Joint determination of factors influencing adoption of climate smart agricultural practices in modern rice production in Southwest Bangladesh: a multivariate probit approach

1 Introduction

Agriculture remains central to livelihoods, food security, and employment in the developing world and continues to drive poverty reduction and economic growth (ADBI, 2009; Christiaensen and Martin, 2018; Christiaensen et al., 2021; Christiaensen et al., 2011; ILO, 2024; Ivanic and Martin, 2018; Yeboah and Jayne, 2018). Most farms are small and family operated, and smallholders account for the majority of producers in South Asia, where rural households farm for both cash income and subsistence (Lowder et al., 2016; Lowder et al., 2021; Ricciardi et al., 2018). The output gains of the past half century rested on the intensive input use of the Green Revolution, which raised food availability through modern high yielding varieties¹, chemical fertilizers, agrochemicals, mechanization, and expanded irrigation (Gollin et al., 2021; Pingali, 2012), but intensive cultivation also degraded soils, lowered groundwater tables, and weakened resilience of agroecosystems (Kopittke et al., 2019; Springmann et al., 2018; Tilman et al., 2002). Climate change sharpens this tension. Rising temperatures and greater rainfall variability already depress crop yields and threaten food security (Lobell et al., 2011; Ortiz-Bobea et al., 2021; Ray et al., 2019; Wheeler and von Braun, 2013; Zhao et al., 2017), and coastal, deltaic, and wetland farming systems face the greatest exposure because high hazard levels coincide with limited capacity to absorb shocks (Edmonds et al., 2020; Prestele and Verburg, 2020; Szabo et al., 2016). Governments must therefore raise agricultural productivity while protecting the resource base that sustains future production.

Climate smart agriculture (CSA) has emerged as the leading framework for this task. It seeks to raise productivity, strengthen the adaptive capacity of farmers, and, where feasible, reduce greenhouse gas emissions (FAO, 2013; Lipper et al., 2014; Lipper and Zilberman, 2018; Totin et al., 2018). CSA is a portfolio of practices rather than a single technology and spans crop diversification, agroforestry, integrated soil fertility management, small scale irrigation, and integrated pest management (Aryal et al., 2018; Campbell et al., 2014; Makate et al., 2019). Because practices complement or substitute one another, adoption decisions across the portfolio are interdependent rather than separable. Adoption is also a decision under risk. Smallholders are typically risk averse and face weather, pest, salinity, and price shocks that they can neither anticipate nor insure against (Feder, 1980; Hansen et al., 2019; Just and Zilberman, 1983; Komarek et al., 2020). Where credit and insurance markets are incomplete, the threat of low consumption after a failed harvest deters investment in higher return but riskier technologies (Cole and Xiong, 2017; Dercon and Christiaensen, 2011; Emerick et al., 2016; Rosenzweig and Binswanger, 1993). Climate smart practices are therefore instruments of risk management, since adaptation lowers the sensitivity of farm income to climatic shocks and a diversified bundle spreads exposure across distinct sources of risk (Asfaw et al., 2018; Di Falco et al., 2011; Di Falco and Veronesi, 2013; Issahaku and Abdulai, 2020; Tesfaye et al., 2021). Farmers consequently assemble bundles instead of adopting practices in isolation, and they respond more strongly where incentives are favorable and institutions function well (Besley and Case, 1993; Feder et al., 1985; Foster and Rosenzweig, 2010; Ruzzante et al., 2021; Suri and Udry, 2022; Takahashi et al., 2020).

Bangladesh exhibits these tensions at the national scale. More than four fifths of its farmers cultivate less than one hectare of land, yet agriculture remains a major source of income, employment, and food supply (BBS, 2024; Tajuddin et al., 2017), and rice anchors both the farming economy and national food security (Prince et al., 2023). The spread of modern varieties expanded output and reduced hunger, but the same intensification has degraded soils, reduced fertility, and depleted groundwater, most severely where dry season irrigation draws on shallow aquifers (Alauddin and Quiggin, 2008; Kafiluddin and Islam, 2008; Mainuddin and Kirby, 2015; Rahman and Thapa, 1999; Shamsudduha et al., 2009; Timsina et al., 2018). Comparable resource stress marks intensive cereal systems elsewhere in Asia and Africa (Arslan et al., 2015; Getnet et al., 2014; Matimolane, 2018; Rahman et al., 2021). Bangladesh ranks among the most climate vulnerable countries in South Asia (Eckstein et al., 2021; Kreft et al., 2017), and climatic extremes already disrupt household food consumption and welfare (Islam et al., 2022; World Bank Group, 2021; World Food Programme, 2021). The southwest coastal region faces recurrent floods, cyclones, tidal surges, and advancing salinity intrusion (Chen and Mueller, 2018; Clarke et al., 2015; Dasgupta et al., 2015; Saha et al., 2019; Shameem et al., 2015), and these hazards have raised crop losses and production risk (Hanif et al., 2015; Mazumder, 2024; Patwary, 2016). Even so, uptake of climate adaptive practices remains partial and selective across many high risk areas (Baffour-Ata et al., 2026). This gap between exposure and adaptation motivates the central question of the paper, namely how climatic risk, institutional support, and household characteristics jointly shape the adoption of climate smart practices.

Studies consistently show that CSA adoption is associated with improved food security, greater resilience, and better rural livelihoods, particularly in the coastal and wetland areas of the country (Hasan et al., 2018; Islam and Farjana, 2024; Majumder et al., 2024; Mazumder, 2024). They also identify education, farming experience, landholding, livestock ownership, access to credit, extension services, training, and market access as important factors shaping the adoption decisions of farmers (Billah and Hossain, 2017; Saha et al., 2019; Sarker et al., 2022; Saroar and Filho, 2016). A recent review of the global evidence similarly finds that adoption is influenced by local agroecological conditions and the institutional context in which farmers operate (Ma and Rahut, 2024). Several limitations in this evidence base motivate the present study. First, most adoption studies estimate a single equation model for each practice, which ignores correlated unobserved factors and the complementarities that govern portfolio adoption and can bias inference when choices are interdependent. Second, the literature treats adoption as a binary outcome and reveals little about how far households extend across the portfolio, although the breadth of adoption bears directly on resilience. Third, integrated micro econometric evidence remains scarce for modern rice systems in the salinity affected areas, precisely where the tensions among productivity, sustainability, and climate effects are most acute.

Given this backdrop, the specific objectives of the present study are (1) to identify the climate smart agriculture practices adopted by modern rice farmers in southwest Bangladesh, (2) to test whether adoption decisions across the five climate smart agriculture practices are statistically interdependent, (3) to measure the breadth of climate smart agriculture adoption using a normalized adoption count, and (4) to identify jointly the socioeconomic, institutional, and climatic factors associated with the adoption of climate smart agriculture practices.

The analysis draws on a survey of primary data from 493 farm households in three districts of southwest Bangladesh covering the gradient of salinity zones in the coastal belt. The paper makes four contributions. First, the study moves beyond the binary treatment of adoption by measuring the breadth of adoption, defined as the number of climate smart practices adopted by a household, which provides a clearer picture of how widely CSA practices are integrated into rice-based farming systems. Second, it examines the joint adoption of five climate smart practices within a unified framework and recovers complementarities and trade-offs that single equation models cannot capture. Third, it provides integrated micro econometric evidence on the adoption of CSA in modern rice production for the salinity affected coastal belt of southwest Bangladesh, an important yet understudied climate vulnerable region. Fourth, the study provides an integrated empirical perspective on CSA adoption that may be relevant to climate-vulnerable rice-producing regions in Asia and Africa facing similar combinations of smallholder farming, climatic stress, and institutional constraints, while interpreting the empirical relationships as conditional associations within the study context.

The remainder of the paper proceeds as follows. Section 2 reviews conceptual and empirical literature and develops the conceptual framework that guides the analysis. Section 3 describes the study area, the data, and the empirical strategy. Section 4 presents and discusses the results. Section 5 concludes and draws policy implications.

