Theoretical and empirical strategy
This study derives its theoretical foundation from the production theory, which establishes a physical relationship between agricultural output and a set of inputs using a given technology. In this study, we include adaptation strategies as farm inputs to mitigate the effects of climate change on land productivity. We express maize yield as a function of a set of farm inputs, including adaptation strategies as in (1):
$$ Yiel{d}_i=\alpha +\gamma {X}_i+\varphi ADAP{T}_i $$
(1)
where Yieldi denotes maize output per ha, Xi is a set of farm inputs, ADAPTi indicates a bundle of adaptation strategies adopted by farmers. α, γ and φ represent parameters to be estimated. We can estimate Eq. (1) using the ordinary least squares (OLS). However, the inclusion of the adaptation strategies would generate biased estimates if Eq. (1) is estimated with the OLS. The reason is that the adoption of adaptation strategies among farmers is not a random process but self-selection, thereby creating selection bias in the empirical analysis. Extant empirical studies have used some econometric approaches to address this selectivity bias issue (Owusu et al. 2011; Donkor et al. 2018; Donkor and Owusu 2019). The most common self-selection bias approaches are the instrumental variable approach, difference in difference and the propensity score matching (PSM) technique. The limitation with the instrumental variables approach is identifying appropriate instruments. The difference in difference method is applicable when the data is longitudinal. The PSM approach employed in the present study is able to control for observable characteristics but cannot account for unobservable factors (Owusu et al. 2011; Donkor et al. 2018; Donkor and Owusu 2019).
Farmers tend to use more than two adaptation strategies to mitigate against climate change effects. It is therefore important to explicate the factors that influence farmers’ adoption of adaptation strategies. The adaptation strategies include organic fertilizers, changing planting dates and planting short duration crop varieties. We conceptualize the adoption process of adaptation strategies using the random utility theory. The random utility theory states that every individual is a rational decision-maker, maximizing utility relative to his or her choices. Based on this theory, we assume that given a set of adaptation strategies, farmers choose adaptation strategies that maximize their utility:
$$ E\left[U\left({A}_1\right)\right]>E\left[U\left({A}_2\right)\right] $$
(2)
where U(A1) is the utility derived from adopting an adaptation strategy A1 and U(A2) is the utility derived from an adopting adaptation strategy A2. The farmer i decides to adopt J adaptation strategy if the perceived utility from strategy J is greater than the benefit from adopting adaptation strategy, k, i.e.
$$ {U}_{ij}\left({\varpi}_j{X}_j+{\varepsilon}_j\right)>{U}_{ik}\left({\varpi}_k{X}_k+{\varepsilon}_k\right),k\ne j $$
(3)
where Uij and Uik are the perceived utilities derived by farmer i from adaptation strategies j and k, respectively; Xi is a vector of explanatory variables that influence the choice of the adaptation strategy; ϖj and ϖk are parameters to be estimated; εj and εk are the error terms.
The economic problem discussed above concerns a decision to choose amongst different alternatives, and in the present study, we employ the multinomial logit (MNL) model. Specifically, we investigate the factors influencing the choice of the use of organic fertilizers, change of planting dates and planting of short duration crops as climate change adaptation strategies in maize production in southern Mali. The probability that a farmer i with Xi characteristics chooses an adaptation strategy j in the MNL framework can be specified as:
$$ {P}_{ij}=\Pr \left({Y}_i=1\right)=\frac{e^{X^{\prime}\varpi }}{1+\sum \limits_{j=1}^J{e}^{X^{\prime}\varpi }},\kern0.75em j=1,...,j $$
(4)
where ϖ is a vector of parameters that satisfies ln(Pij/Pik) = X′(ϖj − ϖk). The multinomial logit (MNL) requires that the assumption of independence of irrelevant alternatives (IIA) is satisfied. Given that the estimated coefficients of the MNL are computed relative to the base variable, the direct interpretations of the signs and the magnitudes become difficult. We therefore compute the marginal effects of changes in the explanatory variables on the probability of choosing each of the climate change adaptation strategy (Wooldridge 2010). The marginal effects of the explanatory variables from the MNL model are computed as:
$$ \frac{\partial {P}_{ij}}{\partial {X}_k}=\left({\varpi}_{jk}-\sum \limits_{j=1}^J{P}_j{\varpi}_{jk}\right) $$
(5)
.
