of foreign exchange earnings; it is the major source of raw material and capital for investment and provides large market (Diriba, 2020). However, the contribution of the sector to the country’s GDP is declining overtime. In addition, a rise in food demand and a subsequent hike in food prices, which were the main components of the consumer price index, have necessitated an improvement in productivity of the sector. Ethiopian government is applying different measures to improve the productivity including the expansion of the adoption of new or improved agricultural technologies. In the country, even though the adoptions of improved agricultural technology are intensely recommended to improve agricultural productivity, the adoption rate is remains very low (Admassie & Ayele, 2010; Mohammed, 2014). Even though the adoption of improved agricultural technologies has resulted in the improvement of the livelihood of the rural community and economic growth; lack of structural transformation, severe poverty, and food insecurity problems combined with instability and conflicts have made the road difficult. In this kind of scenario, the adoption of improved technologies especially for the production of staple crops is fundamental to the transformation of rural area, to achieve food security, and reduce poverty (Teklewold et al., 2013). One of the staple crops, where different improved technologies are being applied is wheat.

Wheat is the second-most important crop, taking 14% of the total calorie intake in the country, which is preceded only by maize (19%) and is ahead of teff (10%), sorghum (11%), and enset (12%; Wakeyo & Lanos, 2019). The country has an enormous potential for wheat production, and it is the largest wheat producer in Sub-Saharan Africa (Tadesse et al., 2019), even though the country is net importer of wheat. Around 5 million farming households’ life is based on wheat, which they grew it on around 1.7–1.8 million ha, annually (CSA (Central Statistical Authority), 2021). Oromia region is the highest producer of wheat (56.8%), followed by Amhara (31.4%), Southern Nations and Nationalities (7.46%) and Tigray (3.87%) of the total wheat of the country. On the other hand, the productivity of wheat varies from 33 quintals per hectare in Oromia, to 29 quintals in SNNP, to 28.3 in Amhara and to 21.9 in Tigray.

As the above data show, Amhara region is one of the major wheat producing regions in the country next to Oromia region, but its productivity is lower than both Oromia and Southern nation’s nationalities and peoples (SNNP) regions. The reasons for this low productivity in wheat are minimal application of agricultural technologies, traditional farming system, rain feed agriculture, and vulnerability to climate change. Similarly, North Shewa zone is one of the wheat producing zones of Amhara region, and the zone has a higher productivity than the regional average with 30.6 quintal per hectare (CSA (Central Statistical Authority), 2021). There are different studies that are conducted on the determinants of adoption and impacts of agricultural technology on productivity of wheat at national and regional level.

The studies by Tamirat et al. (2020), Yirga et al. (2013), Demissie Zeleke et al. (2021), Wudu (2017), and Siyum et al. (2021) focus on the determinants of the adoption of improved wheat technologies in different parts of the country. While, other studies by Hagos (2016), Ketema and Kassa (2016), Abate et al. (2018), Daniel and Belay (2018), and Tesfaye et al. (2016) have examined the impact of technology adoption on the productivity of wheat and other crops. The results of these studies show that household, economic, and institutional factors (including household heads sex, level of education, active labour force members, total land holding size, access to credit, participation in training, exposure to mass media, and frequent contact of extension agents) were the main determinants of adoption. In addition, the studies reveal that adoption of improved agricultural technology significantly improves the productivity of wheat and other crops.

The contribution of this paper to the existing literature is three fold. One, many studies that were conducted in the area focused on other crops than wheat (for example; Ahmed, 2015; Natnael, 2019; Shako et al., 2021). Second, other studies examine the impact of single technology adoption on wheat productivity (for example; Hagos, 2016; Mulugeta & Hundie, 2012). Thirdly, majority of previous studies have employed PSM, double-hurdle, ordered probit and OLS estimation techniques (for example; Hagos, 2016; Mulugeta & Hundie, 2012; Siyum et al., 2021; Tesfaye et al., 2016). But these models and estimation techniques are subject to self-selection bias, endogeneity problems and inadequate counterfactuals. For that reason, this study employs multinomial endogenous switching regression model, which has the potential in solving the above problems. In addition, the researchers use the adoption of fertilizer and/or improved seed adoptions because the adoption of these technologies is widely used and have higher adoption rate in the study area compared to other technology packages. Therefore, the study aims to examine the impact of agricultural technology adoption on wheat productivity taking North Shewa Zone, Amhara Region, Ethiopia, as a case study.

The rest of the paper is organized as follows: part 2 provides a brief description of the methodology, part 3 presents results and discussions, and part 4 presents the conclusion and recommendation of the study.

2. Methodology

2.1. Study area description

North Shewa zone is one of the 12 administrative zones of the Amhara region, and the capital city of the zone is Debre Berhan, which is found about 130 km north of Addis Ababa. The absolute location of the study area is latitude 80 43’ 06”—100 43’ 58” N and longitude 380 39’ 50’ “—400 06” 32’ ‘ E. The zone is bordered on the south and the west by the Oromia Region, on the north by South Wollo zone, on the northeast by the Oromo nationality special zone, and on the east by the Afar Region. Agriculture is the main economic activity and livelihood of people in the zone (North Shewa Zone Administration office, 2020).

