The complex relationship between neighbourhood types and migrants’ socio-economic integration: the case of urban China

Abstract

Migrants are important parts of China’s urban population, but still limited knowledge is known about the contextual determinants of their socio-economic integration in Chinese cities. Based on a large national micro-level data extracted from the 2014 China Migrant Dynamic Survey, we extend the literature on migrants’ socio-economic integration by examining the relationships between neighbourhood types and different dimensions of socio-economic integration. Our multi-dimensional index of socio-economic integration covers economic, socio-cultural and identity perspectives. The results show that migrants demonstrate significantly higher levels of overall socio-economic integration when living in formal neighbourhoods (composed of commercial properties, work unit and affordable housing), compared with those residing in informal neighbourhoods (e.g. urban villages). Moreover, migrants living in commercial property neighbourhoods are best integrated economically, and those living in affordable and work unit housing neighbourhoods are more socio-culturally integrated than others. Inner city old housing neighbourhoods are positively associated with socio-cultural integration, compared with urban villages, while both neighbourhood types are negatively correlated with identity integration. Our findings remain robust after controlling for potential endogeneity bias caused by migrants’ self-selection into different neighbourhood types. We conclude that neighbourhood types do matter for migrants’ socio-economic integration but there is large heterogeneity of the correlations across different migrant groups.

Appendices

Appendix 1: The PSM methods used in the study

Scholars define the propensity score as the conditional probability of a given set of observation covariates assigned to the treatment group (Rosenbaum and Rubin 1983).

$${\text{p}}\left( {\text{X}} \right) = \Pr \left[ {B = 1|X} \right] = E\left[ {B|X} \right]$$

(2)

In formula (2), X is the multidimensional eigenvector of the control group, and B is the core dependent variable, which equals 1 if migrants live in formal neighbourhoods and 0 otherwise. If we can get a propensity score, we can estimate the ATT by comparing the potential differences between the experimental and control groups as follows (Becker and Ichino 2002),

$$\begin{aligned} {\text{ATT}} & = {\text{E}}\left[ {Y_{1i} - Y_{0i} |B_{i} = 1} \right] \\ & = {\text{E}}\left\{ {E\left[ {Y_{1i} - Y_{0i} |B_{i} = 1,p\left( {X_{i} } \right)} \right]} \right\} \\ & = {\text{E}}\left\{ {E\left[ {Y_{1i} |B_{i} = 1,p\left( {X_{i} } \right)} \right] - E\left[ {Y_{0i} |B_{i} = 0,p\left( {X_{i} } \right)} \right]|B_{i} = 1} \right\} \\ \end{aligned}$$

(3)

Here Y_1i and Y_0i respectively indicate the potential results of the experimental group and the control group. Following previous studies (Becker and Ichino 2002; Dehejia and Wahba 2002), we use the following Logit model to get the propensity score.

$${\text{p}}\left({{\text{X}}_{\text{i}}}\right)={ \Pr }\left( {{\text{B}}_{\text{i}}=1|{\text{X}}_{\text{i}}}\right)=\frac{{{ \exp }\left({{\upalpha X}_{\text{i}}}\right)}}{{1+{ \exp }\left({{\upalpha X}_{\text{i}}}\right)}}$$

(4)

Here, X is the same as above. It is assumed to affect migrants’ tendency of living in different neighbourhoods. α is a coefficient vector. Then, we use three matching methods to estimate the ATT, i.e. Nearest neighbor matching, Radius matching and Kernel matching. The Nearest neighbor matching method is the most commonly used matching method. It finds the individual with the smallest difference of PS value in the control group with the experimental group as its comparison object, which can be represented as follows,

$$C\left( i \right) = \mathop {\hbox{min} }\limits_{j} p_{i} - p_{j}$$

(5)

However, the Radius matching method is to set the radius beforehand, and to find all the control samples in the unit circle within the set radius range. The radius is positive. As the radius decreases, the matching requirements become more stringent. The formula is shown as follows,

$$C\left( i \right) = \left\{ {p_{j} |p_{i} - p_{j} < r} \right\}$$

(6)

In the above two formulas, C refers to the control group. We let T represent the experimental group, ${\text{Y}}_{\text{i}}^{\text{T}}$ and ${\text{Y}}_{\text{j}}^{\text{C}}$ respectively represent the observation results of the experimental group and the control group. Meanwhile, let the control group’s number matched with observation i ∈ T by ${\text{N}}_{\text{i}}^{\text{C}}$ and define the weights ${\text{w}}_{\text{ij}} = \frac{1}{{{\text{N}}_{\text{i}}^{\text{C}} }}$ if j ∈ C(i), otherwise, w_ij = 0. In addition, suppose that the experimental group has N^T observations. We can estimate the ATT following (Becker and Ichino 2002).

$$\tau^{M} = \frac{1}{{N^{T} }}\mathop \sum \limits_{i \in T} Y_{i}^{T} - \frac{1}{{N^{T} }}\mathop \sum \limits_{j \in C} w_{j} Y_{j}^{C}$$

(7)

Then, we use M as the matching method, ${\text{w}}_{\rm j} = \mathop \sum \limits_{\rm i} {\text{w}}_{\rm ij}$. Assuming that the weights remain unchanged, the validity of the neighborhood is independent. We can estimated the variance of τ^M as the following,

$$Var(\tau^{M} ) = \frac{1}{{N^{T} }}Var\left( {Y_{i}^{T} } \right) + \frac{1}{{\left( {N^{T} } \right)^{2} }}\mathop \sum \limits_{j \in C} \left( {w_{j} } \right)^{2} Var\left( {Y_{i}^{C} } \right)$$

(8)

The third method (the Kernel matching method) is slightly different from the first two. It is to construct a virtual object to match the treatment group. The principle is to average the weight of the existing control variables, and the value of the weight is inversely related to the PS value difference between the treatment group and the control group. The ATT of this method is estimated as following:

$$\tau^{K} = \frac{1}{{N^{T} }}\mathop \sum \limits_{i \in T} \left\{ {Y_{i}^{T} - \frac{{\mathop \sum \nolimits_{j \in C} Y_{j}^{C} G\left( {\left( {p_{j} - p_{i} } \right)/h_{n} } \right)}}{{\mathop \sum \nolimits_{k \in C} G\left( {\left( {p_{k} - p_{i} } \right)/h_{n} } \right)}}} \right\}$$

(9)

In formula (9), G(·) means the Gaussian kernel function, and bandwidth parameter is h_n.

$$\frac{{\mathop \sum \nolimits_{j \in C} Y_{j}^{C} G\left( {\left( {p_{j} - p_{i} } \right)/h_{n} } \right)}}{{\mathop \sum \nolimits_{k \in C} G\left( {\left( {p_{k} - p_{i} } \right)/h_{n} } \right)}}$$

It is a consistent estimate of Y_0i with counterfactual results.

Appendix 2

See Tables 6, 7, 8, 9 and Fig. 3.

Table 6 Indicators of migrants’ integration in urban society

Table 7 Result of rotated component matrix in integration of migrants

Table 8 The logit model of living in a formal neighbourhood

Table 9 Test results of balanced hypothesis