3.1. Sampling, Data, and Measurements

The data for this study was obtained through a recent farm household survey we conducted, in partnership with PAS in the North East and Upper East Regions of Ghana. As stated earlier, PAS, AGRA, and NRGP have under different projects at various locations within the study area, primed farmers to use organic fertilizer to sustain soil health. Except for the Bunkpurugu area where there is no PAS extension agent, we followed the operational areas of PAS to sample the study participants. Prior to sampling, a discussion was held with MoFA (Bunkpurugu area) and the PAS extension agents to identify the types of organic fertilizer and related practices the farmers currently use.

We selected 504 households across 52 communities clustered around 3 PAS and 1 MoFA extension zones through a multistage cross-sectional sampling approach. At the first stage, we purposively selected the PAS and MoFA extension zones to ensure equal chances of including farmers the various organic fertilizer types in the sample. PAS and MoFA extension agents in the selected zones provided lists of farming communities from which we randomly drew 30% of the communities to represent each zone. At the community level, farmer-group and opinion leaders helped enumerators to identify and compile lists of organic fertilizer adopter and non-adopter households. Households were then randomly selected from the lists to represent each subsample group. Farmers who applied, at least, one ton per acre of any biomass (e.g., crop residue, animal manure, compost, domestic/urban waste, agro by-products, excreta slurry, or a combination of these) constituted the organic fertilizer user group (adopters)Footnote ⁷.

The data was collected through a face-to-face personal interview (PI) with sample participant farmers. Appendix 4A shows the number of communities and sampled households across the study locations. Since the main objective is to evaluate the effects of organic fertilizer on food access and labor use, the questionnaire elicited data on these as the outcome variables of organic fertilizer adoption. For food access, we considered two indicator variables critical in farm household settings: per capita food expenditure and food stability through self-provision (indicated by FG) ^{Footnote 8} . Per capita food expenditure was measured in Ghana Cedis based on market value of adult equivalent units (AEU) ^{Footnote 9} of food consumed, while household FG was captured as the duration of the period (in months) a household had difficulties securing sufficient foodstuff (Tambo and Wünscher, Reference Tambo and Wünscher2017). We used the AEU for per capita food expenditure because it is more directly linked to food access status of subsistent farm households and is less prone to measurement errors than the income-based indicators (Tambo and Wünscher, Reference Tambo and Wünscher2017). Following Tambo and Wünscher (Reference Tambo and Wünscher2017) and Martey (Reference Martey2018), our questionnaire used a 7-day recall period to identify food items consumed by a household and thus captured the corresponding expenditure. Non-purchased food items were valued at current market prices. Regarding labor use, farmers reported the number of worker-days used per acre during the most recent crop season. This was subdivided into male and female worker-days ^{Footnote 10} used in order to examine gender-based effects of organic fertilizer adoption on labor.

Following the strands of literature on farm technology in general (e.g., Di Falco et al., Reference Di Falco, Veronesi and Yesuf2011; Fuglei and Bosch, Reference Fuglie and Bosch1995; Manthenge et al., Reference Mathenge, Smale and Olwande2014; Sheferaw et al., Reference Shiferaw, Kassie, Jaleta and Yirga2014; Tambo and Wünscher, Reference Tambo and Wünscher2017; Teklewold et al., Reference Teklewold, Kassie, Shiferaw and Köhlin2013) and specifically, organic fertilizer adoption (e.g., Chen et al., Reference Chen, Zeng, Xu and Fan2018a, Reference Chen, Camps-Arbestain, Shen, Singh and Cayuela2018b; Martey, Reference Martey2018; Ullah et al., Reference Ullah, Shah, Ali, Naz, Mahar and Kalhoro2015; Waithaka et al., Reference Waithaka, Thornton, Shepherd and Ndiwa2007), the survey also captured data on six categories of farmer background factors (explanatory variables). These moderator variables are farmer/household characteristics, resource constraints, and plot-specific factors. Other groups of explanatory variables include social capital, governance and institutional variables, information access factors, and environmental shocks. Empirically, the literature has proven that these variables significantly influence farmers’ decision to adopt farm technology (Ayalew et al., Reference Ayalew, Chamberlin and Newman2020; Shiferaw et al., Reference Shiferaw, Okello and Ratna Reddy2009; Teklewold et al., Reference Teklewold, Kassie, Shiferaw and Köhlin2013). Table 1 shows the description and summary statistics, including mean differences between adopters and non-adopters regarding for all explanatory variables used in the study.

Table 1. Summary statistics of variables

3.2. Theoretical Model and Estimation Strategy

Assuming farm households aim at maximizing access to food through organic fertilizer farm enterprise under optimal input use. We assume further that they are risk-neutral, and consider only the net benefits to be derived between organic fertilizer use and otherwise when deciding to adopt it. The decision task then is to choose the option that maximizes their welfare, subject to household resource endowment or constraints. Subsequently, the observed inputs used and resultant welfare indicators (food expenditure and FG) are outcomes of the binary adoption decision a household made (Sanglestsawai et al., Reference Sanglestsawai, Rejesus and Yorobe2015; Teklewold et al., Reference Teklewold, Kassie, Shiferaw and Köhlin2013).