2.2. Data and sampling description

The study is based on cross-sectional data collected from farming households in the North Shewa Zone. This study used both primary and secondary data sources. Primary data were collected through structured questionnaires. Whereas, the secondary data were collected from documented and published sources such as books, journal articles, meeting minutes, and North Shewa Agricultural Office reports. The study used the multistage sampling technique so as to select sample households. Firstly, from the total districts of the zone, four districts specifically Minjar Shenkor, Angolela Tera, Moretna Jiru, and Menz Gera were selected purposively because they have high potential to agricultural practices including wheat production and topographical similarity. Following this, according to North Shewa Zone Administration office (2020), there are a total of 117,149 farming households in the selected districts. Taking this, Vogel (1986) and Malhotra (2012) advised that for a large homogenous population, which is between 35,001 and 150,000, the researcher can draw a maximum sample of 800 samples. As a result, 800 households were drawn as sample size in this study. We then randomly selected 30 kebeles from the entire kebeles in the selected districts and finally selected each sample respondent from each selected kebeles using a simple random sample. Due to lack of information, four observations have been deleted, and in addition, 103 sampled households were dropped because they don’t produce wheat. Therefore, 693 households were included in the analysis.

2.3. Methods of data analysis

After the data collection, for data analysis, the researchers have employed both descriptive and econometric methods. The descriptive analysis includes mean and standard deviation and is mainly used to obtain a better understanding of the household demographic features, socio-economic, and institutional features of the farmers. From the econometric methods, a multinomial endogenous switching regression model is employed to scrutinize the impact of adoption of alternative agricultural technologies on wheat productivity.

2.4. Conceptual framework

Theoretically, it is supposed that farmers adopt improved agricultural technologies that can maximize their utility. This is to mean that farmers will only adopt a farm technology if it adds more to farm profit than farm operational cost, given resource constraints. Thus, adoption occurs if the net benefit (i.e. utility, productivity in this case) of the chosen package is higher than the benefit of the other alternatives. However, the utility that is gained from adopting agricultural technology is not observed, what is observed is only the farmers’ choice of technology, one can assume a random utility model which states conditional probability choice given farmers’ choice (Verbeek, 2008). In this study, the impact of farmers’ adoption of alternative agricultural technologies on wheat productivity is modelled using wheat farm yield as outcome variable. To estimate the impact, this study employed an endogenous switching regression model. Because the model has a potential in accounting impact evaluation estimation challenges like endogeneity problem,¹ sample selection bias,² and inadequate counterfactuals,³ the model also allows interaction between adoption of the various packages and other covariates in the productivity functions.

2.5. Multinomial endogenous switching regression (MESR)

In measuring the impact of adoption of alternative agricultural technology adoptions on wheat productivity, the study applied multinomial endogenous switching regression model. The specification of the model used here is adapted from the works of Teklewold et al. (2013), Danso-Abbeam and Baiyegunhi (2018), Kassie et al. (2018), and Belay and Mengiste (2021). The multinomial endogenous switching regression model is estimated in two stages. In the first stage, multinomial logit model (MNL) is used to model farmers’ adoption of the alternative technologies. In the second stage, the impact of each technology packages on wheat productivity is estimated using the ordinary least square (OLS) with selectivity correction computed from the first stage, as the technique proposed by (Dubin & McFadden, 1984). The study employed the inverse-probability-weighted regression adjustment (IPWRA) as robustness check in the setting of the propensity score framework to compute the effect of each alternative technology packages on the outcome variables.

Suppose that a rational farm household i with the primary objective of maximizing productivity, Qi by comparing the benefits s/he enjoys from adopting z alternative alternatives. The rational farmer will choose package j, over any alternative package z if the net benefit is positive, implies Q_ij > Q_iz, j ≠ z. The adoption equation for the farm households multiple choices can be stated as:

(1) $Q_{ij}^{\ast }={\beta }_k{\mathrm{X}}_i+{\mu }_{ij}$(1)

Where $Q_{ij}^{\ast }$ is the latent variable defining the expected net benefits a farmer derives from adopting package k, X_i represents observed covariates (demographic, socioeconomic, institutional, and farm-specific, among others) and ${\mu }_{ij}$ is an error term accounting for unobserved characteristics. Let $A$ be an index that indicates the choice the farmer has made, such that:

(2) $A=\left\{\begin{array}{c}1if\\ .\\ .\\ .\\ kif\end{array}\right.\begin{array}{c}{Q}_{i1}^{\ast }>0\mathrm{m}\mathrm{a}\mathrm{x}\left(Q_{iz}\right)or\\ .\\ forallz\ne k\\ .\\ Q_{iJ}^{\ast }>0\mathrm{m}\mathrm{a}\mathrm{x}(Q_{iz}^{\ast })or\end{array}\begin{array}{c}{\epsilon }_{i1}<0\\ .\\ .\\ .\\ {\epsilon }_{iJ}<0\end{array}$(2)

Where ${\eta }_{ij}$ is the expected difference in productivity between alternative technology packages $j$ and $z$. Hence, $i^{th}$ farm household will adopt alternative technology package $j$ if and only if package k gives the greatest expected benefit than any other packages. Thus, if ${\eta }_{ij}=\mathrm{m}\mathrm{a}{\mathrm{x}}_{z\ne j}(Q_{ij}^{\ast }-Q_{iz}^{\ast })>0$.