2 Literature review

2.1 Technology adoption under risk and the economics of climate smart agriculture

The adoption of agricultural innovations in developing countries unfolds through a gradual process of learning, experimentation, and adjustment rather than through a single discrete choice (Besley and Case, 1993; Feder et al., 1985; Foster and Rosenzweig, 2010). Human capital anchors this process because education strengthens the ability of farmers to obtain, interpret, and apply technical information (Lin, 1991), while social networks and extension services govern how quickly that information travels (Conley and Udry, 2010). Risk is equally fundamental. Smallholders are generally risk averse and face production risk from weather, pests, salinity, and prices, and this risk conditions the technologies they are willing to try (Feder, 1980; Just and Zilberman, 1983). Where credit and insurance markets are incomplete, the prospect of low consumption after a failed harvest deters investment in higher return but riskier technologies, so households hold safer portfolios that sacrifice expected income for stability (Dercon and Christiaensen, 2011; Rosenzweig and Binswanger, 1993). Climate variability magnifies these pressures and raises the value of adaptation, which lowers the sensitivity of farm income to climatic shocks, and a portfolio of complementary strategies raises farm returns by more than any single practice adopted in isolation (Di Falco et al., 2011; Di Falco and Veronesi, 2013). Climate smart agriculture translates these insights into an integrated framework that seeks to raise productivity and farm incomes, strengthen resilience, and reduce greenhouse gas emissions where feasible (FAO, 2013, 2016, 2022; Lipper et al., 2014). Its breadth has drawn criticism because CSA cannot be treated as a single universally applicable package (Chandra et al., 2018; Taylor, 2018). The framework is better understood as a set of context-specific options whose climate smart status should be assessed against the agroecological context in which they are applied rather than assumed from their labels (Thornton et al., 2018). It is therefore treated here as a portfolio of complementary practices spanning crop diversification, agroforestry, integrated soil fertility management, small scale irrigation, and integrated pest management, in which the returns to any one practice may rise with the adoption of others (Campbell et al., 2014; FAO, 2013; Mzyece and Ng'Ombe, 2020; Partey et al., 2018; Wilson and Lovell, 2016).

2.2 Determinants of the adoption of climate smart agriculture

Empirical studies from Sub Saharan Africa and South and Southeast Asia point consistently to four groups of determinants, namely demographic and human capital characteristics, farm and resource endowments, institutional and informational access, and climate related exposure and perceptions (Antwi-Agyei et al., 2021; Etwire et al., 2013; Khatri-Chhetri et al., 2017; Musafiri et al., 2022; Sisay et al., 2023; Waaswa et al., 2022). Better educated farmers acquire and apply technical knowledge more easily and therefore adopt more readily, farming experience generally promotes adoption, and the influence of age varies across the life cycle (Asfaw et al., 2012; Feder et al., 1985; Lin, 1991). Farm size and asset ownership ease scale and liquidity constraints and widen the range of feasible management options (Feder, 1980; Just and Zilberman, 1983; Kassie et al., 2015), while livestock ownership favours practices that exploit complementarities between crop and livestock production, particularly those that improve soil quality (Mujeyi et al., 2021; Teklewold et al., 2013). Institutional access matters at least as much. Credit relaxes the liquidity constraints and consumption risk that deter investment in capital intensive technologies (Asfaw et al., 2012; Dercon and Christiaensen, 2011), extension and training reduce the uncertainty surrounding unfamiliar practices (Conley and Udry, 2010; Khatri-Chhetri et al., 2017), and membership in farmer organizations widens the networks through which information and technologies spread (Kassie et al., 2013; Osorio-García et al., 2020; Tadesse et al., 2021). Climatic experience offers a further pathway. Exposure to droughts, floods, salinity intrusion, pest outbreaks, and rainfall variability, together with perceptions of long term climatic change, can raise both the perceived need for adaptation and the expected benefits of climate smart practices (Abid et al., 2015; Belay et al., 2017; Debela et al., 2025; Di Falco et al., 2011). Awareness alone, however, rarely translates into adoption. Farmers respond most effectively when favorable perceptions are matched by access to finance, extension, and reliable information, and institutional and resource constraints continue to bind even among farmers who recognize climatic risk (Antwi-Agyei et al., 2021; Durgude et al., 2026; Sisay et al., 2023; Wakweya, 2023). A related literature links adoption to gains in productivity, food security, household income, and resilience (Shahzad and Abdulai, 2021; Tsige et al., 2024; Zheng et al., 2024), whereas the present study concentrates on the factors that shape adoption decisions rather than on the outcomes that follow them.

A recurring challenge in this literature is that institutional factors may not be fully exogenous. Farmers who are more interested in new practices may be more likely to seek credit and advisory services, while extension agencies and subsidy programs may target more progressive or motivated farmers. Thus, observed relationships between institutional support and adoption may reflect both service provision and differences in farmers’ likelihood of accessing them (Abdulai and Huffman, 2014; Wooldridge, 2010). A similar concern applies to climate-risk perceptions, as farmers already engaged in adaptive practices may pay greater attention to climatic risks, creating a potential two-way relationship between perceptions and adaptation (Arbuckle et al., 2015). These considerations highlight the need to interpret the estimated relationships carefully.

2.3 Joint adoption, complementarities, and adoption intensity

Farmers seldom adopt climate smart practices one at a time. They assemble combinations that complement or substitute for one another according to production conditions and resource constraints (Kangogo et al., 2021; Mazhar et al., 2021; Teklu et al., 2023). Estimating each adoption decision independently therefore discards the common unobserved factors that shape several choices at once and misses the interdependence that characterizes adoption behavior. The multivariate probit model estimated through simulated maximum likelihood accommodates this interdependence because it allows the error terms across adoption equations to be correlated while permitting each practice to respond to its own explanatory factors (Cappellari and Jenkins, 2003; Teklewold et al., 2013). Multinomial and ordered response models have served as alternatives for examining combinations of practices and variation in adoption intensity (Kassie et al., 2015; Khonje et al., 2018; Manda et al., 2016).

Latent class models provide another approach when the objective is to identify distinct, unobserved patterns or groups of adoption among farmers. However, their primary focus is on identifying latent adoption profiles rather than estimating practice-specific determinants and interdependence among individual adoption decisions, as in the present study (Vermunt and Magidson, 2002). Evidence on individual practices reinforces the portfolio view. Crop diversification, small scale irrigation, and integrated pest management deliver distinct benefits for livelihood stability, water security, and sustainable pest control, yet they are rarely implemented alone and are instead embedded in broader bundles of mutually reinforcing technologies (Barmon et al., 2025b; Kopper and Ruelle, 2022; McCord et al., 2015; Umer et al., 2024). A further strand distinguishes the initial decision to adopt from the intensity of subsequent use and shows that binary measures reveal little about how deeply technologies become embedded in farming systems (Kolady et al., 2021; Wassihun et al., 2025).

2.4 Climate smart agriculture in Bangladesh

Evidence from Bangladesh mirrors these international patterns while reflecting the distinctive agroecology of the country. Studies consistently associate adoption with improved food security, greater resilience, and stronger rural livelihoods in climate vulnerable coastal and wetland regions (Hasan et al., 2018; Islam and Farjana, 2024; Majumder et al., 2024; Mazumder, 2024). They also document widespread climate awareness among farmers and identify education, farming experience, landholding, livestock ownership, credit, extension, training, and market access as important correlates of adoption (Billah and Hossain, 2017; Saha et al., 2019; Sarker et al., 2022; Saroar and Filho, 2016). Agroforestry has attracted particular attention for its contribution to carbon sequestration, biodiversity conservation, and microclimatic regulation (Ali et al., 2024; Hanif et al., 2015; Riyadh et al., 2021), while location specific innovations such as floating agriculture and sorjan cultivation can raise cropping intensity and resilience in deltaic environments even though their uptake remains uneven and context dependent (Kabir et al., 2022; Kundu et al., 2024). The climate smart potential of irrigation requires careful consideration in the Bangladesh context. Although dry-season irrigation has increased rice production, intensive groundwater abstraction in some areas has contributed to declining water tables and may increase the risk of saline upconing in coastal regions. Thus, the adaptation benefits of irrigation depend not only on the use of irrigation itself but also on the source of water and the technology used to deliver it (Mainuddin and Kirby, 2015; Qureshi et al., 2015; Shamsudduha et al., 2009). In contrast, surface-water-based and conjunctive irrigation systems can support dry-season production in southwestern Bangladesh while placing less pressure on groundwater resources (Krupnik et al., 2017). These findings sit within a wider agricultural economy in which crop diversification and value chain development shape household income and livelihoods (Barmon et al., 2025a; Prince et al., 2023).

