Data

The data used for the analysis are a nationally representative sample of rice growers collected during 2014 by the International Center for Tropical Agriculture (CIAT) in collaboration with the Centro de Investigación Agrícola Tropical of Santa Cruz, Bolivia, across the three rice-producing regions in the country. The sampling design followed a multistage procedure distributed among 98 communities, consisting of 802 small- and medium-scale farmer households that provided the complete information required under the proposed model.

Dependent variables

As mentioned earlier, rice is not a native crop of Latin America, and all varieties grown by rice growers in Bolivia are genetically improved. However, the modern varieties released since 2004 currently occupy 45.6 percent of the total rice area. The other 54.4 percent of Bolivia's rice area uses old varieties mainly introduced from other countries before the foundation of the national rice breeding program. The rate of adoption of the different MIVs is shown in Table 2. Two MIVs dominate the preference of Bolivian rice growers: MAC 18, which resulted from the collaboration of the Bolivian rice program and FLAR, and IAC 101, which was bred by the rice breeding program of the State of São Paulo (Brazil), selected, and then released in Bolivia. These two varieties have adapted well to both irrigated and favored rainfed conditions in Bolivia. MAC 18 also offers a slightly higher quality of its grain, which is appreciated in the market. Despite being available to farmers for some time and expected to significantly improve productivity either in input-intensive (and mechanized) or manual production systems (Châtel et al. Reference Châtel, Guimarâes, Ospina, Rodríguez, Lozano, Degiovani, Martínez and Motta2010), the rest of the modern varieties have not achieved a considerable rate of adoption. This is consistent with literature findings that productivity gains may not necessarily cover the additional costs from adopting new crop varieties, hence limiting their appeal to most farmers (Pretty, Ruben, and Thrupp Reference Pretty, Ruben, Thrupp and Uphoff2002). Besides, materials that have long been used in local conditions are likely to be better adapted to their specific agroclimatic needs than their more recent competitors (Hoffman, Probst, and Christinck Reference Hoffman, Probst and Christinck2007).

Table 2. Adoption^a of modern improved rice varieties in Bolivia

^a Farmers may adopt more than one rice variety; hence, the reported values of adoption do not add up to 100 percent.

Source: Elaborated by the authors based on survey data.

Although rainfed systems dominate rice production in Bolivia, large machinery for planting and harvesting is available to rice growers. The possibility of mechanization among rainfed rice growers refers to the concept of favored rainfed systems (Lynch and Tasch Reference Lynch and Tasch1981). Fertilization in rice production mainly relies on the use of mineral supplements with a range of combinations of nitrogen (N), phosphorus (P), and potassium (K). The most available fertilizer formulations in local markets offer N, P, and K concentrations ranging from 21 to 46 percent, 20 to 45 percent, and 50 to 60 percent, respectively (Viruez and Taboada Reference Viruez and Taboada2013). Chemical pest control is virtually a binary decision between products containing either cypermethrin or cyfluthrin to combat both the fall armyworm and Tibraca limbativentris. In contrast, most weed controls (herbicides) are made of bispyribac-sodium, imazapyr, imazapic, and penoxsulam.

We expect that rice growers in Bolivia have up to five technologies available that may increase their rice productivity compared with traditional rice practices, but this may also imply some degree of complementarity. These technologies, the adoption of MIVs, machinery used at seeding or harvesting, and the use of fertilizers, chemical pesticides, and herbicides on rice plots, constitute the dependent variables of our model. Our definition of “modern” improved variety considers varieties released in Bolivia since 2004 (eight years after establishing the rice improvement program). These are up to 12 modern materials, including MAC 18, IAC 101, Azucena, SACIA FL-39, and SACIA FL-40. Other varieties, including Bluebonnet, are considered old types. Although these technologies could be expected as (nearly) perfect complements from an agronomic perspective, the observed extent of adoption (Figure 1) suggests that farms make only a partial adoption of such practices. Such a pattern is understandable for resource-limited households, whose decisions on production and consumption are highly interdependent. The latter follows the concept of nonseparability of decisions of agricultural households (Bardhan and Udry Reference Bardhan and Udry1999). It is reasonable for these farms to adopt a part of an improvement package to gain at least a part of the potential benefits from adoption. Households may possess attributes that should be explored to understand better who are more likely to adopt each technology of interest. Also, it is worth noting that although these technologies may be labeled as input-intensive, they are not necessarily incompatible with future efforts in expanding the uptake of sustainable practices (Wainaina, Tongruksawattanab, and Qaima Reference Wainaina, Tongruksawattanab and Qaima2016).