To specify the adoption model, let Q _ja and Q _jn represent the net outcome farmer j derives from organic fertilizer adoption and non-adoption, respectively. The farmer adopts organic fertilizer if Q _ja >Q _jn (i.e., the benefit from adoption exceeds that from non-adoption). Also, let $$U_j^*$$ (the difference between Q _ja and Q _jn ) be a latent criterion variable on which farmer j bases his/her decision. Further, let D _j indicates the observed binary decision such that $$U_j^*$$ = Q _ja −Q _jn >0 when D _j = 1 (i.e., if farmer j adopts organic fertilizer) and Q _ja −Q _jn ≤ 0 when D _j = 0 (i.e., when farmer j does not adopt organic fertilizer). By expressing $$U_j^*$$ as a function of observed farmer characteristics denoted Z, we obtain a latent variable (adoption or selection) model that sorts farmers into adopters and non-adopter as (Di Falco et al., Reference Di Falco, Veronesi and Yesuf2011):

(1) $$U_j^* = \gamma '{Z_j} + {\varepsilon _j}{\mkern 1mu} {\text{with}}{\mkern 1mu} {D_j} = \left\{ \begin{gathered} 1\,\,\,\,\,\,if\;U_j^* > 0 \hfill \\ 0\,\,\,\,\,otherwise \hfill \\ \end{gathered} \right\}$$

where Z _j is a vector of household and farm characteristics; $$\;\gamma $$ is a vector of parameters to be estimated, while ϵ _j is a random term with zero mean and variance $$\sigma _\varepsilon ^2$$ , capturing measurement errors and effects of unobserved factors (Abdulai, Reference Abdulai2016). Even though we cannot observe $$U_j^*,{\rm{\;}}$$ it can be indexed by the observed adoption decision D _j of farmer j, such that the probability of adoption is specified as:

(2) \begin{gathered} \Pr ({D_j} = 1) = \Pr (U_j^{\text{*}} \gt 0) = \Pr ({\varepsilon _j} \gt - {\text{ }}\gamma '{Z_j}) \\ {\text{ = 1}} - F( - \gamma '{Z_j}) \\ \end{gathered}

where F is the cumulative distribution function of ϵ. Since D _j is the binary decision with value 1 or 0, equation (2) is an adoption model and can be estimated consistently by a standard probit estimator if ε_j is assumed normal. We can assess the effect of adoption on any of the outcomes (Q _j ) by regressing Q _j on the binary adoption decision, D _j, and other explanatory variables as below (Tambo and Wünscher, Reference Tambo and Wünscher2017);

(3) $${Q_j} = {\bf\it X}_j'\beta + \omega {D_j} + {\mu _j},$$

where X _j is a vector of household/farm characteristics such as age, gender, educational attainment of the household head, household size, resource endowment, social network, geographical location, and inputs/production characteristics; ω measures the effect of organic fertilizer adoption on outcome Q _j ; µ captures measurement errors as well as the effects of unobserved factors, with j indexing individual farmers (Abdulai and Huffman, Reference Abdulai and Huffman2014; Tambo and Wünscher, Reference Tambo and Wünscher2017).

Equation (3) could be estimated by standard regression (ordinary least squares) if adoption were randomly assigned. In this observational case, however, adopters selected themselves into adoption probably because their personal characteristics like innate managerial and technical abilities support both the decision to adopt and the outcomes, such that they are also the ones who obtain higher outcomes even if they do not adopt (Fuglie and Bosch Reference Fuglie and Bosch1995). It means adopters systematically differ from non-adopters in terms of their background characteristics. Such a systematic difference obscures the true effect of adoption if ω is estimated by through equation (3) (Tambo and Wünscher, Reference Tambo and Wünscher2017). For this cross-sectional evaluation, where we have no counterfactual information for observed outcomes, using either a propensity score-matched (PSM) sample or adopting the ESR approach allows us to control for biases arising from systematic difference (Abdulai and Huffman, Reference Abdulai and Huffman2014). However, the PSM-based model assumes conditional unconfoundedness and thus becomes an invalid approach, where there is insufficient overlap in characteristics of adopters and non-adopters. Even in cases where it is applicable, the probit model coefficients usually, obtained after the PSM are not true coefficients of determination of adoption and cannot be interpreted as such. Since we wished to identify factors determining organic fertilizer adoption among the farmers in this case, we had to apply the ESR (Abdulai, Reference Abdulai2016; Abdulai and Huffman, Reference Abdulai and Huffman2014).

