Economic Modelling 102 (2021) 105557

Contents lists available at ScienceDirect

Economic Modelling
journal homepage: www.journals.elsevier.com/economic-modelling

Wages, labor quality, and FDI inflows: A new non-linear approach☆

Lei Hou a , Qi Li b,c , Yanfei Wang d , Xintong Yang c, ∗
a Institute of World Economics and Politics, Chinese Academy of Social Sciences, Beijing, 100732, China
b
Department of Economics, Texas A&M University, College Station, TX, 77840, United States
c
International School of Economics and Management, Capital University of Economics and Business, Beijing, 100070, China
d Business School, Renmin University of China, Beijing, 100872, China

A R T I C L E I N F O A B S T R A C T

JEL classification: Although theory predicts that locations with lower wages will attract more FDI inflows, the empirical results
F21 from traditional linear regression models are rather mixed. We consider the possible impact of labor quality,
J31 along with wages, and build a simple FDI location choice model to capture the potential nonlinear relationship
among the three. Further, we propose a partially linear panel data model to investigate the wage effect on FDI
Keywords: location choices. Using macroeconomic province-level data from China between 1993 and 2018, we find that
FDI the marginal effect of wage is generally a decreasing function of labor quality. This implies that facing low
Wage labor quality, FDI firms prefer locations with high wages that ensure high labor quality; while facing high labor
Labor quality
quality, their preferences for low labor costs dominate. Our nonlinear approach reconciles the mixed results in
Partially linear regression
the empirical literature on how wages affect FDI inflows and the role of labor quality in attracting FDI.

1. Introduction indicator of local labor quality. Local labor quality not only determines
the equilibrium wage, it also influences output capacity. We show that
Traditional FDI theory proposes that firms expand production with some regularity assumptions, there exists an optimal labor quality
abroad based on two motives: to pursue low labor costs and to access level such that when the labor quality falls below this level, increasing
high market demand, that is, vertical and horizontal FDI, respectively wages in a province will increase the FDI probability in that province.
(Yeaple, 2003; Zhang and Markusen, 1999). A positive effect of market When the labor quality is higher than this level, increasing wages in the
demand on FDI inflows has been well-established in the empirical stud- province will decrease the FDI probability. The intuition is that when
ies. However, the evidence on low labor costs is ambiguous.1 The effect labor quality is low, the wage can be regarded as an effective indicator
of “wage”, a widely used measurement of labor costs in general, could of labor quality; therefore high wages attract FDI inflows. When labor
be either significantly negative or positive, and even insignificant.2 One quality is high enough, the labor cost effect dominates, and a high wage
explanation for the positive marginal effect of wage on inward FDI is will impede FDI inflows. Given the roles of wage as labor cost and
that a high wage indicates high labor quality (Wei et al., 1999; UNC- quality, the average marginal effect of wage on FDI inflows may be
TAD, 1999). Because FDI firms usually enjoy a technology advantage, ambiguous.
they may prefer a location with high-quality labor ready to adapt their We further formalize this notion by estimating the effect of wage
technology. This research is mainly concerned with how to evaluate the and labor quality on FDI inflows. In particular, we focus on the 26-year
wage effect on FDI inflows while reconciling the role of labor quality. period from 1993 to 2018 across 29 Chinese provinces. The develop-
We first develop a simple model of FDI location choice in which ment experience of China from the early 1990s to the 2010s provides a
wage is not only a measurement of local labor cost but also an important good opportunity for examining this issue. China mainly attracted ver-

☆
The authors’ names are listed in alphabetical order.
∗
Corresponding author.
E-mail address: [email protected] (X. Yang).
1
Nielson et al. (2017) reviewed 83 studies on FDI location choices that include wage variables. They find a negative and positive relationship between wages and
FDI location choices, accounting for 49% and 17% respectively. Meanwhile, 34% of these studies find no support for a negative association as theory predicts.
2
See Noorbakhsh and Paloni (2001), Head and Mayer (2004), among others. See Liu et al. (2010) for a review of wage and firm location choices.

https://doi.org/10.1016/j.econmod.2021.105557
Received 19 December 2019; Received in revised form 15 April 2021; Accepted 21 May 2021
Available online 1 June 2021
0264-9993/© 2021 Elsevier B.V. All rights reserved.
L. Hou et al. Economic Modelling 102 (2021) 105557