If the error term, ${\mu }_{ij}$ has an identical and Gumbel distribution (i.e. under the assumption independence of irrelevant alternatives (IIA) hypothesis), then the probability that an i^th farm household will select package k can be expressed by multinomial logit model and specified as:

(3) $P_{ik}=Pr({\epsilon }_{ij}<0\left|X_i\right.)=\frac{\mathrm{e}\mathrm{x}\mathrm{p}\left(X_i{\beta }_k\right)}{{\sum }_{z=1}^k\mathrm{e}\mathrm{x}\mathrm{p}\left(X_i{\beta }_z\right)}$(3)

Table 1 presents farmers’ choice of alternative technology package. The adoption in this study is defined as adoption of fertilizer and/or improved seed by the farm household’s’ in their crop land. It equals choice 2 if a farmer has adopted the fertilizer, equals 3 if a farmer has adopted the improved seed, equals 4 if the farmer has adopted a combination of fertilizer and improved seed technology, and 1 if a farm household’s’ didn’t adopt any of the above technologies. Table 1 shows that, from the total 693 sampled farm households, about 23.38% are non-adopters (F₀I₀), whereas 26.55% and 6.35% of them adopted fertilizer (F₁I₀) and improved seed (F₀I₁), respectively, and finally, about 43.72% of them have adopted a combination of fertilizer and improved seed (F₁I₁) simultaneously.

Table 1. Farmers’ choice of alternative technology package and their frequency

Choice	Binary package	Fertilizer		Improved seed		Frequency	Percentage
		Fo	F1	Io	I1
1	F0I0	√		√		162	23.38
2	F1I0		√	√		184	26.55
3	F0I1	√			√	44	6.35
4	F1I1		√		√	303	43.72

Note. Each element in the combination is a binary variable and for fertilizer (F) and improved seed adoption (I), and the subscripts represent 1 = adoption and 0 = non-adoption.

After the specification of the selection equation, the outcome equation which is the impact of agricultural technology adoption on wheat productivity⁴ is estimated by using multinomial endogenous switching regression model. As depicted in Table 1, the base category, non-adoption (F₀I₀) is represented by k = 1, while the remaining combination alternatives are represented by k = 2; 3 & 4 where at least one option is adopted. The outcome equation for each possible regime is specified as:

(4) $\left\{\begin{array}{c}Regime1:\\ .\\ .\\ .\\ Regimek:\end{array}\right.\begin{array}{c}{Q}_{i1}={\beta }_1{Z}_i+{\psi }_{i1}ifj=1\\ .\\ .\\ .\\ Q_{ik}={\beta }_k{Z}_i+{\psi }_{ik}ifj=k\end{array}$(4)

Where Q_i’s are the outcome variables, which represent wheat productivity for non-adopter and adopters, $Z_i$ denote a set of explanatory variables that influence wheat productivity; ${\beta }_i\prime \mathrm{s}$ are vectors of parameter to be estimated, and ${\psi }_i^{\prime }s$ represents error terms with zero mean and constant variance assumptions. If the error terms of equations (1)(1) $Q_{ij}^{\ast }={\beta }_k{\mathrm{X}}_i+{\mu }_{ij}$(1) and (4(4) $\left\{\begin{array}{c}Regime1:\\ .\\ .\\ .\\ Regimek:\end{array}\right.\begin{array}{c}{Q}_{i1}={\beta }_1{Z}_i+{\psi }_{i1}ifj=1\\ .\\ .\\ .\\ Q_{ik}={\beta }_k{Z}_i+{\psi }_{ik}ifj=k\end{array}$(4) ) are not independent and identically distributed, a consistent OLS estimate of parameters needs the addition of the selection correction terms of the alternative choices in equation (4)(4) $\left\{\begin{array}{c}Regime1:\\ .\\ .\\ .\\ Regimek:\end{array}\right.\begin{array}{c}{Q}_{i1}={\beta }_1{Z}_i+{\psi }_{i1}ifj=1\\ .\\ .\\ .\\ Q_{ik}={\beta }_k{Z}_i+{\psi }_{ik}ifj=k\end{array}$(4) . Given this, consistent estimates can be found by adding the selection correction terms (mills ratio) generated from the selection equation as follows:

(5) $\{Regime1:...Regimek:Q_{i1}={\beta }_1{Z}_i+{\gamma }_1{\lambda }_1+{\vartheta }_{i1}if\text{j}=1...Q_{ik}={\beta }_k{Z}_i+{\gamma }_k{\lambda }_k+{\vartheta }_{ik}if\text{j}=k$(5)

