Strawberry production in Brazil

Strawberry production in Brazil is important in economic, social and environmental terms. Strawberry is broadly known and accepted by consumers (Antunes and Reisser Junior, Reference Antunes and Reisser Junior2008), and is also a profitable crop in Brazil (Antunes and Peres, Reference Antunes and Peres2013; Fagherazzi et al., Reference Fagherazzi, Grimaldi, Kretzschmar, Molina, Gonçalves, Antunes, Baruzzi and Rufato2017). In 2016, strawberry was cultivated on 4300 ha producing 155,440 t with a gross production value of R$1.2 billion (to facilitate comparison, the average daily exchange rate was 3.22 Brazilian Real (R$) per US$1 from July 27, 2016 to August 25, 2016 according to the Central Bank of Brazil at http://www4.bcb.gov.br/pec/taxas/port/ptaxnpesq.asp?id=txcotacao) (Fagherazzi et al., Reference Fagherazzi, Grimaldi, Kretzschmar, Molina, Gonçalves, Antunes, Baruzzi and Rufato2017).

In social terms, strawberry production is labor intensive (about four persons per ha), which creates jobs, and an important source of income for small-scale family farmers, who currently produce most of the strawberry in Brazil and in our study area, Brazlândia, DF (Lopes et al., Reference Lopes, Silva, Nascimento, Ramos, Pereira and Carneiro2005; EMATER-DF, 2016).

In environmental terms, the conversion to strawberry OP may have important consequences for environmental sustainability and food safety in Brazil. Strawberry CP, which accounts for 99% of the land used for strawberry production (Darolt, Reference Darolt2008), requires frequent applications of pesticides and intensive production practices to control arthropod pests and disease (de Carvalho, Reference de Carvalho2005), which result in negative effects on the environment and on the health of farm workers and consumers. The National Health Surveillance Agency (Anvisa) through its Programme of Pesticide Residue Analysis in Food has consistently found strawberry as one of the foods most contaminated with pesticide residues above the Maximum Residue Limit (MRL) allowed by Brazilian law (Anvisa, 2011; Oshita and Jardim, Reference Oshita and Jardim2012).

Study area and production system

Brazlândia is 475 km² with 53,800 inhabitants, and has many rural communities, including Assentamento Betinho, Barreiro, Disterro, Bucanhão, Cascalheira, Chapadinha, Alexandre de Gusmão, Morada dos Pássaros, Rádiobras, Rodeador and Torre, where the strawberry farmers in our sample are located.

Brazlândia is the seventh largest producer of strawberries in Brazil and the largest one in the Central West region (EMATER-DF, 2017; Fagherazzi et al., Reference Fagherazzi, Grimaldi, Kretzschmar, Molina, Gonçalves, Antunes, Baruzzi and Rufato2017). The growers of Brazlândia produced 7400 t of strawberry valued at R$62M on 200 ha of land, which is about 5% of the land and production in Brazil in 2016. As it is only 45 km from the capital Brasília, a city with a population of 2.9 million people, its strawberry production fulfills from 70 to 100% of the strawberry consumed in Brasília depending on the month of the year (Lopes et al., Reference Lopes, Silva, Nascimento, Ramos, Pereira and Carneiro2005).

Strawberries in Brazlândia are mostly produced in soil in an open-field, annual hill system with plastic mulch under conventional or organic systems using drip or sprinkler irrigation or both (Lopes et al., Reference Lopes, Silva, Nascimento, Ramos, Pereira and Carneiro2005; Falcão et al., Reference Falcão, Lacerda, Carvalho Mendes, Leão and Carmo2013). There are growers who combine drip irrigation with at least 2 weeks of sprinkler irrigation to ‘wash’ the plants so to reduce mite eggs (Lopes et al., Reference Lopes, Silva, Nascimento, Ramos, Pereira and Carneiro2005; Henz, Reference Henz2010). Almost all production practices, including soil preparation, transplanting, mulching, runner removal, weed and pest control, harvesting, and packing are primarily done by hand without machinery. Therefore, conversion to organic practices seems to be feasible from a labor and technology perspective.