tical FDI, especially around the 1990s, because of its large labor force, of local labor quality on FDI as Gao (2005) and Iwai and Thompson
relatively low labor costs, and local governments’ favorable policies (2012).4
toward investment. During that period, many foreign firms in devel- Compared to the existing literature, this paper makes two contribu-
oped countries either shifted their domestic and overseas production tions. First, it uses a nonlinear function to capture the potential influ-
factories to China or established new factories in China, then exported ence of labor quality in estimating the effect of wages on FDI. In prior
their manufactures back to the developed countries. This promoted work, labor quality entered FDI location choices merely as a linearly
China’s processing trade development and explains the simultaneous additive explanatory variable, as in Gao (2005). In our study, apart
high rate of growth in its import and export levels. At the same time, from its direct impact on FDI, labor quality has an effect through wages,
China witnessed a large increase in labor quality due to the enhanced thus indirectly affecting FDI location choices. Our partially linear esti-
education level of its labor force, which intensifies China’s advantage mation is able to identify the nonlinear wage effect on FDI concealed
as an ideal FDI destination. However, China’s labor cost also rose dra- by linear regressions. This nonlinearity may explain the inconclusive
matically. From 2000 to 2012, China’s average real wage quadrupled. results on the wage effect on FDI inflows in existing empirical stud-
As a consequence, in recent years many multinational corporations ies. Second, we introduce a novel nonlinear methodology. Generally,
chose to shift their production lines to other Asian and Latin American a traditional parametric method adopted to estimate the average wage
developing countries with lower labor costs, or simply to close their effect is more suitable when wages affect FDI inflows homogeneously
Chinese factories. Therefore, both labor costs and labor quality may across the entire FDI distribution. Our study shows that when dealing
have played a role in determining China’s FDI inflows over these two with large differentiation across regions and over time, as in China, the
decades. average wage effect estimation may mask variations in the wage effect
We first estimate a linear panel data model with province fixed at the aggregate level. Thus, the nonlinear estimation has an advantage
effects, but this fails to capture either the effect of wages or the effect of in making clear the nonlinear wage effects in the data. From this per-
labor quality on FDI inflows. From a geographic perspective, there are spective, the nonlinear approach that we propose is useful when data
huge regional disparities between coastal and inland areas in China, in have large differentiation across space and over time, which may have
economic development levels, factor endowments, preferential policies, more general applications in economic studies.
and the ability to attract FDI.3 Thus, the effect of local wages and labor The rest of this paper is organized as follows. Section 2 presents
quality on FDI inflows may differ across regions. With our linear panel a simple theory to show that the wage effect on FDI location choices
data model, we therefore perform subsample analyses across three dif- depends on local labor quality. Section 3 describes the data and vari-
ferent time periods and three geographic regions in China. These results able constructions. Section 4 presents the estimation results of the lin-
suggest that the effects of wages and labor quality on FDI inflows are ear panel data model with province fixed effects. We demonstrate the
heterogeneous, both temporally and spatially. Thus, a linear regression existence of a nonlinear relationship among FDI inflows, wages and
model may not be suitable for estimating the effect of wages on FDI labor quality. Section 5 presents our proposed partially linear panel
location choices. data model, taking the nonlinear relationship into account, and pro-
In order to capture the nonlinear effects alluded above, we propose vides our empirical results. Section 6 concludes. A numerical example
a partially linear specification to model the interdependent effect of that supports the theoretical model, and details about the econometric
wages and labor quality on FDI. We find that, in general, the estimated method, are relegated to the Appendices.
marginal effect of wages is a decreasing function of labor quality. When
labor quality is relatively low, wages have a positive effect on FDI. This 2. Theoretical model
implies that FDI firms prefer locations with high wages as they pur-
sue high labor quality. As labor quality increases, the negative cost The theory builds upon Liu et al. (2010). A foreign firm chooses to
effect of the high wage on FDI kicks in. When labor quality exceeds locate where its profits are maximized, and it seeks to invest somewhere
a threshold value, the wage effect on FDI becomes significantly nega- in China. Its production technology uses local labor inputs, a vector of
tive; this negative effect is intensified as labor quality increases. More- goods and services.5 Note that labor input is a function of labor quality,
over, we find that the marginal effect of labor quality on FDI inflows i.e., L(z). Thus, by choosing labor quality in a province, the foreign firm
depends on the wage level. When the wage level is relatively low, labor indirectly chooses the labor force in that province. Controllable profits
quality will significantly attract more FDI. When the wage increases to for the firm if it invests in province j can be written as:
a certain level, high labor quality will attract less FDI or even deter ( ( ) ) ( ) ( )
𝜋j = Q L zj , xj , zj − wj zj L zj − pxj xj , (2.1)
FDI inflows.
Recent empirical studies have investigated various novel location- where L(zj ) denotes the local labor inputs in province j, xj denotes the
specific attributes, other than the market size and labor costs, in affect- input of goods and services in province j, zj denotes the local labor qual-
ing FDI, including logistic infrastructure, economic institutions, immi- ity in province j, Q(·) denotes the production technology, wj (zj ) denotes
gration, historical conflicts, local labor market flexibility, etc. (Blyde
and Molina, 2015; Ascani et al., 2016; Tomohara, 2017; Gao et al., 4
Liu et al. (2010) points out that unobserved location-specific attributes play
2018; Rong et al., 2020). Given the indefinite role of labor costs in
a role in estimating the wage effect on FDI location choices. Our paper confirms
attracting FDI in existing empirical studies, our paper revisits the ques-
their conjecture by introducing labor quality into the estimation of wage effects
tions of how labor costs affect FDI and what role the local labor quality on FDI. Gao (2005) investigates the impact of labor quality on FDI inflows in
may play. From this perspective, our paper is closely related to the stud- China, and confirms labor quality as an established determinant in attracting
ies estimating the wage effect on FDI as Liu et al. (2010), and the effect FDI. Our paper differs from Gao (2005) by using a partially linear model esti-
mation. We find an indirect effect of labor quality on FDI through wages, in
addition to a direct effect of labor quality on FDI through a linearly additive
form as in Gao (2005). In addition, our results support the hypothesis proposed
by Iwai and Thompson (2012) that there exists a take-off point in labor quality
for low labor cost developing countries to turn from potential FDI destination
candidates into real ones. Using the Chinese data, we empirically identify such
a turning point in labor quality such that above this point, a province’s low
labor cost is an advantage in attracting FDI; below this point, the province’s
low labor cost may deter instead of attracting FDI.
3 5
For regional disparity in attracting FDI, see Gao (2005) and Amiti and The input vector of goods and services might contain imported goods, but
Javorcik (2008), among others. this does not matter for our main focus.

2
L. Hou et al. Economic Modelling 102 (2021) 105557

the labor wage level in province j, pxj denotes the price level of goods
and service input in province j.
Taking F.O.C. of 𝜋 j in (2.1) with respect to labor inputs L(zj ), we
find that the optimal labor input and wage level need to satisfy:

wj (zj ) = QL (L(zj ), xj , zj ). (2.2)

Further, based on equation (2.2), we have:

w′j (zj ) = QLL (L(zj ), xj , zj )L′ (zj ) + QLz (L(zj ), xj , zj )

Taking the derivative of 𝜋 j in (2.1) with respect to labor quality zj , we

have:
𝜕𝜋j
= QL (L(zj ), xj , zj ) · L′ (zj ) + Qz (L(zj ), xj , zj ) − wj (zj )L′ (zj ) − w′j (zj )L(zj )
𝜕 zj

= (QL (L(zj ), xj , zj ) − wj (zj )) · L′ (zj ) + (Qz (L(zj ), xj , zj ))(−w′j (zj )L(zj ))

= Qz (L(zj ), xj , zj ) − w′j (zj )L(zj )

Fig. 1. Illustration of how wage effect on FDI choice depends on labor quality.
where the third line is derived from equation (2.2). Hence, there exists
𝜕𝜋
a level of labor quality zj such that 𝜕z j = 0:
j distribution with p.d.f. f (𝜖) = exp (−𝜖 − exp(−𝜖)) and c.d.f. F (𝜖) =
exp (− exp(−𝜖)), we have:
wj (zj )L(zj ) = Qz (L(zj ), xj , zj ).
′
exp(𝜋k )
Lemma 1. Assume that Qzz (L(z), x, z) < QLL (L(z), x, z) · Probk = ∑ . (2.3)
exp(𝜋j )
j ∈J
(L′ (z))2 + w″ (z)L(z) and QLL (L(z), x, z)L′ (z) + QLz (L(z), x, z) > 0.
𝜕𝜋 Equation (2.3) shows that the probability that province k is chosen pos-
When zj < zj , 𝜕wj > 0, i.e., increasing wage level in province j will increase
j itively depends on controllable profits 𝜋 k . Then, we reach our main
𝜕𝜋
profits in province j. When zj > zj , 𝜕wj < 0, i.e., increasing wage level in proposition:
j
province j will decrease profits in province j. Proposition 1. When zk < zk , 𝜕Prob 𝜕 wk
k > 0, i.e., increasing wage level in

Proof. Whenever zj = zj , 𝜕𝜕𝜋z = 0. To see it achieves the local maximum, 𝜕 Prob

province k will increase the FDI probability in k. When zk > zk , 𝜕w k < 0,
j k
we take the second derivative of 𝜋 j with respect to zj , and have: i.e., increasing wage level in province k will decrease FDI probability in k.

