In this study, the authors assume that a rational farmer would adopt the strategy if it maximises utility. Thus, farming households maximise utility by adopting adaptation strategies that yield higher returns (Kaphaika et al., 2023; Mgomezulu et al., 2023). However, it is imperative to note that utility is unobservable, and it behaves randomly under varying alternatives. Specifically, in discrete choice theory, the utility that a household derives from an adaptation strategy is not directly observable to the researcher. Instead, it is treated as a random variable, consisting of a deterministic component (which depends on observed household, institutional, and agro-ecological factors) and a stochastic component (which captures unobserved factors and measurement errors).

We assume an individual farming household, i, from a sample of N households chooses from a given set of adaptation strategies j = 1, 2, 3, namely (1) Irrigation systems; (2) Zero tillage; and (3) Water harvesting. Further, we assume that each farming household attaches a utility U ij to each adaptation strategy depending on institutional and agro-ecological factors ( η ij ) and household factors ( h i ). Therefore, the utility derived by an individual farming household i, from adopting practice j can be presented as follows:

Where η ij constitutes factors such as access to extension services, changes in rainfall and changes in temperature, and h i constitutes factors such as age, gender, income level, household size among others.

Further letting D ij represent a discrete choice variable for each of the adaptation strategies, and assuming the absence of mutual exclusivity in the choices made by farming households, then D ij takes the value 1 if a household chooses adaptation strategy j and zero otherwise. The corresponding probability can be presented as follows;

From Equation 3, the utility for the given choice of adaptation strategy is greater than the utility derived from the other adaptation strategies. Since the social, institutional, and agro-ecological features of the farming household are quite observable, the utility function (Equation 4) can then be modelled as follows:

Where V_ij = δ_jX_ij is the representative farming household utility, and the X_ij is the vector of observed variables relating to household, institutional, and agroecological characteristics. Ε ij is the stochastic error term that captures the unobservable attributes like the farmer’s personal motivation, and δ_j is the vector of unknown parameters which are to be estimated.

In examining the determinants that influence adoption of different climate adaptation strategies, this study utilises the Sustainable Livelihoods Framework (hereafter, SLF) (Chambers and Conway, 1992). This framework provides a holistic understanding of the livelihood activities, assets and institutions that are available and utilised by resource-poor and marginalised households for development and poverty alleviation. Considering that Malawian smallholder farmers are resource-poor and trapped in transient poverty (Machira et al., 2023; Muyanga et al., 2020), this framework will be useful to understand the strategies that are used to better their welfare. It’s imperative to mention that the SLF comprises five assets, also called the asset pentagon, and these include financial capital, human capital, natural capital, social capital and physical capital. Different studies (Asif and Gill, 2025) that have also examined the determinants of climate change adaptation strategies across the globe, explicitly or implicitly using the SLF. In this study, we modified the SLF to capture the relationship between climate vulnerability, the adaptation strategies and the influencing factors as shown in Figure 2.

Greene (2018) and Wooldridge (2013) emphasize that more than two empirical choices can be estimated using a multivariate probit model. However, these models assume independence of irrelevant alternatives, which implies that the likelihood of choosing A or B will not change if another option C is introduced (Arad et al., 2023; Mittal and Mehar, 2016). Other scholars have stated that the error terms of the options should not be correlated. In our study, the farmers were subjected to interrelated choices, as adopting one adaptation strategy would not restrict the adoption of the other strategy. This entails that the error terms for the choices were correlated, and thus invalidates the multinomial probit model. It’s worth noting that different studies have modelled adoption using a multivariate probit model (Assaye et al., 2023; Donkoh et al., 2019; Kanyamuka et al., 2020; Khonje et al., 2022). Following Greene (2018), the multivariate probit model (Equation 5) in this study was specified as:

Where Y_im indicates the adaptation strategies consistently adopted by the smallholder farmer. The farmer in this case consistently adopts if Y im * > 0; X im presents the vector of socioeconomic and institutional factors. β presents the vector of parameters that were estimated. It should further be noted that the error term ε _im follows a multivariate normal distribution. This implies that the residuals have a zero expected value (E ( ε ) = 0) and a variance–covariance matrix (Equation 6), V, which has ones on the main diagonal and correlations ρ_jk = ρ_kj in the off-diagonal (Wooldridge, 2013).