Strawberry growers in Brazlândia plant vegetatively propagated transplants imported either from other states/countries, produced by local nurseries, or produced by local growers (Lopes et al., Reference Lopes, Silva, Nascimento, Ramos, Pereira and Carneiro2005; Antunes and Peres, Reference Antunes and Peres2013). Plants are transplanted from March to April (dry season) on previously prepared beds that are 1–2 m wide, with 2–3 rows per bed at 30 cm plant spacing for a total of 60,000–72,000 plants per hectare. Beds are covered with plastic mulch (30 micron polyethylene canvas) 30–35 days after planting, which requires a four-person crew (Lopes et al., Reference Lopes, Silva, Nascimento, Ramos, Pereira and Carneiro2005). Strawberries are harvested from June through September with peak harvest in July and are almost entirely marketed as fresh fruit in local or regional domestic markets, in 300 g plastic containers packed in 4-container cardboard boxes (Lopes et al., Reference Lopes, Silva, Nascimento, Ramos, Pereira and Carneiro2005; Henz, Reference Henz2010).

Survey design and data

We designed a survey to investigate producer and farm characteristics, quantities and costs of inputs, and revenues associated with OP and CP of strawberry among strawberry farmers in Brazlândia, DF, Brazil. Prior to the final survey design, we met several times with extension personnel from Brazlândia and the central extension office in Brasília to identify and adjust questions on grower and farm characteristics, quantities and costs of inputs, and revenues from organic and conventional strawberry production in Brazlândia. These extension personnel have had many years of experience interacting with strawberry growers in the study area, and know many of them personally, including all of the organic strawberry growers in the area. We tested the draft survey with 12 strawberry growers in Brazlândia, and based on these results modified the questions to remove ambiguities. The survey design related to this study is provided in Table A1 in the Appendix, both in the original Portuguese and translated into English.

With the assistance of the Brazlândia extension office, we identified all 188 strawberry growers in the area. These growers had been classified by area of strawberry production and production system. The production system was an open-field, annual hill system but under conventional or organic production. The organic growers (The organic producers in the sample are all certified according to the Brazilian legislation for organic agriculture in place that consists of the Law number 10,831 for Organic Farming from 23rd December of 2003, and two Normative Instructions from the Ministry of Agriculture (IN17/2014 from June 17, 2014, and IN46/2011 from October 6, 2011) that cover crop production, animal husbandry, food processing and handling and labeling. This legislation states that all farms selling or labeling products as organic should be certified as such. Agriculture that meets OP standards, but is not subject to organic inspection, certification and labeling is referred to as ‘non-certified organic agriculture’ as distinguished from ‘certified organic agriculture.’) were previously identified by the Brazlândia extension office, and were supplemented by two additional farmers who we identified during the interview process. In a sister study (Andow et al., Reference Andow, Resende Filho, Carneiro, Lorena, Sujii and Alves2017), we separated conventional growers into two subgroups, transitional (A grower is considered transitional if he/she has a high diversity of crop production and has used at least two other agroecological practices, such as relying on biological control, using living windbreaks or organic fertilizers (Carneiro, Reference Carneiro2014).) and conventional, but found no difference between these subgroups. As our preliminary analysis also found no differences, the two subgroups were combined in this study.

We classified growers by area as ≤1 ha (small-scale), >1 and ≤2 ha (medium scale) and >2 ha (large-scale). We included all organic and large-scale conventional growers in our sample and randomly sampled 51.5% of medium-scale and 55.5% of small-scale conventional growers. If a producer was no longer in strawberry production, we selected a random replacement, when possible. Only nine growers refused to answer the survey, and they were replaced when possible, for an 85.2% response rate, which makes non-response bias probably absent. The data were collected between May and July, 2015 through face-to-face interviews by one person (DRL). The final sample size was 86 growers (Table 1).