Empirically, we express the adoption of the climate change adaptation strategies as a function of household characteristics, plot-level characteristics, institutional characteristics, perception variables and a location dummy using the MNL model:
$$ {\displaystyle \begin{array}{l} ADAP{T}_{im}={\varpi}_{0m}+\sum \limits_{k=1}^5{\varpi}_{km} HHCHRCT{S}_{ikm}+\sum \limits_{k=6}^8{\varpi}_{km} PLOTLEVE{L}_{ikm}+\sum \limits_{k=9}^{12}{\varpi}_{km} INSTITUTIONA{L}_{ikm}\\ {}\kern3.75em +\sum \limits_{k=13}^{15}{\varpi}_{km} PERCEPTIO{N}_{ikm}+{\varpi}_{16m} LOCATIO{N}_{im}+{\varepsilon}_{im}\ \end{array}} $$
(6)
where ADAPTim denotes m adaptation strategies (Organic fertilizers = 1, planting of short-duration maize varieties = 2, changing planting dates = 3).
HHCHRCTSikm is a set of household characteristics such as the farmer’s age (years), education (number of years of formal education), household size (number of family members in the household), number of livestock owned, and experience in maize farming (years). PLOTLEVELikm represents a vector of plot-level characteristics, including distance from home to the farm in kilometres (km), plot ownership and plot size in hectares (ha). IINSTITUTIONALikm denotes a set of institutional characteristics, including frequency of extension contact, access to technical training, access to credit, participation in off-farm employment and membership of farmer association. PERCEPTIONikm denotes a set of perception variables, including short onset of rainy season, early onset of dry season and decrease in rain frequency. LOCATIONikm denotes a location dummy. ϖ0m and ϖkm are vectors of parameters to be estimated using the simulation maximum likelihood approach. εim represents the error term.
We expect household, farm, institutional and location-specific characteristics to influence the farmers’ use of adaptation strategies. The age of the farmer, number of years of schooling, household size, experience, land ownership, access to credit, frequency of extension contact and wealth of the farmer are expected to exert positive effects on the probability of adopting climate change adaptation strategies, whereas perceptions of farmers on the onset of late rainy season, early dry season and off-farm employments are expected to exhibit negative effects. We hypothesize that local climatic conditions and district level climate variables would have mixed effects on the adoption of climate change adaptation strategies. In the next section, we discuss the propensity score matching approach employed in the present study to estimate the impacts of the climate change adaptation strategies on productivity and food security status of the sampled farming households in southern Mali.
Propensity score and treatment effects
We analysed the impacts of the climate change adaptation strategies on maize yield and household food security with the propensity score matching (PSM) approach (Sianesi 2004; Owusu et al. 2011; Donkor et al. 2018; Donkor and Owusu 2019). The propensity score p(Zi) is defined as the conditional probability of adoption of any of the three adaptation strategies considering the pre-adoption characteristics (Becker and Ichino 2002) as:
$$ p\left({Z}_i\right)\Pr \left({L}_i=|{Z}_i\right)=E\left({L}_i|{Z}_i\right)p\left({Z}_i\right)=F\left\{h\left({Z}_i\right)\right\} $$
(7)
where Li = (1, 0) is an indicator of adoption of an adaptation strategy, Zi denotes a vector of pre-adoption characteristics and F(⋅) can be a normal or logistic cumulative distribution. The propensity score is predicted with either the logit or probit model (Sianesi 2004). The parameter of interest in PSM evaluation methodology is the average treatment effect on the treated (ATT). Given the propensity score, p(Zi), the ATT effect is evaluated as:
$$ ATT=E\left[E\left({Y}_i^1|{L}_i=1p\left({Z}_i\right)\right)\right]E\left[E\left({Y}_i^0|{L}_i=1p\left({Z}_i\right)\right)|{L}_i=1\right] $$
(8)
where \( {Y}_i^1 \) and \( {Y}_i^0 \) are the two counterfactual outcomes of adoption and non-adoption of climate change adaptation strategy. In the present study, we employed matching algorithms such as the nearest neighbour, kernel-based and radius matching.