𝜕 2 𝜋j We express the main idea of Proposition 1 above using the follow-

ing Fig. 1. The intuition is that when labor quality is low, wage can
𝜕 z2j
be regarded as an effective indicator of labor quality; therefore, high
= QzL (L(zj ), xj , zj ) · L′ (zj ) + Qzz (L(zj ), xj , zj ) − w′ j (zj )L′ (zj ) − w″j (zj )L(zj ) wages attract FDI inflows. When labor quality is high and exceeds a
threshold value, the labor cost effect of wages dominates, and the high
= Qzz (L(zj ), xj , zj ) − QLL (L(zj ), xj , zj ) · (L′ (zj ))2 − w″j (zj )L(zj ) wage impedes FDI inflow.

Given Qzz (L(z), x, z) < QLL (L(z), x, z) · (L′ (z))2 + w″ (z)L(z), we have 3. Data and variables
𝜕 2 𝜋j
𝜕 z2j
< 0. This implies that when zj = zj , firm’s profit achieves a max-
The theoretical model in Section 2 considers all FDI firms as a
𝜕𝜋 𝜕𝜋
imum, i.e., when zj < zj , 𝜕z j > 0, and when zj > zj , 𝜕z j < 0. Given representative firm that makes the location choice about which Chi-
j j
nese province to invest in.6 The model’s main proposition predicts
QLL (L(z), x, z)L′ (z) + QLz (L(z), x, z) > 0, we have w′ (zj ) > 0. Hence,
𝜕𝜋 𝜕𝜋 that how the wage level in province k affects the probability of the
we have when zj < zj , 𝜕wj > 0, and when zj > zj , 𝜕wj < 0. Q.E.D. representative FDI firm investing there depends on the labor quality
j j
level in province k. This wage effect is a decreasing function of local
We use production function Q(L(z), x, z) and do not specify its func- labor quality in province k. Under the representative-firm model, the
tional form in order to make our theory more general and more appli- more likely a province is chosen as the FDI destination, the more FDI
cable. To be more convincing, we provide a numerical example in inflows into this province. To investigate the role of wage and labor
the Appendix to show that assumptions Qzz (L(z), x, z) < QLL (L(z), x, z) · quality in determining FDI location choices, i.e., FDI flows into differ-
(L′ (z))2 + w″ (z)L(z) and QLL (L(z), x, z)L′ (z) + QLz (L(z), x, z) > 0 hold ent provinces, we use a panel dataset covering 29 Chinese provinces,
and how to apply our general theory.
Denote the real profits of the firm in province j as Πj = 𝜋 j − ej ,
where ej is an idiosyncratic cost shock in province j. Province k is chosen 6
Since FDI firms in the theoretical model are homogeneous, the aggregation
if the firm’s real profit from investing there is maximized, i.e., of homogeneous firms is equivalent to the case where the decision of a single
firm is multiplied by the number of firms. Hence, the key results of the theo-
Probk = Pr {max(Π1 , … , ΠJ ) = Πk } retical model with a representative firm will keep unchanged after aggregation.
Moreover, it is important to notice that the role of the theoretical model is to
= Pr {Πk ≥ Π1 , Πk ≥ Π2 , … , Πk ≥ ΠJ } inspire our empirical analysis with a nonlinear estimation approach, but not a
= Pr {𝜋k + 𝜖k ≥ 𝜋1 + 𝜖1 , 𝜋k + 𝜖k ≥ 𝜋2 + 𝜖2 , … , 𝜋k + 𝜖k ≥ 𝜋J + 𝜖J } direct test of the model, since the real data comes from the complex world that
may not follow the data generating process as our theoretical model suggests.
= Pr {𝜖1 ≤ 𝜋k + 𝜖k − 𝜋1 , 𝜖2 ≤ 𝜋k + 𝜖k − 𝜋2 , … , 𝜖J ≤ 𝜋k + 𝜖k − 𝜋J } Nevertheless, we expect that “a useful model” can provide some valid explana-
tions to the observed data. Hence, we adopt a simple theoretical setting with
Let J denotes the set of all provinces in China. If {𝜖j }j∈J are indepen- homogeneous firms, which does not only serve to explain the observed data to
dently and identically distributed according to a Type 1 Extreme Value certain extents, but also enables us to focus on the novelty in empirical work.