Where ${\gamma }_i$ is the covariance between the error terms $\mu$ of equation (1)(1) $Q_{ij}^{\ast }={\beta }_k{\mathrm{X}}_i+{\mu }_{ij}$(1) and $\psi$ of equation (4)(4) $\left\{\begin{array}{c}Regime1:\\ .\\ .\\ .\\ Regimek:\end{array}\right.\begin{array}{c}{Q}_{i1}={\beta }_1{Z}_i+{\psi }_{i1}ifj=1\\ .\\ .\\ .\\ Q_{ik}={\beta }_k{Z}_i+{\psi }_{ik}ifj=k\end{array}$(4) , and ${\lambda }_k$ is the inverse Mills ratio that is computed from the multinomial logit model in equation (3)(3) $P_{ik}=Pr({\epsilon }_{ij}<0\left|X_i\right.)=\frac{\mathrm{e}\mathrm{x}\mathrm{p}\left(X_i{\beta }_k\right)}{{\sum }_{z=1}^k\mathrm{e}\mathrm{x}\mathrm{p}\left(X_i{\beta }_z\right)}$(3) as follows:

(6) ${\lambda }_k=\sum _{z\ne k}^K{\rho }_k\left(\frac{{\hat{P}}_{iz}\mathrm{l}\mathrm{n}\left({\hat{P}}_{iz}\right)}{1-{\hat{P}}_{iz}}+ln\left({\hat{P}}_{ik}\right)\right)$(6)

Where $\rho$ is the correlation coefficient of the three error terms $\mu ,$ $\psi$ and $\vartheta$. The error terms have expected mean value of zero. Practically, there is a possibility of heteroscedasticity in generating the regressor in the model, hence; standard errors are bootstrapped to account the problem.

According to Chamberlain and Griliches (1975), a system of equations does not necessarily require instrumental variables for identification. However, for consistent estimates of the parameters, Teklewold et al. (2013) suggested the significance of including selection instruments in the alternative choice model, which is specified in equation (4)(4) $\left\{\begin{array}{c}Regime1:\\ .\\ .\\ .\\ Regimek:\end{array}\right.\begin{array}{c}{Q}_{i1}={\beta }_1{Z}_i+{\psi }_{i1}ifj=1\\ .\\ .\\ .\\ Q_{ik}={\beta }_k{Z}_i+{\psi }_{ik}ifj=k\end{array}$(4) . According to Di Falco et al. (2011), for the ESR model to be adequately identified, it is recommended to use exclusion restriction test. This test is conducted by excluding variables that directly influence the adoption equation but not the outcome equation. Accordingly, this study used variables of extension visit, farmers’ cooperative, distance from the road, and from the market for selection instruments.

2.5.1. Conditional expectations and treatment effects of adoption

The multinomial endogenous switching regression model is used to compute both the average treatment effect on the treated (ATT) and the untreated (ATU). This can be done by a simple comparison of the expected values of the outcome of the treated (adopters) and untreated (non-adopters) in actual and counterfactual situations. Following, Teklewold et al. (2013), Danso-Abbeam and Baiyegunhi (2018), and Kassie et al. (2018); the conditional expectations for the outcome variables in both the observed and their counterfactual scenarios can be specified as follows:

Actual expectations observed in the sample:

(7a) $E\left(Q_{ik}/j=k\right)={\beta }_k{Z}_{ki}+{\sigma }_k{\lambda }_k$(7a)

(7b) $E\left(Q_{i1}/j=1\right)={\beta }_1{Z}_{1i}+{\sigma }_1{\lambda }_1$(7b)

Counterfactual expected outcomes (not observed in the sample):

(7c) $E\left(Q_{ik}/j=1\right)={\beta }_1{Z}_{ki}+{\sigma }_1{\lambda }_k$(7c)

(7d) $E\left(Q_{i1}/j=k\right)={\beta }_k{Z}_{1i}+{\sigma }_k{\lambda }_1$(7d)

The average treatment effect on the treated (ATT) is calculated as the difference between eq. (7a)(7a) $E\left(Q_{ik}/j=k\right)={\beta }_k{Z}_{ki}+{\sigma }_k{\lambda }_k$(7a) and eq. (7c(7c) $E\left(Q_{ik}/j=1\right)={\beta }_1{Z}_{ki}+{\sigma }_1{\lambda }_k$(7c) ); and specified as:

(8) $ATT=E\left(Q_{ik}/j=k\right)-E\left(Q_{ik}/j=1\right)=Z_{ki}({\beta }_k-{\beta }_1)+{\lambda }_k\left({\sigma }_k-{\sigma }_1\right)$(8)

Similarly, the average treatment effect on the untreated (ATU) is the difference between equation (7b)(7b) $E\left(Q_{i1}/j=1\right)={\beta }_1{Z}_{1i}+{\sigma }_1{\lambda }_1$(7b) and equation (7d)(7d) $E\left(Q_{i1}/j=k\right)={\beta }_k{Z}_{1i}+{\sigma }_k{\lambda }_1$(7d) and can be specified as:

(9) $ATU=E\left(Q_{i1}/j=1\right)-E\left(Q_{i1}/j=k\right)=Z_{1i}({\beta }_k-{\beta }_1)+{\lambda }_k\left({\sigma }_k-{\sigma }_1\right)$(9)

The difference between eq. (8)(8) $ATT=E\left(Q_{ik}/j=k\right)-E\left(Q_{ik}/j=1\right)=Z_{ki}({\beta }_k-{\beta }_1)+{\lambda }_k\left({\sigma }_k-{\sigma }_1\right)$(8) and eq. (9)(9) $ATU=E\left(Q_{i1}/j=1\right)-E\left(Q_{i1}/j=k\right)=Z_{1i}({\beta }_k-{\beta }_1)+{\lambda }_k\left({\sigma }_k-{\sigma }_1\right)$(9) provides in transitional heterogeneity (TH) that shows whether the effect of adoption is higher or lower for the adopters than the non-adopters.