Figure 1. Number of Technologies Adopted by Interviewed Bolivian Rice Farmers. Source: Elaborated by the authors based on survey data.

Econometric approach

In this study, we consider a farmer household deciding whether or not to adopt a given technology. With no loss of generality, suppose that if household i chooses to adopt the m th technology, it reaches a utility of $V_{im}^1$, while it would reach a utility of $V_{im}^0$ otherwise. Therefore, they adopt the technology if and only if

(1)$$A_{im}^\ast{ = } V_{im}^1 -V_{im}^0 = {x}^{\prime}_i{\rm \beta }_m + {\rm \varepsilon }_{im} > 0, \;$$

where the ɛ_im terms are unobservable shocks to the farmer utility for each m technology, and x _i are observed attributes of the household. Households are to choose whether to adopt each of the m = 1, 2, …, M nonexclusive technologies. While we do not observe $A_{im}^\ast$ or $V_{im}^h$, we do observe the adoption decision of each household as an index function, where $A_{im} = 1( A_{im}^\ast > 0)$. Relaxing the assumption that the unobservable shocks are independent, and assuming that they are jointly and normally distributed, leads to the multivariate Probit (MVP) regression model (Cameron and Trivedi Reference Cameron and Trivedi2005; Greene Reference Greene2012), where

(2)$$A_{im} = \Pr ( {A_{im}^\ast > 0} ) = \Pr ( {-{\rm \varepsilon }_{im} < {{x}^{\prime}}_i{\rm \beta }_m} ) = {\vector \Phi }_M^m ( {{{x}^{\prime}}_i{\rm \beta }_m} ).$$

Let us model the joint probability of adoption for M technologies as

(3)$$\Pr ( {A_1 = 1, \;\;A_2 = 1, \;\ldots , \;A_M = 1} ) = \int_{-\infty }^{{x}^{\prime}{\rm \beta }_1} {\int_{-\infty }^{{x}^{\prime}{\rm \beta }_2} { \ldots \int_{-\infty }^{{x}^{\prime}{\rm \beta }_M} {{\rm \phi }_M( {{\boldsymbol t}, \;\;{\rm \Sigma }} ) \;{\rm d}{\boldsymbol t}, \;} } } $$

where ϕ_M is the standardized multivariate normal probability density function, and Σ is the symmetric covariance matrix of the error vector ɛ, in which Σ_mn = 1 if m = n, and Σ_mn = ρ_mn = ρ_nm when m ≠ n. Estimating the M-dimensional normal integrals is feasible by the Geweke, Hajivassiliou, and Keane (GHK) simulated-likelihood approach (Boersch-Supan and Hajivassiliou Reference Boersch-Supan and Hajivassiliou1990; Cameron and Trivedi Reference Cameron and Trivedi2005; Greene Reference Greene2012). The decision to model adoption at the household levelFootnote ¹ follows that 95 percent of the interviewed households managed a single plot. Barely over 1 percent managed three or more plots. Thus, a household-level analysis does not conflict with similar studies as those of Donkoh, Azumah, and Awuni (Reference Donkoh, Azumah and Awuni2019), Gebremariam and Tesfaye (Reference Gebremariam and Tesfaye2018), and Zeng et al. (Reference Zeng, Tian, He and Zhang2020).

In addition to this, we model the extent of adoption as the number of technologies effectively implemented by the household (Wollni, Lee and Thies Reference Wollni, Lee and Thies2010), setting up an ordered Probit regression model that also resembles a random utility model. Here, farm household i decides to adopt a total of c _i technologies (c _i = 0, 1, 2, 3, 4, 5) based on the utility

(4)$$U_i = h( {{{x}^{\prime}}_i{\rm \theta }} ) + \eta_i.$$

Although utility cannot be observed, as in the MVP setting, we observe the number of adopted technologies. We assume that a household adopts an additional technology if the utility of using it exceeds the utility of not adopting it. Moreover, the decision of extent of adopted technologies responds to utility levels defined across thresholds λ such that