3
L. Hou et al. Economic Modelling 102 (2021) 105557

autonomous regions and municipalities (henceforth “provinces”) over a data availability, the time span of this variable is from 2002 to 2018.
26-year period from 1993 to 2018.7 ,8 Our sample starts in 1993 because In addition to these explanatory variables, we include two time-dummy
FDI inflows expanded rapidly after Deng Xiaoping’s visit to southern variables to control for the impacts of China’s accession to the WTO
China in 1992. The data we use are taken from the regional database of in 2001 and the Global Financial Crisis in 2008. The detailed variable
the National Bureau of Statistics of China. definitions and descriptive statistics are shown in Table 1.
In our estimation, dependent variable is the FDI inflows into vari- Figs. 2–4 show the time-series plots of log FDI inflow, log wage, and
ous provinces each year, measured by the logarithm of FDI firms’ total labor quality for every province separately. The log FDI inflow (lfdi) and
investment. We include wage, labor quality, GDP, degree of openness, log wage (lwage) for each province exhibit a clear upward trend. More-
government expenditure, infrastructure, industry compositions, and the over, the proportions of the number of people who newly graduate from
population structure as the main explanatory variables in our regres- high school to the total population (lbr_qlty) for most of the provinces
sions. In empirical FDI studies, wage and GDP are well-documented are about 2‰ before year 2000, sharply increasing to a peak value
variables used to control for labor costs and local market demand.9 around 6‰–8‰ in year 2010, and then decreasing slightly. Because
Labor quality, calculated as the proportion of the number of people who most of the high school graduates between years 2000 and 2010 were
newly graduate from high school to the total population, is included to born in the 1980s, this sharp upward trend in the measure of labor qual-
control for not only the local human capital level, but also the local ity matches the occurrence of the third baby boom in China. The mild
availability of human capital.10 Degree of openness and government decline in the measure of labor quality may be caused by the aging pop-
expenditure, two distinctive variables for China, are also included to ulation.,12, 13 In addition, the percentage of total exports and imports in
capture their role in attracting FDI.11 To capture different levels of GDP (open) is less than 50%, and even close to zero, for most of the
infrastructure, we use the average kilometers of roads (including both provinces except these seven: Beijing, Tianjin, Shanghai, Jiangsu, Zhe-
highway and railway) per squared kilometer. To control for industry jiang, Fujian, Guangdong. The economy of these seven provinces relies
compositions across provinces, we use three variables: the percentage heavily on international trade. For all seven, the percentages increase
of primary sector in GDP, that of secondary sector in GDP, and the ratio to a different extent after China’s entrance into the World Trade Orga-
of GDP in tertiary sector to that in others. To control for the aging of nization in 2001.
population that could be a potential confounding factor on labor qual-
ity, we include as a control variable the ratio of population above 65 4. Estimation of traditional linear panel data models
to that between 15 and 64 in our regression. Due to the limitation of
In this section, we present and discuss the estimation results of lin-
ear panel data models with province fixed effects. Then, we explore
7 the heterogeneous effect of wage on FDI inflows over time and across
The empirical application with firm-level data is appropriate for a
heterogeneous-firm model as in Liu et al. (2010). In comparison, our the-
regions. We aim to demonstrate the existence of a possible nonlinear
ory does not highlight the heterogeneous behaviors of firms, but rather the relationship among FDI inflows, wages, and labor quality and thus a
role of wage and labor quality in determining the firms’ FDI flows into differ- potential misspecification of using a linear regression model.
ent provinces from the macro persepctive. Hence, a representative-firm model
serves our focus well, and the provincial-level data is sufficient for our applica- 4.1. Estimation results of linear panel data models
tion.
8
The 29 provinces included in our sample are Anhui, Beijing, Fujian, Gansu,
Given the definitions of variables in Section 3, we first consider a
Guangdong, Guangxi, Guizhou, Hainan, Hebei, Heilongjiang, Henan, Hubei,
traditional linear panel data model with province fixed effects and a
Hunan, Jiangsu, Jiangxi, Jilin, Liaoning, Inner Mongolia, Ningxia, Qinghai,
linear time trend:
Shandong, Shanghai, Shanxi, Shaanxi, Sichuan, Tianjin, Xinjiang, Yunnan, and
Zhejiang, excluding Chongqing and Tibet. lfdiit =𝛽1 lwageit −1 + 𝛽2 lbr_qltyit −1 + 𝛽3 lwageit −1 × lbr_qltyit −1
9
Note that in our study we use the average wage of urban employees. Urban
wage is widely adopted for studying FDI-related topics (See Liu, Lovely and + w′it−1 𝜸 + 𝛿1 aftr_WTOt + 𝛿2 aftr_GFCt + 𝜇i + 𝜆t + 𝜖it . (4.1)
Ondrich (2010), Gao (2005) and Du et al. (2008) for example). This is because
on the one hand, foreign investments are located mainly in the urban areas; on The dependent variable lfdiit is log FDI inflow of province i at year t. It
the other hand, people are mostly self-employed in the rural areas, and there is is important to note that FDI inflows at year t can effect the explana-
a lack of definite wage statistics for rural areas. tory variables contemporaneously. For example, FDI inflows in year t,
10
We use a flow-variable measure instead of a stock-variable measure (the deemed as a part of provincial GDP in year t, could stimulate labor
share of people in population with a high school degree, for example). Both demand and affect other explanatory variables in the same year. To
types of variables were used to measure labor quality in the existing literature. alleviate this simultaneity problem, our main explanatory variables are
For example, Du et al. (2008) used the flow-variable measure, the proportion
one-period lagged. The variables of interest, lwageit −1 and lbr_qltyit −1 ,
of the number of students who are enrolled in higher education institutions
are log wage and labor quality respectively. The effects of wage and
to its total population, whereas Gao (2005) used the stock-variable measure,
the proportion of the population with at least high school education. Given
labor quality on FDI may be intertwined, and we include a term inter-
the relatively long sample period in our paper, we choose the flow-variable acting wage with labor quality (lwageit −1 × lbr_qltyit −1 ) to capture such
measure, the number of people who newly graduate from high school every an interactive effect. The control variables in the vector wit −1 include
year, because it is much longer than the stock-variable measure from our data
source. Furthermore, we choose the proportion of high school graduates, not
12
the proportion of college graduates and above, as the measure of local labor With our measure of labor quality, we consider various dimensions of
quality because it better reflects the local education level. If we considered human capital changes, not only the changes in the average skill level, but also
the share of college graduates and above, we might risk underestimating labor the changes in the availability of the labor force that apply their skills. Hence,
quality in the provinces with fewer colleges and universities. the decline in our measure of labor quality is caused by the latter dimension of
11 Note that we use degree of openness as a policy variable to control for
human capital changes, which is the aging population and hence the decline in
the policy effects. Some existing literature – see Kang and Lee (2007), Du et the availability of labor force. We include an additional variable that measures
al. (2008), Liu et al. (2010) among others – uses a dummy variable to indicate the population structure to separate the population aging effect from our labor
whether this province has Special Economic Zones (SEZs) to control for a policy quality variable.
13
effect. However, in our sample period, this dummy variable does not have time Note that the decline in our flow-variable measure of labor quality does
variations. Therefore, to consider the policy effect on FDI inflows, we use degree not necessarily imply the decline in the stock-variable measure of labor quality,
of openness, which is a highly relevant variable for the preferential policies and but rather indicates the slower increase in the stock-variable measure of labor
has enough variations across provinces and over time. quality.

4
L. Hou et al. Economic Modelling 102 (2021) 105557

Table 1
Variable definitions and descriptive statistics.
Variable Definition Obs. Mean SD. Min. Max.
fdi The total investment of FDI firms (Million RMB Yuan in 2005) 754 462560.9 844306.3 694.0 9271464.0
wage The average wage of urban employee (RMB Yuan in 2005) 754 23838.2 18029.7 3769.3 110831.8
lbr_qlty The annual high school graduates over population (Unit: ‰) 754 4.395 1.951 1.093 8.666
gdp GDP (100 Million RMB Yuan in 2005) 754 9341.9 10995.4 177.5 72801.3
open The total export and import over GDP (Unit: 10%) 754 3.000 3.792 0.168 22.029
govern Local government fiscal expenditure over GDP 754 0.176 0.092 0.049 0.627
infra The density of road and railway (Unit: km∕km2 ) 754 0.626 0.486 0.018 2.379
pct_primary The percentage of primary sector in GDP (Unit: %) 754 14.547 8.161 0.291 36.445
pct_secondary The percentage of secondary sector in GDP (Unit: %) 754 44.413 7.992 16.545 59.397
rt_tertiary The ratio of GDP in tertiary sector to that in others 754 0.757 0.454 0.381 4.914
rt_ab65 The ratio of population above 65 to that between 15 and 64 (Unit: %) 493 12.765 2.673 6.951 22.689
aftr_WTO Indicator of China’s accession to WTO after 2001 754 0.654 0.476 0.000 1.000
aftr_GFC Indicator of Global Financial Crisis after 2008 754 0.423 0.494 0.000 1.000

All the variables in nominal term are deflated with local Consumer Price Index taking 2005 as the base year. The number of provinces is
n = 29, and the number of years is T = 26. The time span of the sample is from 1993 to 2018. For the variable rt_ab65, the time span is from
2002 to 2018. The after WTO dummy is defined as aftr_WTOt = 1 for t ≥ 2002, and it equals to zero otherwise. The after GFC dummy is
defined as aftr_GFCt = 1 for t ≥ 2009, and it equals to zero otherwise.

