1. Introduction

It is projected that climate change will influence the timing and duration of seasonal precipitation in South Asia, and will contribute to a decline in water availability for rice cultivation in the region (Lobell et al., Reference Lobell, Schlenker and Costa-Roberts2011; Kim et al., Reference Kim, Ko, Kang and Tenhunen2013; Burchfield and De La Poterie, Reference Burchfield and De La Poterie2018). In Sri Lanka, where rice is both the staple food of the population and the primary crop grown by farmers, reductions in water availability for rice cultivation has serious impacts on farmers' welfare and on national food security (UNESCAP, 2010; Weerakoon et al., Reference Weerakoon, Mutunayake, Bandara, Rao, Bhandari and Ladha2011). Reductions in precipitation are of particular concern in Sri Lanka's dry zone, which accounts for two-thirds of Sir Lanka's total land and over 70 per cent of paddy production in the country (De Silva et al., Reference De Silva, Weatherhead, Knox and Rodriguez-Diaz2007).

The adverse impact of low rainfall and low water availability was highlighted during the major drought event that affected multiple farming seasons in Sri Lanka between 2016 and 2017. Reduced rainfall in the primary agricultural season (maha) of 2016 and the secondary yala season of 2017 severely affected water availability for agricultural production. The World Food Programme estimated that as a result of the drought, reservoirs were on average at just 18 per cent of their capacity and 45 per cent of communities reported that their closest reservoirs were empty (WFP, 2017). This led to a significant drop in crop production and a rapid increase in food insecurity among rural households. In total, 900,000 households were negatively affected by the drought (WFP, 2017).

The Overarching Agricultural Policy in Sri Lanka recognizes the importance of adapting and building resilience to climate events such as drought in order to achieve national development and food security objectives (Government of Sri Lanka, 2021). A key policy thrust within the agricultural policy framework is to address emerging climate change impacts by supporting the adoption of suitable agricultural strategies and practices by farmers. Detailed empirical evidence on the impacts of adopting climate adaptive agricultural practices under conditions of climate shocks, and the socio-economic and institutional factors that influence their adoption, is required to help guide efforts for translating these policy objectives into effective actions.

In the economic literature, structural Ricardian models are often used to understand how farmers adapt to climate change and the implications of this adaptation on farm outcomes (Seo et al., Reference Seo, Mendelsohn and Munasinghe2005; Kurukulasuriya and Mendelsohn, Reference Kurukulasuriya and Mendelsohn2008; Seo and Mendelsohn, Reference Seo and Mendelsohn2008; Di Falco and Veronesi, Reference Di Falco, Veronesi and Yesuf2013; Di Falco, Reference Di Falco2014). This approach utilizes a two-stage framework, where the first stage models farmers' selection of multiple adaptation strategies, and the second stage models the impacts of adopting the considered practices on farm-level outcomes (Di Falco, Reference Di Falco2014). Using this empirical framework, economists have provided important insights on the stand alone and complementary impacts and adoption drivers of adopting climate adaptive practices (Hassan and Nhemachena, Reference Hassan and Nhemachena2008; Di Falco et al., Reference Di Falco, Veronesi and Yesuf2011; Kurukulasuriya et al., Reference Kurukulasuriya, Kala and Mendelsohn2011; Asfaw et al., Reference Asfaw, Di Battista and Lipper2016; Abidoye et al., Reference Abidoye, Kurukulasuriya, Reed and Mendelsohn2017; Gorst et al., Reference Gorst, Dehlavi and Groom2018; Etwire et al., Reference Etwire, Fielding and Kahui2019). In this article, we contribute to the literature on the impacts of climate shocks and adaptive practices by using a normalized inverse probability weights procedure to inform a simultaneous estimation of a weighted system of partially recursive equations. Our approach expands the existing literature on climate adaptation through a mediation analysis, which allows us to disentangle the direct impact of climate adaptive practices from the indirect impacts associated with reductions in sensitivity to climate shocks.

