Does real wage converge in China?

Abstract

This study adopts flexible Fourier unit root test proposed by Enders and Lee (Oxf Bull Econ Stat 74(4):574, 2012) to explore real wage convergence in China in the process of market-oriented reform. We find that our approximation has higher power to detect U-shaped and smooth breaks than linear method if the true data generating process of relative real wage is in fact a stationary non-liner process. The empirical results show that property of stationarity of relative real wage varies across different regions. Specifically, stationarity prevails in eastern and western regions, implying the regional development strategies launched there enhance labor migration, commodity trade and thus wage convergence. In contrast, almost all provinces in northeastern and central regions show non-stationarity, indicating the strategies there are less effective in promoting wage convergence. These results have practical policy implications for China’s regional development and income equality.

Appendix

In the paper, we capture the structural breaks via Data Generating Process with Fourier form, and the results in Table 2 show that the breaks are gradual. Here, we provide the result of Bai and Perron test (see Table 3) for comparison.

Table 3 Bai and Perron break test of relative real wage

The result of Bai and Perron break test is displayed in Table 3. According to Bai and Perron (1998)’s definition about statistics, Table 3 shows there are abrupt structural changes in the relative real wage for 12 out of 31 provinces. However, Table 2 in the paper shows there are gradual structural changes in the data for 29 provinces. That implies the data we used are better fitted by Ender and Lee’s model. To facilitate regional comparison, it also seems more proper to capture structural breaks with Fourier function in our study.

Besides, there might be size distortion in Bai and Perron test here. In the break test, it is important to select the value of $\varepsilon$ first to avoid size distortion. Bai and Perron (1998) point out that: “ With a sample of $T = 120, \varepsilon = 0.15$ should be enough for heterogeneity in the errors or the data. If serial correlation is allowed, $\varepsilon = 0.20$ may be needed”. In other words, generally $h$ should be 18 at least. In our case, we choose the biggest value (0.25) for $\varepsilon$ and acquire the results in Table 3. As $T=33$, the minimum length of a segment $h$ only equals to 8, much smaller than the one recommended by Bai and Perron (1998, 2003). That implicates size distortion probably exists here and the result in Table 3 lacks robustness. However, there is not such restriction about sample size in Enders and Lee (2012) test.

Consequently, it seems more reasonable to capture structural changes in the data with Enders and Lee (2012)’s method.