Sustainable intensification under resource constraints: Estimating the heterogeneous effects of hybrid maize adoption in Nepal

ABSTRACT Managerial practices for farming-system intensification have received increased focus in research-and-development (R&D) initiatives. These technologies are proven to close the yield gaps in researcher-managed field trials and are recommended for farmer’s adoption. However, not all farmers have the technical, financial, and social capital to adopt and benefit from these recommended technologies. Is the current level of productivity enhancement achieved by smallholder system intensification sufficient to sustain rural livelihoods? To this end, the study assessed the impacts of hybrid maize (Zea mays L.) adoption on productivity and livelihoods in the mid-hill region of Nepal. Smallholders in the study region face severe shortages of labor, improved cultivars, and inorganic fertilizers, resulting in very low yields and profitability. We find that maize hybrid adoption increased crop productivity by 109%, making the crop profitable for smallholders and enhancing the per capita food expenditure by 20%. Nevertheless, these benefits were unevenly distributed: relatively small farms (≤0.3 ha) achieved greater gains in productivity and livelihood per land unit from hybrid maize adoption, but only larger farms (>0.3 ha) enjoyed the aggregate livelihood benefits of the technology. System intensification gains economic relevance because of the severe scarcity of resources, whereas the resource scarcity itself determines the economic relevance of system intensification, presenting a paradox. Increasing market access to material inputs did not significantly alter the observed patterns. More studies are required on the relationship between farm size and the livelihood impacts of sustainable intensification to facilitate R&D targeting and ensure inclusive development.


Introduction
The sustainable intensification of agriculture (SIA) is broadly defined as the process of producing more output from the same land area without damaging the environment while improving livelihoods and food security (Rudel 2020;Godfray and Garnett 2014). The enhancement of system productivity through the efficient use of marketed inputs and natural resources forms the core of many SIA approaches (Haile et al. 2017). Some of the technologies useful for closing yield gaps through system intensification rely on the timely availability of quality external inputs, for example, herbicides for weed management in conservation agriculture (CA) systems (Bouwman and Giller 2020). Insufficient market access to critical inputs in many regions of the Global South, therefore, often impedes technological change (Qian, Ola, and Benjamin 2020).
There is rich socioeconomic literature on the adoption and impacts of SIA technologies in maize systems, with several dozen papers published annually (Garcia and Krishna 2021;Christoph and Krishna 2020). However, most of these studies are confined to sub-Saharan Africa (SSA), and there is a significant information gap on the status and challenges faced by the maize farmers of South and Southeast Asia. In the recent past, many Asian countries have expanded maize production and diversified its use from food to the feed and fuel sectors, and maize has assumed the status of a cash crop in this region (Shiferaw et al. 2011). System intensification with hybrid maize technology necessitates well-developed input and output supply chains (Mango et al. 2018;Alia 2017), and the timely availability of quality inputs, especially inorganic fertilizers, labor, and credit, is a prerequisite to ensuring an efficient production process (Alene and Hassan 2006; Ghimire and Huang 2015). However, the agriculture of several developing countries of South and Southeast Asia in general, and Nepal in particular, is characterized by poorly developed seed systems (Gauchan 2019;Spielman and Kennedy 2016), a constrained supply of inorganic fertilizers (Ward et al. 2020), small and fragmented farms (Niroula and Thapa 2007), and input shortages, resulting in high cultivation costs (Paudel et al. 2019). These factors could limit the scope of the sustainable intensification of maize systems using hybrids as a technology component.
Compared to open-pollinated varieties (OPVs), hybrid maize possesses a high genetic yield potential due to heterosis (Flint- Garcia et al. 2009). Hybrid vigor drops with the reuse of farm-saved seeds, resulting in a perpetual market demand for seed that attracts private seed companies (Spielman and Kennedy 2016;Raghu et al. 2015). Although hybrid maize is promoted and widely cultivated across several developing countries as a sustainable intensification technology due to its high land productivity potential and high nutrient-use efficiency (Devkota et al. 2016), the relative benefits of the technology are dependent upon the local agro-climatic and market conditions (Kathage et al. 2015;Alene and Hassan 2006). Because not many socioeconomic evaluations of maize production technologies have been carried out in South Asia in the last decade (Krishna et al. 2019;Garcia and Krishna 2021), we hardly know anything about the relative advantage of adopting hybrid seeds in countries like Nepal. Some of the experimental trials conducted in Nepal confirmed that maize farmers could improve land productivity by switching from local OPVs to hybrids (Devkota et al. 2016(Devkota et al. , 2015. However, these experimental trials are inadequate for shedding light on the biophysical and socioeconomic impacts of the technology across diverse farm types and production environments. Against this backdrop, the present study examined the adoption and impacts of hybrid maize among smallholder farmers of the mid-hills of Nepal, where maize grain is primarily used for household consumption. We hypothesize that the impacts of the technology on productivity and farmers' livelihoods are heterogeneous due to the severity of resource constraints and differential access to input markets. The unique situation of Nepalese agriculture that shapes farmers' adoption of new technologies is discussed in the next section (Section 2). The data collection methods, summary of data used for the empirical analysis, and analytical framework are provided in Section 3. The empirical results are presented and discussed in Section 4. The last section (Section 5) concludes the study and provides policy recommendations.