3.3. Empirical Specification of ESR Model

In the ESR framework, equation (2) sorts farmers into adopters and non-adopters, having different outcome regimes. Because we observed either Q _ja or Q _jn for every farmer, there is no counterfactual information on the outcome. Thus, the ESR approach specified potential outcome equations conditional on adoption decision $${D_j}$$ as:

(4a) $${Q_{ja}} = {\bf\it X}_{ja}'{\beta _a} + {u_{ja}}\;\;\;\;\;if\;{D_j} = 1$$

(4b) $${Q_{jn}} = {\bf\it X}_{jn}'{\beta _n} + {u_{jn}}\;\;\;\;\;if\;{D_j} = 0$$

With a continuous outcome variable Q, linear regressions could be applied to estimate β _a and β _n . However, OLS estimator will give biased βs if the error term ϵ _j of equation (2) correlates with those (u _ja and u _jn ) of the outcome equations (4a) and (4b). In other words, OLS cannot be applied when there is selection bias to be addressed. Bias is addressed by including selectivity correction terms in the outcome equations to capture the effect of selection (Abdulai and Huffman, Reference Abdulai and Huffman2014; Khonje et al., Reference Khonje, Manda, Alene and Kassie2015; Lokshin and Sajaia Reference Lokshin and Sajaia. “2004; Shiferaw et al., Reference Shiferaw, Kassie, Jaleta and Yirga2014; Tambo and Wünscher, Reference Tambo and Wünscher2017).

Error terms u _ja and u _jn correlating with ϵ _j implies that the expected values of u _ja and u _jn, conditional on the sample selection (adoption decision) are not zeros. But we can assume that ϵ _j has a variance $$\sigma _{{\varepsilon _j}}^2$$ = 1, and the correlation coefficients between u _j s and ϵ _j sum up to zero, meaning that they have a trivariate normal distribution with a zero-mean vector and a variance–covariance matrix ( $${\rm{\Sigma }}$$ ) as defined below (Manthenge et al., Reference Mathenge, Smale and Olwande2014):

\begin{align}\Sigma = \left[ {\matrix{ {\sigma _{{\varepsilon _j}}^2} & {{\sigma _{{u_{ja}}{\varepsilon _j}}}} & {{\sigma _{{u_{jn}}{\varepsilon _j}}}} \cr {{\sigma _{{\varepsilon _j}{u_{ja}}}}} & {\sigma _{{\mu _{ja}}}^2} & . \cr {{\sigma _{{\varepsilon _j}{u_{jn}}}}} & . & {\sigma _{{u_{jn}}}^2} \cr } } \right],\end{align}

where $$\sigma _{{u_{ja\;}}}^2$$ = var (u _ja ), $$\sigma _{{u_{jn}}}^2$$ = var(u _jn ), $${\sigma _{{u_{ja}}{\varepsilon _j}}}$$ = cov(u _ja , ϵ _j ) and $${\sigma _{{u_{jn}}{\varepsilon _j}}}$$ = cov(u _jn , ϵ _j ). The cov(u _ja, u _jn ) is undefined since we cannot observe Q _ja and Q _jn simultaneously from any farmer (Maddalla, 1983). The expected values of u _ja (E(u _ja |D_j = 1)) and u _jn (E(u _jn |D _j = 0)) contain the selection bias and are statistically equivalent to the product of the error covariances ( $${\sigma _{{u_{ja}}{\varepsilon _j}}}$$ and $${\sigma _{{u_{jn}}{\varepsilon _j}}}$$ ) and inverse Mills ratios (IMRs) for study participants computed from probabilities obtained by (2) at $$\gamma '{Z_j}$$ . Thus, we calculate the IMRs ( $${\lambda _{ja}}$$ and $${\lambda _{jn}}$$ ) as:

(5a) $${\lambda _{ja}} = {{\phi (\gamma '{Z_j})} \over {\Phi (\gamma '{Z_j})}}{\rm{if}}\,{{\rm{D}}_j} = 1$$

(5b) $${\lambda _{jn}} = {{\phi (\gamma '{Z_j})} \over {1 - \Phi (\gamma '{Z_j})}}{\rm{if}}\,{{\rm{D}}_j} = 0$$

where ϕ is the standard normal probability density function and Ф represents cumulative density function of equation (2).

Including $${\lambda _{ja}}$$ and $${\lambda _{jn}}$$ in equations (4a) and (4b), respectively, gives the switching outcome equations that correct for selection bias as follows:

(6a) $${Q_{ja}} = {\bf\it X}_{ja}{\beta _a} + {\sigma _{{\mu _{ja}}{\varepsilon _j}}}{\lambda _{ja}} + \;{\xi _{ja}}\;\;\;if\;{D_j} = 1$$

(6b) $${Q_{jn}} = {\bf\it X}_{jn} {\beta _n} + {\sigma _{{\mu _{jn}}{\varepsilon _j}}}{\lambda _{jn}} + \;{\xi _{jn}}\;\;\;if\;{D_j} = 0,$$

where $${\sigma _{{\mu _{ja}}{\varepsilon _j}}}$$ and $${\sigma _{{\mu _{jn}}{\varepsilon _j}}}$$ become coefficients of the selection control terms λ _ja and λ _jn , capturing effects of all unobserved selection variables on outcomes. $${\xi _{ja}}$$ and $${\xi _{jn}}$$ become the standard error terms with zero expectations. Equation (6) gives more consistent and efficient estimates if estimated simultaneously with the adoption equation (2), using the full information maximum likelihood (FIML) ^{Footnote 11} estimator rather than Maddala’s (Reference Maddala1983) two-step approach (Lokshin and Sajaia, Reference Lokshin and Sajaia. “2004; Tambo and Wünscher, Reference Tambo and Wünscher2017).