Fig. 2. Log FDI (lfdi) series of 29 provinces in China (1993–2018).

log GDP, degree of openness, government expenditure, infrastructure, In equation (4.1), the fixed effects 𝜇 i capture the impacts of
the percentage of primary sector in provincial GDP, that of secondary province-specific time invariant factors on FDI inflows, and 𝜆t is the lin-
sector in provincial GDP, the ratio of GDP in tertiary sector to that in ear time trend that captures changes in the domestic and global macroe-
others, and the ratio of population above 65. conomic environment. In the model, we use a linear time trend instead
of year fixed effects for two reasons. First, for our panel dataset, the time

5
L. Hou et al. Economic Modelling 102 (2021) 105557

Fig. 3. Log wage (lwage) series of 29 provinces in China (1993–2018).

span of 26 years is quite long considering there are only 29 provinces. term 𝜀it .15
Including both province and year fixed effects would consume too many According to the classical theory, FDI prefers locations with lower
degrees of freedom from our model and result in the invalidity of statis- wage costs, higher labor quality, larger market demand, a greater
tical inference. Second, by visual inspection of the time-series plots, we degree of openness, and better availability of infrastructure. In China,
find that most of the variables exhibit similar linear upward trends. the government may supply infrastructure, finance, or preferential poli-
Based on these reasons, and the fact that unobserved heterogeneity cies beneficial to attracting FDI, or it may deter FDI due to strin-
across provinces is much more pronounced than over time, we make gent administration and government-expenditure-induced misalloca-
the next best choice to include province fixed effects and use the linear tions. Thus, the coefficient of government expenditure could be either
time trend to capture changes in the macroeconomic environment. To positive or negative. Considering the fact that FDI firms tend to invest
remedy the insufficiency of the linear time trend in capturing important in the manufacturing and service sectors in China, we would expect, ex
events — China’s accession to the WTO in 2001 and the Global Finan- ante, that provinces with a large primary sector are at a disadvantage
cial Crisis (GFC) in 2008 — we include two time dummy variables, the for attracting FDI inflows, and therefore that the coefficient associated
after-WTO dummy (aftr_WTOt ) and the after-GFC dummy (aftr_GFCt ), with the predictor percentage of primary sector in GDP (pct_primary)
as additional regressors.14 would be negative while the coefficients associated with the predic-
Moreover, we assume that all of the explanatory variables are only tors percentage of secondary sector in GDP (pct_secondary) and ratio of
contemporaneously exogenous. The error term 𝜀it is assumed to have an GDP in tertiary sector to that in others (rt_tertiary) would be positive.
unknown serial correlation within individuals and an unknown cross- We would also expect, ex ante, that provinces where the population is
sectional correlation within time. Correspondingly, we use the within- of advanced age would see their workforce shrink and thereby be less
group method to estimate the parameters in the model. For statistical attractive for FDI, hence the coefficient on the predictor ratio of popu-
inference, we use the two-way cluster-robust standard errors that are lation above 65 (rt_ab65) would be negative. In addition, China’s entry
robust to both temporal and cross-sectional dependence of the error into the WTO implies more openness to FDI; thus the sign of the time
dummy for joining the WTO is positive. Moreover, the Global Finan-
cial Crisis in 2008 deterred global FDI flows; thus the sign of the time
14
The after-WTO dummy is defined as aftr_WTOt = 1 for t ≥ 2002, and it
equals to zero otherwise. The after-GFC dummy is defined as aftr_GFCt = 1 for
t ≥ 2009, and it equals to zero otherwise. 15
It is proposed by Cameron et al. (2011) and Thompson (2011).

6
L. Hou et al. Economic Modelling 102 (2021) 105557

Fig. 4. Labor quality (lbr_qlty) series of 29 provinces in China (1993–2018).

dummy for post-crisis is negative. The expected signs of all variables the ratio of population above 65 in Model 4 is significantly positive.
are summarized in the second column of Table 2. A possible explanation of this result may be the reverse causality from
Table 2 reports the estimation and related hypothesis testing results FDI to the aging of population: Locations with more FDI are usually
of the three linear fixed-effects panel data models. In Model 1, we more developed regions with higher living costs, which restrains fertil-
only include the two variables of interest, log wage (lwage) and labor ity and therefore are more prone to population aging. Because we only
quality (lbr_qlty), as the explanatory variables in the regression. After use the ratio of population above 65 as a control variable to account for
controlling for province fixed effects and linear time trend, the coef- the potentially omitted factor of population aging, the estimated coeffi-
ficients of both variables are positive but statistically insignificant. cient of the ratio of population above 65 may not reflect its causal effect
After we include the interactive term between wage and labor qual- on FDI.
ity (lwage × lbr_qlty) in Model 2, none of the coefficients of log wage,
labor quality, and their interaction are significant. Moreover, the joint
effects of log wage and labor quality on log FDI are all statistically 4.2. Heterogeneity over time and across regions
insignificant, because the three Wald tests fail to reject the nulls that
the overall and joint effects of log wage and labor quality are zero. One possible explanation for these insignificance results using the
Because the data for the ratio of population above 65 (rt_ab65) is only whole sample is that the effects of log wage and labor quality are
available between 2002 and 2018, we consider two models with control very heterogeneous across time and/or provinces. So, in addition to
variables. Model 3 includes all control variables but the ratio of popu- using the whole sample, we investigate the effects of log wage and
lation above 65 and has a larger sample from 1993 to 2018. Among the labor quality on FDI using subsamples, defined by different time peri-
three variables of interest, only the coefficients of labor quality and the ods or geographic locations. First we create three subsamples for three
interaction regressor are marginally significant; the three Wald tests time periods: from 1993 to 1997; from 1998 to 2008; and from 2009
still fail to reject their nulls respectively. In Model 4, we include the to 2018. The average labor quality values of the 29 provinces over
ratio of population above 65 to the regression with all other controls the 26 years, as shown in Fig. 5, exhibit drastically different patterns.
and the sample period is from 2002 to 2008. The empirical results of During 1993 to 1997, the labor quality values are low, around 2‰
Model 4 are similar to those of Model 3: all the Wald tests in the two for all provinces in China. In the next 11 years, China experienced
models fail to reject the nulls. Note that the estimated coefficient of rapid growth in labor quality, rising from 2‰ to its peak value of
6.4‰ in the year 2008. From 2009 on, average labor quality values