3. Empirical results and discussion

3.1. Descriptive statistics

The descriptive statistics of the explanatory variables for the four alternative technology packages considered in this study are presented in Table 2. The explanatory variables’ mean value of non-adopters (F₀I₀) is used as a base category to compare with mean values of alternative adopters (F₁I_0, F₀I₁ and F₁I₁). The summary shows that the average mean comparison test between adopters and non-adopters is significantly higher for adopters. The mean values of explanatory variables are also significantly different crosswise the different adopters. For instance, the summary shows that most of the adopters are male headed households, have higher access to credit service, extension visit, off-farm activities, and membership to farmer cooperatives, which are very important for agricultural technology adoption. Moreover, the average family size and livestock asset are higher for adopters. In addition, on average, the adopter households are located near to the market and urban centers than their counterparts significantly.

Table 2. Description and descriptive summary of variables used for the regressions

Dependent variable for the outcome equation: Wheat Productivity (measured in Quintal per hectare)
Explanatory Variables	Description and measurement	Expected Sign	Adopters			Non-Adopters
			F1I0	F0I1	F1I1	F0I0
			Mean	Mean	Mean	Mean
Sex	Sex of the household head = 1 if male; 0 if female	±	.929 (.018)***	.909 (.043)	.920 (.015)***	.888 (.024)
Education	Education level of the household head = 0 if illiterate; 1 if grade 1–8; 2 if grade 9–12; 3 if above grade 12	+	.576 (.045)***	.840 (.107)	.831 (.042)	.851 (.057)
Access to credit	=1 if the farm household had taken loan; 0 otherwise	+	.614 (.035)***	.272 (.067)***	.557 (.028)**	.469 (.039)
Off-farm Employment	=1 if the farm household had participated; 0 otherwise	±	.635 (.035)***	.5 (.076)	.696 (.026)***	.549 (.039)
Saving	=1 if the farm household had saving; 0 otherwise	+	.760 (.031)	.795 (.061)	.867 (.019)***	.728 (.035)
Extension visit	=1 if household had extension visit during their practice; 0 otherwise	+	.940 (.017)***	1.00 (0.00)***	.953 (.012)***	.876 (.025)
Farm cooperative	=1 if the household is member of farm cooperative; 0 otherwise	+	.847 (.026)**	.818 (.058)	.917 (.015)***	.796 (.031)
Age	Age of the household head, measured in years	±	44.08 (10.92)**	42.11 (12.10)	43.97 (10.89)**	42.11 (11.58)
Family size	The total number of household size	±	5.14 (1.98)***	4.27 (1.82)	5.19 (2.06)***	4.59 (1.82)
TLU	Total Livestock herd size of the household measured in Tropical Livestock Unit (TLU)	±	6.64 (4.25)***	5.94 (4.17)	6.32 (4.23)***	5.17 (3.59)
Farm size	Cultivated wheat farm size of the household in hectare	+	.512 (.322)***	.593 (.321)***	.950 (2.82)	.836 (1.55)
Distance from market	Distance from home to the nearest market (in kilometer)	-	7.15 (6.40)***	6.85 (7.01)***	7.91 (7.81)***	17.84 (11.18)
Distance from all-weather road	Walking distance from home to all road centers (in kilometer)	-	5.12 (5.78)***	5.83 (7.03)***	6.14 (7.52)***	13.84 (10.32)

Note. Standard errors in parentheses

Mean comparison test is used to compare the means of explanatory variables between non-adopters and adopters of each packages of alternative technology, and the signs **, *** signifies significance level at 5% and 1%, respectively.

3.2. Econometrics analysis

3.2.1. Determinants of adoption of agricultural technology

The study applies a maximum likelihood estimation technique for the purpose of estimating the multinomial logit functions, and the results are presented in Table 3. For the purpose of effective estimation of the model, several diagnostics tests were conducted. The results of the Wald test accepts the alternative hypothesis that all regression coefficients are jointly different from zero (P > chi² = 0.000). The Hausman test result for test of IIA assumption displays that all the alternative packages are distinguishable with respect to the variables in the model, as presented in Appendix B. During the estimation, robust regression is used to control the problem of heteroskedasticity and non-normality.