The FIML ESR model is identified through non-linearity of $${\lambda _{ja}}$$ and $${\lambda _{jn}}$$ (Lokshin and Sajaia, Reference Lokshin and Sajaia. “2004), but identification is better when, at least, one variable (exclusion variable) affecting the adoption decision D _j but not the outcomes is included in equation (2) (Tambo and Wünscher, Reference Tambo and Wünscher2017). Following Di Falco et al. (Reference Di Falco, Veronesi and Yesuf2011) and Tambo and Wünscher (Reference Tambo and Wünscher2017), we conducted falsification tests (see Appendix 4C) ^{Footnote 12} for several potential instrumental variables to identify relevant exclusion variables. The tests revealed that the distance between household’s location and the nearest agricultural extension office is the only valid instrument for use in our case. After estimating equations (2) and (6) simultaneously, we derived conditional expectations of the outcomes as follows:

Adopters with adoption (actual/observed outcome for adopters):

(7a) $$E({Q_{ja}}{\rm{|}}{D_j} = 1;{\bf\it X}{\rm{)} = {{\bf\it X}_{ja}}{\bf\it \beta} _a} + {\sigma _{{\mu _{ja}}{\varepsilon _j}}}{\lambda _{ja}},{\rm{\;\;\;}}$$

Non-adopters without adoption (actual/observed outcome for non-adopters):

(7b) $$E( {Q_{jn} {\mid} {D_j} = 0;{\bf\it X}{\rm{\;}}} ) = {{\bf\it X}_{jn}}{\bf\it \beta} _n + {\sigma _{{\mu _{jn}}{\varepsilon _j}}}{\lambda _{jn}},{\rm{\;\;\;}}$$

Adopters had they not adopted (counterfactual outcome for adopters):

(7c) $$E ( Q_{jn} {\mid }{D_j} = 1;{\bf\it X}{\rm{\;}} ) = {{\bf\it X}_{ja}}{\bf\it \beta} _n + {\sigma _{{\mu _{jn}}{\varepsilon _j}}}{\lambda _{ja}},{\rm{\;\;}}$$

Non-adopters had they adopted (counterfactual outcome for non-adopters):

(7d) $$E( Q_{ja}{\mid}{D_j} = 0;{\bf\it X} ) = {{\bf\it X}_{jn}}{\bf\it \beta} _a + {\sigma _{{\mu _{ja}}{\varepsilon _j}}}{\lambda _{jn}},{\rm{\;\;\;}}$$

From these expectations, we calculate the average effect of adoption on adopters (AEAA) as:

(8) $$\eqalign{ {\rm{AEAA}} = \left( {{\rm{7a}}} \right) - \left( {{\rm{7c}}} \right) = E({Q_{ja}}|{D_j} = 1;{\bf\it X}) - E({Q_{jn}}|{D_j} = 1;{\bf\it X}) \cr \,\,\;\quad\quad = {{\bf\it X}_{ja}}({\beta _a} - {\beta _n}) + {\lambda _{ja}}({\sigma _{{u_{ja}}{\varepsilon _j}}} - {\sigma _{{u_{jn}}{\varepsilon _j}}}), \cr} $$

where X _ja (β _a – β _n ) captures the expected change in adopters’ mean outcome if they have the characteristics similar to non-adopters.

AEAA estimates obtained from FIML are consistent but not efficient if there is effect heterogeneity (Wooldridge, Reference Wooldridge2015). Therefore, we checked the robustness of the AEAA estimates by comparing them with their precise but inconsistent counterparts obtained from endogenous adoption effect (eteffect) model, using the control function approach. Detail econometric results of the eteffect model are not presented in this paper but are available upon request. Further, we also analyzed farmer-level pre-adoption and transitional adoption effect heterogeneity, following the procedure outlined by Di Falco et al. (Reference Di Falco, Veronesi and Yesuf2011). However, for policy purposes, it is essential to examine heterogeneity across relevant farmer groups. This could be done by estimating and comparing the local average effects of adoption (LAEA) within different quantiles of the outcomes (Issahaku and Abdulai, Reference Issahaku and Abdulai2019). This implies discrete measures of heterogeneity for groups. Yet, effect heterogeneity could be continuous; meaning that discrete measures would not reflect its true nature (Xie et al. Reference Xie, Brand and Jann2012). Thus, following Xie et al.’s (Reference Xie, Brand and Jann2012) approach, we expressed farmer-specific effect of adoption (effect of adoption on adopters (EAA) and effect of adoption on non-adopters (EAN)) as a function of the probability to adopt organic fertilizer. We then visualized the trends of EAA and EAN by two-way scatter plots fitted with local polynomials of degree 1. (Xie et al., Reference Xie, Brand and Jann2012). That allows us to observe how effect heterogeneity (the gap between polynomials lines) differs across levels of probability to adopt.

