Global efforts for documenting the contribution of research to agricultural development, food security, poverty alleviation and other outcomes have mainly focused on major cereal food crops such as wheat, rice, and maize [1–7]. As a result, little attention has been given to the quiet revolution that has taken place due to wide diffusion of improved varieties of food legume crops. Over 99% of lentil area in western Bangladesh is cultivated with short-duration improved lentil varieties (ILVs), which are grown between irrigated rice crops and on unirrigated lands during the dry season [8]. Little is known about the impacts of these varieties. This article documents plot-level impacts of these lentil varieties in Bangladesh where “plot” is defined as a small agricultural land covered by a single type of vegetation. Sometimes, a given agricultural field with a single owner can be divided into two or more plots cultivated with two or more types of vegetation.

Lentil is the most important pulse crop in Bangladesh where it serves as human food, animal feed and fodder, and as a raw material for agro-processing industries [9]. It is rich in protein (25.7–33.4%) and several essential micronutrients [10–12]. Moreover, because it fixes nitrogen, legume production can improve soil fertility [13].

Area under lentils in Bangladesh grew from 62,000 ha in 1961 to 295,000 in 1979. Prior to the 1980s, lentil was grown during the dry season as few alternatives existed to follow the rainy season (Aman) rice harvest. In the 1980s, two forces converged, and lentil area declined. First, aggressive expansion of irrigation allowed irrigated dry season (Boro) rice to displace lentil production. Second, traditional lentil varieties became vulnerable to two important diseases—Stemphylium blight (Stemphylium botryosum) and rust (Uromyces fabae) which began to spread in the early 1980s.

In response to the disease problem, the Pulses Research Center (PRC) of Bangladesh Agricultural Research Institute (BARI) and Bangladesh Institute of Nuclear Agriculture (BINA) individually and in collaboration with the International Center for Agricultural Research in the Dry Areas (ICARDA) began breeding disease-resistant lentil varieties. In the mid-1990s, two improved varieties (BARI Masur-3 and BARI Masur-4) with resistance to these diseases were released by BARI. Despite their reduced disease susceptibility, lentils remained at an economic disadvantage to the dry season Boro rice, particularly on irrigated lands. Area planted to lentils continued to decline to 90,000 ha in 2013. Meanwhile, BARI and BINA along with ICARDA continued development and release of ILVs which, in addition to disease resistance, are high yielding and more remunerative. As of 2013, eight modern varieties of lentil had been released. The new varieties were also early maturing making it possible to grow them during the short window between rice crops. As a result, area under lentil started to increase reaching 156,000 in 2018—still much lower than 1979 levels but 71% higher than in 2013 [14].

While the biophysical benefits of legumes are well documented, mostly using on-station experimental data, the literature on the socio-economic impacts of improved varieties of legume crops in general and lentils in particular is scant. In a study from two districts in Western Bangladesh [15], fit a Cobb-Douglas production function, estimated net margins of 27,838 (US$339.62) per ha and concludes that adoption of ILVs is profitable. Another study [16] also show that adoption of improved faba-bean varieties in Morocco leads to higher adoption of legume-based rotations and to higher yields and gross margins. A study in Ethiopia [17] documents significantly higher yields for two ILVs. It has also been document that adoption of improved bean varieties has a positive impact on dietary diversity and reduces food insecurity, which they suspect is likely coming from the income effect resulting from the high yielding properties of these varieties [18].

This study carries out an analysis of plot-level, per unit area impacts of adoption of ILVs on yield and gross margins (defined in this paper as revenue minus cost of all inputs except land). By providing options for rotation, relay- and inter-cropping, legumes are vital components in diversification of Bangladesh’s predominantly rice-based cropping system and for enhancing yields of subsequent (for rotation and relay) and accompanying (for intercropped) cereals.

Plot-level evaluation of economic benefits of adoption of rotation, inter cropping, and relay cropping, especially associated with the introduction of ILVs, is best achieved by considering all benefits and costs from all crops during the entire cropping cycle [16]. While such an analysis is a subject of future research, due to data limitations, the analysis in this study is limited to only plot-level impacts on lentil yields and gross margins. Findings will be useful for researchers, policy makers, development practitioners, and extension personnel in Bangladesh and other countries in South Asia.

The agroclimatic conditions of Bangladesh are generally suitable for growing leguminous crops. Food legumes are traditionally cultivated under rainfed conditions, usually with minimum or no inputs other than seeds, land, and labor. Legumes are relatively short duration crops. Lentil is cultivated during winter (November-March), requires minimal tillage and fewer agricultural operations including weeding and no or minimal irrigation [19].