Our analysis makes use of a unique dataset of 1,100 rice producing households in the Anurādhapura district of Sri Lanka covering the agricultural seasons 2017–2018. With these data, we assess the impacts of six different climate adaptation practices, which are further disaggregated by the agricultural field types (upland/lowland) and agricultural seasons (maha/yala seasons) that they are implemented in. Because the survey reference period coincides with an exceptionally dry period, we are able to disentangle the causal impacts of adopting these practices on farmers' sensitivity to water stress, measured as the probability of having experienced crop losses due to wilting; their impacts on farm productivity, proxied by total harvest value; and their impacts on the household's income. This multidimensional approach allows us to identify important trade-offs and complementarities between the reduction in the sensitivity of farm systems to weather shocks and profit maximization objectives that farmers navigate when adopting a particular adaptation practice.Footnote ¹ Finally, we complement the analysis by examining the socio-economic and institutional factors that are associated with the adoption of these practices.

The results show that while a number of the practices considered are effective at reducing sensitivity to water stress, these benefits rarely lead to improvements in agriculture profitability and household income. The reasons are practice-specific, but are linked to high opportunity costs of labour for Sri Lankan farmers, poor development of output markets for crops other than rice, and management constraints that may arise from their adoption. The results suggest that agricultural research and extension should be bundled with farm service and output market development programmes in order to make these climate adaptive practices more productive and welfare enhancing. Bridging the divide between the climate adaptation benefits of the practices and their profitability is essential in order to foster widespread adoption of adaptive practices, and greater overall resilience to climate change in the Sri Lankan agricultural sector.

The remainder of the article is organized as follows. In section 2, background is provided on the study location and the practices under consideration. In section 3, a description of the conceptual model is presented, which is followed by a description of the empirical strategy in section 4. Section 5 provides information on the data set and key variables used for the analysis along with descriptive evidence from the sample population. Sections 6 and 7 present the results from the quantitative analyses considering the impacts on water stress sensitivity and welfare, and the adoption determinants of selected practices, respectively. Finally, in section 8, concluding remarks and policy implications are discussed.

2. Background

3. Conceptual framework: linking water stress sensitivity, household welfare, and adaptation choices

The conceptual framework used for this study accounts for the complex interplay between household-level socio-economic characteristics, the factor intensities and biophysical attributes of the farming practices, and heterogeneous production environments within which they are implemented (upland/lowland fields, maha/yala seasons, and irrigation system).

Our starting point is that climate change in Sri Lanka's rice systems will increase the probability of low rainfall events and the risk of crop loss due to water stress (Madduma and Wickremagamage, Reference Madduma Bandara, Wickremagamage, Herath, Pathirana and Weerakoon2004; De Silva et al., Reference De Silva, Weatherhead, Knox and Rodriguez-Diaz2007). As farmers' perceptions of climate risk change, the expected utility derived from the adoption of practices to mitigate this risk increases (Deressa et al., Reference Deressa, Hassan, Ringler, Alemu and Yesuf2009, Reference Deressa, Ringler and Hassan2010). However, farmers face a utility optimization problem, as they are unable to predict if, in any given season, water stresses will occur. This optimization problem is further confounded by uncertainties and opportunity costs associated with climate adaptation practices themselves and the alternative risk management strategies that households have at their disposal. Therefore, whether or not the adoption of a practice promoted to reduce sensitivity to water stress achieves this objective, relative to conventional practices, depends on a wide range of factors. This includes how well the practice was implemented, the duration of implementation, and the appropriateness of the practice to the local environment, among many others (Molua, Reference Molua2002; Imbulana, Reference Imbulana, Mollinga, Dixit and Athukorala2006; Esham and Garforth, Reference Esham and Garforth2013).

