Propensity Score and Matching Analysis

In an impact assessment study, one of the most difficult issues is the possibility of selection biases. This problem occurs because we would like to know the effect of a treatment on the participants' outcome but cannot observe the outcomes with and without treatment on the same individual at the same time. Simply comparing mean outcomes may not reveal the actual treatment effect, as participants and non-participants typically differ even in the absence of treatment (Caliendo and Kopeining, 2005). For example, contract farmers may differ systematically from non-contract farmers and the above simple mean comparisons may reflect differences in their characteristics rather than the impacts of contract farming. In other words, failure to account for treatment selection biases may lead to biased estimation of the true treatment effect.

The propensity scoring matching (PSM) method (Becker and Ichino, 2002) provides a more refined method of comparing the performance of contract and non-contract farmers by accounting for their inherent differences. The basic concept is to compare contract farmers to non-contract farmers who are similar to contract farmers in all relevant characteristics except the contract. The differences in the outcomes of contract farmers and the selected non-contract farmers can then be attributed to the contract.

The first step of the PSM approach is to estimate farmers' propensity scores based on their basic characteristics (i.e., characteristics that are not affected by the choice of contract). The propensity score of each farmer measures his tendency to join the contract. The magnitude of a propensity score ranges between 0 and 1; the larger the score, the more likely the farmer is to join the contract.

After farmers' propensity scores are estimated, the second step is to divide farmers into groups of similar propensity scores. In addition, each group should be balanced, containing farmers who do not have significantly different characteristics.

After the balanced groups are formed, we can compare the performance of contract and non-contract farmers in each group. As such comparisons are based on stratification control for the differences of farmers' characteristics, the performance differences between contract and non-contract farmers would be more likely caused by contract farming rather than farmers' intrinsic characteristics.

Finally, the performance difference between contract and non-contract farmers can be measured by the weighted average of the contract and non-contract differences in each group, with the number of observations in each group as the weights.

The propensity score approach is used here to compare contract farmers' and non-contract farmers' performance in their commercial operation. The following variables are used in the propensity score estimation: 1) farm size; 2) number of adult family members; 3) ratio of females in the family; 4) value of production assets; 5) value of consumption assets; 6) value of transportation assets; 7) farm distance to highway; and 8) farm distance to market.

Table 5 [ PDF 22.8KB | 1 page ] presents the differences in the performance of contract and non-contract farms, using simple mean and propensity score matching comparisons. The findings of the PSM comparisons are consistent with the results of the simple mean comparisons. They indicate that contract farms have higher revenue, rice price, yield, cash costs, and profit than noncontract farms, and that the results are statistically significant.

The use of PSM to minimize selectivity bias thus suggests that these differences are the result of contract farming rather than the intrinsic characteristics of the sampled households. However, like the simple mean comparison, PSM may misinterpret the treatment effect, because it only controls for observed variables, and hidden self-selectivity bias may remain. As the decision to join the contract is voluntary and is based on individual self-selection, it is possible that contract farmers have systematically different unobserved characteristics from non-contract farmers. For example, farmers' motivation may be an unobserved covariate affecting both their performance and their decision to join the contract. To address these unobservable selection biases, we employ an endogenous switching regression model as described below.

Download this Discussion Paper [ PDF 127.3KB| 24 pages ].

Post a Comment

We welcome your feedback on this publication. Post a comment. ADBI is not obliged to acknowledge or publish comments and may abridge or edit them before web posting.

Comment(s)

There are [0] comment(s) for this entry. Post a comment.