Table 1. Number of growers contacted, sampled and interviewed and response rate.

¹ 12 growers were involved in the test sample, and thus could not be included in the final data, and 13 growers were either no longer producing strawberry, moved from Brazlândia or a relative who managed the farm had been interviewed.

² 2 organic growers were dropped of the sample because their data showed inconsistencies.

Notes: The data were collected between May and July, 2015 through face-to-face interviews by one person (DRL).

Impact evaluation strategies and techniques

Estimation of the impact of conversion to OP on outcome variables of growers based on nonexperimental observational data is not simple. The counterfactual outcome variables for organic growers in case they had chosen to produce conventionally are unobservable, which creates a missing data problem. In an experimental study, this problem would not exist because growers would be randomly assigned to produce conventionally (control group) or organically (treatment group). However, this was not possible in this study.

In observational studies like this one, wherein the assignment of a producer to produce conventionally or organically is not random, growers decide themselves given the information they have. Thus, organic and conventional growers may be systematically different on the basis of their observable and unobservable characteristics. To correct for these potential selection biases in estimating the impact of conversion to OP on outcome variables of growers, we use the propensity score matching (PSM) estimator as proposed by Rosenbaum and Rubin (Reference Rosenbaum and Rubin1983) and ESR estimator brought into the modern literature by Heckman (Reference Heckman1976, Reference Heckman1978).

The Roy–Rubin model based on potential outcomes (Roy, Reference Roy1951; Rubin, Reference Rubin1974) is the standard approach to this problem of inference in an observational study of individuals. The Roy–Rubin model relies on three main pillars (individuals, treatment and potential outcomes) and on the assumption that each individual has a well defined outcome for each treatment level (Caliendo and Kopeinig, Reference Caliendo and Kopeinig2008). Thus, let potential outcomes y _0i and y _1i be realizations of random variables y ₀ and y ₁ where the subscript zero denotes CP and subscript 1 denotes OP for individual i = 1, …, n, such that individual treatment effects, τ_i, is calculated as:

(1)$${\rm \tau} _i \equiv y_{1i} - y_{0i}$$

But in reality, for an organic grower i, only y _1i is observed and, for a conventional grower i, only y _0i is observed, as counterfactual outcomes are unobservable. The impossibility of observing both potential outcomes for the same individual at the same time constitutes the fundamental problem of causal inference in observational studies (Holland, Reference Holland1986) that ultimately is a missing data problem.

This study uses propensity score estimation and matching to overcome this missing data problem by matching the nearest organic and conventional growers according to the propensity score (Amare et al., Reference Amare, Asfaw and Shiferaw2012) to impute the missing potential outcome for each individual in the sample (Uematsu and Mishra, Reference Uematsu and Mishra2012).

We estimated the ATT growers as we are interested in the effect of the conversion to organic on outcomes of growers who might convert (Heckman, Reference Heckman1997) as:

(2)$${\rm ATT} \equiv E(y_1 - y_0 \vert t = 1) = (E(y_1 \vert t = 1) - E(y_0 \vert t = 1)$$

where t receives 1 if the grower is treated, otherwise zero.

The first assumption underlying matching estimators is the conditional independence (CI) assumption, which is also known by several other names, including ignorability (Wooldridge, Reference Wooldridge2010), selection on observables (Fitzgerald et al., Reference Fitzgerald, Gottschalk and Moffitt1998) or unconfoundedness (Rosenbaum and Rubin, Reference Rosenbaum and Rubin1983; Imbens, Reference Imbens2004). A problem, however, is that CI is not a directly testable assumption (Uematsu and Mishra, Reference Uematsu and Mishra2012).

