Síntesis de los efectos de las políticas sobre el cambio tecnológico

Measurement of the productivity and welfare gain of adoption of agricultural technologies based on non-experimental observations is quite complex because of the need to find a counterfactual of intervention. In real life situation, it is impossible to observe the productivity and welfare outcomes of those farmers who adopted the technology had they not adopted during the same time. Furthermore, technology is not randomly distributed to the two groups of the households (adopters and non-adopters), but rather the households themselves decide to adopt or not to adopt based on several observable and non-observable characteristics. Therefore, adopters and non-adopters may be systematically different (Asfaw, 2010).

Against the above backdrop, impact evaluation using simple mean comparisons of welfare outcome variables (Ahimbisibwe and Mungatana, 2012;

Wellard et al 2015) may lead to erroneous results because the adopters and non-adopters may not be the same prior to the intervention such that the expected difference in outcome variables between the two groups may not solely be due to

adoption of the improved technology. Asfaw, (2010) and Kuntashula and Mungatana (2013) contend that the difference in farmer welfare between the two groups in the absence of technology adoption can be attributed to selection effect. Therefore the observed difference in welfare due to uptake of improved technologies includes the difference attributed to the selection effect or bias. It is argued that since the counterfactual of adopters is not known, it is difficult to estimate the magnitude of selection bias. By extension therefore, it is difficult to know the extent to which selection bias makes up the observed difference in outcomes between the adopters and non-adopters.

According to Asfaw (2010), the simplest approach to examine the impact of adoption of improved technologies on welfare outcomes would be to include on the welfare equation a dummy variable equal to one if the farm-household adopted the new technology, and then, to apply ordinary least squares. This approach, however, might lead to biased estimates because it assumes that adoption of improved technology is exogenously determined while it is potentially endogenous. The decision to adopt or not is voluntary and may be based on individual self-selection.

Farmers that adopted may have systematically different characteristics from the farmers that did not adopt, and they may have decided to adopt based on expected benefits. Unobservable characteristics of farmers and their farms may affect both the adoption decision and the productivity and welfare outcomes, resulting in inconsistent estimates of the effect of adoption of the agricultural technology on household productivity and welfare. For instance, if only the most skilled or motivated farmers choose to adopt and we fail to control for skills and motivation, then according to Asfaw (2010), we will incur an upward bias. The solution is to explicitly account for such endogeneity.

Selection bias is controlled for through the creation of the counterfactual or a situation the adopting farmer would have experienced had they not adopted during the same simultaneous period. Using data from a cross sectional survey like the one that was conducted for this study, a counterfactual can be created through different approaches which include randomization in treatment assignment, Propensity Score Matching, using Instrumental Variable (IV) method, and Endogenous Switching Regression Modeling.

To start with, the selection effect disappears if treatment assignment is completely random (Asfaw et al., 2011; Asfaw, 2012; Taylor et al., 2012). The aim of randomisation is to make sure that the farms adopting the improved technologies and those not, have an equal probability of adopting the technology. This is because randomization eliminates the economic decisions that drive the treatment choice. In this study, this would imply that if participation in AIS initiatives or adoption of improved certified cassava seed is completely random, then the problem of selection effect disappears. However, this hypothetical situation cannot be achieved for this ex-post study because there was no control during the dissemination of the technology. Even still, it would be difficult and bordering on ethical issues to only give the technology to a selection of farmers and intentionally leave out others as controls in a situation where everyone is striving to benefit from the technology so as to get out of poverty. Thus the only seemingly plausible way out would be to explicitly account for such endogeneity using simultaneous equation models as suggested by Hausman, (1978).

However, the simultaneous equation modeling approach becomes problematic because it is inappropriate to use a pooled sample of adopters and adopters (with a binary indicator equaling to one for adoption and zero for non-adoption). This is because the approach would assume that technology adoption has an average impact over the entire sample of farmers, by way of an intercept shift, or that it raises the productivity of factors of production, by way of slope shifts in the outcome functions (Alene & Manyong, 2007).

Secondly, impact evaluations (Heckman et al., 1998; Blundel and Dias, 2000) can use matching methods to randomise the farmers and thereby create a plausible counterfactual. The matching technique works by creating randomness in treatment assignment on the assertion that if untreated individuals (non-adopters) have the same probability of participation as treated households (adopters), then the average welfare outcome estimates for the non-adopters becomes a good approximation of adopters’ productivity and welfare outcome estimates had they not adopted (Madola, 2011). The Propensity Score Matching technique corrects the estimation of treatment effects by controlling for confounding factors based on the premise that the bias is reduced when the comparison of outcomes is performed using treated and

control subjects who are as similar as possible in all ways before the treatment (Becker and Ichino, 2002).