Due to their ability to fix atmospheric nitrogen (35–100 kg/ha/annum), lentils can play an important role in enhancing the sustainability of rice-based cropping systems [14]. In Bangladesh, lentils are grown in several ways—as a sole, mixed, or inter crop. While the most common cultivation pattern is sole cropping of short-duration varieties, sizeable area is also under intercropping with crops such as wheat, mustard, linseed, and sugarcane. Intercropping and mixed cropping are age-old practices, particularly in the North and North-western parts of Bangladesh [20]. The practice has developed as an informal insurance against a crop failure. Relay cropping in transplanted rice plots is common in low-elevation plots receiving water runoff from areas with higher elevation.

Between 1991 and 2015, BARI and BINA released 15 ILVs, developed individually or in collaboration with ICARDA (Table 1). Of these, in 2015, only eight were cultivated by farmers and these covered over 99% of national lentil area [8]. Due to their shorter duration, the introduction of the ILVs has made rice-based relay cropping more attractive. In this system, one crop is interplanted with a second crop as it approaches maturity. The practice is now common in rain-fed agro-ecosystems in Bangladesh. This system is suitable for post-monsoon cropping where a second crop in succession to rice is grown taking advantage of residual soil moisture after the rice harvest. The second crop is cultivated without tilling and with no application of fertilizers; seeds are broadcast about 2–3 weeks before the rice harvest [21]. Higher yields of rice were also found when grown after lentils [22]. Production of lentils also has the benefit of enhancing agrobiodiversity by breaking the pattern of rice mono-cropping and spreading labor and land use into previously idle periods.

Some economic benefits of new varieties are from higher yields and/or lower costs, both of which contribute to lower per-unit cost of production. Statistical identification of the effect of an adopted crop variety using observational data is known to be a challenge because adoption is a choice, and the evaluation must consider multiple sources of confounding. To overcome this problem, we employ a multivariate analysis where the outcome Y (yield or gross margins per ha) is regressed on a dummy variable T (taking a value of 1 if the farmer plants ILVs and 0 otherwise) and other household, farm, and farmer characteristics X including inputs as follows: Y=θ+αT+γX+ε (1)

In this regression, omitted variables (such as differences in land quality, farmer motivation, skills, IQ among others) can affect the adoption decision (T) and the outcome variable (Y)—causing correlation between the error term and T–violating one of the assumptions of Ordinary Least Squares (OLS) regression.

Here, an instrumental variables regression (IV) approach is used. IV methods are commonly used in studies when observational data are used [23–26]. With experimental data, the confounders can be controlled for through the experimental design but use of observational data requires special care. An IV approach requires instrument(s) which are correlated with the decision to adopt but uncorrelated with the unobserved factors influencing the outcome [27–29]. Suppose there is endogeneity between the treatment variable T and the outcome variable Y in Eq 1. Suppose also that Z is a matrix of exogenous covariates that include valid instruments for X. Then the IV model can be described by Eqs 1 and 2, with the latter known as the selection equation.

Where Π is a vector of coefficients, ε and μ are the error terms and there can be overlap in X and Z. The fact that Cov(ε,μ) = σ_με≠0 means that OLS estimation will lead to biased estimates of the coefficient; σ_με is a measure of the level of endogeneity between the treatment and outcome. Two-stage least squares (2SLS) is used to estimate Eqs 1 and 2 and recover α, the parameter of interest. In the first stage, due to the binary nature of the dependent variable in the selection equation and to account for possible causes of endogeneity between the treatment and outcome variables, the probit model is estimated first to generate the predicted probabilities as is the case with standard Stata commands such as ivreg. As the ivreg Stata command by default treats all independent variables in the selection equation as instruments, we have estimated the model in two steps using the predicted probabilities of adoption from the first step. While the parameter estimates in the second step are still unbiased, the standard errors are not valid and hence bootstrapping gives asymptotically valid approximations of the distribution of errors [30,31]. Therefore, we report coefficient estimates and bootstrapped standard errors.

A critical determinant of the quality of the analysis is the presence of a credible instrument. We use an indicator variable reflecting whether a farmer received government/project support for growing lentils at least one year before the survey was carried out in 2015 (support to grow lentil) as an instrument. Even if a farmer cultivates lentils in multiple years, support for growing lentils was provided only once and it was given in the form of one or more of low interest credit, free lentil seed (1–2 kg certified seed of lentils) and fertilizers (adequate to grow the seed). Many farmers in Bangladesh either don’t apply or apply low amounts of fertilizers to legume crops. Therefore, in the government support project, fertilizer is given to demonstrate to farmers that adding phosphorus fertilizers in legume production is beneficial. The support was made available for all farmers and hence there was no criteria to discriminate between farmers. Therefore, the support was randomly given to anyone who asked for it.