Moreover, even if the adoption of a practice does reduce the sensitivity of a system to water stress, this may not necessarily contribute to improvements in productivity and economic welfare gains relative to non-adopters (Reardon et al., Reference Reardon, Stamoulis, Balisacan, Cruz, Berdegué and Banks1998; Barrett et al., Reference Barrett, Bezuneh and Aboud2001, Reference Barrett, Sherlund and Adesina2008). There are several reasons for this. First, the choice to adopt one practice over others entails trade-offs between the allocations of production factors and their opportunity costs. For example, the choice to adopt a labour-intensive adaptation practice, such as building erosion control structures, will divert labour away from other income opportunities; therefore the opportunity costs of this investment choice are potentially high (Deininger et al., Reference Deininger, Jin and Sur2007). If the positive effect on water stress sensitivity is not sufficient to compensate for reductions in off-farm income resulting from this investment choice, the net income effect of the practice will be negative. Second, trade-offs can also exist between the overall impact of a practice on productivity and its impact on reducing water stress sensitivity. For example, some practices can reduce losses from wilting, but may also reduce overall productivity by lower planting densities or increased weed pressure relative to alternative practices. Finally, practices that entail changes in cropping systems, for example the cultivation of crops less exposed to water scarcity, may reduce sensitivity to water stress, but also expose farmers to more thinly traded, less competitive market conditions than those in rice markets.

Given these reasons, the empirical approach adopted for this study distinguishes between the productivity and welfare impacts derived indirectly through a reduction in water stress sensitivity from those obtained directly through productivity gains and improvements in factor allocations. The combination of these two impact pathways shapes the net impact of the practice. Disentangling these two impact pathways provides insights into the complementarities and trade-offs between the objectives of increasing the climate resilience and improving household livelihood conditions.

The final element of our conceptual framework seeks to understand factors associated with the adoption of the adaptive practices under consideration. In the context of partial or incomplete markets, where production choices are linked to consumption outcomes, as is the case for many producers in Sri Lanka, investments that reduce risk are often prioritized over profitability maximizing activities (Holden and Binswanger, Reference Holden, Binswanger, Lutz, Binswanger, Hazell and McCalla1998). This is further conditioned by a range of socio-economic and institutional factors, which affect households' ability and willingness to cope with production-related risks, and their capacity to allocate production factors to a practice, relative to alternative investment options. As a result, our empirical approach must account for this heterogeneity, in order to reduce concerns over endogeneity due to self-selection into the treatment and to identify the constraints and the barriers to the adoption of practices.

These characteristics include farmers' human capital endowments (education) and physical assets (land, wealth and livestock), which influence the propensity of households to adopt practices with different capital, land or labour factor intensities. In addition, variations in off-farm income earning opportunitiesFootnote ⁴ and their associated effects on the opportunity costs for household labour are likely to be important (Deininger et al., Reference Deininger, Jin and Sur2007). Farm households that derive a large share of their income from off-farm sources are in a better position to invest in capital-intensive farm technologies and practices and are relatively less prone to adopt relatively labour-intensive practices. Moreover, access to off-farm income may, in principle, help to spread the livelihood risks associated with climate- or market-induced volatility in the farm sector. The type and share of irrigated land controlled by a household is also a potential determinant in the choice of adaptation strategy, as this mediates the relative risk of water stress that a household is exposed to. Finally, access to institutional support systems such as input subsidies, concessionary production loans, and insurance is also likely to shape heterogeneous adaptive strategies among the farmers. These programmes mediate farmers' risk exposure and thus their propensity to adopt risk mitigating practices.

4. Estimation strategy

The estimation procedure used in this analysis relies on an inverse weighted probability simultaneous equations model (Hirano et al., Reference Hirano, Imbens and Ridder2003; Bang and Robins, Reference Bang and Robins2005). This approach is expected to address selection bias through a doubly robust estimation procedure of the effects of adopting a specific adaptation strategy on sensitivity to water stresses, measured as the probability of experiencing crop wilting; a productive indicator, namely the total value of the harvest (net and gross); and a welfare indicator, namely the gross household income, measured as the sum total value of the crops sold plus off-farm income, and income from cash and commodities received by the household from third parties. During the agricultural season considered for this study, the Anurādhapura district was hit by a severe drought which affected water availability and created conditions for agronomic water stress in all the farm households involved in the rice sector. As a result, water stress is considered exogenous to all farms and farm plots in the district. However, whether or not a plot has experienced yield losses as a result of this widely covariant shock depends on the sensitivity of the production system to water stress. We define this sensitivity as a latent variable, which is determined by individual-level and plot-level attributes, such as access to irrigation and position on the irrigation canal, as well as by the adoption of the adaptive strategies considered in the analysis.