2.5 Research gaps and the contribution of the study

Bangladesh evidence leaves four gaps that the present study is designed to close. First, much of the literature draws on small and localized samples and examines individual practices in isolation, which constrains its ability to explain how farmers within one production system assemble several practices into a coherent portfolio (Hasan et al., 2018; Saha et al., 2019; Saroar and Filho, 2016). Second, adoption is typically measured as a binary outcome, so a household that adopts a single practice cannot be distinguished from one that implements a broad and complementary bundle (Billah and Hossain, 2017; Sarker et al., 2022). Third, few studies combine objective indicators of climatic exposure, such as experience of drought, salinity, and pest pressure, with measures of institutional access within one estimation framework, which leaves the interaction between environmental risk and institutions insufficiently understood in the national context. Fourth, existing work examines agriculture broadly or concentrates on specific innovations, and integrated micro econometric evidence on the joint adoption of multiple practices within modern high yielding rice systems in the salinity affected southwest remains limited (Kabir et al., 2022; Kundu et al., 2024; Mazumder, 2024).

The present study responds to each gap. It estimates the adoption of five climate smart practices simultaneously within a multivariate probit framework, which identifies complementarities and trade-offs that single equation models applied to Bangladeshi data cannot capture. The specification combines objective climatic risk indicators with institutional access variables and thereby weighs environmental exposure against institutional support within one analytical framework. A normalized adoption count complements the practice-specific analysis by capturing heterogeneity across households in the breadth of CSA adoption. The measure provides a descriptive summary of the number of practices adopted by each household, while the econometric analysis examines factors associated with adoption of each practice and accounts for the interdependence among adoption decisions.

2.6 Conceptual framework

Figure 1 presents the conceptual framework that organizes the analysis and links the theoretical discussion to the empirical specification developed in Section 3. A farmer compares the expected utility from adopting a given practice under climate risk with the utility expected from continuing without it and adopts the practice when the expected net benefit is positive. Four broad groups of observable characteristics enter this decision, namely demographic and human capital characteristics, farm and resource endowments, institutional and market access, and climatic exposure and perception. The framework also recognizes that the five CSA adoption decisions may be interrelated. The multivariate probit model allows the error terms across the five adoption equations to be correlated, thus accounting for interdependence among adoption decisions that may not be captured by the observed explanatory variables. Positive correlations provide evidence consistent with complementarity between practices, whereas negative correlations provide evidence consistent with substitutability. The five binary decisions together define the household’s joint adoption portfolio, while the number of practices adopted provides a measure of the breadth of CSA adoption. As shown in Figure 1, the breadth measure is a normalized adoption count ranging from 0 to 1 and is used to describe heterogeneity in the extent of household engagement with the five practices. The framework further recognizes potentially reciprocal relationships between CSA adoption and some explanatory factors. Farmers who adopt CSA practices may subsequently be more likely to obtain credit, receive advisory services, or access market information, while adoption may also increase their attention to climatic risks. These potential feedback relationships are shown by dashed arrows in Figure 1, as they cannot be separately identified using the available cross-sectional data. The estimated relationships are therefore interpreted as conditional associations rather than definitive causal effects.

3 Methodology

3.1 Description of the study area

Southwest Bangladesh ranks among the farming regions of South Asia that are most heavily exposed to climatic stress, and it offers a well-suited setting for studying the adoption of climate smart agriculture practices under production risk. The three study districts lie within coastal saline belt and its inland margin, an area that covers roughly one fifth of the national territory and more than 30 percent of the net cultivable land, and that sustains a population of about 40 million people who depend largely on agriculture for their livelihoods (BBS, 2024; SRDI, 2010). Salinity is the defining agroecological constraint of the region. According to the Soil Resource Development Institute, the area of coastal Bangladesh affected by salt rose from about 0.83 million hectares in 1973 to 1.06 million hectares in 2009, an increase of roughly 28 percent, and the affected area extends across 49 subdistricts in about 20 districts (SRDI, 2010). Recurrent cyclones, tidal surges, and rising sea levels reinforce this intrusion of salt and expose rice production systems to repeated shocks (Kreft et al., 2017; Shameem et al., 2015). These conditions depress yields, raise production risk, and constrain the intensification of rice cultivation (Islam et al., 2022; Saha et al., 2019), which makes the region an instructive and policy relevant frontier for examining how farmers combine adaptive practices under a changing climate.

The study was carried out in three districts of the southwest, namely Satkhira, Khulna, and Jashore. These districts were purposively selected to capture the salinity and agroecological gradient that characterizes the region. Satkhira and Khulna sit within the coastal saline belt and are strongly shaped by tidal action and soil salinity, whereas Jashore lies further inland, is markedly less saline, and supports higher cropping intensity (Shameem et al., 2015; SRDI, 2010). This gradient allows the sample to capture variation in climatic exposure, soil salinity, and farming systems while holding the broader context of rice production constant. The southwest is also a recognized center of CSA research in Bangladesh, and much of the existing evidence on adaptation in coastal and saline farming systems has been generated in these very districts (Islam and Farjana, 2024; Kundu et al., 2024; Majumder et al., 2024; Mazumder, 2024; Saha et al., 2019; Sarker et al., 2022). The location of the three study districts is shown in Figure 2.

One village was selected purposively from each district for detailed enumeration, namely Binerpota in Satkhira, Bamundia in Khulna, and Kopalia in Jashore. The selection was based on three criteria, each assessed using district-level agricultural office records. First, rice had to account for the largest share of cultivated land in the village, ensuring that the sample represented modern rice-based farming systems rather than mixed or predominantly horticultural systems. Second, the village had to have documented exposure to climatic shocks during the previous decade, defined as at least one officially reported episode of cyclone damage, tidal inundation, salinity intrusion, or drought affecting the village. Third, the village had to have documented experience with at least three of the five climate smart practices examined in this study, based on extension programme records or the presence of a demonstration plot. This criterion was intended to ensure that the selected villages included exposure to a sufficiently broad range of the practices examined in the joint adoption model. Agriculture and inland fisheries are the principal livelihood activities at all three sites. Rice and fish dominate local production, and the cultivation of vegetables throughout the year is widely practiced on dikes within integrated systems that combine rice and fish. Cropping patterns follow the salinity gradient. Jashore supports double and triple cropping of paddy, wheat, maize, potato, brinjal, sesame, pointed gourd, onion, chilli, and jute, whereas cultivation in Khulna and Satkhira is governed by coastal processes and soil salinity, so that farmers grow watermelon, tomato, cucumber, onion, brinjal, pumpkin, and leafy vegetables alongside both traditional and improved rice varieties. Selected demographic and land use characteristics of the three villages are summarized in Table 1. Because the selected villages had documented engagement with CSA practices, the observed adoption rates may be higher than those that might be found in randomly selected villages in the southwest.

Table 1. Description of the study villages.

Source: Field survey, 2025.

3.2 Sampling design

The sampling design combined purposive selection of study sites with probability sampling of farm households within those sites. These stages support different types of inference. First, the three districts were purposively selected to represent the salinity and agroecological gradient described in Section 3.1. Second, one village was purposively selected from each district based on the three criteria described above. Third, farm households were selected through simple random sampling from complete village household lists obtained from the district agricultural offices, which served as the sampling frames. Thus, random sampling was applied only within the three selected villages and not at the district or regional level. This design therefore supports household-level estimates that are representative of the sampled village populations under the probability sampling design and provides empirical evidence on CSA adoption within the selected climate-vulnerable rice-producing systems.