Resource constraints and the history of hybrid maize dissemination in Nepal
Despite being an agrarian economy, with two-thirds of its population dependent on agriculture for livelihood (MoAD 2017; World Bank 2020), food insecurity is rampant in Nepal. Two-thirds of Nepal's districts face food shortages every year (Joshi, Conroy, and Witcombe 2012), and a quarter of the population lives in absolute poverty (NPC 2017). Among the three distinct agro-ecological zones of the country, food insecurity is most pronounced in the mid-hill and mountain regions . Low agricultural productivity due to the low adoption of modern production technologies has been identified as the primary reason for food insecurity and high poverty in the region (Subir, Mishra, and Giri 2019;Ghimire and Huang 2015). The adoption of technologies, such as hybrid maize, has the potential to improve food security and reduce rural poverty in Nepal by increasing crop productivity and profitability.
Maize is one of Nepal's major cereal crops, where it is grown for food, feed, and fodder (Bahadur et al. 2014;Tiwari, Virk, and Sinclair 2009). The average domestic consumption of maize grain as food in Nepal is about 3 kilograms per person per month (Peter, Pe˜a-Rosas, and Maria Nieves 2014), and a substantial portion is also dedicated to livestock production (Paudel et al. 2014). The use of maize grain differs widely across agro-ecological regions. In the mid-hills and mountains, it is mainly used for human consumption, unlike in the Terai region, where maize is primarily used for industrial purposes (e.g., for making poultry feed) (Govind et al. 2015;CBS 2015). Domestic demand cannot be met with the current level of maize production in Nepal. The average maize yield of Nepal is 2.6 tons/ha (as of 2018), which is less than half the global average (FAO 2020; MoAD 2017). Low crop productivity and high demand necessitate the import of maize (Timsina, Nath Ghimire, and Lamichhane 2016).
The history of hybrid maize adoption is relatively recent in Nepal. In 2003, the National Maize Research Program (NMRP) released the first maize hybrid ("Gaurav Hybrid") (SQCC 2013), recommended for cultivation in the lowland agroecology (altitude <600 m). Since then, seven hybrids have been developed and released by the NMRP in collaboration with the International Maize and Wheat Improvement Center (CIMMYT). However, hybrid seed production has taken place at a slow pace in the country, with a limited quantity of seed commercially available in the market (Adhikari 2014). Furthermore, "Gaurav Hybrid" did not become popular with maize farmers (SQCC 2013). Between 2010 and 2018, about 61 hybrid maize cultivars were registered by the Government of Nepal (SQCC 2019), of which only eight were developed domestically. With an increasing demand for hybrid maize seed by farmers and a shortage of domestic supply, in 2011, the government liberalized the hybrid maize seed market, which presented an opportunity for regional seed companies to enter the Nepalese market and formally register maize hybrids. 1 The reduction of dependence on cereal imports, including maize, has gained high political importance, as evidenced in the Government of Nepal's Agriculture Development Strategy  and Seed Sector Development Strategy (SQCC 2013). 2 These strategies emphasize the intensification of internal productivity through developing and deploying high-yielding hybrids, improvements in crop management, and the efficient use of fertilizers. However, the country's ability to intensify maize productivity is affected by the constrained access to production inputs ( Figure 1). Firstly, the Nepalese agriculture sector is suffering from an acute labor shortage due to the increasing trend of labor out-migration, which has increased five-fold in recent years from the year 2000 ( Figure 1a). This accelerating trend of domestic labor loss has sharply increased rural wages (more than doubled between 2002 and 2017) ( Figure 1b). Rising rural wages due to labor shortages have sharply increased production costs of all crops. Secondly, the average household landholding size has been reduced by 36% (Figure 1e), and the per capita landholding reduced by 31% during the last three decades (Figure 1d). Thirdly, due to a nascent national seed system, the country relies heavily on imported hybrid maize seed, while the demand for hybrid maize seed has expanded from 20 tons in 2008 to 1,410 tons in 2017 ( Figure 1c). This high dependence on the import of hybrid seed has created a trade deficit of almost US$ 4 million ( Figure 1f) from the maize seed sector alone. Moreover, the price of the imported hybrid seed is high, and resource-constrained farmers may be unable to purchase a sufficient quantity of seed, limiting the scope of maize system intensification. Finally, due to underdeveloped industries, Nepal currently imports all inorganic fertilizers from other countries. Although the overall import of inorganic fertilizers, such as urea, di-ammonium phosphate (DAP), and potash, has drastically increased during the last two decades (Figure 1g-i), it can only fulfill <50% of the current domestic demand.