The CI assumption says that after accounting for covariates in vector X, which is the set of characteristics used in propensity matching, the choice to convert or not can be taken as independent of the potential outcomes (y ₀, y ₁), which implies E(y ₁|t) = E(y ₀|t) = 0 (mean independence of outcomes from treatment). In other words, any remaining difference in the outcome variable y after matching can be solely attributed to organically or conventionally producing (Imbens, Reference Imbens2004), which can then be taken as purely random among matched observations (Becker and Ichino, Reference Becker and Ichino2002). Therefore, under the CI assumption, as E(y ₁|X,t) = E(y ₁|X) and E(y ₀|X,t) = E(y ₀|X), ATT may be obtained from matched observations as:

(3)$${\rm ATT} = E(y_1 \vert {\bf X}) - E(y_0 \vert {\bf X})$$

The second assumption underlying matching estimators is the overlap assumption. The overlap assumption means that each grower in the sample must have a positive probability of being assigned to producing organically so that for each possible X in the population, 0 < Pr(t = 1|X) < 1. Rosenbaum and Rubin (Reference Rosenbaum and Rubin1983) call the combination of the CI and overlap assumptions strong ignorability. As there is no specification test for the overlap assumption, we checked for violation of this by looking at the estimated probability densities to see if they have too much mass around 0 or 1 (Busso et al., Reference Busso, DiNardo and McCrary2014), but did not find evidence of violation.

The third assumption underlying matching estimators is that the observations are independent and identically distributed (i.i.d.), by which the outcome and treatment status of each individual are unrelated to the outcome and treatment status of all the other individuals in the sample. This assumption is fulfilled for this study because we used stratified random sampling.

Matching growers based on X might be difficult when the set of covariates is large (Amare et al., Reference Amare, Asfaw and Shiferaw2012) because of the problem of high dimensionality. However, Rosenbaum and Rubin (Reference Rosenbaum and Rubin1983) showed that if it is sufficient to estimate effects by adjusting for a vector of covariates, X, then one can use the estimated probability of treatment conditioned on X, $\widehat{{\rm p}}_i = \widehat{{{\rm pr}}}(t = 1 \vert{\bf X}_i)$ to match growers.

To avoid the curse of dimensionality, we follow Rosenbaum and Rubin (Reference Rosenbaum and Rubin1983) and instead of matching on each of the covariates in X, we match along $\widehat{{\rm p}}_i$, which is a single index variable that summarizes the covariates in X. In so doing, we use the NNM estimator derived by Abadie and Imbens (Reference Abadie and Imbens2006, Reference Abadie and Imbens2011) to match growers based on PS that we call the propensity score matching (PSM) estimator or PSM/NNM estimator. The NNM estimator based on the PSM estimator selects the individuals who are the closest matches to impute the missing potential outcome for each individual in the sample using the PS conditional on the vector of covariates X_i, $\widehat{{\rm p}}_i$. In other words, the PSM estimator matches samples based on a single continuous covariate, $\widehat{{\rm p}}_i$, such that the difference between the observed outcome and the imputed potential outcome is an estimate of the individual-level treatment effect. More recently, Abadie and Imbens (Reference Abadie and Imbens2016) derived a method to estimate the standard errors of the NNM estimator based on PS that we used in this study.

First step of the PSM estimator

The first step of the PSM estimator consists in estimating the probabilities that growers choose OP conditional on the vector of observable covariates, $\widehat{{\rm p}}_i({\bf X})$. We used a probit selection model to obtain those estimates as we assume the grower i decision to produce organically or conventionally is driven by the stochastic process (4):

(4)$$t_i\;{\rm = 1(}t_i^{^\ast} \gt {\rm 0)}\,{\rm \equiv}\, {\rm 1(}{\bf X}_i{\rm ^{\prime}}\beta \;{\rm +}\; {\rm \epsilon} _i\gt {\rm 0)\;} $$

where t _i is a binary variable that receives 1 if grower i were to choose to produce organically, which means $t_i^* \gt 0$, and that receives 0 if grower i were to choose to produce conventionally, which means $t_i^* \le 0$, where $t_i^* \;{\rm \equiv}\; {\bf X}_i{\rm ^{\prime}\beta \;+ \;} {\rm \epsilon} _i$ is a latent variable (non observable), X_i is the vector of observed covariates for grower i, β is a vector of parameters and ε_i is the realization of an i.i.d. normal error ε.