3.3 Sample size determination

The required sample size was determined with the standard formula of Cochran (1963) for estimating a population proportion, given in Equation 1.

n 0 = Z 2 pq e 2 (1)

In this expression Z = 1.96 corresponds to the 95 percent confidence level, p is the expected adoption proportion, q = 1 − p, and e = 0.05 is the acceptable margin of error. The value p = 0.15 was adopted as an informed prior reflecting the relatively low prevalence of the least common CSA practice reported in earlier survey evidence from coastal Bangladesh, where the least frequently adopted practices were observed among fewer than one in five farmers (Billah and Hossain, 2017; Saha et al., 2019; Sarker et al., 2022). This value was chosen because the sample-size calculation was intended to accommodate the least common practice in the adoption portfolio rather than the average prevalence across practices.

In a five-equation system, the precision of the analysis may be constrained by the least frequently observed outcome. Using p = 0.15 and q = 0.85 gives an initial requirement of about 196 households. Because the sampling frame is finite, this initial figure was adjusted with the finite population correction shown in Equation 2.

n = n 0 1 + n 0 − 1 N (2)

The correction was applied separately within each village, with N equal to the number of farm households in that village, which yields the village level minimums reported in Table 2. To improve the precision of the multivariate estimates, which involve a large set of covariates across five simultaneous equations, and to permit disaggregated analysis across villages and practices, the achieved sample exceeds the corrected minimum in every village. In total, 493 farm households were surveyed and allocated across the three villages, with 152 households in Binerpota, 171 in Bamundia, and 170 in Kopalia (Table 2).

Table 2. Distribution of the sample size.

Source: Field survey, 2025.

A single-proportion sample-size calculation does not, by itself, establish that the achieved sample is adequate for a five-equation multivariate probit model. The adequacy of the sample is thus assessed using three additional criteria. First, descriptive precision is evaluated under the least favorable assumption. Applying the finite population correction to the combined sampling frame of 2,775 households, the pooled sample of 493 households gives a margin of error of approximately 4.0 percentage points at the 95 percent confidence level when p = 0.50, which maximizes sampling variance. At the village level, the corresponding margins of error are 6.8, 7.0, and 6.8 percentage points for Binerpota, Bamundia, and Kopalia, respectively. Second, the information available for the binary regression models is considered. Each equation includes 22 covariates and an intercept, resulting in approximately 21.4 observations per estimated parameter. The smaller outcome category contains between 158 observations for integrated pest management and 188 for integrated soil fertility management. Thus, the events-per-variable ratio ranges from 7.2 to 8.5. Although these values are slightly below the conventional threshold of 10, previous research shows that this threshold is not a rigid requirement and that its relevance depends on factors such as outcome balance and covariate strength (Peduzzi et al., 1996; van Smeden et al., 2019). Third, the balance of the binary outcomes is considered. In the present study, adoption prevalence ranges from 0.618 to 0.679 across the five practices, indicating relatively balanced outcome categories and therefore a comparatively favourable information structure.

3.4 Survey design, fieldwork, and data quality

The analysis draws on primary data from a household survey conducted between July and October 2025. A structured questionnaire was designed to elicit household adoption of the five CSA practices together with the demographic, farm, institutional, and climatic variables potentially associated with that adoption. The instrument was first prepared in English and then translated into Bengali, which ensured accurate communication with respondents and reduced the response bias that can arise from differences in language. The Bengali version was independently back translated into English by a second translator who had not seen the original questionnaire. The two English versions were then compared item by item, and differences in the wording of adoption and perception questions were resolved before fieldwork (Brislin, 1970). The survey gathered detailed information on household demographics, landholding, asset ownership, livestock, crop production, perceptions of climate change and climate variability, adoption of CSA practices, access to institutional support and agricultural services, participation in training, and contact with extension agents. The questionnaire was piloted with 30 farm households from a nearby village excluded from the final sample. Based on the pilot, adoption questions were revised to use the specific operational criteria reported in Table 3, and the recall period for climatic shocks and institutional contact was limited to the preceding cropping season to improve response accuracy. The asset module was also revised to use an inventory-based approach with prevailing local prices rather than open-ended valuation. Six enumerators received four days of training on the CSA practices, adoption criteria, interview procedures, and response checking. Face-to-face interviews were conducted with the household head or the member primarily responsible for cultivation decisions. A supervisor reviewed completed questionnaires daily and revisited a random 10 percent of households to verify key information. The completed questionnaires were double entered, reconciled, and checked for range and logical consistency before analysis. Of the 526 households approached, 493 completed the interview, giving a response rate of 93.7 percent. Perception-based measures were checked against reported crop failure and the village-level salinity classification recorded by the District Agricultural Office. Because farmers’ perceptions of soil conditions may differ from objective measures, these variables are interpreted as farmers’ own assessments relevant to their adoption decisions rather than as substitutes for laboratory measurements (Berazneva et al., 2018). As all information was collected from the same respondent in a single interview, common method variance cannot be completely ruled out. To reduce this risk, the adoption and perception questions were placed in separate sections of the questionnaire and different response formats were used (Podsakoff et al., 2003).

Table 3. Operational definitions of the five climate smart practices and the basis for treating each as climate smart in the study area.

Source: Authors’ construction based on the survey instrument and on FAO (2013), adapted to the rice based systems of the coastal southwest.

3.5 Operational definitions of the five climate smart practices

A household was classified as an adopter of a practice only when it met the operational criterion specified in Table 3, based on the relevant period specified for each practice. The criteria were developed from the CSA Sourcebook and adapted to the rice-based systems of the coastal southwest in consultation with district agricultural officers (FAO, 2013). Table 3 also explains why each practice is considered climate smart in this specific agroecological context, since the climate smart status of a practice depends on how and where it is applied rather than on its label alone (Chandra et al., 2018; Taylor, 2018; Thornton et al., 2018). The climate smart classification of small scale irrigation depends on how it is practiced in this setting. Irrigation is considered climate smart when it helps reduce dry-season and drought risks while using water sources that do not place excessive pressure on groundwater resources or increase the risk of saline upconing. In the study villages, irrigation mainly involved low-lift pumps using surface canals and shallow tubewells, consistent with the conjunctive systems documented for the southwest (Krupnik et al., 2017; Qureshi et al., 2015).

3.6 Analytical techniques

3.6.1 Measuring the breadth of CSA adoption

The breadth of CSA adoption is measured by a normalized adoption count that records how many of the five practices a household has adopted. Consistent with the portfolio view of CSA, the measure summarizes the extent to which households combine complementary practices rather than rely on a single technology (Kassie et al., 2015; Khonje et al., 2018; Teklewold et al., 2013). Each practice is coded as a binary indicator that equals 1 if household i has adopted it and 0 otherwise, and the adoption count for household i is defined in Equation 3.

B i = 1 K ∑ k = 1 K CS A ik (3)

In this expression CSA_ik denotes the adoption status of practice k by household i, and K = 5 is the total number of practices considered. The measure equals 0 when a household adopts none of the practices and 1 when it adopts all five, and it moves in steps of 0.2. The measure thus provides an equally weighted record of the number of practices adopted. Equal weighting is commonly used in studies of multiple technology adoption because it provides a simple and transparent measure that can be compared across households (Kassie et al., 2015; Teklewold et al., 2013). The adoption count characterizes how widely households engage with the CSA portfolio, which addresses the third objective, whereas the determinants of each individual practice, and the interdependence among these practices, are analyzed through the multivariate probit model, which retains the five binary practices as its dependent variables, in line with the second and fourth objectives. The two instruments are complementary rather than competing. The count answers how much of the portfolio is adopted, while the multivariate model identifies which factors are associated with each practice and how the practices relate to one another.

3.6.2 Preliminary analysis of associations among CSA adoption decisions

Before estimating the joint model, the pairwise associations among the five binary adoption indicators are examined using two complementary measures reported in Section 4.3. The phi coefficient (φ), which is equivalent to the Pearson product–moment correlation for two binary variables, measures the association observed directly in the sample. Its magnitude can be affected by the marginal adoption rates of the two practices and therefore does not directly represent the association between their underlying latent adoption propensities. The tetrachoric correlation provides a complementary measure by estimating the correlation between two continuous latent variables that are assumed to be jointly normally distributed and observed through binary threshold outcomes (Bonett and Price, 2005; Drasgow, 1986). The tetrachoric correlations therefore provide a descriptive indication of the association between the underlying adoption propensities and provide preliminary evidence on the interdependence among the practices (Dorfman, 1996; Kolenikov and Angeles, 2009). Reporting both measures allows the observed and latent associations to be considered separately. The two measures are evaluated using their respective statistical procedures. The significance of φ is assessed using the Pearson chi-square test of independence for the corresponding 2 × 2 contingency tables, while the significance of the tetrachoric correlations is assessed using their asymptotic standard errors.