Data
The basis of this empirical study is a farm-household survey dataset collected from Nepal's mid-hill region's villages during October-November 2017. Face- to-face interviews using a structured questionnaire were implemented with a computer-assisted personal interview (CAPI) software, with several validation rules to minimize data entry errors and survey time. The questionnaire elicited information on the household's socioeconomic status, cropping systems, inputs for maize cultivation and outputs, and sources of household income and expenditure.
The data came from six districts from the mid-hills of Nepal: Doti, Surkhet, Palpa, Nuwakot, Kavre, and Illam, which were purposively selected based on the area under maize cultivation. The location of the selected districts is shown in Figure 2. The district-level maize acreage was quantified in consultation with the District Agriculture Development Offices, key informants, and agricultural input dealers (e.g., agro-vets, who sell maize seed to farmers). According to the Agriculture Knowledge Centers of the Government of Nepal, about 6.02% of the maize area was cultivated with hybrid seeds in the selected districts (Appendix Table A1). In each district, the area under maize and the extent of hybrid maize adoption in sub-districts (Village Development Committees or VDCs) were estimated based on the area under maize cultivation and an initial estimation of hybrid maize adoption among farmers. A total of 34 VDCs were purposively selected, and 731 maize-growing farm-households were randomly selected from these VDCs for the survey. Here, we define hybrid maize adopters as farmhouseholds that cultivated hybrid maize in any of their plots in 2017. 3 About 43% of sample households were adopters of the technology, and their over-representation was necessary for impact estimation, which was ensured through the purposive selection of VDCs.

Empirical framework
As economic agents, farmers are often assumed to be resource-constrained and rational, attempting to maximize their yields and profits through the judicious use of scarce resources like land, labor, and material inputs. Under this assumption, farmers adopt hybrid maize technology based on whether the expected benefit from adoption is higher than the non-adoption status quo (OPV cultivation) (Abdoulaye, Wossen, and Awotide 2018;Jaleta et al. 2018;Mishra et al. 2016;Mishra, Khanal, and Pede 2017;Mishra et al. 2018;Shiferaw et al. 2014). For the empirical analysis, we restrict our analysis to monetary benefits, although non-monetary benefits (e.g., consumption utility) can also play a crucial role in determining farmers' adoption decisions (Krishna et al. 2013).
Let us assume that Ŷ is the difference in the net gain in the outcome variables between hybrid maize adopters and non-adopters. Then, b Y > 0 implies that the adoption of hybrid maize is more beneficial to the farmer than non-adoption. However, Ŷ cannot be observed directly and can only be expressed as the function of observed farm-level socioeconomic attributes in a latent model (details of the empirical framework are provided as Supplementary Materials Text). However, estimating the causal effect of hybrid adoption on the selected outcome indicators (e.g., maize productivity) is difficult due to the likelihood of an endogeneity problem. Finding the "true causal effect" of the technology adoption on key outcome indicators requires controlling observed and unobserved sources of heterogeneity between technology adopters and non-adopters (Wooldridge 2010;Angrist and Pischke 2009). Technology adopters and non-adopters may differ in their inherent individual skills and abilities. Failure to account for these heterogeneities may bias parameter estimates and result in false inferences. In this regard, the use of the ordinary least squares (OLS) method that can control only the observed heterogeneities while estimating the effect of technology adoption may lead to biased estimates. In this study, we used an endogenous switching regression (ESR) with a selection instrument to account for both sources of heterogeneity. To estimate the causal effect using ESR, the selection instrument should affect outcome indicators -maize productivity, gross margin, and per capita food expenditure (PCFE) -only through hybrid adoption.
With the help of the instrumental variable, the ESR addresses the problem of endogeneity by estimating the selection equation with a binary adoption variable (first stage) and the outcome equation with a continuous variable (second stage) simultaneously, employing the full information maximumlikelihood estimation approach (Lokshin and Sajaia 2004).