4.1 Addressing self-selection to measure the effect of practices on water stress sensitivity

This analysis controls for endogeneity related to the adoption of practices through a propensity score method (Rosenbaum and Rubin, Reference Rosenbaum and Rubin1983). In our framework, each treatment regime has been defined with a binary variable T which is equal to 1 if the household adopts the strategy and 0 otherwise. Participation in the treatment is estimated using the vector of pre-exposure characteristics, W. In particular, we estimate the weights used to mitigate the selection bias taking advantage of the vector of observable characteristics, W, which includes human capital endowments (proxied by the household head education), physical assets (proxied by the amount of the available land used for cultivation, a dummy variable identifying the household's sole ownership of the largest field, a wealth index including all the agricultural assets, and a dummy variable taking value one if the household owns livestock), household's available workforce and off-farm income earning opportunities (proxied by family size and a dummy variable identifying off-farm employment), type and share of irrigated land controlled by a household (proxied by a dummy variable identifying the irrigation system type and a continuous variable indicating the share of land under irrigation), access to institutional support systems (such as input subsidies, concessionary production loans, and insurance) and a number of dummy variables indicating the availability of information related to the improved seeds and other innovative technologies. Although it is not possible to rule out the possibility that the selection is also based on other unobservable characteristics, the doubly robust procedure relaxes the concerns about estimating unbiased result even though the selection model has not been perfectly specified. Another (somewhat related) issue that is not possible to rule out with the identification strategy used for this study is the existence of reverse causality among the choice of the production system (the adoption of specific technologies and/or management practices) and the sensitivity to water stresses. In fact, household farmers could adopt practices or management practices in response to the onset or anticipation of water stresses.Footnote ⁵ However, the concerns are relaxed by the fact that the infrastructural and institutional constraints which determine the adoption of a specific production system in the Sri Lanka rice sector framework are quite irreversible in the short term and make the households' choice extremely unlikely to be modified in response to and/or anticipation of sporadic water stresses.Footnote ⁶

Once the selection model has been estimated, the predicted probabilities have been inverted and normalizedFootnote ⁷ to obtain a vector of weights, w, for the sub-sample of households on common support.

Formally the probability of treatment given the pre-exposure covariates is:

(1)\begin{equation}e(W )= P({T = 1|W} )= E\{{I({T = 1} )|W} \}= E({T|W} ).\end{equation}

This has been first modelled using a binomial logit function such that:

(2)\begin{equation}P({T = 1|W} )= ({W, \beta } )= \frac{{exp ({{\beta_0} + {W^T}{\beta_1}} )}}{{1 + \textrm{exp}({{\beta_0} + {W^T}\beta_1} )}}.\end{equation}

The intuition behind this approach consists of estimating the average treatment effect, $\mathrm{\Delta \;\ }$, through the difference of the inverse propensity score weighted averages between treated and control groups (Lunceford and Davidian, Reference Lunceford and Davidian2004):

(3)\begin{equation}\widehat {{\Delta_{IPW}}} = {n^{ - 1}}\sum _{i = 1}^n \frac{{{T_i}{Y_i}}}{{e({W_i}, \widehat {\beta )}}} - {n^{ - 1}}\sum_{i = 1}^n \frac{{({1 - {T_i}} ){Y_i}}}{{1 - e({W_i}, \widehat {\beta )}}},\end{equation}

where ${Y_i}$ is the outcome variable of the unit of analysis, i, and T is the treatment status.