3.6.3 Specification of the multivariate probit model

Farmers typically adopt several CSA practices at the same time, and these decisions are likely to share unobserved determinants such as managerial ability, risk preferences, and unmeasured plot and locality characteristics. Estimating a separate univariate probit for each practice would ignore this dependence across equations and would yield inefficient and potentially misleading estimates (Belderbos et al., 2004; Greene, 2012). The multivariate probit (MVP) model overcomes this limitation by estimating the adoption equations jointly while it allows the error terms to be freely correlated across practices (Cappellari and Jenkins, 2003; Young et al., 2009). The approach is now well established in the literature on the adoption of multiple technologies and of climate smart agriculture (Kassie et al., 2015; Rahman and Chima, 2014, 2016; Teklewold et al., 2013).

The empirical model follows from the theory of adoption under risk set out in Section 2, in which a risk averse farmer evaluates each practice by comparing the benefit expected from adoption with the benefit expected from the existing practice under an uncertain climate (Di Falco et al., 2011; Feder, 1980; Just and Zilberman, 1983). Consider first a farmer who faces a single practice, and let U^A_ij and U^N_ij denote the utility that farmer i derives from adopting and from not adopting practice j. Each utility depends on a vector of observed household, farm, institutional, and climatic attributes X_i and on a random disturbance that captures unobserved influences, as in Equation 4.

U ij A = X i β j A + ε ij A , U ij N = X i β j N + ε ij N (4)

A rational farmer adopts practice j whenever the utility obtained from adoption exceeds the utility obtained from non adoption. Because utility is not observed, adoption is analyzed through the latent net benefit of adoption Y_ij*, which is the difference between the two utilities defined in Equation 5.

Y ij ∗ = U ij A − U ij N = X i β j + ε ij (5)

In this expression β_j = β^A_j − β^N_j is the parameter vector that summarises the net influence of the attributes on the propensity to adopt, and ε_ij = ε^A_ij − ε^N_ij is a stochastic error term with zero mean. The latent variable Y_ij* is continuous and unobserved, whereas the analyst observes only the binary adoption outcome Y_ij, which records whether the net benefit is positive, as shown in Equation 6.

Y ij = { 1 if Y ij ∗ > 0 0 if Y ij ∗ ≤ 0 (6)

When the error term ε_ij follows a standard normal distribution, the probability that farmer i adopts practice j equals the standard normal cumulative distribution function evaluated at the index X_iβ_j, as in Equation 7.

Pr ( Y ij = 1 ∣ X i ) = ∫ − ∞ X i β j ϕ ( t ) dt = Φ ( X i β j ) (7)

Here Φ(·) denotes the cumulative distribution function of the standard normal. Equation 7 is the familiar single equation probit model, and it describes the adoption of one practice considered in isolation (Greene, 2012).

Because farmers assemble a portfolio of practices rather than adopt a single technology, the single equation model is generalized to a system of J correlated equations that are estimated jointly (Young et al., 2009). Let farmer i face J = 5 binary adoption decisions, and let each decision be governed by its own latent net benefit, so that the system takes the form given in Equation 8.

Y ij ∗ = X i β j + ε ij , Y ij = { 1 if Y ij ∗ > 0 0 if Y ij ∗ ≤ 0 , j = 1 , … , J (8)

The five latent equations correspond to crop diversification (Y_i1), agroforestry (Y_i2), integrated soil fertility management (Y_i3), small scale irrigation (Y_i4), and integrated pest management (Y_i5). All five equations share the same vector of explanatory variables X_i, which spans household demographics and human capital, farm and resource endowments, institutional and market access, and factors related to climate and risk. The interdependence of the adoption decisions is introduced through the joint distribution of the error terms rather than through the observed regressors, and this feature is the defining property of the multivariate probit model (Belderbos et al., 2004; Cappellari and Jenkins, 2003).

The vector of disturbances (ε_i1, …, ε_iJ) is assumed to follow a multivariate normal distribution with a mean vector of zeros and a symmetric covariance matrix that is normalized to a correlation matrix, as summarized in Equation 9.

( ε i 1 , ε i 2 , … , ε iJ ) ∼ N J ( 0 , Σ ) , E ( ε ij ) = 0 , Var ( ε ij ) = 1 , cov ( ε ij , ε ik ) = ρ jk (9)

The variances are normalized to unity because the scale of each latent index is not separately identified, so that the parameters of interest are the slope coefficients and the correlations that lie outside the main diagonal. The correlation matrix Σ collects these associations and takes the form displayed in Equation 10.

Σ = [ 1 ρ 12 ρ 13 ρ 14 ρ 15 ρ 21 1 ρ 23 ρ 24 ρ 25 ρ 31 ρ 32 1 ρ 34 ρ 35 ρ 41 ρ 42 ρ 43 1 ρ 45 ρ 51 ρ 52 ρ 53 ρ 54 1 ] (10)

The elements of particular interest are the terms outside the main diagonal, ρ_jk, each of which measures the correlation between the unobserved determinants of practices j and k. By the symmetry of the matrix these terms satisfy ρ_jk = ρ_kj, so that the model contains J(J − 1)/2 = 10 distinct cross equation correlations. A positive value of ρ_jk indicates that the unobserved factors which raise the probability of adopting one practice also raise the probability of adopting the other, which is the empirical signature of complementarity, whereas a negative value indicates substitutability (Belderbos et al., 2004; Teklewold et al., 2013).

3.6.4 Estimation, marginal effects, and inference

The model is estimated by full information maximum likelihood. Conditional on the observed choices, the contribution of farmer i to the likelihood is the probability of the particular combination of adoption decisions that the farmer is observed to make, which is a J variate normal probability evaluated over the relevant orthant of the error distribution (Cappellari and Jenkins, 2003; Young et al., 2009). Defining the sign indicator q_ij = 2Y_ij − 1, so that q_ij = 1 when practice j is adopted and q_ij = −1 when it is not, the likelihood contribution of farmer i takes the form given in Equation 11.

L i = Φ J ( q i 1 X i β 1 , q i 2 X i β 2 , … , q iJ X i β J , Σ ∗ ) (11)

In this expression Φ_J(·) denotes the J variate standard normal cumulative distribution function, and the elements of the transformed correlation matrix Σ* are ρ^*_jk = q_ijq_ikρ_jk. The sample log likelihood is the sum of the logarithms of these individual contributions across all farmers, and it is maximised with respect to the slope parameters and the correlation coefficients. Because the J variate normal integral has no closed form, it is approximated with the smooth recursive simulator of Geweke, Hajivassiliou, and Keane, which yields an unbiased estimate of the integral and thereby permits simulated maximum likelihood estimation of the full model (Cappellari and Jenkins, 2003).

Two probabilities summarize the behavior implied by the estimated model. The marginal probability that a farmer adopts a given practice, obtained by integrating out the remaining equations, is the univariate normal probability shown in Equation 12.

Pr ( Y ij = 1 ) = Φ ( X i β j ) , j = 1 , … , J (12)

The joint probability that a farmer adopts the entire portfolio of practices at once follows from the J variate normal distribution and is given in Equation 13.

Pr ( Y i 1 = 1 , Y i 2 = 1 , … , Y iJ = 1 ) = Φ J ( X i β 1 , X i β 2 , … , X i β J ∣ Σ ) (13)

Analogous expressions describe the joint probability of any other combination of adoption and non adoption, and it is this feature that allows the model to characterize the full pattern of portfolio choice rather than the adoption of each practice in isolation (Young et al., 2009).