The instrumental variable used in this study is the number of years (duration) of availability of hybrid maize seed at the local input dealers. Local availability of the technology determines its adoption but might not affect the outcome variables directly. We assume that other farmers in the village gradually started adopting hybrid maize with the availability of the technology (hybrid seed at the village dealers) after witnessing the productivity gains in adopted farms. Nevertheless, one may assume that the duration of hybrid seed availability at the local dealers might not affect maize productivity, gross margin, and PCFE unless the technology is adopted. However, it is imperative to rule out the possibility that general economic development in the village does not determine both input supply chain formation and farmer income (Christoph and Krishna 2020). We included regional dummies and distance to input markets from households to capture the differences in the general economic development of the locality. Furthermore, we verified the suitability of our instrument by conducting a simple falsification test, following Di Falco et al. (2011). A good instrumental variable would be strongly associated with hybrid adoption but not with the outcome indicators for non-adopters. This instrument falsification test suggested that our instrument satisfied the exclusion restriction for productivity and PCFE but not for profitability (Appendix Table A2).
We estimated the average treatment effects on the treated (ATT) for the hybrid maize adopters and the average treatment effects on the untreated (ATU) for the non-adopters from ESR models. Furthermore, following Di Falco et al. (2011), we estimated the heterogeneity effects; hybrid maize adopters may have a different socioeconomic profile than non-adopters, shaping the impact magnitude. Further details of the empirical framework are provided as Supplementary Materials Text.

Descriptive statistics
The summary statistics of the input and output variables disaggregated by hybrid maize adopters (n = 311) and non-adopters (n = 420) are presented in Table 1, while the detailed maize enterprise budget is presented in Appendix  Table A3. The seed cost for the hybrid adopters was about five times higher (NPR 8,665 or US$ 83 per ha) than for non-adopters (NPR 1,731 or US$ 17 per ha), and this could be one of the major limiting factors for hybrid maize dissemination among poor farm-households of the mid-hills of Nepal. Hybrid maize adopters also applied inorganic fertilizers at a significantly higher rate. Labor costs were also high, but the expenditure for land preparation (particularly the cost of tillage operations) was lower for adopters. The lower land preparation cost for adopters could be associated with mechanized tillage instead of human or animal traction. Due to higher expenditure on material costs and human labor, the total variable cost for hybrid adopters was significantly higher: on average, NRP 74,299 (US$ 714) per hectare for adopters against NRP 60,380 (US$ 581) per hectare for nonadopters. On the other hand, maize productivity was 93% higher (at 4,370 kg per ha) for hybrid maize adopters than for non-adopters (at 2,261 kg per ha). Albeit with higher variable costs, gains in crop productivity enabled hybrid adopters to secure a significantly higher gross margin (NPR 28,773 or US$ 277 per ha) than non-adopters, who were losing money (gross margin NPR -2,930 or US$ -28 per ha) on average. However, due to the small farm size (0.27 ha), the average incremental household income gain from hybrid adoption was relatively modest (NPR 7,769 or US$ 75). Adopting farmhouseholds had a 29% higher PCFE (NPR 13,093 or US$ 126) than nonadopting ones (NPR 10,175 or US$ 98). 4 Selected socioeconomic attributes of hybrid maize adopters and nonadopters are presented in Table 2 and further details in Appendix Table  A4. The average farm size for maize growers in the study area was 0.42 ha, and the farm size was not statistically different between adopters and nonadopters. However, on average, the adopters had a slightly larger area under maize cultivation (0.27 ha) than the non-adopters (0.25 ha). The hybrid adopters had more years of experience in cultivating maize. Among the adopters, the percentage of male household heads (87%) was higher among adopters than that of non-adopters (82%), and they were located closer to input markets (5.67 km) than the non-adopters (9.53 km). Closeness to markets could be one of the main drivers of the adoption of hybrid maize and the increased use of material inputs, including inorganic fertilizers. Moreover, the number of years since maize hybrid was first introduced in the village agro-vet (private input dealers) stores was significantly higher for hybrid maize adopters than for non-adopters. Finally, geographic differences Notes: For all the socio-economic variables included in the analyses, please see Appendix (Table A4). ***, ** , and * indicate that the difference between adopters and non-adopters is significantly different at 1%, 5%, and 10% levels respectively. † NPK includes the amount of nitrogen, phosphorus, and potassium applied through different forms of fertilizers such as Urea, DAP, and Potash. Exchange rate 1 US{{footpara}}amp; #x00A0;= NPR 104, during the survey year 2017 (NRB, 2019).
