Study area

The study was conducted in the Gert Sibande District Municipality in Mpumalanga Province of the Republic of South Africa. Gert Sibande District is a Category C municipality and is the largest of the three districts in the province, measuring 31,841 km² in area, which covers almost half (40%) of Mpumalanga Province's total land mass of 76,495 km². The district, which accounts for about 6.5% of South Africa's land surface, consists of seven local municipalities: Govan Mbeki, Chief Albert Luthuli, Msukaligwa, Dipaleseng, Mkhondo, Lekwa and Dr Pixley ka Isaka Seme. Figure 1 shows the district and its local municipalities. The major towns are Amersfoort, Amsterdam, Balfour, Bethal, Breyten, Carolina, Charl Cilliers, Chrissiesmeer, Davel, Ekulindeni, Embalenhle, Empuluzi, Ermelo, Evander, Greylingstad, Grootvlei, Kinross, Leandra, Lothair, Morgenzon, Perdekop, Secunda, Standerton, Trichardt, Volksrust, Wakkerstroom, eManzana and eMkhondo (Piet Retief).

Fig. 1. Map of Gert Sibande District, Mpumalanga Province of South Africa.

This area was chosen because of its high concentration of subsistence farmers which accounted for about 2000 farmers (Mngqawa et al., Reference Mngqawa, Mangena-Netshikweta and Katerere2016). The area between Carolina, Bethal and Ermelo produces the most sheep and wool in South Africa, which indicates the region's significance to agriculture in South Africa. Also, the area is known for its high maize production. Despite its size, in 2018, Gert Sibande District contributed only 2.06% to the GDP of South Africa, or 27.68% to the Mpumalanga Province's total GDP of R 363 billion, ranking it the lowest relative to all the regional economies in the Mpumalanga Province (SA Stat, 2018). The district was part of the mapping area where CA adoption was introduced, and it is one of the top three maize producing regions in South Africa, accounting for 23.5% of the total maize production in South Africa (Greyling and Pardey, Reference Greyling and Pardey2019). The major economic sectors are mining, agriculture, energy and manufacturing.

Sampling technique and data collection

Farm-level cross-sectional survey data were collected between December 2019 and August 2020 from smallholder maize farmers in the Gert Sibande District of Mpumalanga Province, South Africa, using a semi-structured survey questionnaire validated by two agricultural economists. A test of reliability was performed on the questionnaire to establish its use. The questionnaire contained logic flow questions aimed at farmers' socio-economic characteristics, farm-based characteristics, CA adoption and knowledge about CA. The survey was conducted through 40 min face-to-face interviews with the help of four trained enumerators who received training on CA and related information. The questionnaire was translated into local languages that the farmers can understand for better interpretation and accuracy. Consent from the farmers and approval from the ethical department of the University of South Africa was obtained before the commencement of the questionnaire administration, dated 19 February 2020. A representative sample size (250 maize farmers) was determined from the population of 710 smallholder maize farmers, using Slovin's formula, given in Equation (1), after which a total number of 250 questionnaires were administered to the maize farmers in the district, using a proportionate random sampling technique. Following Oduniyi (Reference Oduniyi2018), this was achieved by adopting a quantitative model, as presented below:

(1)$$n = \displaystyle{N \over {1 + N{( e ) }^2}}$$

where n is the sample size; N equals the total population of maize farmers in the seven local municipalities across the district; e equals maximum variability or margin of error, estimated at 5% (0.05); 1 equals the probability of the event occurring, and 250 equals the number of respondents sampled or sample size. Table 1 shows the distribution of sample size collected according to each municipality or stratum. The data collected were analyzed using STATA 15.

Table 1. Sample size taken in each municipality (stratum)

Source: Author's computation (2021).

A conceptual framework and empirical/estimation techniques

A smallholder farmer's decision to adopt CA practices is a behavioral response, where farmers choose or adopt components of the CA practices that increase their utility or expected profit subject to constraints. The constraints are where the polychotomous adoption decision depends on several factors, such as socio-economic characteristics, available CA information and the benefits of CA. The adoption of any of the components of CA is modeled within the random utility framework (Ali and Abdulai, Reference Ali and Abdulai2010; Kassie et al., Reference Kassie, Marenya, Tessema, Jaleta, Zeng, Erenstein and Rahut2018). A farmer may decide to adopt either a single CA practice or a combination of CA practices, such as MT variety, CD or a combination of minimum tillage and crop diversification (MTCD) if the expected profit from adoption is higher than the expected profit from non-adoption. See Table 2, which explains the different CA choices.

Table 2. Adoption of conservation agricultural (CA) strategies and sample distribution

Source: Author's computation (2021).

Thus, the adoption of CA practices is based on individual choices and may be correlated with unobservable characteristics that could also affect farmers' income. However, a simple comparison of income groups may lead to an inaccurate result. To account for both endogeneity and sample selection, a MESR (multinomial endogenous switching regression) framework was used. More specifically, following Dubin and McFadden (Reference Dubin and McFadden1984), we apply a selective corrected MESR treatment effect approach and use the method by Bourguignon et al. (Reference Bourguignon, Ferreira and Walton2007) to correct for selection bias.