With J = 5 practices the model yields the ten cross equation correlations displayed in Equation 10, and their joint significance provides the formal test of interdependence. The null hypothesis that the adoption decisions are independent corresponds to the restriction that every correlation outside the main diagonal is zero, and it is assessed with a likelihood ratio test of the hypothesis that all ρ_jk = 0. Rejection of this null implies that the adoption decisions are interdependent and that the multivariate probit model is preferred to a set of five independent univariate probits (Cappellari and Jenkins, 2003; Teklewold et al., 2013). The estimated correlations, the Wald statistic for the joint significance of the slope coefficients, the log pseudolikelihood, and the likelihood ratio test of independence are reported in Section 4.4.

Because probit coefficients convey only the direction and the statistical significance of an effect, average marginal effects are computed to quantify the change in the probability of adopting each practice that is associated with a change in each explanatory variable, and these effects are reported alongside the coefficients in Table 4 (Greene, 2012; Rahman and Chima, 2014). Average marginal effects (AMEs) are calculated for each of the 493 households using their observed covariate values and then averaged across the sample. For continuous and count variables, marginal effects are calculated from the derivative of the predicted probability, while for binary variables they are calculated as the discrete change in predicted probability when the indicator changes from 0 to 1, holding other covariates at their observed values (Williams, 2012). Standard errors of the AMEs are calculated using the delta method and reported in parentheses in Table 4. Inference is based on Huber and White robust standard errors, which are consistent in the presence of heteroskedasticity of unknown form. Cluster robust standard errors are not employed because the study design comprises only three villages, whereas reliable cluster robust inference generally requires a substantially larger number of clusters (Cameron and Miller, 2015).

Table 4. Multivariate probit estimates of the joint adoption of the five CSA practices.

Model diagnostics. Wald χ² = 201.020. Log pseudolikelihood = −1505.447. Likelihood ratio test of independence χ²(10) = 43.700 with p < 0.01. Prob > χ² = 0.000. Number of observations = 493.

Source: Authors’ calculations, 2026.

Coef. denotes the estimated probit coefficient and AME denotes the average marginal effect. Robust Huber and White standard errors are reported in parentheses beneath every estimate, and the standard errors of the average marginal effects are obtained by the delta method. Average marginal effects are computed as sample averages over all 493 households, as the discrete change from 0 to 1 for the binary regressors and as the partial derivative for the continuous and count regressors. Total agricultural land is measured in hectares and total asset value in BDT 100,000. Standard errors for the estimated cross equation error correlations are reported in parentheses. *, , and * denote statistical significance at the 10, 5, and 1 percent levels respectively.

3.6.5 Interpretations of the estimates and treatment of endogeneity

The estimates reported in Section 4 are interpreted as conditional associations rather than causal effects. This follows from the cross-sectional design, which does not establish temporal ordering, and from the possibility that some explanatory variables are jointly determined with CSA adoption. Farmers who intend to adopt may be more likely to seek credit, subsidies, extension services, and market information, while such services may also be directed toward more progressive farmers (Abdulai and Huffman, 2014). Similarly, farmers who already adopt CSA practices may pay greater attention to climatic risks and thus report stronger perceptions of climate change, salinity, or pest pressure (Arbuckle et al., 2015). These relationships cannot be separately identified with the available data. Because the survey does not provide a credible exclusion restriction suitable for identification, instrumental-variable or endogenous-switching approaches are not pursued (Wooldridge, 2010). The estimated marginal effects therefore describe differences in adoption probabilities associated with observed household characteristics and should not be interpreted as causal effects. The empirical specification is further assessed using diagnostic checks. Pairwise phi and tetrachoric correlations provide preliminary evidence of interdependence among the five adoption decisions, while the likelihood-ratio test of the joint null that all cross-equation error correlations are zero provides formal evidence for joint estimation. Multicollinearity is assessed using variance inflation factors, and the model is checked for convergence and numerical stability.

3.7 Description and measurement of the variables

The dependent and explanatory variables are defined in line with the study objectives, the conceptual framework of technology adoption under climate risk, and the availability of data. Full definitions and units of measurement are given in Table 5.

Table 5. Description and summary statistics of the variables.

Source: Authors’ calculations based on field- survey data, 2025.

Number of observations is 493. Monetary values are in Bangladeshi taka (BDT), where 1 USD was approximately 123.82 BDT in July 2026.

The dependent variables are the five binary CSA adoption indicators, namely crop diversification, agroforestry, integrated soil fertility management, small scale irrigation, and integrated pest management, each equal to 1 if the household satisfies the operational criterion set out in Table 3 and 0 otherwise. These practices were selected because they are widely promoted, readily observable, and well suited to rice systems in the coastal southwest (Kassie et al., 2015; Saha et al., 2019; Sarker et al., 2022; Teklewold et al., 2013).

The explanatory variables fall into four groups, each grounded in the adoption literature. The first group captures the demographic and human capital characteristics of the household, and it comprises the age, gender, and education of the household head, family size, the number of working members, farming experience, and the main occupation. Education and experience shape the capacity of a household to acquire, interpret, and apply technical information and rank among the most consistent correlates of technology adoption, whereas the effects of age and household composition are typically less uniform (Feder et al., 1985; Foster and Rosenzweig, 2010; Lin, 1991). The second group represents farm and resource endowments, which include farm size, total asset value, livestock ownership, perceived soil fertility, perceived soil salinity, irrigation frequency, and previous crop failure. These variables proxy productive capacity, wealth, liquidity, and the biophysical conditions that determine the feasibility and the expected returns of adoption (Arslan et al., 2015; Kassie et al., 2015; Teklewold et al., 2013). The third group measures institutional and market access. It covers access to credit, extension services, and market information, together with the receipt of agricultural subsidies and the distance to the nearest market, which proxies transaction costs and market integration. Credit relaxes liquidity constraints, extension and market information lower the cost of learning about new practices, and subsidies reduce the cost of adoption, all of which are repeatedly found to shape the uptake of CSA (Conley and Udry, 2010; Di Falco et al., 2011; Mazhar et al., 2021; Musafiri et al., 2022). The fourth group comprises variables related to climate and risk, namely household experience of drought and of pest infestation and the perception of climatic change over the long term. These variables capture exposure to production risk and the perceived need for adaptation, and they are expected to raise the likelihood of adopting practices that reduce risk (Abid et al., 2015; Belay et al., 2017; Di Falco et al., 2011).

3.8 Multicollinearity diagnosis

Before estimation, the explanatory variables were screened for multicollinearity with variance inflation factors (VIFs), which are reported in Table 6. All VIF values lie well below the conventional threshold of 10, above which multicollinearity is generally treated as a serious concern (Gujarati, 2004). The maximum value is 1.53 and the mean is 1.13, so the regressor matrix is well conditioned. The absence of problematic linear dependence among the regressors supports the stability and the reliability of the estimated coefficients.

Table 6. Collinearity statistics.

Source: Authors’ calculations, 2026.

4 Results and discussion

4.1 Descriptive statistics

Table 5 summarizes the profile of the sampled households across demographic attributes, farm resources, institutional access, and exposure to climatic stress. Three features stand out. First, adoption of all five practices is common and varies within a narrow band, from 0.618 for integrated soil fertility management to 0.679 for integrated pest management. These rates are higher than those reported in some earlier studies of coastal Bangladesh, as the study villages were selected partly based on their documented engagement with CSA (Billah and Hossain, 2017; Islam and Farjana, 2024; Mia and Roy, 2025; Saha et al., 2019; Sarker et al., 2022). Second, the sample reflects the structure of smallholder rice farming in coastal Bangladesh, with modest landholdings, limited formal schooling, and a predominance of male headed households. Third, exposure to production risk is pervasive, since a large share of households report soil salinity, recent crop failure, drought, and perceived changes in local climatic conditions. This combination of widespread environmental stress and constrained productive resources frames the central tension that the empirical analysis addresses. Rising exposure to climatic risk may strengthen the incentive to adopt climate smart practices, while limited land, finance, and information may restrict the capacity of many households to act on that incentive (Antwi-Agyei et al., 2021; Sisay et al., 2023). The full magnitudes are reported in Table 5.