4.1. Data Summary Statistics

Table 1 presents the summary statistics of the data we used in the study. Column three shows the pooled sample means, while columns four to six display subsample means for adopters and non-adopters and their differences, respectively. Mean-difference tests showed significant difference between adopter and non-adopter households regarding key socioeconomic variables, including gender, education, household size, and livestock size. Adopters also differ from non-adopters in terms of information access and social leverage profiles, such as extension visits, distance to the nearest input market, and membership of a farmer-based organization. Other characteristics distinguishing adopters from non-adopters are soil quality status, type of soil tillage practiced, and experience regarding production shocks such as drought and waterlogging. These differences mean that the randomized selection of communities during sampling did not yield a sample with participants of similar characteristics across the adoption regimes. This indicates that the observed characteristics and, probably, other unobserved factors determine self-selection of farmers into adoption as well as the outcomes.

The summary statistics of the outcome variables reported at the bottom of Table 1 show that the average per capita (adult equivalence) food expenditure is about 598GHS (Ghana Cedis) per annum. An average sample household experiences about 1 month of foodstuff shortages, usually from late April to late May. This period seems shorter (0.74 months) for adopters than for non-adopters (1.5 months), although the difference is not statistically significant. Regarding labor use, an average sample farmer deploys 11.6 worker days of female labor and 14.14 worker days of male labor, totaling 25.74 worker days per acre of maize. Organic fertilizer adopters, however, use significantly (at 1%) more labor than non-adopters.

Figure 1 shows kernel density distributions of the four outcome variables. These distributions highlight the systematic differences between adopters and non-adopters, affirming the validity of ESR model. Since the ESR command we used relies on a linear function of outcome variable, we carried out Box–Cox’s functional form tests to identify and appropriately specify each outcome equation. The test reveals (see appendix 4B) that food consumption expenditure follows a log-linear function, while FG follows a multiplicative inverse (reciprocal) function. Female labor and male labor each support a linear identity process.

Figure 1. Kernel density distribution of outcome variables by adoption status.

4.2. Determinants of Organic Fertilizer Adoption

An initial probit model for organic fertilizer adoption showed that the explanatory variables in our empirical model jointly determine adoption (LR $${\chi ^2}$$ (34) = 545.02 [0.000], Pseudo $${R^2}$$ = 0.78) and correctly predict about 94% of observed organic fertilizer adoption. The results for this first-stage adoption (selection) model are presented in Table 2. The excluded variable (i.e., distance to extension office) is statistically significant, proofing its validity as an instrument for identifying the empirical model. The coefficients of the other variables remained almost the same when the model is estimated simultaneously with any of the outcome’s model. The error term of the adoption model correlates with that of FG and female labor use equations.

Table 2. First-stage FIML estimates of organic fertilizer adoption probit model

*** p<0.01, ** p<0.05, * p<0.1 and [ ] = p> $${\chi ^2}$$ . Robust standard errors in parentheses. Other variables controlled for are meanstrans, plotsize, walkdisttoplot,watrlog, SoilStat, capitalexp herbvaluacre, seedgrade, off-farmincom, and OffFarmRes.

Organic fertilizer adoption is significantly explained by: (a) household characteristics (gender, age, and educational attainment of household head, household size, and female-to-male adults ratio); (b) resource endowment (tropical livestock units [TLUs], arable farmland available, and land tenure); (c) market, information, and social leverage constraints (extension visit, distance to input market, farmer group membership, and market relationships), and (d) husbandry practices and plot-specific production shocks (tillage type, mineral fertilizer use, pest stress, disease occurrence, drought, and waterlogged conditions). Finally, the probability of adoption also correlates positively with PAS extension clusters (zone) location relative to the Bunkpurugu zone.

Female-headed households are more likely to adopt organic fertilizer than male-headed ones. Age of the household head shows a quadratic relationship with the likelihood of organic fertilizer adoption. From youth, it increases with the likelihood of adoption, but the relationship tends to decrease for farmers beyond prime age, as indicated by the quadratic age term (age²). Unlike the mixed findings by Martey (Reference Martey2018), this study finds that education is associated with an increased likelihood of adoption. This supports previous findings (e.g., Kassie et al. (Reference Kassie, Teklewold, Marenya, Jaleta and Erenstein2015) that better-educated farmers are able to access information required to understand farm technology; in this case, the complex relationship between soil health and organic practices and, therefore, are more likely to adopt than less the educated ones. Household size is significant and positively related to adoption, meaning that large households are more likely to adopt than their small counterparts with few members. This is in line with the assertion that organic fertilizer use is labor-demanding, and therefore, households who are capable of meeting the requirement will, more likely, adopt the technology (Chikowo et al., Reference Chikowo, Zingore, Snapp and Johnston2014). However, households with more female farmworkers are less likely to adopt the input’s use than those having more male farmworkers.