The joint probabilities of adopting strategies are presented as follows,

These are assumed to form M-variate normal probabilities and are predicted using a likelihood function presented as follows;

Where,

Thus, ρ_jm is the correlation coefficient among the pairs of the error terms of the equations, ϵ_j and ϵ_m. A correlation coefficient of greater than zero implies that the smallholder farmers indeed do not make independent decisions in adopting the three practices. Nonetheless, if ρ_jm = 0 then the multivariate probit model collapses to independent probit models as it implies independence in alternatives. Nonetheless, since farmers are in this scenario subjected to interrelated choices, the multivariate probit model is a better fit than the multinomial probit or logit model (Rahman and Chima, 2015). It’s imperative to mention that the model estimation and all other analyses were done using Stata 17 software.

Table 1 reports the variables of the study, including their expected signs. The variables used in this study are supported by literature. Noteworthy, X 1 to X 7 are household determinants, X 8 is the institutional determinant and X 9 to X 10 are agro-ecological determinants.

Table 4 presents the pairwise correlation matrix on the study dependent variables. The matrix was necessary to confirm if the climate change adaptation strategies are correlated. As indicated in Table 4, from the significance of the chi-square and p-values (Chi²(3) = 109.86, p-value = 0.000), the correlation coefficients indicate that the adaptation strategies are correlated, supporting the use of MVP.

The study also assessed the presence of multicollinearity by using the conditional index methodology. According to Shrestha (2020), multicollinearity is present if a condition index is greater than 15. As evidenced in Table 5, the condition index for this study is 11.46, which is less than 15. This confirms that multicollinearity is not a problem in this study.

From the study, male smallholder farmers were found to significantly and positively adopt irrigation (β = 0.462, ME = 0.103, p < 0.000) and zero tillage adaptation strategies (β = 0.217, ME = 0.061, p < 0.000). However, the study noted that being male had a negative influence on using water harvesting (β = −0.165, ME = −0.039, p < 0.05). Our findings concur with the findings of other scholars (Fredriksson and Gupta, 2023; Hailemariam et al., 2024) but contrast with Hirpa Tufa et al. (2022). Zero tillage and irrigation are high-risk technologies that also require huge investments. On the other hand, water harvesting is a low-risk adaptation technology and requires little investment. Notably, this is a human asset aspect. These gendered disparities evidence that men generally have better access to resources and decision-making power in Malawi agriculture, similarly echoed by Mgalamadzi et al. (2024).

In this study, it was found that age, which is a human asset, strongly predicts adoption patterns. Elderly farmers were less likely to use irrigation (β = −0.535, ME = −0.126, p < 0.000) but significantly more likely to adopt zero tillage (β = 1.866, ME = 0.304 p < 0.000) and water harvesting (β = 0.543, ME = 0.203, p < 0.000). This aligns with studies showing that older farmers may avoid labour-intensive strategies like irrigation due to physical constraints (Samuel, 2025), while opting for less labour-demanding methods such as zero tillage and water harvesting.

Education shows mixed effects. Farmers with primary education were significantly less likely to adopt irrigation (β = −0.546, ME = −0.072, p < 0.000) and, to a lesser extent, zero tillage (β = −0.372, ME = −0.058, p < 0.05). Secondary education significantly increased irrigation adoption (β = 1.730, ME = 0.367, p < 0.05). However, tertiary education level was associated with increased adoption of irrigation (β = 1.964, ME = 0.321, p < 0.05), and zero tillage (β = 1.822, ME = 0.212, p < 0.05). In Malawi, agriculture is largely subsistence, characterised with limited knowledge, and coupled with high risks, but low returns. The aforementioned traditional perspectives impose difficulties to adopt modernised technologies such as irrigation and zero tillage, which is more pronounced in primary level farmers. On the contrary, secondary and tertiary education shifts the perspective of a farmer from subsistence to commercial farming, and it provides contextualised training land management for employing adaptation strategies. As such, it is not surprising to note the use of irrigation and zero tillage among them. The findings of this human asset are consistent with Maguza-Tembo et al. (2017).

Marital status played a significant role in adaptation decisions. Married households were significantly more likely to adopt irrigation (β = 1.152, ME = 0.323, p < 0.000), zero tillage (β = 1.142, ME = 0.106, p < 0.01), and water harvesting (β = 0.509, ME = 0.068, p < 0.000) relative to unmarried households. This supports the notion that marriage facilitates joint decision-making and labour pooling (Doss and Quisumbing, 2020). Divorced households, conversely, were less likely to adopt irrigation (β = −0.344, ME = −0.102, p < 0.000), highlighting the potential vulnerability of fragmented households. Marital status can be kindled as bonding social capital that improves trust for collective action. This has also been emphasised by Karakara (2025) that married households are more likely to implement activities that reduces their vulnerability to agricultural risk.