According to Kumar et al. (Reference Kumar, Mishra, Saroj and Joshi2019), MESR has the advantage of evaluating both individual and alternative combinations of practices. MESR also captures both self-selection bias and the interactions between choices of alternative practices (Wu and Babcock, Reference Wu and Babcock1998; Mansur et al., Reference Mansur, Mendelsohn and Morrison2008) as well as unobserved heterogeneity associated with economic evaluations of the non-random adoption of CA practices. The model involves two stages. In the first stage, the maize farmers' choice of combination of CA practices was modeled using a multinomial logit selection model while identifying the interrelations among the CA choices. In the second stage, the impacts of each choice of CA practices on the outcome variable were estimated using ordinary least squares (OLS) with a selectivity correction term from the first stage. Similarly, for identification, a variable named the number of extension visits was used as an instrumental variable.

Multinomial logit (adoption) selection model

Following Kumar et al. (Reference Kumar, Mishra, Saroj and Joshi2019) and Danso-Abbeam and Baiyegunhi (Reference Danso-Abbeam and Baiyegunhi2018), it is assumed that smallholder maize farmer i aims to maximize his/her net returns, π_i, by comparing the positive return from different choices of CA practices, k, (k = 1,2,3,4) over any alternative practice, m, which is given as

(2)$$\eqalignno{& {\rm \pi }_{ik} > {\rm \pi }_{im}\;m\ne k, \;\;{\rm or\ equivalently}\;\Delta {\rm \pi }_{im} = {\rm \pi }_{ik}-{\rm \pi }_{im} > \!0\;m\ne k. \cr & {\rm \pi }_{ik} ^{\;\; \ast} = X_i{\rm \beta }_k + {\rm \varepsilon }_{ik} }$$

where π_ik ^∗ is a latent variable defining the expected net benefits a maize farmer derives from the adoption of CA choices k, which is determined by both observed and unobserved characteristics. X_i represents a vector of observed exogenous or covariates variables, and ɛ_ik is an error term accounting for unobserved characteristics.

Assuming J is the index that signifies maize farmers' choices of CA practices k, such that

(3)$$J = \left\{\matrix{ 1 \quad if \, \pi_{i1}^\ast \gt\! 0 \max ( {\pi_{im}} )\ or \ \eta_{i1} < \!0 \cr for \, all \, m \ne k \cr K \quad if \, \pi_{ik}^\ast \gt\! 0 \max ( {\pi_{im}^\ast } )\ or \ \eta_{iK} < \!0}\right.$$

In Equation (2) above, the index function suggests that maize farmers will adopt CA practice k, if they derive or expect a greater benefit or net returns from other CA practices m.

$${\rm Thus}, \;\;\eta _{ik} = \max \;( {{\rm \pi }_{ik}^{\rm \ast } - {\rm \pi }_{im}^\ast } ) \gt\! 0, \;\;m\ne k$$

Assuming that the error term (ɛ) is identical, and Gumbel is distributed independently, then the probability that maize farmer i with characteristics X_i will adopt CA practice k can be expressed by the use of a multinomial logit model (McFadden, Reference McFadden and Zarembka1973).

(4)$$P_{ik} = \Pr ( {\eta {}_{ik} < \!0/Z_i} ) = \displaystyle{{\exp ( {Z_i\beta_k} ) } \over {\sum {_{m = 1}^K } \exp ( {Z_i\beta_m} ) }}$$

The parameters of the latent variable model are then estimated with a maximum likelihood function.

Multinomial endogenous switching regression (MESR)

The use of MESR was explored in the second stage where the relationship between the outcome variables and a set of exogenous variables is assessed for the adopted CA practices. The model estimated the parameters of the impacts of CA choice sets k (MT₀CD₀ = non-adopters as reference category; MT₁CD₀ = adopters of MT practices; MT₀CD₁ = adopters of crop rotation practices; MT₁CD₁ = adopters of both MT practices and crop rotation practices) on the outcome variables. The outcome equation for each possible regime j is given as

(5)$$\left\{\matrix{{regime}\ 1\ \colon&Y_{i1} = \, \beta_1X_i \, + \, \varepsilon_{i1}&if\ j \, = \, 1 \cr & \colon\cr {regime}\ k \ \colon & Y_{ik} = \, \beta_k X_i \, + \, \varepsilon_{ik}& if\ j \, = \, k}\right.$$

where Y_ik represents the outcome variable of the ith maize farmer associated with the selected regime k, X_i represents a vector of explanatory variables or exogenous covariates, β is the vector of parameters, and ɛ_ik and ɛ_{i 1} are random disturbance terms. Y_ik is observed if, and only if, CA practice k is adopted, which occurs when π_ik ^∗ > max_{m ≠k} (π_im ^∗). However, if the ɛ_k and ɛ₁ are not independent, OLS estimates obtained from Equation (4) will be biased. Thus, consistent estimation of X_i requires the inclusion of the selection bias correction terms of the alternative CA practices m in Equation (4). It is assumed that the linear assumption of the DM (Dubin and McFadden, Reference Dubin and McFadden1984) is given as