Measurements of outcome indicators
Constant changes in climatic conditions negatively affect the maize production, food systems and food security of smallholder farmers. Therefore, climate change adaptation strategies are expected to increase maize yield and reduce food insecurity levels of households in rural areas. As noted by Di Falco et al. (2011), the higher the probability of adoption of climate change adaptation strategies, the more likely household-level food insecurity would decrease. As already indicated, we use maize yield and household food insecurity as the outcome indicators in the present paper. We measure maize yield as the grains weight (kilograms) per harvested area (hectare). Various household food insecurity indicators exist in the empirical literature. In the present study, we use the Household Food Insecurity Access Scale (HFIAS) to measure the food insecurity status of the maize farm households. The HFIAS comprises a series of questions for the food insecurity (access), which allows predictable responses to be gathered through inquiry as outlined on the food insecurity scale (Leroy et al. 2015). The Household Food Insecurity Access Scale score for each household is computed by adding the coded frequency of experience for all the questions as:
$$ HFIAS\left(0-27\right)=\left({Q}_{1a}+{Q}_{2a}+{Q}_{4a}+{Q}_{5a}+{Q}_{6a}+{Q}_{7a}+{Q}_{8a}+{Q}_{9a}\right) $$
(9)
where a denotes the coded frequency of experience and Q represents the various questions we asked the farmers regarding their household food insecurity. Generally, the higher the score, the higher the food insecurity status of the household. Households with HFIAS score between 0 and 4 are classified as food secure and those with HFIAS score above 4 are classified as food insecure.
Source of data and data description
The present study was conducted in two districts—Koutiala and Bougouni—in the Sikasso region of southern Mali. Maize is produced in all the districts in the Sikasso region of Mali (Fofana et al. 2011). As shown by Fig. 1, Koutiala and Bougouni rank second and third in terms of maize production in the region, and they contribute about 14% (56,714 t) and 12% (47,653 t), respectively, to the total maize output. This shows that the two districts are among the most important maize producing areas in the Sikasso region. Koutiala covers an area of about 18,000 km2 with a population of 575,253 inhabitants. It is located in the Sudano-Sahelian zone between 12° 38′ N and 5° 66′ W with a wet season of around 3–4 months, and a dry season for the rest of the months. The average annual rainfall is around 895 mm, with a maximum temperature of about 38 °C. The size of Bougouni is about 20,028 km2, with a population of 459,509 inhabitants. This district is situated between 11° 24′ N and 7° 35′ W in the Sudano-Guinea zone. The duration of the rainy season is about 4–5 months with the rest being mostly dry. The maximum temperature is around 33 °C, and the average rainfall is of about 1100 mm.
The study employed primary data collected using a multi-stage sampling approach. First, the two districts were purposively selected due to high maize production and about 20% and 40% frequency of occurrence of drought risk in the districts (Fofana et al. 2011). Second, three out of the six agricultural subsections from Koutiala District and four out of the eight in Bougouni District were purposively selected. The agricultural subsectors represent about 50% of the total agricultural subsection in each district. Third, two communities were randomly selected from each of the agricultural subsection, making a total of 14 randomly selected communities for the study (i.e. six from Koutiala and eight from Bougouni). During the final stage, 155 maize farmers from Koutiala and 153 from Bougouni were randomly sampled, making a total sample size of 308 farm households for the study (see Table 1). The sample size for the study was determined using the formula, \( n=\frac{N}{1+N{e}^2} \) proposed by Yamane (1967), where n denotes the required sample size, N denotes the total population and e represents the margin of error. Using the level of confidence of 92% and a margin of error of 8% (0.08), we obtained a sample size of 308 for the survey. The margin of error of 8% enabled us to obtain a representative sample. It also assisted us to avoid large sample size, which would have led to inefficient data collection due to inadequate financial resources and time constraints.
Table 1 Sampled communities With the aid of structured questionnaires, we collected data on personal, household, farm and institutional characteristics, climate change adaptation strategies and perceptions of maize farmers on the adaptation strategies and food security status of the farm households using the Household Food Insecurity Access Scale (HFIAS).