were also observed for the adoption of hybrid maize. There was a higher rate of adoption in the central hills, as compared to the eastern and mid-western hill districts of Nepal. Table 2 also contains a summary of variables, such as a farmer's age, education, caste, and primary occupation, across the two adoption categories. In the adopter group, the share of households belonging to the socially nonmarginalized castes (such as Chhetry and Brahmin) was high at 61%. Among non-adopters, the difference between the share of households belonging to marginalized castes (such as Dalits and Janajatis; 52%) and that of nonmarginalized castes (48%) was statistically insignificant. There was no notable difference in perceived labor scarcity among adopters and non-adopters, but hybrid adopters were found to secure labor by providing higher wages for farming operations. A significantly higher percentage of adopters reported difficulty in finding draft animals for agricultural land preparation, leading to the increased use of mechanized tillage. Although there were no detectable differences in off-farm income, a higher percentage of adopters' farms had concrete houses, indicating that relatively wealthy households adopted hybrid maize (Table 1).

Adoption of hybrid maize in the mid-hills of Nepal and its implications
The key results from the ESR model estimates for maize productivity, profitability, and PCFE are presented in Table 3, and the full models in Appendix Tables A5-A7. The ESR method employed jointly estimated the selection equation in the first stage and the outcome equation in the second stage, as specified in the empirical framework section. The empirical estimation of the selection equation can be interpreted as that of normal probit coefficients. In Appendix Tables A5-A7, we also provide the OLS coefficients, in which the coefficient of hybrid maize adoption was found to be positive and statistically significant. Nevertheless, the correlation between the error term in the selection equation and the outcome equation was different from zero in the ESR model on maize productivity, indicating the existence of selection bias in using OLS (Lokshin and Sajaia 2004). The correlation between the gross margin and the PCFE in the ESR framework, however, was not statistically significant, indicating the absence of selection bias in these models. We kept the ESR estimates, however, and derived treatment effects to gain insights into the heterogeneous effects of hybrid maize adoption with respect to key socioeconomic variables, viz., landholding size, and market access.
Household heads with a higher number of years of experience in maize cultivation, farms where a higher rate of inorganic fertilizers in maize was applied, and households that paid high agricultural labor wages had a greater probability of adopting maize hybrids. Farm households that possessed concrete houses increasingly adopted hybrids. Farmers with better market access were found to have a higher chance of adoption. The instrumental variable -the duration of availability of hybrid maize seed with the local trader -was also statistically significant and positive. Finally, spatial heterogeneity in hybrid maize adoption was evident from the significant regional dummies in the selection models (Appendix Tables A5-A7). We do not overinterpret the adoption estimates, as our primary goal is to examine the impacts of adoption, and the selected models with binary dependent variables have severe limitations with regard to capturing the complex decisionmaking process concerning technology adoption (Glover, Sumberg, and Andersson 2016; Garcia and Krishna 2021).   Tables A5-A7. ***, ** , and * indicate that coefficients are significantly different at 1%, 5%, and 10% levels, respectively. Standard errors of coefficients in parentheses. # This particular selection model is taken from ESR model on maize productivity. The selection model estimated with other outcomes are slightly different.
The rest of Table 3 and Appendix Tables A5-A7 include the key estimates on maize productivity per Ha, gross margin per ha, and PCFE for hybrid maize adopters and non-adopters, respectively. Farm size was negatively associated with maize productivity, indicating the higher productivity of small farms compared with large farms. In the literature, there is plenty of evidence for highly productive small farms, owing to better crop management (Carter 1984;Bardhan 1973;Ramesh, Prasanna, and Singh 2017;Yiyun et al. 2018). However, in our case, most sample farms were in the marginal farm category, with 93% farms operating under 1 ha, and hence better crop management could not be the only reason for the high productivity of smaller farms. Furthermore, the coefficient of farm size was positive (0.14) for non-adopters in the model, with the gross margin per ha as the dependent variable. Similarly, farm size was positively associated with PCFE among adopters: with a 1% increase in farm size, the PCFE of hybrid maize adopters increased by about 0.24%. Land was not only a factor of production but also an asset facilitating access to working capital, and hence had a conflicting role in enhancing productivity and profitability. Furthermore, the smallest farmers of the Nepal hills might be able to acquire production inputs in small doses, as required, increasing not only maize productivity but also the variable cost of production.