Table 3. Maximum likelihood estimates for the multinomial logit model

Base category: Non-adopters	Adopters of
Variables	F1I0	F0I1	F1I1
	Coefficient (SE)	Coefficient (SE)	Coefficient (SE)
Sex	0.129 (0.424)	−0.030 (0.597)	−0.075 (0.393)
Age	0.005 (0.012)	0.015 (0.018)	0.009 (0.011)
Family size	0.070 (0.072)	−0.108 (0.113)	0.107 (0.067)
Education (base: illiterate)
Grade 1–8	0.011 (0.275)	0.336 (0.427)	0.412 (0.158)**
Grade 9–12	0.813 (0.467)*	0.284 (0.581)	0.126 (0.374)
Grade >12	0.228 (0.068)***	0.404 (0.045)***	0.049 (0.758)
Off-farm employment	0.294 (0.263)	−0.304 (0.361)	0.531 (0.247)**
TLU	0.050 (0.023)**	0.061 (0.049)	0.013 (0.003)***
Saving	0.245 (0.321)	0.723 (0.477)	0.705 (0.313)**
Distance to market	−0.101 (0.024)***	−0.128 (0.025)***	−0.104 (0.019)***
Distance to main road	−0.052 (0.026)**	−0.018 (0.027)	–0.022 (0.018)
Extension visit	0.953 (0.462)**	0.958 (0.316)***	1.275 (0.459)***
Credit access	0.329 (0.151)**	1.033 (0.403)**	0.082 (0.230)
Farm cooperative	0.067 (0.370)	−0.520 (0.525)	0.221 (0.395)
Farm size	1.330 (0.300)***	−0.571 (0.355)	0.091 (0.082)
Constant	0.343 (0.856)	−0.423 (1.100)	−1.375 (0.822)*

Note. ***significant at 1% level; **significant at 5% level; *significant at 10% level; Standard errors in parentheses.

Table 3 shows that the educational level of the household head have a positive and significant effect on the adoption of all packages (F₁I₀, F0I₁, and F₁I₁), meaning that having an educated household head increases the chance of adopting the combination of improved wheat seed and/or inorganic fertilizer. This is because education enables to obtain, analyze, and appraise information on new or improved technologies, market opportunities and its implied benefit. This is consistent with Wudu (2017) and Tamirat et al. (2020). Tropical livestock unit (TLU) is also found to have a positive and significant influence on the adoption package of fertilizer (F₁I₀) and improved seed (F₀I₁), representing that households who have a livestock asset are more expected to adopt F₁I₀ and F₀I₁ than the non-adopters. Since farmers who own a flock of livestock helps to have an additional source of income and serve as a source of organic fertilizer input. This is in line with (Tesfaye et al., 2016). The coefficient of wheat farm size is positive and significant for the adoptions of package fertilizer (F₁I₀), showing that having a large farm size increases the probability of adopting fertilizer technology than non-adopters. The possible explanation is that farm households who have large farm size can produce more than household consumption and the surplus amount of production is sold and this generates a certain amount of cash that can be used to purchase agricultural inputs. The finding is consistent with (Wudu, 2017).

Engagement in off-farm works is found to have a direct effect on the adoption of a combined fertilizer and improved seed (F₁I₁), which implies that farmers participating in off-farm activities are more likely to adopt (F₁I₁) than the counterfactual. This is because engagement in off-farm activities can generate additional income, and it can be used to solve liquidity problem that the farm household’s face while intending to purchase agricultural technologies. This finding is similar with (Wudu, 2017). The coefficient of saving is positive and significant on the adoption package of full technology (F₁I₁); showing that households who had saved money are more expected to adopt farm technologies than their counterparts. This is for the reason that saving assists as a means of overcoming liquidity constraint and as a means of buying inputs for agricultural production. This is in line with (Belay & Mengiste, 2021). The result also shows that access to credit increases the likelihood of the adoption of F₁I₀ and F₀I₁, suggesting that farm households who get credit access are more likely to adopt farm technologies. Those farmers who have credit access significantly ease their liquidity limitations they face while they want to purchase new or improved farm technologies (Belay & Mengiste, 2021; Tamirat et al., 2020).

As expected, the result for the extension visit significantly and positively affects the adoption of all packages, meaning that farmers who have an extension visit during their adoption practice of agricultural technologies are more likely to adopt farm technologies than their counter parts. This is because that extension agent assists the farm households to get the required information (characterization), application, and benefits of agricultural technologies. The result is similar with (Hagos et al., 2016 and Wudu, 2017). Lastly, the results also reveal that distance to the market and main road significantly affects the adoption of agricultural technologies, indicating that households that live near to the market places and urban centers are expected to adopt farm technologies than the non-adopters. Since farmers may possibly have higher access to information on new or improved farm technologies, information on input (output) prices, and also could lead to timely adoption, and lower production cost, and hence are more likely to adopt, which is in line with the results of (Belay & Mengiste, 2021; Wudu, 2017).

3.2.2. Determinants of wheat productivity

The paper deliberates the impact of agricultural technology adoption on wheat productivity, estimated using multinomial endogenous switching regression model. Using this model, the study estimates the outcome equation (wheat productivity), and the results are presented in Table 4. The result shows that the variables like education level of the household, off-farm work participation, wheat farm size, livestock in terms of TLU, and credit access are significantly associated with wheat yield. More specifically, being an educated farmer helps to produce more wheat yields; farm households who actively engage in off-farm work increases wheat productivity through generating additional incomes; farm households who have high wheat farm size, a flock of livestock asset, and having access to credit services produces more wheat yields than their counterparts. The results show that some variables have a different effect between adopters and non-adopters. These differences in the sign of coefficients reflect the existence of heterogeneity among adopters and non-adopters (Di Falco et al., 2011). The inverse mills ratios are also significant in the estimation results of the outcome equations, indicating there is self-selection bias in technology set choice, proofing that using the ESR model is appropriate. The falsification test results show that (P = 0.378), the selected instruments are valid and the model is adequately identified, as they are highly insignificant at 1% level, see Appendix A.