7
L. Hou et al. Economic Modelling 102 (2021) 105557

Table 2
Estimation results of fixed effects models.
Expected sign Fixed effects
Model 1 Model 2 Model 3 Model 4
log wage (lwage) – 0.210 0.132 −0.298 −0.317
(0.306) (0.324) (0.348) (0.451)
labor quality (lbr_qlty) + 0.005 −0.416 −0.584† −0.179
(0.023) (0.292) (0.319) (0.345)
interactive term between wage and labor quality (lwage × lbr_qlty) 0.043 0.060† 0.022
(0.030) (0.033) (0.034)
log GDP (lgdp) + 0.385 −0.182
(0.349) (0.410)
total export and import over GDP (open) + 0.023 0.009
(0.019) (0.011)
local government fiscal expenditure over GDP (govern) +∕− 0.935 −0.060
(0.798) (0.862)
density of road and railway (infra) + 0.149 0.143
(0.153) (0.146)
percentage of primary sector in GDP (pct_primary) – −0.011 0.010
(0.016) (0.017)
percentage of secondary sector in GDP (pct_secondary) + 0.008 0.022∗
(0.011) (0.011)
ratio of GDP in tertiary sector to that in others (rt_tertiary) + 0.325∗∗∗ 0.649∗∗∗
(0.091) (0.101)
ratio of population above 65 (rt_ab65) – 0.035∗
(0.015)
after-WTO time dummy (aftr_WTO) + −0.055 n.id.
(0.100)
after-GFC time dummy (aftr_GFC) – −0.345∗∗ −0.285∗
(0.123) (0.132)
Wald statistics and p-values testing exclusion of variables
lwage, lwage × lbr_qlty = 0 3.089 3.475 0.537
[0.213] [0.176] [0.765]
lbr_qlty, lwage × lbr_qlty = 0 2.087 3.448 3.066
[0.352] [0.178] [0.216]
All lwage, lbr_qlty and related variables = 0 0.610 3.232 3.488 3.274
[0.737] [0.357] [0.322] [0.351]
Linear time trend Yes Yes Yes Yes
Province FE Yes Yes Yes Yes
Adjusted R2 0.692 0.696 0.716 0.655
Time span 1993–2018 1993–2018 1993–2018 2002–2018
Sample size nT 754 754 754 493

In the top panel, the numbers in the parentheses are two-way cluster-robust standard errors. In the middle panel, the numbers in the brackets
are the p-values of the test statistics. The number of knots (spline base functions) are selected by AIC. The asterisks in the superscript denote
the significance level, †p < 0.10; ∗ p < 0.05; ∗∗ p < 0.01; ∗∗∗ p < 0.001.“n.id.” means the variable is not identified in the model.

dropped a bit, becoming stable around 6‰. In addition, we create

three additional subsamples corresponding to three geographic loca-
tions: the 11 provinces of East China (Beijing, Tianjin, Hebei, Liaoning,
Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, Hainan); 8
provinces of Middle China (Shanxi, Jilin, Heilongjiang, Anhui, Jiangxi,
Henan, Hubei, Hunan); and 10 provinces of West China (Inner Mon-
golia, Guangxi, Sichuan, Guizhou, Yunnan, Shaanxi, Gansu, Qinghai,
Ningxia, Xinjiang). These three regions have different natural envi-
ronment endowments and socio-economic attributes.16 Consequently,
wages and labor quality may play different roles in attracting FDI
inflows in each region.
The results in Table 3 suggest that the effects of log wage and labor
quality on log FDI change over time. In the early time period from
1993 to 1997, log wage has a marginally significant positive effect on
log FDI that diminishes as labor quality increases. In the first column,
the p-value of the first Wald statistic, which tests whether the coeffi-
cients of log wage and its interaction with labor quality are both zero,

Fig. 5. Average labor quality of 29 provinces in China (1993–2018).

16
In addition to the heterogeneity revelation, another benefit of performing
the subsample analysis across regions is to control for the bias that might arise
from using variables of the urban labor force to represent the labor characteris-
tics of a province. In our subsample analysis, within each region the provinces is 5.7%; the coefficient of log wage (lwage) is positive, and that of the
share similar socio-economic attributes; therefore, the bias is weakened to a interactive term between wage and labor quality (lwage × lbr_qlty) is
certain extent. negative. During the second time period from 1998 to 2008, both log

8
L. Hou et al. Economic Modelling 102 (2021) 105557

Table 3
Estimation results of FE models using different time periods.
Time periods
1993–1997 1998–2008 2009–2018
log wage (lwage) 3.353∗ 0.236 −0.766 −0.847
(1.405) (0.539) (0.976) (0.933)
labor quality (lbr_qlty) 6.110† 0.128 −1.486 −1.549
(3.235) (0.727) (1.418) (1.399)
interactive term between wage and labor quality (lwage × lbr_qlty) −0.663† −0.015 0.153 0.159
(0.358) (0.079) (0.133) (0.132)
log GDP (lgdp) −1.687 0.971 −0.843∗ −0.717†
(1.172) (0.823) (0.410) (0.434)
total export and import over GDP (open) −0.008 0.070∗ −0.051† −0.041
(0.037) (0.030) (0.030) (0.031)
local government fiscal expenditure over GDP (govern) −2.634 0.316 −0.662 −0.582
(5.459) (2.278) (1.358) (1.333)
density of road and railway (infra) −3.592 −0.320† 0.860 0.826
(2.398) (0.165) (0.561) (0.550)
percentage of primary sector in GDP (pct_primary) −0.318∗ −0.042 0.001 0.004
(0.140) (0.034) (0.027) (0.029)
percentage of secondary sector in GDP (pct_secondary) −0.285∗ 0.015 0.031∗ 0.033∗
(0.117) (0.030) (0.016) (0.015)
ratio of GDP in tertiary sector to that in others (rt_tertiary) −7.325∗ −0.047 0.469 0.510
(3.146) (0.460) (0.392) (0.378)
ratio of population above 65 (rt_ab65) 0.019
(0.017)
after-WTO time dummy (aftr_WTO) n.id. 0.045 n.id. n.id.
(0.128)
after-GFC time dummy (aftr_GFC) n.id. n.id. n.id. n.id.
Wald statistics and p-values testing exclusion of variables
lwage, lwage × lbr_qlty = 0 5.716† 0.310 1.554 1.548
[0.057] [0.856] [0.460] [0.461]
lbr_qlty, lwage × lbr_qlty = 0 3.585 0.042 8.720∗ 8.766∗
[0.167] [0.9792] [0.013] [0.012]
All lwage, lbr_qlty and their interaction = 0 5.810 0.310 11.094∗ 10.683∗
[0.121] [0.958] [0.011] [0.014]
Linear time trend Yes Yes Yes Yes
Province FE Yes Yes Yes Yes
Adjusted R2 0.046 0.558 0.699 0.700
Number of provinces 29 29 29 29
Sample size nT 145 319 290 290
In the top panel, the numbers in the parentheses are two-way cluster-robust standard errors. In the middle panel, the numbers
in the brackets are the p-values of the test statistics. The asterisks in the superscript denote the significance level, †p < 0.10;
∗ p < 0.05; ∗∗ p < 0.01; ∗∗∗ p < 0.001. “n.id.” means the variable is not identified in the model.