The choice of the instrument is justified on grounds that, though not exclusive, certified seeds predominantly include latest improved varieties. We argue that this support, because it was provided in years past, does not have direct effect on current outcomes—yield and gross margins per ha except through its effect on adoption. Whether endogeneity is a problem, and if so, the validity of the instrument are tested empirically in the results section. Selection bias can also be an issue because farmers may self-select in or out of the treatment (adoption of the improved lentil varieties) for which a decision has to be made whether to correct for it. However, the IV approach helps to control biases arising from self-selection [32,33], omitted variables [32] and due to inclusion of variables which we should not have controlled for [34]. IV method’s key advantage compared with other statistical methods used to analyze observational data is its ability to isolate exogenous effects of an intervention by excluding endogenous self-selection effects without having to directly measure such self-selection [35].

In order to be able to fit a variant of the Cobb-Douglas production function (including farmer characteristics as inputs) for yield, all continuous variables (such as yield, gross margins, lentil area, all quantities of inputs and age, education and experience of the household head) are logged. To overcome the problem of zero values, a constant of 1 was added to all values before the transformation. Several factors such as the amount of inputs applied are important in determining yield and, in turn, gross margins per ha. Therefore, all these were included as explanatory variables.

Before proceeding with estimation, diagnostic specification tests were conducted. First, we applied the Durbin [36] and Wu–Hausman [37,38] test statistics both of which rejected (at p<0.01) the null hypothesis that the treatment is exogenous–showing that endogeneity is a problem. Hence, the use of a model that corrects for endogeneity is necessary. Then, the correlation between the instrument and the endogenous variable (adoption) was examined. The two variables have a correlation of 0.16, significant at 0.01 level. The yield and gross margins equations were also estimated using OLS to generate residuals. The correlations between the residuals from the yield and gross margins equations and the support dummy variable (the instrument) were 0.04 and 0.05, respectively, both statistically insignificant showing that the instrument is valid. We also followed [39] and carried out a falsification test, which showed that the instrument had a positive and significant (p<0.01) effect on the adoption decision but no significant effect (p>0.1) on both per ha yield and gross margins of the non-adopters. Unless there is another variable in the equation which is highly correlated with the instrument which might cause the instrument to be insignificant (which is not the case in our regression), an insignificant coefficient on the instrument is an indication of the absence of endogeneity.

With a single instrument, there is no way to test the exclusion restriction. As a result, we are forced to assume that the model is identified. Even with two or more instruments, one can put some level of assurance by testing for overidentification restrictions, but these tests are all based on an assumption that the model itself is identified, which cannot be tested, because it involves the structural error term. Therefore, while we have eliminated the obvious cases of failures, we would like to caution the reader that our instrument can still fail.

As proposed by [40], the Stock and Yogo minimum eigenvalue statistic which is reported after the estimation of the Two Stage Least Squares Estimation (2SLS) was used to test if the instrument is weak. With an F-statistic value of 21.4 which is far greater than the critical values of the minimum eigenvalues, we reject the null hypothesis that the instrument is weak. After careful investigation of the diagnostic results, we conclude that our model is correctly specified, and the instrument is theoretically plausible, empirically valid, and strong enough to adequately identify the effects of adoption of ILVs on the outcome variables.

This study focuses on ten major lentil-growing districts in western Bangladesh, constituting about 74% of total national lentil area. A complete sampling frame listing all lentil growing farmers in all ten districts in western Bangladesh was not available, making use of simple random sampling infeasible. Elicitation of adoption estimates from a panel of experts showed that the ILVs were cultivated by over 80% of farmers in Western Bangladesh. Therefore, power analysis was then used to determine the minimum sample size (MSS) required to ensure at least 95% confidence and 2% precision levels for our estimates to capture between 20% and 80% adoption levels. The MSS was determined to be 864 but to compensate for possible non-response and missing data and to ensure adequate statistical power, we increased the sample size to 1,000.

The sample was drawn using a multi-stage stratified sampling technique. First, all ten districts were purposively included. Second, a random sample was drawn to select 20 sub-districts (locally called Upazillas) and 52 villages. The sample was distributed among the 20 Upazillas and 52 villages in proportion to the total number of lentil growers in each district, and sub district while the sample size per village was fixed between 17 and 19 (Table 2).

An average of 20 households per village were randomly selected from a master file containing the list of all lentil growing households in the village. The questionnaire, enumerated to the person most knowledgeable about lentil cultivation, covered household demographic and economic conditions, asset ownership and other relevant factors. Information on lentil farming was obtained by asking detailed questions on varieties planted (farmer recall), input use, management practices, yields and factors of production for all lentil plots. Community-level information on access to infrastructure, farm services, extension, etc. was obtained from a separate village-level survey.