Among the resource endowment factors, the number of TLUs a farmer owns has a significant positive relationship with organic fertilizer adoption. Since organic fertilizers are not commercially available, farmers who have substantial livestock herds may have better access to it in the form of animal manure than those without livestock. Hence, they are more likely to use organic fertilizer than those without livestock. The number of acres of arable land (farmland) a household has seems to reduce the probability of using organic fertilizer. On the contrary, ownership (tenure) of the organic fertilizer plot on which organic fertilizer is applied is associated with increased tendency to use organic fertilizer. Organic fertilizers release nutrients and build organic matter content of soil slowly, with the yield benefits accruing in the medium to long-run time span (Waithaka et al., Reference Waithaka, Thornton, Shepherd and Ndiwa2007). Thus only the farmers who own plots on which organic fertilizer is applied are sure to derive the benefits themselves. Hence they are more motivated and likely to invest in organic fertilizer than those farming on rented plots (Abdulai, Owusu and Goetz, Reference Abdulai, Owusu and Goetz2011; Chen et al., Reference Chen, Zeng, Xu and Fan2018a; Jacoby et al., Reference Jacoby, Li and Rozelle2002; Jacoby et al., Reference Jacoby and Mansuri2008; Kousar and Abdulai, Reference Kousar and Abdulai2015).

Of the information and social leverage factors, the distance between a farm household and local input market, the number of extension visits, and membership of farmer-based organization(s) significantly increase the probability of organic fertilizer adoption. This is in line with previous studies that extension services, through information and skill enhancement, enable farmers to adopt new technologies (Abdulai, Reference Abdulai2016; Abdulai and Huffman, Reference Abdulai and Huffman2014; Issahaku and Abdulai, Reference Issahaku and Abdulai2019). Improved access to information through interaction with extension agents and fellow farmers increases awareness about the need to use organic soil amendments, hence, the probability of adoption. Distance to the nearest input market has a similar relationship with organic fertilizer use, probably by limiting access to the commercial alternative (i.e., mineral fertilizer), and thus, inducing the farmers to adopt organic fertilizers, which they can produce locally by themselves. On the contrary, the number of trusted marketing relations (traders) negatively correlates with adoption. Access to traders seems to enhance farmers’ capacity to access and use mineral fertilizers (Teklewold et al., Reference Teklewold, Kassie, Shiferaw and Köhlin2013). In summing up the results of the social leverage factors, it is reasonable to assert that organic fertilizer is a traditional resource-poor soil management technique rather than a modern technology to the sample farmers.

Regarding husbandry practices, minimum tillage is related to an increased probability of using organic fertilizer. That is because farmers in the area generally do not practice minimum tillage, except when they have to apply organic fertilizer by the Zai pit method. On the contrary, the quantity of mineral fertilizer a farmer applies tends to relate with reduced likelihood of organic fertilizer adoption, indicating farmers might have perceived the inputs as substitutes (Bellwood-Howard and Al-hassan, Reference Bellwood-Howard and Al-hassan2016). Experience in production shocks such as droughts and pest stress encourage adoption, while previous crop disease incidence seems to discourage adoption, even though statistically not very significant. This result is consistent with Teklewold et al.’s (2013) findings about factors affecting adoption of sustainable agricultural practices in Ethiopia.

Lastly, the location of a farm household, relative to extension zone 0 (Bunkp., reference zone), significantly determines organic fertilizer adoption. Households in zone 1 (Langbinsi area), zone 3 (Garu-Tempane area), and zone 2 (Manga-Basua area) are more likely to use organic fertilizer than their counterparts in zone 0. This could be attributed to better access to extension through organic fertilizer interventions by PAS and its affiliates.

4.3. Factors Explaining Household Food Expenditure, Food Gap, and Gender-Based Labor Use

Table 3 presents the results of second-stage ESR outcome model for adopters and non-adopters, respectively. At the bottom of the table are correlation coefficients (ρ _jaϵ and ρ _jnϵ ) between error the terms of the adoption equation and that of the outcome regime equations. Statistically significant values for these coefficients indicate a considerable self-selection of farmers into the outcome regimes based on some unobserved factors that influence both the adoption and the outcomes. Self-selection could lead to biased estimates for organic fertilizer effects, had it not been accounted for (Tambo and Wünscher, Reference Tambo and Wünscher2017). The ρ _jnϵ between adoption model and non-adopter regimes and the Wald test of independence of the equations are also statistically significant, meaning that outcomes are not independent of adoption decision. Thus, we had to apply the ESR to control for selection bias in order to obtain accurate estimates of adoption effects on FG and female labor. A positive ρ _jnϵ (in the non-adopter regime) between the reciprocal of FG and the adoption model suggests that non-adopters’ decision not to adopt leads them to reduced FG. Similarly, the negative correlation (ρ _jnϵ ) with the female labor equation indicates a reduction in female labor use under non-adoption.

