We next check the threshold constancy by employing the test statistic F2 proposed in Section 2, and the p-value is 0.079 when output is treated as a threshold variable while 0.721 when the threshold variable is literary. Therefore, we conclude that there might exist an unstable threshold effect when output is treated as the threshold variable. By ignoring potentially significant time-varying features in the threshold, the estimated slopes in sub-samples are seriously biased as shown in Section 3. Hence, the parameter estimation based on the proposed approach can be expected to be more robust.