Institutional support emerged as a decisive factor. Specifically, access to extension services consistently had positive coefficients across all three adaptation strategies ( β = 0.573, p < 0.000) for irrigation, ( β = 1.693, p < 0.000) for zero tillage, ( β = 0.781, p < 0.000) for water harvesting. This indicates a strong positive direction of influence on the adoption propensity. The marginal effects indicate a 13.0 percentage point higher probability for irrigation, 48.5 percentage points for zero tillage, and 26.8 percentage points for water harvesting when households access extension services. Within the Sustainable Livelihoods Framework, extension services enhance both human capital through technical knowledge transfer and social capital by creating institutional linkages, which explains their outsized effect on information-intensive practices like zero tillage. These findings align with Mapanje et al. (2021) and Pangapanga-Phiri and Mungatana (2021), who found extension as a key determinant of climate-smart technology adoption among Malawian smallholders, and with Assaye et al. (2023), who documented similar strong positive marginal effects in Ethiopian interrelated conservation practices.

Farmers perceiving temperature changes were less likely to adopt irrigation (β = −0.626, ME = −0.192, p < 0.000) but more inclined towards zero tillage (β = 0.249, ME = 0.062, p < 0.000) and water harvesting. This may stem from water conservation and cost minimisation perspectives. This is especially true when a farmer considers that a temperature change will be accompanied with increased water evaporation from the farm. Hence, a higher likelihood of adopting zero tillage to prevent direct effect on the soil. Similarly, utilising irrigation will prove to be ineffective, and it will incur a sunk cost to the farmer. This perception is part of the human asset, and it has direct synergies with the financial assets of the farmer. It’s imperative to mention that our findings are also in line with other studies in empirical literature (Abid et al., 2019; Nyang’au et al., 2021; Reddy et al., 2022).

A perception that rainfall will either decrease or increase significantly influenced farmers to adopt irrigation (β = 0.585, ME = 0.129, p < 0.000) and water harvesting (β = 0.724, ME = 0.201, p < 0.000). This is especially true when a farmer’s perception exposes his vulnerability to climate, which translates to soil erosion, food insecurity and increased poverty. As a result, to minimise the occurrence of the aforementioned extreme climate events, farmers would either irrigate or collect more underground water. Different studies have also emphasised that rainfall perceptions influence the adoption of adaptation strategies (Asfaw et al., 2019; Al Mamun et al., 2021). This result has significant implications for bridging and bonding social networks that are available to farmers. Noteworthy, the perceptions emanating from fellow farmers, extension workers, community farmer clubs and other community-based organisations can influence the farmers to adopt the adaptation strategies.

Household resource endowments positively influenced the adoption of zero tillage (coefficient = 0.520, ME = 0.064, p < 0.000) and water harvesting (0.053, ME = 0.720, p < 0.000) technologies in this study. This may stem from the labour-intensive demands associated with zero tillage and water harvesting. Again, household size, which is an indicator of the available human capital at the household level, can be specialised for conservation-oriented practices. Our finding agrees with other scholars such as Jha et al. (2019) and Timothy et al. (2022).

The study found that income level was negatively associated with zero tillage (β = −1.98e-06, ME = −2.69e-07, p < 0.000) and water harvesting (β = −2.82e-06, ME = −8.37e-07, p < 0.000). This is because income level, which is a financial asset at the household level, influences the use of capital-intensive strategies like irrigation. It’s worth mentioning that our findings agree with other scholars such as Marie et al. (2020). This also aligns with the Sustainable Livelihoods Framework, where financial assets reduce liquidity constraints for technology adoption, consistent with Aboye et al. (2023).

In this study, landholding size significantly increased the likelihood of adopting irrigation (β = 0.246, ME = 0.074, p < 0.000), zero tillage (β = 0.517, ME = 0.053, p < 0.000), and water harvesting (β = 0.075, ME = 0.032, p < 0.05). Land size has physical asset characteristics, and it is a crucial determinant to invest, minimise risks and realise economies of scale, as far as adaptation strategies are concerned. Our findings, which are consistent with Pronti et al. (2024) confirms that landholdings enhance flexibility in experimenting with and sustaining adaptation strategies