The PSM method is one of the non-parametric estimation techniques that do not depend on functional form and distributional assumptions as is the case in OLS, Instrumental Variable (IV) and Heckman procedures (Bryson et al., 2002). Mendola (2007) and Magrini and Vigani (2016) argue that imposing any restriction – such as linearity and normal distribution for the error term - on the relationship between outcome variables and their determinants would be a strong assumption if not supported by theory. Further still, it is argued that matching does not impose any exclusion restrictions for identifying the selection process as in the case of IV and Heckman procedure (Magrini and Vigani, 2016). Jalan and Ravallion (2003) advise that finding such a good instrument – especially in cross-sectional datasets - is always complicated and its suitability is not fully testable.

The PSM method is intuitively attractive as it helps in comparing the observed outcomes of technology adopters with the outcomes of counterfactual non-adopters (Heckman et al., 1998). According to Asfaw (2010), the matching method can produce experimental treatment effect results when such data are not feasible and/or available. It also helps to evaluate programs that require longitudinal datasets using single cross-sectional dataset where the former does not exist as is the case in this study. The basic idea of the PSM method is to match observations of adopters and non-adopters according to the predicted propensity of adopting a superior technology (Rosebaum and Rubin, 1983; Heckman et al., 1998; Smith and Todd, 2005; Wooldridge, 2005). The main feature of the matching procedure is the creation of the conditions of randomized experiment in order to evaluate a causal effect as in a controlled experiment.

In agreement with Caliendo and Kopeinig (2005), Kuntashula and Mungatana (2013) summarise that matching is a form of randomisation that assumes away the selection effect by assuming that selection is based on observables. If all observable characteristics can be used to match adopters and non-adopters, then the causal effect of improved technology on farmer welfare indicators can be compared using like or similar groups of farmers. Although matching methods are intuitively easier, the assumption that selection bias is based only on observed characteristics is its

main weakness. Matching cannot account for unobserved factors such as (skill, motivation, ambition and risk taking behavior) influencing adoption of technologies thereby leaving unsolved the problem of endogeneity due to unobservable covariates.

The third impact evaluation approach to solve the selection and endogenous problem is the Instrumental Variable (IV) approach in which a randomly assigned variable (instrument) that would not affect the outcome variable except through its effect on the treatment can be used. According to Kuntashula and Mungatana (2013), this becomes vital when the estimation is concerned with correlation of the treatment variable (e.g. improved certified cassava seed adoption or participation in AIS initiative) with the errors. The instrument should be correlated with adoption of improved certified cassava varieties but uncorrelated with productivity and farmer welfare so that by extension it should not be correlated with the error term.

Instrumental variable estimation is a good identification strategy to estimate causal relationships in theoretical work since there are practical difficulties when applying it to an empirical study. The main weakness with this approach is that it is very difficult to find such an instrument. As earlier mentioned, it is difficult to find good instruments that are not correlated with the endogenous variable or the error term. If the instruments are weakly correlated with these factors, biased and inconsistent estimates may be produced. Even if good instruments are utilized in a model, it is hard to assess the extent to which the treatment of endogeneity affects the magnitude of the outcome estimates.

The fourth approach is called the Heckman two step selection procedure which is closely linked with the Instrumental Variable approach. Heckman (1979) proposed an alternative model that addresses the selection problem arguing that an estimation on a selected subsample results in sample selection bias. Also called the sample selection model, it involves two equations: firstly, the regression equation that considers mechanisms determining the outcome variable and secondly, the selection equation considering a portion of the sample whose outcome is observed and mechanisms determining the selection process. While the first estimates the probability of observing a positive outcome (known as the selection or participation equation), the second estimates the level of participation conditional on observing positive values (known as the conditional equation) (Dow & Norton, 2003). The

model assumes that different sets of variables could be used in the two-step estimations. Kennedy (1998) however states that the Heckman Selection model does not perform well when: the error terms are not distributed normally; the sample size is small; the amount of censoring is small; the correlation between errors of the regression and selection equations is small; and the degree of collinearity between the explanatory variables in the regression and selection models is high.

Finally, the most robust impact estimation approach is called the Endogenous Switching Regression modeling (ESR) approach proposed in Madada and Nelson (1975), Freeman, et al. (2001), Alene and Manyong, (2007), Asfaw (2010) and emphasized in Madola et al, (2011), Kuntashula and Munagata, (2013) and further in Shiferaw et al. (2014), Kassie et al. (2015) and Magrini and Vigani (2016). The ESR uses maximum likelihood estimation (MLE) techniques to predict the potential outcomes the adopter (or non-adopter) of a technology would get in the two regimes of either adopting or not. The model is comprised of the selection equation or the criterion function and two continuous regressions that describes the behaviour of the farmer as he faces the two regimes of adopting the improved technology or not.

According to Freeman et al. (2001) and Alene & Manyong (2007), the ESR accounts for both endogeneity and sample selection and allows interactions between adoption and other covariates in the outcome function.

Based on the above discourse, this study used Propensity Score Matching (PSM) for binary treatment effects complimented with Endogenous Switching Regression modeling (ESR) because of their heightened robustness as detailed in this section. The next section presents a detailed appraisal of previously adduced empirical evidence of productivity and welfare impacts of agricultural technology adoption on smallholder farmers with a view of further crystallizing the research gap.

2.4.2 Empirical considerations of impact of cassava technology adoption