Using the parameters obtained by the probit estimation, we compute propensity scores (i.e., the probability of SRI use) for all plots. To estimate SRI’s impacts on observations of SRI plots and non-SRI plots that have common support, we drop the observations of SRI plots with propensity scores higher than the maximum, or lower than the minimum of those not in SRI. Different matching algorithms exist, each possessing positive and negative attributes (Caliendo and Kopeinig 2008). Among the available options, the simple NNM and kernel matching methods are the ones most commonly used. As the counterfactual of the treatment group (SRI plots), the NNM employs those members of the control group (non-SRI plots) that are closest in terms of covariates or propensity scores. Yet as figure 2 suggests, the distribution of the propensity score of non-users in our sample is highly skewed to the right, while that of the users is skewed oppositely. This raises a concern about matching relatively dissimilar plots under the simple NNM method. Moreover, Abadie and Imbens (2011) show that simple NNM estimators are generally not root-N consistent. We therefore report and interpret the results based on kernel matching, which uses a large number of matches per unit to estimate counterfactual outcomes via nonparametric smoothing techniques, with the weighted averages of all non-SRI plots, where the weight is inversely proportional to the propensity score distance between SRI plots and non-SRI plots. We ultimately drop 12 treated observations that do not meet the common support condition. To estimate the ATT, we employ an Epanechnikov kernel with a bandwidth of 0.06, obtaining standard errors by bootstrapping with 500 replications.