4.1 Empirical Results
The study considered cocoa farmers' who purchased crop insurance as the treatment group, and those who did not purchase are referred to as the control group. In order to have a thorough investigation, we estimate the descriptive statistic in three main categories. Thus, we consider the entire sample, the treatment, and the control group. In addition to providing basic descriptive statistics for each variable employed in the study, we compare the means of farmers who purchased crop insurance against those who did not.
Evidence from the fundamental descriptive analysis indicates the systematic differences between treatment and control group. From the descriptive in Table 3, we can deduce that Income, Age, Marital status, Education, awareness of crop insurance, age of farm, extension officer and savings recorded a significant difference between treatment and control groups. In contrast, other variables did not exhibit any significant difference between the two groups. However, a focus on our interest variable (income) implies that farmers who had the opportunity to participate or purchase crop insurance improved their income compared to farmers who did not purchase crop insurance.
This finding is consistent with Zhao et al. (2016). They identified a difference among farmers who purchase crop insurance to have an average income than farmers who do not pay for it. They, however, likewise disapproved of the findings on the basis that the DiD estimator is biased. They equally attribute the increment in farmers' income to be nominal since it is not inflationary adjusted as in the case of this study. This finding cannot be reliable since it has been confirmed that there is an imbalance among the distributions. Hence, the study needs to employ the propensity score matching method to reduce the bias identified among the treated and control group based on the observable covariates and better compare the groups.
Table 3
|
All
|
Uninsured
|
Insured
|
M(c) – M(t)
|
Variables
|
Mean
|
Std. Dev.
|
Mean
|
Std. Dev.
|
Mean
|
Std. Dev.
|
|t|
|
lnINC
|
7.40
|
.566
|
7.374
|
.528
|
7.450
|
.630
|
-2.633**
|
Gen
|
1.32
|
.468
|
1.309
|
.463
|
1.347
|
.477
|
-0.922
|
Edu
|
1.88
|
.823
|
1.817
|
.831
|
2.020
|
.791
|
-2.858**
|
Hsz
|
5.80
|
1.799
|
5.807
|
1.741
|
5.796
|
1.919
|
0.070
|
AgeF
|
7.23
|
4.420
|
7.005
|
4.220
|
7.679
|
4.786
|
-1.754*
|
Age
|
43.99
|
6.405
|
43.931
|
6.402
|
44.117
|
6.425
|
-0.335
|
MS
|
2.37
|
.968
|
2.275
|
.897
|
2.561
|
1.077
|
-3.429***
|
Fexp
|
14.06
|
5.914
|
14.205
|
5.976
|
13.760
|
5.788
|
0.865
|
Fsz
|
11.95
|
4.699
|
12.069
|
4.927
|
11.699
|
4.191
|
0.905
|
ACIN
|
1.58
|
.494
|
1.629
|
.484
|
1.481
|
.501
|
3.416***
|
AC
|
1.67
|
.469
|
1.668
|
.472
|
1.672
|
.471
|
-0.104
|
EO
|
2.13
|
1.078
|
2.230
|
1.093
|
1.923
|
1.017
|
3.296***
|
lnSAV
|
6.43
|
.420
|
6.379
|
.453
|
6.512
|
.349
|
-2.779**
|
lnOTI
|
7.26
|
.554
|
7.249
|
.588
|
7.278
|
.479
|
-0.555
|
Source: Field Survey, 2019 |
4.3 The impact of crop insurance on cocoa farmers’ income
Figure 2 below displays the distribution of the estimated propensity score standard in the region of common support for cocoa farmers' who purchased crop insurance from the year of its implementation and cocoa farmers who did not purchase crop insurance since its implementation.
Before a PSM estimation, it is necessary to check for possible violations of some conditions; a typical example is the overlap assumption. There is a need for both treated and control groups to satisfy the common support condition. Thus, according to Caliendo and Kopeining (2008), both groups must be within the standard support region. Evidence from the visual assessment of Fig. 2 above on the density distribution between the groups indicates that the treated and control group are within the region of common support. From Fig. 2 above, the upper region in red implies the distribution of the treated group, whereas the bottom in blue represents the control group. The y-axis denotes the propensity scores for the two groups. By implication, each cocoa farmer had a positive probability of being a buyer of crop insurance or not a buyer of crop insurance. It validates the common support assumption that necessitates each cocoa farmer who purchased crop insurance to have a corresponding non-crop insurance buyer as a match (Austin, 2011; Ashimwe, 2016).
Matching is an essential estimation for treatment effect to remove over bias and estimate the treatment effect using observational data (Baser, 2006). The goal of propensity score estimation is to balance the covariate distribution in the treated and control groups (Rosenbaum & Rubin, 1983). Propensity score matching employs a predicted probability of the treated and control grouped in this case, cocoa farmers who purchased crop insurance and did not purchase crop insurance, respectively. In order to maintain consistent matching between the results, four different techniques for the estimation of the propensity score matching was employed, namely, Nearest Neighbor (NNM), Radius (RM), Kernel (KM) and Local linear regression matching (LLRM). With the NNM method, the study seeks to order treated and control groups randomly. Then select the first cocoa farmers who purchased crop insurance and find one cocoa farmers who did not purchase crop insurance with the closest propensity score (LaLonde, 1986; Baser, 2006).