wage and labor quality become statistically insignificant. In the second FDI inflows into the middle region of China.
column, all three Wald tests fail to reject the nulls that the overall and The heterogeneity of the marginal effects of log wage and labor
joint effects of log wage and labor quality are zero. quality on log FDI inflows over time and across regions hints at the
Finally, in the last time period from 2009 to 2018, labor quality existence of possible nonlinear relationships among the three. As the
turns out to be a salient factor and the joint effect of log wage and time period and region changes, the values of labor quality and log
labor quality on log FDI is significant, regardless of whether the ratio of wage also change dramatically. The fact that the marginal effects of log
population above 65 (rt_ab65) capturing the degree of population aging wage and labor quality on log FDI inflow depend on the values of labor
is included in the model or not.17 However, the sign of the coefficients quality and log wage implies that, all other things being equal, the rela-
changes while remaining insignificant. In the third and fourth columns, tionship among the three variables could be nonlinear. Therefore, the
the p-values of the second and the third Wald statistics are all around estimation results of the linear panel data models may not fully reveal
1%; the coefficients on the predictors labor quality (lbr_qlty) and log the relationship among the three.
wage (lwage) become negative and the coefficient on the predictor that
is an interactive term between wage and labor quality (lwage × lbr_qlty)
is positive. 5. Estimation of a partially linear model
Similarly, the results in Table 4 imply that the effects of log wage
and labor quality on log FDI are quite heterogeneous across differ- 5.1. Specification
ent regions of China. In the east and west regions, neither log wage
nor labor quality has a significant effect on log FDI. However, in the To account for the existence of possible nonlinear relationships
middle region, the coefficient of log wage is significantly negative, among log FDI inflow, log wage and labor quality, we propose a par-
which implies that a lower wage level generally would attract more tially linear panel data model with province fixed effects and a linear
time trend:

17
Because the values of the ratio of population above 65 (rt_ab65) is available lfdiit = g (lwageit −1 , lbr_qltyit −1 ) + w′it −1 𝛾 + 𝛿1 aftr_WTOt
from 2002 to 2018, we include it as a control variable only in the model using
subsample with time period from 2009 to 2018. + 𝛿2 aftr_GFCt + 𝜇i + 𝜆t + 𝜖it . (5.1)

9
L. Hou et al. Economic Modelling 102 (2021) 105557

Table 4
Estimation results of FE models using different locations.
Geographic locations
East Middle West
log wage (lwage) 0.152 −0.945∗∗∗ −0.251
(0.446) (0.251) (0.524)
labor quality (lbr_qlty) −0.383 −0.588 −0.534
(0.268) (0.361) (0.550)
interactive term between wage and labor quality (lwage × lbr_qlty) 0.036 0.063 0.064
(0.026) (0.038) (0.059)
log GDP (lgdp) −0.152 1.407∗∗ 0.493
(0.434) (0.518) (0.578)
total export and import over GDP (open) 0.024 0.153∗∗ −0.071
(0.021) (0.058) (0.092)
local government fiscal expenditure over GDP (govern) −2.977 2.348 −0.025
(2.465) (1.709) (0.575)
density of road and railway (infra) 0.178 0.176 −0.110
(0.287) (0.202) (0.299)
percentage of primary sector in GDP (pct_primary) −0.014 0.066 0.133
(0.026) (0.060) (0.085)
percentage of secondary sector in GDP (pct_secondary) −0.002 0.082 0.142†
(0.023) (0.070) (0.084)
ratio of GDP in tertiary sector to that in others (rt_tertiary) 0.261 3.597† 5.398†
(0.199) (2.000) (3.091)
ratio of population above 65 (rt_ab65)
after-WTO time dummy (aftr_WTO) −0.069 0.098 0.024
(0.139) (0.096) (0.097)
after-GFC time dummy (aftr_GFC) −0.122 −0.455∗∗ −0.442∗
(0.165) (0.151) (0.199)
Wald statistics and p-values testing exclusion of groups of variables
lwage, lwage × lbr_qlty = 0 2.120 23.864∗∗∗ 1.252
[0.347] [0.000] [0.535]
lbr_qlty, lwage × lbr_qlty = 0 2.120 2.758 1.938
[0.346] [0.252] [0.379]
All lwage, lbr_qlty and their interaction = 0 3.144 23.888∗∗∗ 1.943
[0.370] [0.000] [0.584]
Linear time trend Yes Yes Yes
Province FE Yes Yes Yes
Adjusted R2 0.760 0.860 0.664
Number of provinces 11 8 10
Time span 1993–2018 1993–2018 1993–2018
In the top panel, the numbers in the parentheses are two-way cluster-robust standard errors. In the middle
panel, the numbers in the parentheses are the p-values of the test statistics. The asterisks in the superscript
denote the significance level, †p < 0.10; ∗ p < 0.05; ∗∗ p < 0.01; ∗∗∗ p < 0.001. The variable rt_ab65 is not
included in the three models due to severely decreasing the sample sizes if otherwise.

The specification of our proposed empirical model (5.1) is the same as B-splines to approximate the unknown smooth function.18
that of the traditional linear fixed effects model (4.1) except that log
wage (lwageit −1 ), and labor quality (lbr_qltyit −1 ) have a joint nonlinear 5.2. Empirical results
effect on log FDI (lfdit ), characterized by an unknown smooth function
g(lwageit −1 , lbr_qltyit −1 ). We only allow wages and labor quality to have In combination with the ANOVA decomposition of smooth functions
the joint nonlinear effects on FDI. That is a trade-off between model par- in (5.2), our proposed partially linear panel data model (5.1) can be
simony and flexibility. It is known that the semi-parametric approach written explicitly as
ensures the simplicity of estimation and allows for some economically
meaningful flexibility of model specification, suitable for our purposes. lfdiit = f1 (lwageit −1 ) + f2 (lbr_qltyit −1 ) + f3 (lwageit −1 , lbr_qltyit −1 )
In this study, we focus on the intertwined effect of wage and labor qual-
+ w′it−1 𝜸 + 𝛿1 aftr_WTOt + 𝛿2 aftr_GFCt + 𝜇i + 𝜆t + 𝜖it (5.3)
ity. Therefore, we only allow the effect of wage and labor quality to
depend on each other, and we let other control variables keep a linear where the vector wit −1 includes the control variables, f1 (lwageit −1 )
relationship, as in the existing literature. and f2 (lbr_qltyit −1 ) representing the nonlinear effects of log wage
Based on ANOVA decomposition of smooth functions, we can and labor quality on log FDI inflow alone respectively, and
decompose the unknown smooth function g(x, z) as follows, f3 (lwageit −1 , lbr_qltyit −1 ) captures the interactive nonlinear effects of the
two variables. The construction of the three functions, f1 , f2 and f3 ,
g (lwageit −1 , lbr_qltyit −1 ) = f1 (lwageit −1 ) + f2 (lbr_qltyit −1 ) using cubic splines follows our discussion earlier in this section.