4.2 Breadth of CSA adoption

To look beyond the adoption of individual practices, the normalized adoption count measures the number of practices that each household combines. The distribution in Table 7 shows that very few households adopt none of the practices and that adoption is concentrated among households that combine three or four practices. The mean adoption count is 0.64, equivalent to 3.22 of the five practices, with a standard deviation of 0.20 and a median of 0.60. Households adopting three and four practices account for 36.51 and 33.87% of the sample, respectively, together representing 70.38%. A further 7.71% adopt all five practices, while only 0.61% adopt none. This descriptive pattern indicates that most households operate with a moderate portfolio of climate smart practices rather than relying on a single technology, which is consistent with evidence that smallholders assemble adaptation measures into bundles (Kassie et al., 2015; Teklu et al., 2023; Wassihun et al., 2025).

Table 7. Distribution of the normalized adoption count.

Source: Authors’ calculations, 2026.

4.3 Association among the adoption decisions

Before interpreting the joint model, the pairwise associations among the five adoption decisions are examined as preliminary evidence of interdependence. Table 8 reports both measures discussed in Section 3.6.2. The phi coefficients below the main diagonal show the observed associations between the binary adoption indicators, while the tetrachoric coefficients above the diagonal estimate the associations between their underlying latent propensities under the bivariate normal assumption. The two measures have the same signs throughout, with tetrachoric correlations generally larger in absolute value, reflecting the effect of binary measurement on observed correlations. The discussion therefore focuses on the tetrachoric estimates as the descriptive counterpart of the cross-equation error correlations reported in Table 4.

Table 8. Tetrachoric and phi correlations among the five CSA practices.

Source: Authors’ calculations, 2026.

Tetrachoric correlations are reported above the main diagonal and phi coefficients below it. The phi coefficient is the Pearson product moment correlation applied to two binary indicators and describes the association observed in the sample. The tetrachoric coefficient estimates the correlation between the underlying latent propensities under the bivariate normal assumption imposed by the multivariate probit model. The significance of each phi coefficient is assessed with the Pearson chi square test of independence applied to the corresponding two by two table, and the significance of each tetrachoric coefficient is assessed against its asymptotic standard error. Number of observations is 493. *, , and * denote statistical significance at the 10, 5, and 1 percent levels respectively.

The associations are modest in magnitude and mixed in sign, which indicates that the practices are neither independent nor uniformly complementary. Two patterns are evident. Integrated soil fertility management is positively and significantly associated with crop diversification and with agroforestry. These practices share soil and land management objectives and tend to reinforce one another, which is consistent with evidence that soil improving measures form a complementary cluster within the broader portfolio (Kassie et al., 2015; Teklewold et al., 2013). The strongest associations, however, are negative. Small scale irrigation is negatively and significantly correlated with crop diversification, with integrated soil fertility management, and with integrated pest management. This pattern can be interpreted as evidence of potential substitution among some practices, but the underlying mechanism cannot be established from the correlation structure alone. Limited land, labour, and liquidity may cause irrigation to compete with diversification and soil and pest management for available resources, while saline conditions may also encourage farmers to focus on particular water or soil management strategies. These explanations are consistent with the wider literature (Karlan et al., 2014; Suri, 2011). The remaining associations are small and statistically insignificant, providing little evidence of strong pairwise association among the outcomes.

Two implications follow. First, the coexistence of positive and negative associations demonstrates that adoption decisions are shaped by common unobserved factors while remaining far from perfectly correlated, which is precisely the setting for which the multivariate probit model is designed. Second, the substitution revealed by the negative correlations is consistent with the breadth results reported in Section 4.2, since the five practices do not form a single complementary package and universal adoption of the entire portfolio is not to be expected. The estimated cross equation error correlations reported in the lower panel of Table 4 display the same qualitative structure as the tetrachoric correlations. Positive error correlations link integrated soil fertility management to crop diversification and to agroforestry, while negative error correlations link small scale irrigation to crop diversification, to integrated soil fertility management, and to integrated pest management. The likelihood ratio test rejects the null hypothesis that these error correlations are jointly zero, which provides formal support for joint estimation in place of a set of separate univariate probit models (Cappellari and Jenkins, 2003).

4.4 Results of the multivariate probit model

Table 4 reports the multivariate probit estimates. The model fits the data well. The Wald test rejects the null hypothesis that the coefficients are jointly zero at the 1 percent level, and the likelihood ratio test of independence rejects the null hypothesis of zero cross equation error correlations, which confirms that the five adoption decisions are statistically interdependent and are appropriately estimated within a single framework rather than separately. The discussion focuses on average marginal effects, which quantify the change in predicted adoption probability associated with a change in each explanatory variable. The table reports the average marginal effect for each regressor, with the corresponding robust standard error in parentheses beneath each estimate. This presentation allows the magnitude, direction, and statistical precision of each estimated association to be assessed directly. Because the data are cross sectional and several institutional and perception variables may be jointly determined with adoption, the marginal effects are interpreted as conditional associations rather than causal effects (Wooldridge, 2010). The analysis therefore emphasizes the direction, the magnitude, and the pattern of the estimated associations across practices rather than attributing causation to any single factor.

4.4.1 Demographic and human capital factors

Demographic and human capital characteristics matter selectively rather than uniformly. Age and farming experience are not significantly associated with the adoption of any practice, which suggests that neither age nor accumulated farming experience shows a statistically significant conditional association with adoption once the other observed factors in the model are considered. This finding is consistent with a literature in which the effect of age is inconsistent across settings and in which experience raises adoption only when it is complemented by knowledge and support (Feder et al., 1985; Foster and Rosenzweig, 2010). Education displays the most consistent positive association, since additional years of schooling are associated with a higher probability of adopting agroforestry, integrated soil fertility management, small scale irrigation, and integrated pest management. This result accords with the central role of human capital in the adoption of knowledge intensive technologies, because education strengthens the capacity to obtain, interpret, and apply technical information (Hasan et al., 2018; Huffman, 2001; Lin, 1991; Saha et al., 2019). Gender is associated with crop diversification, since male headed households are more likely to adopt it, a pattern that is consistent with well documented gender differences in access to land, labor, and inputs that shape the feasibility of alternative production strategies (Doss and Morris, 2001). Household labor endowments operate through particular practices, since family size and the number of economically active members are positively associated with integrated soil fertility management. The relatively labour-intensive nature of this practice may partly explain this association, and the relationship may also reflect differences in land or livestock holdings. Households whose main occupation is agriculture are also more likely to adopt agroforestry, which may reflect their stronger involvement in farming and the longer-term nature of agroforestry.

4.4.2 Farm characteristics and asset endowments

Farm resources shape adoption through both scale and complementarities across enterprises. Cultivated area is positively associated with crop diversification, agroforestry, and small scale irrigation, which is consistent with the argument that larger farms can allocate land to alternative practices more flexibly and can better absorb the fixed costs of new technologies (Feder et al., 1985). Livestock ownership is positively associated with agroforestry and with integrated soil fertility management, which points to complementarities between crop and livestock enterprises, since livestock supply organic inputs that improve soil fertility. This interpretation is consistent with evidence on the benefits of integrated crop and livestock systems for soil management and resilience (Herrero et al., 2010; Saha et al., 2019). Perceived soil salinity produces the most striking pattern, with associations that correspond to several of the substitution relationships identified in Section 4.3. Salinity is associated with a higher probability of adopting agroforestry, with a marginal effect of 0.164 significant at the 1 percent level, and small scale irrigation, with a marginal effect of 0.146 significant at the 1 percent level, and with a lower probability of adopting crop diversification and integrated pest management, with marginal effects of 0.087 and 0.121 in absolute value, significant at the 10 and 1 percent levels, respectively. The pattern suggests that salinity may be associated with a shift in adoption toward practices that address water and soil constraints. This interpretation is consistent with the salinity conditions documented in coastal Bangladesh (Dasgupta et al., 2015; Saroar and Filho, 2016). However, the observed relationships may also reflect the agronomic suitability of saline plots for diversification or differences in the types of practices promoted by extension programmes in saline areas. Irrigation frequency is positively associated with crop diversification and agroforestry, and previous crop failure is positively associated with integrated soil fertility management, a pattern consistent with the proposition that recent production shocks encourage investment in soil restoring measures (Di Falco et al., 2011). By contrast, total household asset value is not significantly associated with any practice, which suggests that household wealth does not show a statistically significant conditional association with adoption in these models, a result that resonates with evidence that expected returns and comparative advantage, rather than wealth as such, govern adoption (Suri, 2011).