4.2 Estimate the direct, the indirect and the impact of practices on productivity and welfare

Once having estimated the normalized inverse probability weights, the empirical strategy relies on simultaneous estimation of a weighted system of partially recursive equations.Footnote ⁸ Empirically, the re-weighting procedure creates a pseudo-population in which measured confounders can be equally distributed between treatment and comparison groups, thus relaxing concerns of endogeneity. Furthermore, the simultaneous estimation of the two outcome equations is expected to accommodate the correlation among the error terms, and to control for a wide set of additional covariates. Importantly, the doubly robust procedure ensures the consistency of the estimator when either the propensity score or the outcome model are mis-specified, thus addressing a crucial shortcoming of the propensity score models (Robins et al., Reference Robins, Rotnitzky and Zhao1994).

As sensitivity to water stresses is assumed to be a latent variable, $S_i^\mathrm{\ast }$, it is proxied by an observed binary outcome ${S_i}$ which is equal to 1 when farmers harvest an area smaller than the area planted because of wilting and 0 otherwise,

(4)\begin{equation}{S_i} = \begin{cases} 1 & if \, S_i^{\ast } > 0\\ 0 & \textrm{otherwise} \end{cases}.\end{equation}

The weights are integrated into the simultaneous linear estimation of the following system of two equationsFootnote ⁹:

(5)\begin{equation} \begin{cases} S_{I, j} = {\beta_0} + {\beta_1}{T_{ij}} + {\beta_i}{W_i} + {u_{1\; I,j}}\\ {Y_{I,j}} = {\beta_0} + {\beta_1}{{\hat{S}}_{I,j }} + {\beta_2}{T_{I, j}} + {\beta_i}{X_i} + {u_{2 I,j}} \end{cases},\end{equation}

where ${S_{ij\; }}$ represents the observed proxy for sensitivity of the household i due to the adoption of the practices j; ${T_{ij\; }}$ is a dummy variable that is equal to 1 if household $i\;$adopts practice j, and 0 otherwise; ${W_i}$ is a vector of exogenous household and farm characteristics (which includes all the variables shaping the selection into the treatment constituting the vector W in the equation (2)); and ${u_{1\; i,j}}$ represents the random error term that is assumed to be uncorrelated with the explanatory variables, but correlated with the error term ${u_{2\textrm{\; }I,j\textrm{\; }}}$ of the second equation of the system. Moreover, ${Y_{I,j}}$ is the selected outcome for household i depending on the estimated sensitivity to water shock ${\hat{\boldsymbol{S}}_{I,j\textrm{\; }}}$, the direct effect of the adoption of the practice ${T_{I,\; j}}$ and a vector of household characteristics ${X_{2i}}$, including all the household variables included in ${W_i}$ and a number of additional variables which are assumed to influence the outcomes but excluding field level ones (in order to ensure the identification of the system).

This estimation procedure allows for a mediation analysis to disentangle the direct impact of the adoption on productivity and welfare (the partial derivative of the outcome relative to the adoption of the practice $\Delta {Y_j}/\Delta {T_j}$) from the indirect impact through the sensitivity (the product of the partial derivative of the outcome and the partial derivative of the sensitivity $(\Delta {Y_j}/\Delta {\hat{S}_j})\ast \; (\Delta {S_j}/\Delta {T_j})$). Finally, the net effect is obtained by adding the impacts from the direct and indirect channels (i.e., $(\Delta {Y_j}/\Delta {\hat{S}_j})\ast \; (\Delta {S_j}/\Delta {T_j}) + \; (\Delta {Y_j}/\Delta {T_j})$, the partial derivative of the outcome of a reduced form specification).

4.3 Estimate the determinants and the barriers to the adoption of adaptive strategies

In order to investigate the barriers to and the determinants of the adoption of adaptive practices, this analysis takes advantage of a random utility framework. Assuming that the utility difference from the adoption of the practice relative to the alternatives is a latent variable ${U_j}$, farmers select the strategy when the expected utility from adoption is higher than that from alternative strategies $({U_j} > 0)$.