(6)$$( {\delta_{ik}/\varepsilon_{i1}\ldots ..\varepsilon_{iK}} ) = \sigma _k\sum\limits_{m\ne k}^K {r_k( {\varepsilon_{im}-E\,( {\varepsilon_{im}} ) } ) } $$

With this assumption, the MESR in the equation above can be expressed as

(7)$$\matrix{{regime \ 1 \ \colon \quad Y_{i1} = \beta _1X_i + \sigma _1\lambda _1{\rm} + \varphi _{i1}\quad if \ j = 1}\cr{regime \ k \ \colon \quad Y_{ik} = \beta _kX_i + \sigma _k\lambda _k{\rm} + \varphi _{ki} \quad if\ j = k} }$$

where σ_k is the covariance between the error terms δ in Equation (1) and ɛ in Equation (4); φ's are the error terms with an expected value of zero, and λ_k is the Inverse Mills ratio computed from the MESR estimate in Equation (2).

(8)$$\lambda _k = \sum\limits_{m\ne k}^K {\rho _k\left[{\displaystyle{{{\widehat{P}}_{im} \ln ( {{\widehat{P}}_{im}} ) } \over {1-{\widehat{P}}_{im}}} + \ln ( {{\widehat{P}}_{ik}} ) } \right]} $$

where ρ _k defines the correlation coefficient of the three error terms, ɛ, ẟ and φ. However, the heteroscedasticity problem, which could arise from the generated regressor λ_k, was accounted for by the use of bootstrap errors.

Conditional expectations and estimation of average treatment effects

MESR was used to compute the treatment effect. This framework can compute both the average treatment effect on the treated (ATT) and on the untreated (ATU). This was done by comparing the expected outcome value of the treated (each choice of CA adopted) and the untreated (non-adopters) in actual and counterfactual scenarios. Following Di Falco and Veronesi (Reference Di Falco and Veronesi2013), the ATT in the actual and counterfactual situation can be expressed as

Adopters with adoption (actual expectations observed in the sample):

(9a)$$\left\{\matrix{{E \, ( {Y_{i2}/j \, = \, 2} ) \, = \, \beta_2X_{2i} \, + \, \sigma_2\lambda_2 }\cr\colon\cr{E \, ( {Y_{ik}/j \, = \, k} ) \, = \, \beta_kX_{ki} \, + \, \sigma_k\lambda_k}}\right.$$

(9b)$$\left\{\matrix{{ E \, ( {Y_{i1}/j \, = \, 1} ) \, = \, \beta_1X_{1i} \, + \, \sigma_1\lambda_1 }\cr \colon \cr {E \, ( {Y_{i1}/j \, = \, 3} ) \, = \, \beta_3X_{3i} \, + \, \sigma_3\lambda_3}}\right.$$

Adopters, had they decided not to adopt (counterfactual expected outcomes):

(10a)$$\left\{\matrix{E \, ( {Y_{i1}/j \, = \, 2} ) \, = \, \beta_1X_{2i} \, + \, \sigma_1\lambda_2 \cr \colon \cr E\,( {Y_{i1}/j \, = \, k} ) \, = \, \beta_1X_{ki} \, + \, \sigma_1\lambda_k}\right.$$

(10b)$$\left\{\matrix{E \, ( {Y_{i2}/j \, = \, 1} ) \, = \, \beta_2X_{1i} \, + \, \sigma_2\lambda_1 \cr \colon \cr E \, ( {Y_{i2}/j \, = \, 3} ) \, = \, \beta_2X_{i3} \, + \, \sigma_2\lambda_3}\right.$$

Equations (8a) and (8b) represent the real expectations observed in the sample. Similarly, Equations (9a) and (9b) represent the counterfactual expected outcome. In order to compute the average treatment effect on the treated (ATT), the difference between Equations (8a) and (9a) was calculated. This could be expressed as

(11)$$\eqalign{{\rm ATT} = & \, E\,[ {Y_{i2}/j = 2} ] - E \, [ {Y_{i1}/j = 2} ] \cr = & \, X_{2i}\,( {\beta_2-\beta_1} ) + \lambda _2 \, ( {\sigma_2-\sigma_1} ) } $$

Correspondingly, the average treatment effect on the untreated (ATU) is the difference between Equations (8b) and (9b), given as

(12)$$\eqalign{{\rm ATU} = & \, E \,[ {Y_{i1}/j = 1} ] -E \, [ {Y_{i2}/j = 1} ] \cr = & \, X_{1i}\,( {\beta_2-\beta_1} ) + \lambda _1( {\sigma_2-\sigma_1} ) } $$