In the ESR outcome equations, we included some but not all possible indicators of input scarcity. The included ones were found to limit maize productivity. The productivity of hybrid maize adopters who faced difficulties finding human labor was lower by 11%, and of non-adopters by 17% (Appendix Table A5). 5 This variable also had a significant negative effect on the household's PCFE. This finding is consistent with earlier studies in Nepal that showed that labor outmigration and shortages caused a significant delay in crop establishment and farm operations and had a negative effect on crop productivity (Khanal 2018;Uttam et al. 2015;Knerr 2013b, 2013a). Moreover, the marginal effects on productivity of inorganic fertilizers for hybrid maize adopters and non-adopters were similar, although the former applied more fertilizers and had higher nitrogen-use efficiency (Devkota et al. 2016). Finally, an increase in the labor wage rate was found to be negatively associated with maize profitability for both adopters and non-adopters, which was not surprising. However, a positive association between the wage rate and PCFE existed. Most farm households hired out their labor, and a higher wage rate enhanced household income and hence food consumption.
The estimates of the impact of hybrid maize adoption on maize productivity, gross margin per hectare, and PCFE are presented in Table 4, together with the OLS estimates. The marginal effect of the OLS estimates after transformation included a 74% increase in maize productivity, a 34% increase in gross margins per hectare, and a 14% increase in PCFE. The ESR values also showed significantly higher benefits for technology adoption. Here, we reported the ATT and ATU values for adopters and non-adopters separately. The ATT values capture the difference in the outcome variables for hybrid adopters between the current estimates (with the adoption of the technology), and the estimates, had they not adopted it. Similarly, the ATU values show the difference between the current estimates and the outcomes of the non-adopters had they adopted the technology. The estimates showed that the adoption of hybrid maize has a significant and positive impact on maize productivity, gross margins, and PCFE for adopting households. It would also have been beneficial for non-adopters had they adopted maize hybrids.
For adopters, the adoption of hybrid maize enabled the sample households to increase maize productivity by 109% and PCFE by 20%. For non-adopters, adoption would have increased productivity by 66% and PCFE by 64%. Maize cultivation in the study area was not profitable for non-adopters, and this would also have been the case had adopters not adopted the technology. Despite the statistically significant, positive ATT estimates, the crop income accrued to the household due to adoption was not high (NPR 10,622 or US$ 102) for sustaining a rural household due to the small size of farm-holding. This could be one of the reasons for the relatively small increase in PCFE (NPR 1,982 or US$ 19 per household) for adopters. More research is required on whether this increase has helped farmers to cross over the poverty threshold. In the case of non-adopters, however, the use of maize hybrids could help them avoid financial losses from crop production. Hybrid maize adopters and non-adopters were systematically different, as indicated by the transitional heterogeneity (TH) estimates in Table 4, and a direct comparison of the impacts of hybrid maize on adopters and non-adopters without addressing the observed and unobserved heterogeneities would lead to biased estimates. In contrast, the positive and statistically significant values of TH for productivity and gross margin (ATT > ATU) showed that hybrid maize adopters had the potential to realize higher productivity and gross margin. For PCFE, however, ATU > ATT, possibly because non-adopters face negative returns in the absence of the technology.
We checked the robustness of our findings using the Inverse Probability Weighted Regression Adjusted (IPWRA) method. 6 The results are presented in Table 5. Consistent with the results from the ESR, the treatment effects for hybrid maize adoption on maize productivity, gross margin, and PCFE were statistically significant. The IPWRA results show that the adoption of maize hybrids increased crop productivity, gross margins, and PCFE by 1,586 kg/ ha, NPR by 25,972 per ha (US$ 250), and NPR by 1,747 (US$ 17), respectively. However, the impact magnitude was lower than that estimated in the ESR framework, possibly because unobserved heterogeneity was not accounted for in the IPWRA method.
In sum, the OLS and ESR estimates presented in this section clearly indicate the significant positive productivity, gross margins, and welfare (i.e., PCFE) effects of hybrid maize adoption, even if farmers have several resource constraints. While our findings are similar to the studies conducted in other regions (Becerril and Abdulai 2010;Abdoulaye, Wossen, and Awotide 2018;Mathenge, Smale, and Olwande 2014;Ahmed et al. 2017;Manda et al. 2018;Jaleta et al. 2018;Kassie, Jaleta, and Mattei 2014), they also indicate that transition to maize hybrid technology will have modest benefits for smallholder farm-households working under severe resource constraints. Given that the per capita land availability in Nepal is less than 0.12 ha (Figure 1), that maize is only one of several crops cultivated in the farming systems, and that the hybrid maize technology could increase gross margins by US$ 378 per ha (US$ 45 per capita per season; ATT from Table 4), more assessments are required on the potential of the technology to reduce poverty and food insecurity in the mid-hill region of Nepal. The effects of technology on the livelihoods of agricultural laborers should also be examined. Due to the heightened demand for human labor with adoption, the technology may also be beneficial for the labor-providing households of Nepal.