(5)$$c_i = 0, \;\quad {\rm if}\;U_i \le {\rm \lambda }_1, \;$$

$$c_i = k, \;\quad {\rm if}\;{\rm \lambda }_k < U_i \le {\rm \lambda }_{k + 1}\quad {\rm for}\;k = 1, \;\;2, \;\;3, \;\;4, \;$$

$$c_i = 5, \;\quad {\rm if}\;U_i > {\rm \lambda }_5.$$

Finally, if η follows a normal distribution, then

(6)$$\Pr ( {c_i = 0{\rm \vert }x_i} ) = {\vector \Phi }( {{\rm \lambda }_1-{{x}^{\prime}}_i{\rm \theta }} ) , \;$$

$$\Pr ( {c_i = k{\rm \vert }x_i} ) = {\vector \Phi }( {{\rm \lambda }_{k + 1}-{{x}^{\prime}}_i\theta } ) -{\vector \Phi }( {{\rm \lambda }_k-{{x}^{\prime}}_i{\rm \theta }} ) , \;\quad {\rm if}\;k = 1, \;\;2, \;\;3, \;\;4, \;\;$$

$$\Pr ( {c_i = 5{\rm \vert }x_i} ) = 1-{\rm \Phi }( {{\rm \lambda }_5-{{x}^{\prime}}_i{\rm \theta }} ).$$

The sign of θ coefficients in the ordered Probit model indicates the aggregate direction that covariates have in the number of technologies adopted by the household, i.e., whether the explanatory variable increases (decreases) the value of our latent variable U (Cameron and Trivedi Reference Cameron and Trivedi2005). Nonetheless, marginal effects do not necessarily hold the same sign, as covariates may imply more (or less) likelihood of a point outcome.

Explanatory variables and descriptive statistics

Covariates used in this article (Table 3) are consistent with previous adoption studies on rice technologies and follow the underlying theoretical economic models that propose them (Saka et al. Reference Saka, Okoruwa, Lawal and Ajijola2005; Usman, Ango, and Barau Reference Usman, Ango and Barau2013; Yamano et al. Reference Yamano, Arouna, Labarta, Huelgas and Mohanty2016). We have included producers' associations and access to extension services to represent households' access to agricultural information. The model also includes variables related to human capital and income characteristics (of the household head) to control for differences in the dynamics followed toward the adoption decision. Geographic control of distance (in log-Geodesic scale) to San Juan de Yapacaní, one of the leading centers of rice technological diffusion, attempts to account for the level of exposure to the dissemination of rice technologies. Also, we use the farm's size as a proxy for its wealth endowment across all the decisions modeled, while yield priority represents households' varietal preference for higher-yielding varieties. Finally, we included department-fixed effects to account for the differences in labor availability and access to markets, since evidence suggests that changes in these attributes may increase the opportunity cost of technologies, making them less attractive or even nonaffordable (White, Labarta, and Leguía Reference White, Labarta and Leguía2005). Santa Cruz has the most extensive availability in both labor and markets; thus, we choose it as the base category in both regression analyses.

Table 3. Description of variables

Source: Elaborated by the authors.

Descriptive statistics of all the explanatory variables used in the analysis are summarized in Table 4, which compares average values between adopters and nonadopters of the five complementary technologies used in rice production. The distribution of adoption is somehow dissimilar across the different technological options. While roughly 65 percent of the rice growers use pesticide and weed controls, we found adoption levels of MIVs and machinery slightly under 50 percent of the interviewed households. On the other hand, fertilizer use is the lowest adopted rice technology (25 percent). Among adopters, the household heads tend to be slightly more educated and younger than nonadopters and, although female-headed households are unusual, they are marginally fewer among adopters. Households adopting the rice technologies have, on average, more alternative employment options and larger farm sizes. Adopters of technologies seem to have more access to agricultural extension and farmers' organizations. Statistical differences are significant between adopters and nonadopters, particularly for the household head level of education, farm size, belonging to a farmers' association, access to extension services, and the distance to the main technology diffusion center (San Juan de Yapacaní). Notice how the difference of average distance to diffusion centers can be a scale factor of two when comparing adopters and nonadopters. In any case, relying on alternative income sources is an evenly distributed condition between adopters and nonadopters. On the other hand, having a female household head reports only a statistical difference when comparing adopters and nonadopters of herbicides.

Table 4. Descriptive statistics

¹ Yield priority is disaggregated only for MIVs, following the rationale in Table 2.

^aNonadopters, ^badopters, and ^creporting means and percentages for continuous and binary variables, respectively.

Reporting significance of variable's mean difference between adopters and nonadopters of each technology: ***p < 0.01, **p < 0.05, *p < 0.1.

Source: Elaborated by the authors based on survey data.