[Objective] Studying whether the implementation of land supervision is effective in arable land protection and solving the sample selection bias for land supervision improvement.[Methods] Based on province-level panel data from 1999 to 2008, propensity score matching(PSM) was employed to estimate the effect of land supervision on arable land protection.[Results] ① Kernel matching is the best method by balancing test. This study estimates that the loss of arable land due to construction use decreases 8 037.489 hm² per year because of special land supervision in effect and the loss of arable land due to construction use decreases 62 741.880 hm² per year because of regular land supervision in effect by Kernel matching. ② The sample selection of special land supervision is not random, resulting in the effect of special land supervision on cultivated land protection larger in previous studies. ③ The sample selection of regular land supervision is more random, resulting in the sample selection bias smaller and the effect of regular land supervision on cultivated land protection more accurate compared with the special land supervision in previous studies.[Conclusion] The effect of cultivated land protection by regular land supervision is better. In the evaluation of land policy, we need to consider the difference of the relationship between the policy and its object, in order to choose the method used in policy effect evaluation.