Heterogeneous effects of hybrid maize adoption
Technology adoption among maize farmers may have differential impacts across different socioeconomic strata, depending on the degree of resource scarcity that the farmers face. To examine these heterogeneous effects, we stratified the data into different categories by (a) farm-size quartiles and (b) farmers' access to input markets. The results for the heterogeneous effects of hybrid maize adoption across these two categorical variables are presented in Tables 6 and 7.
To assess the impacts of hybrid maize adoption on productivity, gross margin per ha, and PCFE across different farm sizes, we stratified the sample into four farm-size quartiles (Table 6). An inter-quantile ATT comparison shows unique patterns, demonstrating that the livelihood impacts of hybrid seed adoption are crucially dependent on the resource status. The same holds true for ATU. While the impacts of hybrid maize adoption on maize productivity and gross margins (for both ATT and ATU) across all the farmsized quartiles were statistically significant at the 1% level for the adopting farms, the magnitude of the effects of the hybrid technology on productivity and gross margins (NPR per ha) was considerably higher for the first quartile farms; the size of the effect diminished in the third and fourth quantiles. On the other hand, the treatment effects of hybrid maize adoption on PCFE for the adopters' farms were positive and statistically significant only in the third and fourth quartile farms. The non-significant livelihood effects of hybrid maize adoption could be associated with a smaller area allocated to maize in the first and second quartile farms (an average of 0.15 ha) than in the third and fourth quartile (0.35 ha) farms. Irrespective of the high agronomic potential of hybrid maize technology, a certain minimum landholding size is required for its livelihood impacts to be manifested.
To assess the impacts of hybrid maize adoption across different farm sizes and levels of market access, we stratified our data into low and high access to a market with respect to large and small farms ( Table 7). The impacts of hybrid maize adoption on maize productivity and gross margins were statistically significant, irrespective of the level of market access. Hybrid maize adoption did not enhance the PCFE of the smallest farms across both high and low market-access categories. This could, again, be due to the low level of household net economic benefit (i.e., gross margin per household) obtained by the smallest farms. Nevertheless, the PCFE of the largest farms was statistically significant across both the high and low market-access categories. The treatment effect results (ATU) were similar for non-adopters, which suggests that hybrid maize adopters across the small farm-size categories, with or without market access, did not benefit with respect to PCFE. In the context of Nepal, results suggest that farm size is a much more important determinant of livelihood effects than access to an input market.

Conclusion and policy implications
Farmers of rural Nepal face a constrained supply of several marketed inputs and natural resources, making the sustainable intensification of cereal systems a challenging endeavor. By examining hybrid maize adoption as a case of sustainable intensification by the smallholder farmers of Nepal's rainfed mid-hills, we tested the hypothesis that adoption of this technology enhances productivity, profitability, and household welfare, depending on factors such as the scale of operation. The current Nepalese context of rising on-farm wages (due mainly to labor out-migration), fragmented landholdings, high dependency on maize hybrid seed imports, and limited availability of inorganic fertilizers provides ample opportunities to test the aforementioned hypothesis. Our findings revealed that the constrained availability of material inputs such as inorganic fertilizers, on-farm wage rates (labor scarcity), and market proximity were strongly associated with farmers' adoption of maize hybrids. Moreover, the adoption of maize hybrids increased on-farm productivity and profitability overall. However, we also found that hybrid maize adoption had heterogeneous effects, with relatively larger farms (the upper 50%, above 0.3 ha) benefiting from adoption in terms of gains in productivity, profitability, and welfare outcomes. Smaller farms (the lower 50%, below 0.3 ha) did not benefit with respect to welfare outcomes, although they were more productive than the larger farms. Even with increased market access, the scenario remains unaltered. We derived three major policy implications for the smallholder farming systems in Nepal, which are not strictly in line with recommendations from previous studies on cereal system intensification. For example, Spielman et al. (2010) focused on opening up the input market as a precondition for the intensification of cereal systems in Ethiopia. However, our findings suggest a public-private partnership in R&D programs on sustainable intensification. A single and piece-wise intervention, a common method of disseminating proprietary technologies, is insufficient to improve rural livelihoods. R&D programs and government policies should target the diffusion of multiple sustainable intensification technologies, ensuring input access and addressing fundamental resource constraints. While the private sector plays a key role in making the technologies available, facilitating the adoption of multiple interventions requires public R&D and extension support. Only a multi-layering of technologies would offset the adverse effects of resource shortages. The incremental effect of technology combinations was observed earlier by Marenya et al. (2020), who showed that the fertilizer-dominant intensification strategy provided only a 30% incremental yield, whereas the combination of fertilizer, maize-legume diversification, and soil and water conservation provided an 88% incremental yield. It is expected that technology combinations are also likely to have highly positive returns in terms of profitability and rural livelihoods in countries like Nepal. This is our first recommendation derived from the empirical analysis.