Regarding the above, the study employed all the four techniques mentioned above to estimate the differences between the covariates of cocoa farmers. With RM estimation, each treated element was matched only with the control element whose propensity score falls in a predefined neighbourhood of the propensity score of the treated unit (Dehejia & Wahba, 2002; Baser, 2006). According to Baser (2006), the pros of the NR method are that it uses only the number of judgment elements accessible within a predefined radius, hence allowing for the use of extra elements when suitable matches are available and fewer units when they are not. One disadvantage of this method is the decision of precise radius to use compared to KM. All treated elements are matched with a weighted average of all controls, with weights inversely proportional to the distance between the propensity scores of the farmers who purchased and did not purchase the insurance. All cocoa farmers who did not purchase crop insurance contribute to the weights achieving lower variance. According to Caliendo and Kopeining (2008), this can be considered as a plausible counterfactual.
These four methods have evidentially portrayed no systematic differences in the distribution of covariates between treated and control groups. Insignificant p-values of the likelihood ratio and a reduction in bias after matching for the covariates balance tests, according to Rosenbaum and Rubin(1985), should be used to specify the estimation. Caliendo and Kopeining (2008) later posit that variances are predictable before matching; nevertheless, there should be a balance in the treated and control group after matching the covariates, indicating that no significant difference is found. Evidence from the current study shows no significant difference found between the treated and control group after matching.
Table 5 below presents summary statistics, including the standardized mean, median bias coupled with the pseudo-R2. According to Rosenbaum and Rubin (1985), the standardized mean and median bias differences between the treated and control groups should not be more significant than 20%. The difference between the standardized mean and median bias when greater than 20% is considered large. Evidence from Table 5 indicates that among the four matching algorithms employed in the covariate balancing tests, radius matching and kernel-based matching stands out to be the appropriate matching algorithm. This assertion is based on the reduction rate of standardized mean and median bias. The mean bias before matching for all the four matching techniques is 11.5, which has reduced to 3.8, 2.2, 2.1 and 3.6 after matching for the nearest neighbour, radius, kernel-based and local linear regression matching, respectively. The median bias also reduces from 8.5 to 3.3, 2.1, 1.9 and 3.3 for the nearest neighbour, radius, kernel-based, and local linear regression matching. Based on the results, the appropriate matching techniques we adopt is radius and Kernel-based propensity matching. Considering the mean and median biases of the RM and KM are below 10%, indicating a good match between cocoa farmers who purchased and did not purchase crop insurance.
Again, pseudo-R2 of the RM and KM before and after matching indicates the probability that farmers who did not purchase crop insurance are likely to purchase (Sianesi, 2004).
Before matching, the pseudo-R2 was 0.037, but it dropped to 0.001 for RM and KM, which is very low and significant enough to support the assumption that there are no variations in the distribution of variables between the treatment and control groups. As a result, this study asserts that the matching method was able to effectively balance the circulation of covariates between the treated and control groups, and there is a reliable counterfactual on the assertion that this study achieved low pseudo-R2 values, insignificant p-values, low standardized mean bias, and high total bias reduction. These findings indicate that there is no consistent difference in the covariate distribution of income between the treatment and control groups. As a result, the crop insurance intervention in the community could result in a unit difference in farmers' income between the two groups with the possibility of increasing.
Table 5
Summary of covariates balancing tests for Matching Algorithms
Types Matching algorithms
|
Pseudo-R2 Before Matching
|
Pseudo-R2 After Matching
|
P > Chi2
before
|
P > Chi2
After
|
Mean bias before
|
Mean bias After
|
Median bias before
|
Median bias After
|
NNM
|
0.029
|
0.004
|
0.037
|
0.997
|
11.5
|
3.8
|
8.5
|
3.3
|
RM
|
0.029
|
0.001
|
0.037
|
1.000
|
11.5
|
2.2
|
8.5
|
2.1
|
KM
|
0.029
|
0.001
|
0.037
|
1.000
|
11.5
|
2.1
|
8.5
|
1.9
|
LLRM
|
0.029
|
0.040
|
0.037
|
0.999
|
11.5
|
3.6
|
8.5
|
3.3
|
Source: Author’s computation based on survey data (2019) |
The treatment effect (ATT) of crop insurance intervention on cocoa farmers' income are presented in Table 6 below. Table 6 reports the treatment effects based on nearest neighbour, radius and kernel matching algorism. The results for comparison of cocoa farmers’ who purchased and did not purchased crop insurance are statistically insignificant but in the anticipated positive direction for the neighbour, radius and kernel matching estimations. This implies that farmers' participation in crop insurance has a positive impact on farmers’ income. The interpretation of this result would imply that a cocoa farmer who purchases crop insurance tends to earn a higher income than their counterparts who did not purchase crop insurance by 8.3% and 6.8%, respectively. The result is consistent with Zhao et al. (2016) where they established that there is no significant impact of crop insurance on farmers’ income inner Mongolia in China. The estimation also agrees with Varadan and Kumar's (2012) study, which revealed that high use of farm inputs and production risks are absorbed by agricultural insurance. Nahvi et al. (2014) found a significant positive relationship between income and agricultural insurance in Iran. Similarly, Yanuarti et al. (2019) also found a positive impact of crop insurance on Indonesian farmers' income.
Table 6
Impact of crop insurance on cocoa farmers’ income
Dependent variable Farmers’ income
|
Matching Algorithms
|
Treated
|
Control
|
ATT
|
Bootstrap
S.E
|
T
|
NNM
|
194
|
102
|
0.083
|
0.077
|
1.083
|
RM
|
4
|
317
|
0.776
|
0.224
|
3.463***
|
KM
|
194
|
384
|
0.068
|
0.052
|
1.309
|
Source: Author’s computation based on survey data (2019) |