The maximum likelihood estimate for β, which is $\widetilde{{\rm \beta}} $ = arg max lnL($\widetilde{{\rm \beta}} $), is obtained by finding $\widetilde{{\rm \beta}} $ that maximizes the log-likelihood function of the probit model:

(5)$${{\rm In}\;L{\rm (}\widetilde{{\rm \beta}} {\rm ) =} \sum\nolimits_{i \in S} {{\rm In}\;\Phi {\rm (}{\bf X}_i{\rm ^{\prime}}\widetilde{{\rm \beta}} {\rm )}} + \sum\nolimits_{i \notin S} {{\rm In(1} - \Phi {\rm (}{\bf X}_i{\rm ^{\prime}}\widetilde{{\rm \beta}} {\rm ))}}} $$

where Φ(.) is the cumulative normal distribution, i ∈ S denotes a grower i who belongs to set S of growers who have chosen to produce organically, and $i \notin S$ denotes a grower i who does not belong to set S and therefore produces conventionally.

The conditional probability of grower i choosing OP over CP is estimated as:

(6)$$\widehat{{\rm p}}_i\;{\rm =}\; \Phi {\rm (}{\bf X}_i{\rm ^{\prime}}\widehat{{\rm \beta}} {\rm )}$$

Following Uematsu and Mishra (Reference Uematsu and Mishra2012), we used the first step of the PSM estimator to analyze how the covariates in vector X determined the conditional covariate probabilities of choosing to produce strawberry organically.

To eliminate the potential observable sample selection bias, it is important to carefully choose the covariates in vector X. They should be observable characteristics of growers that truly condition the probability that a producer would choose OP over CP, so that conditional probabilities can be estimated by Equation (6) according to the probit selection model.

The covariates we used in the probit model were selected based on empirical findings in the literature (Burton et al., Reference Burton, Rigby and Young2003; Uematsu and Mishra, Reference Uematsu and Mishra2012) and their values were directly obtained or calculated based on growers’ answers to related questions in Table A1 in the Appendix. To represent observable individual characteristics, we used the binary covariate male that receives 1 if a grower is male, and 0 otherwise, educ and educ ² that are the number and squared number of years of formal education a producer declared, age ² that is the squared age of a producer, p34 and p342 that are years and squared years a producer declared s/he has been farming strawberry.

We also included covariates related to farm characteristics in the probit selection model. The first of them was the binary variable emater that receives 1 if a producer declared s/he has used services provided by EMATER (Empresa de Assistência Técnica e Extensão Rural), which is the primary rural extension institution in Brazil, and 0 otherwise. The other two covariates were related to farm characteristics. We used area, which is the total area in hectares that a producer declared to have used to produce strawberry in year 2014, and the covariate p42, which receives values 1, 2, … ,7 according to the category of debt ratio or total debts divided by total assets in years 2013/14, such that, 0–15% (=1), 15–30% (=2), 30–45% (=3), 45–60% (=4), 60–75% (=5), 75–95% (=6) and 90–100% (=7). Thus, the higher the value of p42, the higher is the debt ratio in years 2013/14.

Second step of the PSM estimator

In the second step, the PSM estimator uses the NNM estimator developed by Abadie and Imbens (Reference Abadie and Imbens2006, Reference Abadie and Imbens2011) to generate the set of nearest neighbor indices, Ω(i), for each observation i = 1, … , n, according to Equation (7):

(7)$$\eqalign{& \Omega (i) = \{ j_1,j_2, \ldots j_{m_{_i}} \vert t_{\,j_k} = 1{\rm -} t_i, \cr & \quad \; \vert \widehat{\,p}_i - \widehat{\,p}_{\,j_k} \vert \lt \vert \widehat{\,p}_i - \widehat{\,p}_l \vert, \;t_l = 1 - t_i,\;l \ne j_k} $$

where m _i is the number of elements in the set Ω(i).