4.4.3 Institutional and market related factors

Institutional and market factors display the strongest and most consistent associations with adoption, and they carry the clearest relevance for policy. Access to credit is positively associated with four of the five practices, and the largest difference arises in the probability of adopting small scale irrigation, where the marginal effect is 0.088 and is significant at the 5 percent level. This pattern is consistent with the argument that relaxing liquidity and risk constraints raises investment in technologies that require upfront expenditure, a mechanism that has been demonstrated experimentally for agricultural decisions (Asfaw et al., 2012; Karlan et al., 2014). Agricultural subsidies are positively associated with agroforestry and small scale irrigation, which indicates that public support is most strongly associated with practices that carry high initial costs or benefits accruing over a longer horizon. Access to extension services and access to market information exhibit the largest marginal effects among the institutional and market variables, reaching 0.126 for integrated pest management and 0.113 for small scale irrigation through extension services, both significant at the 1 percent level, with similarly strong associations through access to market information. These results underscore the central role of information and advisory support in reducing the uncertainty that surrounds unfamiliar and management intensive practices, and they align with evidence that information transmission through extension and social learning accelerates technology adoption and diffusion (Conley and Udry, 2010; Genius et al., 2014; Hasan et al., 2018; Majumder et al., 2024). Market access, measured by distance to the nearest market, is negatively associated with small scale irrigation, suggesting that greater distance may increase transaction costs and reduce incentives for irrigation investment. However, credit, subsidy, extension contact, and market information may be jointly related to adoption, as farmers intending to adopt may seek these services and service providers may target farmers who are already more likely to adopt. The estimated effects thus describe differences in adoption probabilities associated with institutional access rather than the causal effects of expanding such access.

4.4.4 Climate risk and perception variables

Climate related exposure and perceptions are systematically associated with adoption, but their influence is selective and operates alongside resources and institutions rather than independently. Farmers who perceive changes in climatic conditions are more likely to adopt integrated soil fertility management, small scale irrigation, and integrated pest management, and the association with small scale irrigation is particularly strong, with a marginal effect of 0.114 significant at the 1 percent level. Direct exposure to shocks reinforces this pattern in a practice specific manner, since the experience of drought is positively associated with crop diversification and small scale irrigation, and the experience of pest infestation is positively associated with integrated pest management. The practice specific alignment between the shock experienced and the measure adopted is the most notable feature of this group, and it is consistent with farmers responding to the particular risks they encounter by adopting the measures that address those risks most directly (Abid et al., 2015; Debela et al., 2025; Di Falco et al., 2011). The direction of the relationship between perception and adoption cannot be established from the cross-sectional data. Farmers’ perceptions may influence adoption, but adoption may also affect how farmers perceive climatic and production risks (Arbuckle et al., 2015). At the same time, the results are consistent with perception and exposure being insufficient on their own. The associations between climatic factors and adoption coexist with strong associations for household resources and institutional access, which is the pattern expected if awareness of climatic risk translates into adoption only when farmers also command the financial, informational, and technical means to act (Antwi-Agyei et al., 2021; Islam and Farjana, 2024; Sisay et al., 2023).

4.5 Model diagnostics and sensitivity assessment

Several diagnostic results support the adequacy and internal consistency of the estimated model reported in Table 4. The variance inflation factors reported in Table 6 have a maximum of 1.53 and a mean of 1.13, indicating that multicollinearity among the explanatory variables is unlikely to materially affect the estimates. The simulated maximum likelihood estimation converged successfully, and the Wald statistic of 201.020 rejects the joint null hypothesis that all slope coefficients are zero at the 1% significance level. The likelihood-ratio test of independence provides further support for the joint modelling approach. The test statistic is 43.700 with 10 degrees of freedom, rejecting the null hypothesis that the ten cross-equation error correlations are jointly zero. This result indicates that the adoption decisions are statistically interrelated and supports the use of a multivariate probit specification rather than five separate univariate probit models. The estimated error correlations also reproduce the general sign and ranking of the tetrachoric correlations reported in Table 8. In addition, the concentration of households at intermediate levels of adoption observed in Table 7 is consistent with the mixture of complementarity and substitution captured by the estimated correlation structure. Thus, the descriptive patterns and the econometric results provide broadly consistent evidence on the interdependence among adoption decisions. Inference is based on Huber and White robust standard errors, and average marginal effects are calculated for all 493 households using the delta method to obtain their standard errors.

5 Conclusion and policy implications

This study examined the joint adoption of five climate smart agricultural practices among modern rice farmers in the salinity affected southwest of Bangladesh. It set out to measure the breadth of adoption across households, to test whether the adoption decisions are interdependent, and to identify the factors associated with adoption. Using primary data from 493 farm households drawn by simple random sampling within three purposively selected villages in three districts spanning the coastal saline belt and its inland margin, the study first summarized the breadth of adoption using a normalized adoption count. Next, the study identified the correlates of the portfolio of CSA practices, namely crop diversification, agroforestry, integrated soil fertility management, small scale irrigation, and integrated pest management, jointly using a multivariate probit model.

Three main conclusions follow from the analysis. First, adoption decisions are interdependent rather than independent. The tetrachoric correlations and the estimated cross equation error correlations display the same structure, and the likelihood ratio test rejects the hypothesis that the practices are adopted independently. Soil improving measures form a complementary cluster, whereas irrigation-based intensification substitutes for diversification and for soil and pest management. This interdependence is consistent with the observation that most households operate an intermediate portfolio of practices rather than the full set, and it confirms that joint estimation is more appropriate than a series of separate models. Second, adoption is associated less closely with demographic attributes or with household wealth than with the combination of climatic exposure and institutional access. Education, farm size, and livestock ownership matter for particular practices, and perceived salinity produces a clear asymmetric pattern, yet access to credit, extension services, and market information emerge as the most consistent correlates of adoption across the portfolio. Third, awareness of climatic risk appears insufficient on its own. Farmers who perceive climatic change and who have experienced drought or pest pressure are more likely to adopt the practices that address those risks directly, and this association is observed alongside, rather than independently of, the financial, informational, and technical means to act.

These findings carry clear policy implications for strengthening climate smart agricultural adoption in the salinity-affected southwest of Bangladesh. The consistent association of credit, extension services, and market information with adoption highlights the importance of strengthening these institutional and informational channels, with extension services linking technical guidance to timely climate and market information and improved access to affordable credit helping address investment requirements. The interdependence among practices further suggests that policy should support complementary practices in a coordinated manner, while allowing flexibility in the CSA portfolio given the substitution between irrigation-based intensification and diversification, soil, and pest management. The results also indicate that climate-risk awareness is more likely to translate into adoption when farmers have the financial, technical, and institutional means to respond. The association of subsidies with small scale irrigation and agroforestry further supports targeted assistance for relatively investment-intensive practices; for irrigation, such support may consider water and energy sources and place greater emphasis on surface-water and conjunctive systems, alongside monitoring of groundwater levels and salinity to support sustainable adaptation (Krupnik et al., 2017; Qureshi et al., 2015).

The study has some limitations that also point to avenues for further research. The cross-sectional design identifies associations rather than temporal changes in adoption, and the study setting and purposive selection of locations define the empirical scope of the findings. Future research using broader samples, panel or quasi-experimental designs, and richer measures of adoption intensity could build on these findings by examining how CSA portfolios evolve and how different combinations of practices translate into productivity, income, and resilience outcomes.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Vijay Singh Meena, ICAR - Mahatma Gandhi Integrated Farming Research Institute, India

Reviewed by: S. R. Bishnoi, Indian Agricultural Research Institute (ICAR), India

A. H. M. Monzurul Mamun, Pabna University of Science and Technology, Bangladesh

Footnotes

The term modern rice production refers to rice based production systems based on modern high yielding rice varieties and typically characterized by the use of chemical fertilizers, agrochemicals, mechanization, and groundwater irrigation, in contrast to traditional low input systems based on local rice varieties (Kabir et al., 2015; Timsina et al., 2018).