Formally, the adoption model is:

(6)\begin{equation}T_{ij}^\ast = {X_i}{\beta _i} + {\upsilon _{ij}}, \quad j = 1, \ldots ,J \quad \textrm{and} \quad i = 1, \ldots ,N,\end{equation}

where $T_{ij}^\ast$ is a latent variable capturing the demand and/or preference of farm household i for strategy j; ${X_i}$ is a vector of field and household sociodemographic, infrastructural and institutional characteristics affecting the adoption of strategy j; and ${\upsilon _{ij}}$ is a stochastic error term (Kassie et al., Reference Kassie, Jaleta, Shiferaw, Mmbando and Mekuria2013). We assume that the latent variable $T_{ij}^\mathrm{\ast }$ is the utility difference between adopting a practice or not, and if the difference is positive the farmer will adopt the practice in question. The latent variable is proxied by the following observed binary outcome ${T_{ij}}$ which is

(7)\begin{equation}{T_{ij}} = \begin{cases} 1 & \textrm{if} \ T_{ij}^\ast > 0\\ 0& \textrm{otherwise}\end{cases} .\end{equation}

Also in this case, the probability of the treatment has been modelled using a binomial logit function such that:

(8)\begin{equation}P({T = 1|X} )= ({X, \beta } )= \frac{{exp ({{\beta_0} + {X^T}{\beta_1}} )}}{{1 + \textrm{exp}({{\beta_0} + {X^T}{{\rm B}_1}} )}}.\end{equation}

5. Data sources and descriptive evidence

The analysis takes advantage of a unique dataset of rice-producing households in Anurādhapura district, Sri Lanka. The data were gathered as part of a joint effort between the Economic and Policy Analysis of Climate Change unit of the FAO-UN and the Environmental and Water Resources Management Division of Hector Kobbekaduwa Agrarian Research and Training Institute (HARTI). A Multistage Stratified Random Sampling procedure was used to ensure the representativeness of the sample at the district level, as well as the proportional random selection of farmers from each of the four irrigation systems in Sri Lanka.Footnote ¹⁰ The number of farm households within each Divisional Secretariat and within each irrigation system was used to draw a proportional random sample of 11 Divisional Secretariats and 110 farmers organizations from which 1,100 households (corresponding to 3,954 seasonal fields) have been interviewed (details on the sample design and the geographic distribution of the households interviewed are in online appendix A).

The dataset is multilevel, and includes modules at the household, individual, field, activity, and crop level. At the field level, detailed information about all the plots owned or used by the household during the 2017/18 agricultural year, including owned cultivated parcels, sharecropped, or rented parcels, and other pieces of land (such as home gardens, orchards, fallow fields, virgin lands) is collected. To capture variations in seasonal practices, the questionnaire contains separate modules for the 2017/2018 maha season, the 2018 intermediate season, and the 2018 yala season. The questionnaire distinguishes lowlands from uplands, and captures specific seasonal information on agricultural activities, from land preparation to harvesting. It also captures annual information about orchards and home gardens activities. Additional modules have been designed to gather information on input and output market behaviours, input use, livestock holdings and sales, institutional access, assets, and off-farm income.

6. Empirical results: impacts on water stress sensitivity and welfare

In this section we examine the impacts of adopting the identified practices on water stress sensitivity, value of harvest and household income, while differentiating between the direct, indirect, and net impacts of the practice on the welfare indicators. Table A3 reports only the estimated marginal effects of each of the selected adaptive strategies at the field level.Footnote ¹⁴ Of the 12 field- and season-specific practices considered, four are found to significantly reduce sensitivity to water stress. Famers retaining residues and enhancing their decomposition rate (i.e., improved residue retention) on lowland fields during maha are about 14.8 per cent less likely to experience production loss due to wilting than non-adopters.

Growing other field crops reduces sensitivity to water by 18.7 per cent on uplands during maha and 9.3 per cent on lowlands during yala. However, farmers cultivating maize on uplands during the maha season are more prone to water stresses (+16.4 per cent). All else equal, using short-duration seed on lowland during the yala season reduces water stress sensitivity by 5.3 per cent on lowland. Despite the positive impacts of the practices on reducing water stress sensitivity, the impact on productivity and welfare is limited.