The variety was also verified using DNA fingerprinting, to ensure that farmer knowledge of the specific variety was not a source of error. The DNA fingerprinting process began with collection of reference seed samples where breeder seed samples were obtained for all released varieties from BARI and BINA. During the household survey, samples of seeds were collected from the 1694 plots planted by the 1000 sampled farmers. Prior to the analysis, a set of five Inter Simple Sequence Repeats (ISSR) and 41 Simple Sequence Repeats (SSR) markers [41,42] were identified from the lentil genome covering different linkage groups and individually tested for polymorphism in the breeder samples. Two ISSR and 20 SSR markers showed significant polymorphism across the Breeder seed samples and were used for varietal identification. For more details about the procedures used for DNA fingerprinting, we refer the reader to [8]. There was an 89% concordance between farmer recall and varietal identification using DNA-fingerprinting–showing reasonably accurate recall [8]. All analysis carried here is based on the DNA-verified varietal identification.

The survey was enumerated in 2015. During a stakeholders meeting, breeders from BARI and BINA, extension personnel from the district offices, and staff of the Bangladesh Agricultural Development Corporation (BADC), the main public seed producing and marketing entity, agreed that only varieties released after 2006 are considered improved. To exclude farmers who are testing the varieties for the first time and might eventually not adopt them, we agreed that adopters must be those who cultivated ILVs for at least 2 years. Therefore, for the purpose of this study, the definition of adoption is use of ILVs that are released after 2006 for at least two years.

About 48.8% of sampled household have adopted ILVs released in or after 2006. With an average of about 45 and 7 years, the age and education of the household heads in adopter and non-adopter groups are not statistically different while the farming experience of the household heads in the adopter households is statistically greater (at 0.01 level). Farmers who received support to grow lentil and those who purchased certified seeds are found to adopt the improved lentil varieties more than those who didn’t receive any support and those who didn’t buy certified seeds. Table 3 provides summary statistics for selected variables including those used in the IV model.

Results of the first stage of the 2SLS estimation of the IV model are presented in Table 4. Parameter estimates of the selection equation (or first stage estimation) show that adoption of ILVs was positively and significantly (all at 0.01 level) affected by the education level and farming experience of a household head, the use of certified seed, whether a plot is located in Kushtia, and Pabna regions, and whether a farmer obtained support to grow lentil. These results are all consistent with theoretical expectations because educated and more experienced farmers are likely to better understand the opportunities and challenges involved in adopting a new variety. Moreover, 99% of certified seeds sold in western Bangladesh in 2015 were for ILVs and hence anyone buying certified seed was highly likely to also be using ILVs. Given that Faridpur (the province which was dropped from the equation) is predominantly irrigated, these lentil varieties which were mainly bred for rainfed areas may not fit well in the region for which farmers located in the relatively drier provinces of Kushtia and Pabna are highly likely to adopt the ILVs. As argued in the selection of the instrument, support for growing lentil is expected to motivate farmers to adopt improved varieties.

Age of household head and whether lentil is planted between two rice crops negatively and significantly affect adoption of ILVs. These results are also theoretically sound as older farmers are often overly attached to their age-old varieties. Plots on which a farmer plans to plant rice before and after the middle crop is less likely to be planted to lentils because jute, mung-bean, and sesame, which are summer crops, can be planted and harvested successfully in the spring/summer season. For a successful rice-lentil-rice system, lentil, which is a winter crop, has to be planted in November and harvested in February making it cumbersome for farmers. In the absence of super-early lentil varieties, farmers prefer to plant jute, mung-bean or sesame after lentil which can be harvested much earlier (by November) and get some rest for themselves in Dec-Jan. That is why area under lentils between two rice crops is much less than that which is under rice-lentils–jute/mung-bean/sesame. Improved lentil varieties called BARI_Masur 7 and 8 released by BARI and ICARDA after 2011 are said to be super-early and are expected to make farmers’ dream of the rice-lentil-rice cropping system a reality.

This paper conducts plot-level impacts of varietal replacement (i.e., replacement of land races and old improved varieties) by most recent improved varieties. For the purpose of this study, adoption is defined as the use, for at-least two years, of lentil varieties that have been released since 2006. This means, the counterfactual group in this study comprises of mostly old, improved varieties which originated from BARI, BINA and ICARDA with only 16 cases of other varieties suspected to be landraces. Therefore, the impacts measured in this study represent the impacts of lentil varietal replacement in Bangladesh. These types of impacts are described by [43] as impacts of Type II technical change in post-green revolution agriculture where farmers regularly replace modern varieties to maintain disease resistance and to take advantage of the ever-higher yields of newer varieties.