We use the teffects psmatch command in Stata (StataCorp, 2015) to execute the entire PSM estimator. Within this command, the number m such that m _i ≥ m can be chosen so that the number of matches for each grower i is at least m growers from the opposite treatment group (t = 1 − t _i). If m is small then unmatched observations may be discarded in estimating ATT but if m is larger more observations are used, which may compromise the quality of match (Uematsu and Mishra, Reference Uematsu and Mishra2012). Thus, we considered m _i = 1,…,5 in our analyses to check for the consistency of the PSM estimates.

Third step of the PSM estimator

In the third step, the PSM predicts potential outcomes $\widehat{y}_{1i}$ and $\widehat{y}_{0i}$ for each grower i = 1,…,n, according to Equation (8):

(8)$$\widehat{y}_{t_i} = \left\{ {\matrix{ {y_i} & {{\rm if}\;t = t_i} \cr {\displaystyle{{{\rm \sum} _{\,j \in \Omega (i)y_j}} \over {m_i}}} & \;\;\;\;{{\rm otherwise}} \cr}} \right.$$

where t _i is 0 if the grower i had chosen CP and 1 if s/he had chosen OP, Ω(i) is the set of growers j = 1, …, m _i from the opposite treatment group of grower i and matched by NNM to grower i.

Finally, ATT is estimated according to Equation (9):

(9)$$\widehat{{{\rm ATT}}} = \displaystyle{{\sum\nolimits_{i = 1}^n {t_i} ({\widehat{y}}_{1i} - {\widehat{y}}_{0i})} \over {\sum\nolimits_{i = 1}^n {m_i}}} $$

ESR estimator

Despite the fact that PSM corrects for observable bias by comparing the difference between the outcome variables of organic and conventional growers with similar observable characteristics, it does not correct for potential selection bias caused by unobservable characteristics of growers. Moreover, if there is selection bias based on unobservable characteristics, the CI assumption is violated.

We follow Amare et al. (Reference Amare, Asfaw and Shiferaw2012) and complement the PSM/NNE analysis with the ESR estimator brought into the modern literature by Heckman (Reference Heckman1976, Reference Heckman1978) to check the consistency of the results (Maddala, Reference Maddala1983; Dutoit, Reference Dutoit2007). We apply the ESR estimator using the etregress command in Stata (StataCorp, 2015).

The ESR estimator offers less flexibility than the PSM estimator because more structure must be imposed on the model when CI is not assumed. We use the ESR estimator that assumes a linear model for potential outcomes and a probit model for treatment assignment, such that:

(10)$$\eqalign{& y_i = {\rm \delta} _0 + {\rm \delta} t_i\;{\rm +}\; {\rm \epsilon} _i \cr & t_i = 1(x_i{\rm ^{\prime}}{\rm \beta} + u_i \gt 0{\rm )}} $$

where δ₀ is the intercept, δ is the parameter whose value is the ATT, ε_i is the error unrelated to treatment that is used to model potential outcomes, X_i is the vector of covariates that are unrelated to the error term u _i that is used to model treatment assignment, and the errors ε_i and u _i are bivariate normal with mean zero and covariance matrix given by:

(11)$$\left[ {\matrix{ {{\rm \sigma} ^{\rm 2}} & {{\rm \rho \sigma}} \cr {{\rm \rho \sigma}} & {\rm 1} \cr}} \right]$$

where σ² is the variance of ε_i, the variance of u _i is normalized as 1 so that the covariance of ε_i and u _i is ρσ.

We estimated the ESR model by full maximum likelihood based on the likelihood function given by Maddala (Reference Maddala1983, p. 122), and conducted a Wald test of the null hypothesis of H ₀: ρ = 0; to test if the correlation between the treatment assignment errors and the outcome errors is zero. We take the non-rejection of the null as evidence in favor of the (CI) assumption.