Retaining the residues on lowland during maha has a positive direct effect on the value of the crop harvest (+28.6 per cent) that almost doubled (+53.5 per cent) considering the net effect (i.e., direct effect on the gross value of the production plus the indirect effect through the reduction of the sensitivity). Such positive impacts persist, after netting out the cost related to input and labour (+27.4 per cent). However, it does not result in a statistically significant impact on the gross household income. The direct effect of this practice on the household welfare is confirmed on the lowland fields cultivated during the yala season that is associated to a productivity increase (both gross +42.9 and net +39.5 per cent) and to a higher household income (+18.1 per cent). However, there is no evidence of an effect through a sensitivity reduction and in the positive impact of implementing this practice regardless of the season considered. It is important to note that conventional residue retention is widespread in Sri Lanka. Therefore, these results show the impact of improved residue management practices relative to conventional residue retention practices.

Cultivating other crops on upland during maha is found to generate positive indirect effects through the channel of the reduction of the sensitivity to water stresses on the value of crops harvested (+23.6 per cent). However, this indirect effect does not produce a statistically significant positive net effect on any outcomes. Weaknesses in markets for other field crops likely explain why reduced sensitivity to water stress does not lead to a net improvement in both the value of crops harvested (gross and net) and the household gross income.

Adopting soil erosion barriers on upland during yala is associated with an increase in the gross value of the harvest (+ 29.8 per cent), which persists after netting out the input costs (+ 29 per cent), but disappears when we consider the household income as outcome. Again, these results are likely to be due to the opportunity cost of the labour applied on upland during the short agricultural season (yala).

Overall, the mismatch between productivity and welfare outcomes is likely due to the labour intensiveness of many of the practices considered, which diverts labour from off-farm income generating activity. This assumption is further supported by the descriptive data presented in table A4, which reports the average number of person days employed on the fields by adaptive strategy. It is apparent that fields on which diversification strategies are adopted employ, on average, the highest number of person days. It is worth highlighting that most of the heterogeneity among different strategies is due to the household labour, which supports the interpretation that high opportunity cost for labour related to the adoption influences the practice's impact on household income.

Another potential bottleneck emerging from the analysis is related to the weakness and the fragmentation of markets for crops other than rice. Sri Lankan farmers cultivate other crops preferably on upland during the maha season and on lowland during the yala. In the previous case, the indirect effect of the practice on the sensitivity to water stresses is transmitted to all the agricultural outcomes but does not produce a significant increase in the gross income. In the latter case, the results point out that the gains in the wake of a shock are eroded by issues with the commercialization of this product. Moreover, when we focus on the cultivation of maize on upland during maha, a greater sensitivity to water stresses associated with this strategy negatively affects farm-level productivity and the profitability.

All together these results suggest that planting other crops is likely to be a residual strategy to preserve the fertility of the soil for the subsequent agricultural season on lowlands and/or a way to produce food for self-consumption rather than an adaptive strategy on uplands.Footnote ¹⁵

These findings highlight the challenges faced in Sri Lanka in terms of addressing emerging climate vulnerabilities, which are likely to become more pronounced in the future. Consistently with other recent contributions on the same topic (e.g., Gorst et al., Reference Gorst, Dehlavi and Groom2018), the results of this analysis highlight that the current practices promoted to reduce drought sensitivity are not very effective in translating the reductions in water stress sensitivity into measurable productivity and welfare gains, particularly in terms of profitability and gross income.

7. Adoption determinants of selected practices

In this section, the factors associated with the adoption of each adaptive strategy at the household level are explored, focusing specifically on those practices that were found to reduce sensitivity to water stress.Footnote ¹⁶ The first notable result in table A5 is that the factors explaining adoption are highly specific to the practice selected.