Table 4. Estimated results of multinomial endogenous switching regression

Variables	Yield/hectareF0I0	Yield/hectareF1I0	Yield/hectareF0I1	Yield/hectareF1I1
	Coefficient (SE)	Coefficient (SE)	Coefficient (SE)	Coefficient (SE)
Sex	−.611 (3.48)	4.51 (3.87)	4.53 (193.6)	−2.51 (3.63)
Age	−.041 (.065)	−.037 (.034)	−.292 (.319)	.053 (.244)
Education	1.13 (1.38)	3.08 (1.81)*	−1.45 (0.86)	4.48 (1.67)**
Family size	−.225 (1.14)	−.658 (.778)	1.171 (1.18)	−.863 (.758)
Saving	−.304 (2.81)	11.07 (8.50)	−9.73 (8.03)	−.971 (3.31)
Off-farm employment	1.84 (1.83)	4.53 (1.17)***	−5.59 (6.78)	1.59 (3.54)
Farm size	1.36 (.621)**	23.94 (8.93)***	6.13 (1.49)***	.359 (.276)
Credit access	4.86 (2.51)**	4.37 (2.51)*	13.96 (204.6)	3.91 (1.44)**
TLU	.277 (.137)**	.611 (.021)***	.419 (0.13)***	.046 (.018)**
_m0		−7.81 (13.23)	−11.05 (185.6)	24.51 (7.30)***
_m1	−1.78 (10.59)		50.54 (42.28)	.389 (9.93)
_m2	1.08 (19.44)	9.59 (18.56)		−22.45 (8.44)**
_m3	−4.16 (13.97)	12.96 (3.59)***	−28.89 (306.18)
Sigma2	98.64 (327.86)	519.71 (102.51)***	3084.3 (208.87)	1116.5 (251.5)***
rho0		−.439 (.563)	−.255 (.635)	.940 (.216)***
rho1	−.230 (.571)		1.16 (.836)	.014 (.448)
rho2	.139 (1.10)	.539 (.068)***		−.861 (.237)**
rho3	−.537 (.771)	.729 (.807)	−.667 (0.21)***
Constant	11.00 (8.67)	36.73 (16.05)*	77.00 (462.01)	41.00 (14.80)

Note. ***significant at 1% level; **significant at 5% level; *significant at 10% level; Standard errors in parentheses.

3.2.3. Impact of agricultural technology adoption on wheat productivity

Table 5 provides the impact of adoption on wheat productivity. The results are estimated by comparing the actual wheat crop yield with the respective counterfactual scenarios. The results show that the adoption of improved wheat seed and/or fertilizer packages grants higher wheat yields as compared with non-adoption. Explicitly, farm households who adopted would have gained lower if they had decided not to adopt and those farm households who did not adopt would have gained higher if they decided to adopt.

Table 5. Estimation of conditional expectations and treatment effect using MESM

Alternative		Decision stage
		To Adopt	Not to adopt	Adoption Effect
Adopters of	F1I0	30.37 (.446)	21.10” (.343)	TT^1 = 9.26***
	F0I1	27.91 (.863)	21.96” (.668)	TT^2 = 5.94***
	F1I1	35.73 (.408)	20.77” (.349)	TT^3 = 14.96***
Non-adopters	F0I0	26.92” (.927)	15.18 (.430)	TU^1 = 11.73***
	F0I0	20.01” (.629)		TU^2 = 4.83***
	F0I0	26.87” (.548)		TU^3 = 11.69***
				TH^1 = −2.46**
				TH^2 = 1.11
				TH^3 = 3.27***

Notes. TT = Adoption effect for adopters, TU = Adoption effect for non-adopters, TH (TT-TU) = transitional heterogeneity

Note. Standard errors in parentheses, and *** and ** denotes significance level at 1% and 5% respectively.

In detail, the highest wheat yield (35.73 quintals) is obtained when adopters adopt a combined adoption fertilizer and improved seed (F₁I₁). The second highest wheat yield 30.37 and 27.91 quintals per hectare are obtained when farm households adopt fertilizer package only and improved seed only, respectively. On the other hand, the average treatment effect on the non-adopters is 15.18 quintals per hectare, but this will increase significantly if they had decided to adopt all the packages.

The transitional heterogeneity effect (TH) for the adoption of fertilizer is negative and significant, which implies that the effect of adoption would be significantly higher for the farm households who had not adopted relative to those adopted, if they had decided to adopt. To sum up, the results of the study validate that the adoption of fertilizer and/or improved seed significantly increased wheat productivity than non-adoption. Moreover, the result shows that adoption of combined technologies grants higher wheat yield than the adoption of a single technology. This finding is in line with the previous works of Hagos (2016), Tesfaye et al. (2016), Abate et al. (2018), Daniel and Belay (2018), and Natnael (2019).

