Informal search, bad search?: the effects of job search method on wages among rural migrants in urban China

Abstract

The use of informal job search method is prevalent in many countries. There is, however, no consensus in the literature on whether it actually matters for wages, and if it does, what are the underlying mechanisms. We empirically examine these issues specifically for rural migrants in urban China, a country where one of the largest domestic migration in human history has occurred over the past decades. We find that there exists a significant wage penalty for those migrant workers who have conducted their search through informal channels, despite their popularity. Our further analysis suggests two potential reasons for the wage penalty: (1) the informal job search sends a negative signal (of workers’ inability to successfully find a job in a competitive market) to potential employers, resulting in lower wages, and (2) there exists a trade-off between wages and search efficiency for quicker entry into local labor market. We also find some evidence that the informal job search may lead to low-skilled jobs with lower wages. We do not find strong evidence supporting alternative explanations.

Appendix: Simulation results for the impacts of misspecification on KV-IV estimates

In our empirical analysis, we follow Millimet and Tchernis (2013) and employ a parametric variant of the KV-IV approach by specifying both the distributional and heteroskedastic error functions. This practice has two distinct advantages as it greatly reduces the computation burden, and it also eliminates the need for continuously distributed exogenous variables in the semi-parametric approach. As argued in Millimet and colleague’s paper, this practice is innocuous. Wooldridge (2010, p.939) and Angrist and Pischke (2009) also note that the consistency of IV estimates does not depend on the correct specification of the first-stage equation. In other words, misspecification of the distribution and heteroskedastic functions should not necessarily impact the KV-IV estimates, provided that the other assumptions hold.

Monte Carlo (MC) simulations corroborate this theoretical point. Specifically, we undertake two sets of MC experiments using simulated data. The finite sample performances of the KV-IV estimator under correct specifications have been shown elsewhere (e.g., Millimet and Tchernis (2013)).

The first set of MC experiments consider only the impacts of the misspecification of the distribution functions, while the second set of MC experiments consider the impact of the misspecification of both the distribution and the heteroskedastic functions. We perform this exercise twice for each case using 1000 simulations of sample sizes 4200 (roughly the size in our application) and 10,000.

The first MC design is based on the following data-generating process:

$$\begin{array}{@{}rcl@{}} y & = & 1 + D + x_{1} + x_{2} + \epsilon \end{array} $$

(10)

$$\begin{array}{@{}rcl@{}} D & = & I(x_{1} + x_{2} - v >0) \end{array} $$

(11)

$$\begin{array}{@{}rcl@{}} v & = & \exp(x_{1}+x_{2})v^{*} \end{array} $$

(12)

$$\begin{array}{@{}rcl@{}} \epsilon & = & 0.2 v^{*} + \eta \end{array} $$

(13)

$$\begin{array}{@{}rcl@{}} & & \eta, x_{1}, x_{2} \sim \mathbb{N}(0,1) \end{array} $$

(14)

$$\begin{array}{@{}rcl@{}} & & v^{*} \sim {\chi^{2}_{1}} \end{array} $$

(15)

Note that Eq. 12 specifies the heteroskedastic function, while Eq. 15 specifies the distribution function (which is chi-squared distributed). The endogeneity arises because of Eq. 13.

For comparison, we choose the best scenario as our benchmark where the correctly specified predicted value, $F\left (\frac {x_{1} + x_{2}}{\exp (x_{1}+x_{2})}\right )$ (where F(⋅) is the cumulative chi-squared distribution with 1 degree of freedom), is used as the IV. This is the best scenario not only because we use the correct specification, but also because we use the true parameter values (as opposed to the estimated ones); this is denoted as True IV. We then compare the benchmark results to the results using our misspecified parametric IV, ${\Phi }\left (\frac {\widehat {\lambda _{1}}x_{1} + \widehat {\lambda _{2}}x_{2}}{\exp (\widehat {\theta _{1}}x_{1}+\widehat {\theta _{2}}x_{2})}\right )$ (where Φ(⋅) is the cumulative distribution function for standard normal variables); this is denoted as Parametric IV.

The second MC design is based on the following data-generating process:

$$\begin{array}{@{}rcl@{}} y & = & 1 + D + x_{1} + x_{2} + \epsilon \end{array} $$

(16)

$$\begin{array}{@{}rcl@{}} D & = & I(x_{1} + x_{2} - v >0) \end{array} $$

(17)

$$\begin{array}{@{}rcl@{}} v & = & (x_{1}+x_{2})^{2}v^{*} \end{array} $$

(18)

$$\begin{array}{@{}rcl@{}} \epsilon & = & 0.2 v^{*} + \eta \end{array} $$

(19)

$$\begin{array}{@{}rcl@{}} & & \eta, x_{1}, x_{2} \sim \mathbb{N}(0,1) \end{array} $$

(20)

$$\begin{array}{@{}rcl@{}} & & v^{*} \sim {\chi^{2}_{1}} \end{array} $$

(21)

Note that the true heteroskedastic function is now different (Eqs. 18 vs. 12). Therefore, for the second MC design, the best scenario case $F\left (\frac {x_{1} + x_{2}}{(x_{1}+x_{2})^{2}}\right )$ (where F(⋅) is the cumulative chi-squared distribution with 1 degree of freedom) is used as the IV. For Parametric IV, we continue to use ${\Phi }\left (\frac {\widehat {\lambda _{1}}x_{1} + \widehat {\lambda _{2}}x_{2}}{\exp (\widehat {\theta _{1}}x_{1}+\widehat {\theta _{2}}x_{2})}\right )$. Note that we now misspecify both distribution and heteroskedastic functions.

The results are presented in Table 13; OLS results are also included for comparison. The simulations indicate that in the presence of endogeneity, OLS is severely biased, which is not surprising. What is surprising is the outstanding performances of the parametric approach, relative to the best scenario. First, in the smaller sample (as ours), the parametric IV, although misspecified, performs extremely well and similarly to the true IV (the best scenario). Second, as the sample size increases, both IVs have an average bias nearing zero, and the difference between two IVs becomes even smaller and barely distinguishable. In sum, these MC results corroborate the theoretical expectations above.

Table 13 Monte Carlo results