The gender and the age of the household head is found to reduce the probability of cultivating other crops on upland during maha, holding other factors constant. In addition, cultivation of other field crops in the upland during maha is positively associated with the household head's level of education. This is likely driven by structural inequalities between men/women and young/old headed households in terms of mobilizing labour and accessing capital required to adopt these practices.

The age of the head of household is also found to be negatively associated with the adoption of other field crops in maha upland and yala lowland fields, while family size is positively associated with the latter practice.

Looking at the household agricultural assets' endowments, diversification on lowland during the short season (yala) is positively associated with the normalized agricultural wealth index. Conversely, the index is negatively associated with the probability of cultivating crops other than rice on upland during the main season (maha), although households cultivating larger fields are more prone to diversify the upland fields during maha.

Taken together, the results suggest that the promotion of crop diversification is often a knowledge-intensive strategy limited by the capacity to mobilize labour, and the nature of the labour force employed. These findings highlight the importance of addressing labour constraints to achieve widespread adoption of these strategies.

Moreover, the asset endowments of households that diversify their cultivation differ with the season/land type considered. During the maha season, the diversification on upland fields is an option for less endowed households to reduce their sensitivity to water stresses. This indicates that wealthier farmers, who may have lower subjective risk levels, are less likely to diversify their production during maha and are more likely to concentrate their efforts on the lowland rice production (Kim et al., Reference Kim, Chavas, Barham and Foltz2014). During the short season (yala), relatively more capital- and labour-endowed households are more likely to diversify their production on lowland field. However, in this case the diversification is not an adaptive strategy to reduce the sensitivity to water shock, but rather a way to preserve the fertility of the field for the next agricultural season instead of leaving it uncovered. Further evidence based on panel data is required to disentangle these puzzling results, although our conjectures are supported by findings on the adoption of mutual combinations of practices (online appendix D), which show that the effect on the sensitivity of the stand-alone adoption of this strategy turns out to be positive when the benchmark is composed of a sub-sample of the population not adopting any practices.

The sole ownership of the largest field reduces the probability of adopting improved practices on lowlands (short-duration rice seeds during the yala season and improved residue retention during the long season). In fact, the farmers renting the fields have more incentives to increase the productivity as the harvest must be sufficient to compensate the rent of the land.

As for the available infrastructure in the field, the larger is the field area covered by agro-wells, the higher is the probability of adopting improved residue retention technique on lowland and cultivating maize on upland during the long season (maha).

Fertilizer and other input subsidies reduce the probability of crop diversification on lowland during yala. Although subsidies are given to both rice and other field crop producers in lowlands, the results suggest that the subsidies are associated with increased incentives for rice cultivation.

An important finding is related to the risk management tools available. Participating in crop insurance schemes increases the probability of adopting short-duration rice seeds on lowland during the yala season and reduces the probability of adopting other risk reducing strategies on these fields, such as diversifying the cultivation during yala or retaining residues and enhancing their decomposition during maha.

As the distance from the fertilizer's retailers increases, the probability of using improved seed on lowlands during yala decreases while that of retaining residues on these fields and enhancing their decomposition during maha increases. For the latter practice, a negative association has been found with the distance from the agrarian service centre. These results are likely to reflect the trade-offs between boosting the productivity through a greater availability of chemical fertilizers and other improved inputs and implementing alternative sustainable management of natural resources, promoted by the Ministry of Agriculture through the agrarian services centres, when these chemicals became less available.

The availability of water, measured as the share of irrigated land and the irrigation type, is associated with the adoption of short-duration rice seeds on lowlands during yala. This relationship is likely driven by the fact that rice production during the yala season is of short duration and is concentrated in irrigated farm systems.

Finally, it is worth highlighting the existence of substantial peer effects within farmers' organizations for all the practices considered. Leveraging these positive peer effects through group-level extension approaches can generate positive adoption impacts.

Overall, the results from the analysis highlight that the adoption of climate adaption strategies is conditioned on the interlinked relationships between factor intensities of the practices and the factor endowments of the households, and complementary/alternative risk management tools available to the household, as well as peer effect within the community.