Table 3. ESR estimates for determinants of household food access and labor use

*** p<0.01, ** p<0.05, * p<0.1. Robust standard errors in parentheses. Other variables included are meanstrans, plot size, walkdisttoplot, disease, Drought, watrlogg, SoilStat, capitalexp herbvaluacre, seedgrade, peststress, OffFarmRes, and offfarmincom. For their coefficients and sig. level, the unedited Stata model outputs are available upon request.

Such kind of biases could exist in the other outcome (logFCE and male labor) equations without being detected, probably due to low statistical power. Thus, empirically we needed to employ the ESR approach in this study (Issahaku and Abdulai, Reference Issahaku and Abdulai2019), even though the insignificant ρ _jaϵ s between adoption and adopter outcome equations suggest that adopters do not do better than a farmer selected at random. The transformed error covariances, lnσ _jaϵ, and lnσ _jnϵ , also reported at the bottom, are highly significant, implying endogenous switching of outcomes between adoption and non-adoption (Tambo and Wünscher, Reference Tambo and Wünscher2017). These coefficients having the same sign, in this case, imply that farmers’ decision to adopt organic fertilizer is not based on any comparative advantage of organic fertilizer. Rather farmers choose fertilizer regime that makes them better off in terms of the outcomes.

The estimates in Table 3 show that gender, age, and education of household heads do not affect food consumption expenditure (logFCE) and FG but they negatively influence labor, especially female labor use. Household size is negatively related to logFCE of both regimes and the FG of non-adopters, but positively associated with female labor use of adopters. This means that large households generally have low FCE, while only non-adopter large households experience higher FG compared with their small counterparts. Large adopter-households use female labor more than small households, probably because they (large adopter-households) have more household labor. Similarly, female-to-male ratio in a household is associated with extended periods of FG and more female labor use. Farm asset value is also significant and positively affects FCE, FG, and female labor of non-adopters, but negatively relates to adopters’ female labor use. TLU statistically relates to only labor use, with opposing signs on adopters and non-adopters. The size of total arable land a household possesses has a similar alternate relationship with outcomes of adopters and non-adopters, even though it is not statistically significant in the case of FCE of adopters, FG of non-adopters, and male labor use of both regimes.

Besides the household characteristics presented above, market and information access variables, including extension visits, membership of farmer-group(s), and relationship with traders also significantly influence outcomes of the regimes differently. The same holds for geographical location (cluster) of households relative to Bunkpurugu area (reference zone). Lastly, we refer the reader to Table 3 for the joint effects of plot-specific variables. Even though some of them have insignificant coefficients, their joint-significance test result (see χ² (37)) is highly significant (at 1%). Based on equations 7a to 7d, we obtained both conditional and unconditional expected outcomes for each adoption regime. We then computed different measures of organic fertilizer effect from the expectations, as presented below.

4.4. Effects of Organic Fertilizer Adoption on Food Security and Labor UseFootnote ¹³

Table 4 shows the expected outcomes, average effect of adoption on adopters (AEAA) ^{Footnote 14} , average effect of adoption on non-adopters (AEAN), as well as adoption effect heterogeneity indexes (BH1, BH2, and AH) for each outcome variable (i.e., under the first column). BH1 and BH2 are the transitional and base heterogeneity, respectively, while AH indicates average effect heterogeneity, showing whether the effect of organic fertilizer is more or less on adopters than non-adopters, in a counterfactual situation that non-adopters did adopt (Di Falco et al., Reference Di Falco, Veronesi and Yesuf2011). In other words, it represents the effect of selection on adoption (Wooldridge, Reference Wooldridge2015).

Table 4. ESR-based expected outcomes; adoption effects and heterogeneity

Standard error in parenthesis. **, *** significant at 5% and 1%, respectively. ^a and ^b are actual expected outcomes for adopters and non-adopter, respectively, while ^c and ^d are their counterfactual outcomes. For food gap, +AEAA means decrease, while the reverse is true for AEAN, since the estimates are in reciprocals.

We obtained alternative estimates for the adoption effects (see appendix 4D) using the endogenous adoption effect (eteffects) model to check the robustness of ESR estimates. The eteffects-based AEAA estimates have the same signs and are close to those from the ESR model even though, except for female labor use, they are not statistically significant. The eteffects model employs the control function estimator, which gives precise but inconsistent estimates because it is sensitive to self-selection and influential observations (Wooldridge, Reference Wooldridge2015). Thus, its estimates are subject to larger standard errors than those of ESR in cases where there is self-selectivity ( $$ln{\sigma _{ja\varepsilon }}/ln{\sigma _{jn\varepsilon }}$$ ) and effect heterogeneities (BH and AH). On the other hand, the ESR model if estimated by the FIML, exploits all available information in the data to fit the model. Hence, the estimates are more consistent than those of the control function estimator (Issahaku and Abdulai, Reference Issahaku and Abdulai2019).