Secondly, policies and R&D strategies aimed at targeting sustainable intensification technologies across farming communities should be developed. Takeshima et al. (2017) showed that an in-depth understanding of the returns from inputs is critical in formulating effective policies for the dissemination of agricultural inputs in developing countries. However, we found that the benefits to the adoption of hybrid maize varied widely across socioeconomic strata. Sample farmers with the smallest landholdings, despite being more productive, were unable to sustain their livelihoods (i.e., PCFE) from maize cultivation alone. More research is needed on the differential access to production resources across social groups and its implications for economic inequality. For example, the nature of our dataset does not allow us to examine the role of social marginalization with respect to gender. Burke and Jayne (2021) showed an association between social marginalization and low input quality: women farmers of Africa were found to be more likely to farm with lower quality seed and less fertilizer on marginal land. In Nepal, more studies need to be carried out on the roles played by male out-migration and the feminization of Nepalese agriculture on productivity and farm income.
Finally, smallholder farmers in the mid-hills of Nepal should be encouraged toward income diversification, and options should be provided to divert them toward alternative farming systems such as the vegetable or dairy sectors for the betterment of rural livelihoods. Despite such system diversification, a small landholding size might sometimes not be sufficient for households to raise themselves out of poverty. There have not been many studies made on the role of non-farm income in rural livelihoods in Nepal. However, studies conducted in other parts of the Global South are also valid here. Holden, Shiferaw, and Pender (2004) observed that unconstrained access to low-wage, non-farm employment could improve household income more substantially than the provision of unconstrained access to credit for the purchase of farm inputs in the Ethiopian highlands; this observation is particularly relevant for the resource-constrained farming conditions of the mid-hills of Nepal. The structural transformation of the country requires more research and policy attention.

Notes
1. An incident of severe crop loss due to spurious maize seed imported from India in 2009 induced a passionate debate on promoting locally improved varieties of crops in Nepal, shaping the seed policies of the government. Farmers from the Terai districts had been buying hybrid maize seeds from the bordering districts of India. In the winter of 2009, however, some of the varieties bought from across the border led to a heavy yield loss for thousands of Nepali farmers, as these varieties could not withstand the cold weather of the region (Adhikari 2014). The resulting uproar prompted the government to regulate non-domestic seed companies, which are now required to conduct multi-location trials for at least two seasons or years to ensure the performance and yield stability of these varieties before they can be registered in the seed quality control section (SQCC) of the Government of Nepal, and the maize hybrid seed sold in the Nepalese market. 2. In the Government of Nepal's National Seed Vision (2013-25), the development of 12 hybrid maize varieties by the end of 2025 is envisaged. To meet this goal, the government actively promotes the development of the private seed sector in the country. 3. Only six sample farmers (0.82%) grew both maize hybrids and OPVs on their farms in 2017 (i.e., partial adoption of the technology). The maize area allocated for hybrid maize was equal or greater than that allocated for non-hybrid maize varieties in all these farms. We considered them as the hybrid maize adopters in this study. 4. The PCFE did not include expenditure on infrequent events such as marriage ceremonies. 5. The marginal effect of a dummy variable with the dependent variables in the log-form is calculated as 100*[exp(Coefficent)-1], following Giles (2011). 6. It should be noted that the IPWRA method captures only observed heterogeneity between hybrid maize adopters and non-adopters.
Development (USAID) and the Bill and Melinda Gates Foundation (BMGF). We gratefully thank the farmers for their time in participating in the survey, Anish Dahal, Bivek Aacharya, Kiran Dahal, and Sameer Shrestha for conducting the interviews with farmers, and Elizabeth Waygood for language editing. This paper reflects the authors' views and not those of USAID, BMGF, CSISA, or authors' organizations.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Notes on contributors
Gokul P. Paudel is ATSAF PhD scholar at the Institute for Environmental Economics and World Trade, Leibniz University Hannover, Germany and International Maize and Wheat Improvement Center (CIMMYT). He holds masters in agricultural economics from Tribhuvan University and a post graduate diploma in advance agricultural studies for agronomist from Israel. His areas of interest and expertise are sustainable intensification of agricultural systems, technology adoption and impact assessment, climate change impacts and adaptation, productivity and efficiency analysis, non-market valuation, and programming, data mining and advance machine learning. Gokul is a programmer and is heavily involved in application of data driven agriculture in smallholders' agricultural systems in South Asia.