3.2.4. Inverse-probability-weighted regression adjustment (IPWRA)

In this paper, we have used IPWRA as robustness check in the setting of the propensity score framework to compute the impact of each alternative technology packages on the outcome variables. The impact of agricultural technology adoption on wheat productivity is presented in Table 6 with the IPWRA estimates. The result shows that adoption of improved wheat seed (F₀I₁) package produces no significant effect on wheat farm yields. But, adopters of fertilizer (F₁I₀) package and fertilizer and improved seed in combination (F₁I₁) generate a large increase of productivity which is about 9.364 and 13.561 quintals of wheat yield per hectare respectively. Moreover, the combined adoption of fertilizer and improved seed (F₁I₁) generates higher gains than adopting each package in isolation. The results of the average treatment effects presented in Tables 5 and 6 show that there are minimal differences of the estimates of the outcome variables in the MESR and IPWRA. This divergence may be due to some sort of differences in unobserved heterogeneity among farm households (Danso-Abbeam & Baiyegunhi, 2018). But, the results of both methods confirm that the adoption of improved seed and or fertilizer packages significantly increases wheat yield.

Table 6. Average treatment effects of adoption using IPWRA

Technology combination option	Treatment effect	Outcome variable
		Wheat yield (quintal/ha)
Fertilizer only (F1I0)	ATT	9.364 (2.64)***
Improved seed only (F0I1)	ATT	2.275 (3.41)
Fertilizer and improved seed (F1I1)	ATT	13.561 (2.94)***
Baseline potential outcome mean	POM	21.00 (2.34)***

Note. Standard errors in parentheses, and *** denotes significance level at 1%.

4. Conclusion

The aim of this paper is to examine the impact of agricultural technology adoption on wheat productivity in north Shewa zone, Amhara region, Ethiopia, by using a household level data and a multinomial endogenous switching regression model. Accordingly, two major conclusions can be drawn from the results of the study. First, the decision to adopt improved wheat seed and or fertilizer technology packages are significantly determined by the education level of the household head, engagement in off-farm works, livestock asset, credit access, saving, farm size, extension contact, and distance from (market and all-weather roads). Second, adopters of improved wheat seed and or fertilizer packages have benefited from increased wheat yield than non-adopters. Thus, the results confirmed that agricultural technology adoption positively affects crop productivity including wheat. These results suggests that policymakers should have to support and promote the adoption of fertilizer and/or improved seed technologies so as to exploit the productivity effects of agricultural technology adoption. Furthermore, future studies should consider Rosenbaum bounds sensitivity analysis to check if there exist hidden biases that are likely to undermine the causal effect.

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

The authors received no direct funding for this research.

Notes on contributors

Mesele Belay Zegeye

Mesele Belay Zegeye is a lecturer at Debre Berhan University, Ethiopia. His research interest includes the effect and effectiveness of policies, poverty, food security, technology adoption, and agricultural innovations and extensions, and agricultural productivity.

Abebaw Hailu Fikrie

Abebaw Hailu Fikrie is a lecturer at Debre Berhan University, Ethiopia. His current research interests are development issues like urban and rural development, agricultural economics, food security, microeconomic, and macroeconomics policy analysis.

Anteneh Bizualem Asefa

Anteneh Bizualem Asefa is a lecturer at Debre Berhan University. His research interest includes macroeconomics policy analysis and international trade.

Notes

1. The problem arises because farmers’ decision to adopt alternative technology packages may be correlated with the outcome variable of interest (Teklewold et al., 2013). Therefore, separate productivity functions have to be specified for each group of complementary options while at the same time accounting for endogeneity.

2. Farmers make their own decisions (self-select) whether to adopt or not to adopt a particular combination of the package. As a result, their decisions are usually affected by unobserved factors such as demographic, socio-economic, and institutional factors. The unobserved factors may have a direct correlation with the variable of interest (Danso-Abbeam & Baiyegunhi, 2018).

3. It is the problem of being unable to assess what the situation would be like if the technology had not been adopted (Belay & Mengiste, 2021). This is because outcomes can only be observed in one state at a time and the counterfactuals for the adoption of each combination alternatives cannot be observed.

4. Measured as quintal per hectare.

Appendix A.

Falsification Test (OLS Regression)

Linear regression

Table

Productivity	Coef.	St.Err.	t-value	p-value	[95% Conf	Interval]	Sig
Distance from main road	−.435	.257	−1.69	.092	−.943	.072	*
Farm cooperative	1.945	3.758	0.52	.605	−5.471	9.361
Distance from market	−.005	.238	−0.02	.984	−.473	.464
Extension	−2.93	5.54	−0.53	.598	−13.862	8.002
Constant	33.743	6.747	5.00	0	20.429	47.058	***
Mean dependent var	30.371		SD dependent var		17.793
R-squared	0.023		Number of obs		184
F-test	1.059		Prob > F		0.378
Akaike crit. (AIC)	1586.254		Bayesian crit. (BIC)		1602.329

*** p < .01, ** p < .05, * p < .1

Appendix B.

Hausman IIA specification Test

Table

	Coef.
Chi-square test value	0.000
P-value	1.000