The estimated average effects of adoption show that food consumption expenditure for adopter households increased by about 0.6 log units (AEAA = 0.595) over what it would be, had they not adopted. This represents about 11% (GHS 30.35) increased consumption under organic fertilizer adoption in adopter households. However, food consumption would have been 2% lower in non-adopter households (GHS 9.46) had they adopted organic fertilizer (AEAN = −0.126 log units). These results, as indicated earlier, imply that while adopters are better off adopting organic fertilizer, non-adopters are also better off not adopting. Yet, adopters are far better off under their decision to adopt than non-adopters are under non-adoption. The base and transitional effect heterogeneity estimates (BH1, BH2, and AH) indicate that adopters systematically differ from non-adopters are regarding their background characteristics. And adoption makes adopters a higher per capita food consumers than non-adopters. This meso-level finding is at variance with Martey (Reference Martey2018) macro-level conclusion that organic fertilizer adoption insignificantly lowers food consumption expenditure. The difference could be due to the fact that the present study consisted more subsistent farm households whose additional farm proceeds from adoption reflect in higher food consumption rather than in other expenditures such as report in Martey (Reference Martey2018). It could also be that the previous study’s estimates were biased through averaging over different agroecological zones.

The FG estimate gives an AEAA of 0.438 reciprocal months of FG, representing 0.69 months (55%) decrease from the mean FG that adopter households would have experienced, had they not adopted. Non-adopter households, on the other hand, have a reciprocal of 0.623 FG months but would have had 0.591 with an AEAN (−0.032) representing a 5% increase in the duration of the FG, had non-adopters adopted organic fertilizer. These results show that organic fertilizer adoption decreases food deficits in adopter households, though they still would have experienced fewer months of food inadequacy than non-adopters, had they not adopted.

Regarding labor use, the AEAA shows that organic fertilizer adoption results in a vast increase (5.9) in female worker days from 6.5/acre to 12.3/acre, representing 90%. For non-adopters, the difference (i.e., AEAN) between expected labor use, if they had adopted, and the actual labor use under non-adoption is 0.56 worker days, representing a 5% increase. Hence the transitional heterogeneity (AH = 5.3 worker days) indicates that the effect is more on adopters than non-adopters. However, the pattern is reversed in the case of male labor use. The AEAA on male labor used by adopters is about 9% (1.3 worker days) more than they would have used if they had not adopted organic fertilizer. In the counterfactual case that non-adopters would have adopted, the AEAN indicates they would have used 7.2 worker days (63%) more male labor. This means that the effect of organic fertilizer adoption on male labor use (as indicated by the negative AH) is less for adopter households than for non-adopters. In general, the results regarding labor use are consistent with Teklewold et al.’s (2013) findings that sustainable intensification practices increase farm labor use, with female workers supplying the extra labor needed. For organic fertilizer in particular, much human effort is required in place of carting equipments to collect, transport, and apply the high tonnage (Xu et al., Reference Xu, Huang, Zhong, Chen and Yu2014). Farmers in the study area generally lack such equipments. Thus, traditionally they employ young women to carry out such operations.

4.5 Organic Fertilizer Effect Heterogeneity

The indexes (BH1, BH2, and AH) in Table 4 are average difference adoption effects on adopters and non-adopters. They indicate the presence and general direction of pre-adoption and adoption effects heterogeneity between adopters and non-adopters. We can observe the heterogeneity trend by analyzing and comparing farmer-specific deviations from their sub-sample means (AEAA and AEAN).

Figure 2 shows scatter plots of EAA, EAN, with fitted polynomial lines representing AEAA (pink) and AEAN (green) for each outcome against the probability of adopting organic fertilizer. The gap between the fitted lines indicates the transitional heterogeneity between adopters and non-adopters of equal probability at every point along the probability continuum. For per capita food consumption expenditure, adoption effect heterogeneity is high between farmers with probability ranging from 0.3 to 0.6, where the majority of non-adopters have a negative AEAN. The heterogeneity between farmers with probabilities beyond 0.6 is low, but positive and appears increasing with the probabilities approaching 0.9. The intuition behind this is that non-adopters with an adoption probability above 0.6 will improve their per capita food consumption if they are facilitated to adopt, whereas the same cannot be said about those with low adoption probabilities.

Figure 2. Adoption effects on food access (a and b) and labor demand (c and d) by adoption regime.

The adoption effect on the FG among adopters decreases from positive to negative as the probability of adoption increases, while that of non-adopters remains negative except for those with probabilities between 0.6 and 0.8. Thus, AH becomes negative at probabilities beyond, where non-adopters have almost zero but higher adoption effects than adopters. In that case, encouraging non-adopters to adopt without first changing ground conditions that select them on non-adoption will yield no benefit. Similar to the trend observed in AH of the FG is that of male labor use but the AH effect in that case is negative at all probabilities of adoption, meaning that non-adopters, if adopt, will use more male labor than adopters. Finally, for female labor use, the effect of adoption is positive for both adopters and non-adopters but higher on the adopters than the non-adopters. However, the difference (AH) between the two remains relatively unchanged at all probabilities of adoption.