+ f3 (lwageit−1 , lbr_qltyit−1 ), (5.2)

18
The construction of the splines in the decomposition (5.2) can be found in
where f1 (lwageit −1 ) and f2 (lbr_qltyit −1 ) are smoothed main effects de Boor (2001) and Wood (2017). Appendix B describes the construction of the
of log wage and labor quality on dependent variable, and nonlinear function and estimation method in detail. For econometric theory on
f3 (lwageit −1 , lbr_qltyit −1 ) captures the smooth interactive effects that partially linear panel data models with fixed effects, please see Li and Stengos
include no component of the form f1 + f2 . We use tensor product cubic (1996), Baltagi and Li (2002), and Huang et al. (2004) among others.

10
L. Hou et al. Economic Modelling 102 (2021) 105557

Table 5
Estimation results of partially linear models.
Partially linear models
Model PL1 Model PL2 Model PL3 Model PL4
f1 (lwage) Included Included Included Included
f2 (lbr_qlty) Included Included Included Included
f3 (lwage, lbr_qlty) Included Included Included
log GDP (lgdp) 0.345 −0.122
(0.303) (0.360)
total export and import over GDP (open) 0.027 0.025†
(0.020) (0.014)
local government fiscal expenditure over GDP (govern) 0.961 0.685
(0.792) (0.695)
density of road and railway (infra) 0.068 0.147
(0.148) (0.144)
percentage of primary sector in GDP (pct_primary) −0.032∗ −0.004
(0.014) (0.012)
percentage of secondary sector in GDP (pct_secondary) −0.006 0.012
(0.011) (0.009)
ratio of GDP in tertiary sector to that in others (rt_tertiary) −0.174 0.132
(0.181) (0.172)
ratio of population above 65 (rt_ab65) 0.045∗∗
(0.015)
after-WTO time dummy (aftr_WTO) −0.100 n.id.
(0.078)
after-GFC time dummy (aftr_GFC) −0.196∗∗ −0.126†
(0.071) (0.068)

Wald statistics and p-values testing exclusion of groups of variables

f1 (lwage) = 0 9.69∗ 2.78 7.246† 8.564∗
[0.021] [0.428] [0.064] [0.036]
f2 (lbr_qlty) = 0 25.3∗∗ ∗ 11.9∗∗ 8.656∗ 5.900
[0.000] [0.008] [0.034] [0.117]
f3 (lwage, lbr_qlty) = 0 365.2∗∗∗ 29.99∗∗∗ 17.140∗
[0.000] [0.000] [0.047]
Linear time trend Yes Yes Yes Yes
Province FE Yes Yes Yes Yes
# of splines 6 15 15 15
Adjusted R2 0.734 0.742 0.747 0.715
Time span 1993–2018 1993–2018 1993–2018 2002–2018
Sample size nT 754 754 754 493

In the top panel, the numbers in the parentheses are two-way cluster-robust standard errors. In the middle panel, the
numbers in the parentheses are the p-values of the test statistics. The asterisks in the superscript denote the significance
level, †p < 0.10; ∗ p < 0.05; ∗∗ p < 0.01; ∗∗∗ p < 0.001. “n.id.” means the variable is not identified in the model.

In Table 5, we report the estimation results of the four partially lin- effects model (Models 3 and 4 in Table 2), we find that the partially
ear panel data models, PL1, PL2, PL3, and PL4. These specifications linear model not only fits the data better but also captures the marginal
are similar to those of Model 1, 2, 3, and 4 shown in Table 2. In Model effect of wages on FDI more effectively than the FE models. Intuitively,
PL1, we consider only additively nonlinear effects of log wage and labor the adjusted R-squared of Models PL3 and PL4 is 74.7% and 71.5% and
quality. By comparison, Model PL2 accommodates the nonlinear inter- that of Models 3 and 4 is 71.6% and 65.5%. Models PL3 and PL4 has
active effects. In Model PL3, we include all the control variables except twelve more regressors than Models 3 and 4 because of its construc-
the ratio of population above 65 (rt_ab65) in the linear part. In the esti- tion of nonlinear functions using splines. Still, the adjusted R-squared
mation of PL1-3, we use the full sample from 1993 to 2008. In Model of Models PL3 and PL4, which measures the goodness of fit after penal-
PL4, we incorporate the ratio of population above 65 as an additional izing the number of regressors, is 3.1% and 6% larger than those of
control variable. Because the values of the ratio of population above Models 3 and 4 respectively.
65 is only available from 2002 to 2018, we have to restrict our sample Also, we conduct a battery of linearity tests, reporting the results in
to the same time period. Our first finding is that the interactive non- Tables 6 and 7. Specifically, the bottom panels of Tables 6 and 7 dis-
linear effects matter. When testing whether the nonlinear interaction play the results of testing the null hypothesis that the three nonlinear
term f3 (lwageit −1 , lbr_qltyit −1 ) is zero in Models PL2, PL3 and PL4, all functions f1 (lwage), f2 (lbr_qlty) and f3 (lwage, lbr_qlty) are jointly linear.
the Wald statistics strongly reject the null that the nonlinear interaction Under this null, the specifications of both Models PL3 and PL4 coin-
is zero.19 In addition, the estimation results of Models PL3 and PL4 sup- cide with those of Model 3 and Model 4 respectively, i.e. the linear
port that log wage has a statistically significant effect on log FDI inflow, fixed effects models. The Wald statistics are 43.74 and 61.28 with the
because the Wald statistics used to test whether f1 (lwageit −1 ) is zero has p-values smaller than 0.001, strongly rejecting the null hypothesis that
p-values of 6.4% and 3.6% respectively. Moreover, the Wald statistics the linear specification of Models 3 and 4 are adequate. More interest-
testing f2 (lbr_qltyit −1 ) = 0 has p-values of 3.4% and 11.7%, showing ingly, the result of linearity test for the single function f3 (lwage, lbr_qlty)
that labor quality is at least marginally significant. also strongly rejects the null of linearity in Models PL3 and PL4. Com-
Comparing the estimation results of our proposed partially linear bining it with the testing results of the null f3 (lwage, lbr_qlty) = 0 in
models (Models PL3 and PL4 in Table 5) to the traditional linear fixed Table 5, we conclude that the interactive effect of wage and labor qual-
ity on FDI does exist and this interaction is truly nonlinear. All of these
empirical results suggest that our proposed partially linear model cap-
19
See the test results in the third row of the middle panel in Table 5.

11
L. Hou et al. Economic Modelling 102 (2021) 105557

Table 6
Linearity tests of nonlinear functions in Model PL3.
Null hypothesis Wald statistic p-value
Linearity tests for single functions
f1 (lwage) is linear. 3.322 [0.190]
f2 (lbr_qlty) is linear. 5.804† [0.055]
f3 (lwage, lbr_qlty) is linear. 27.707∗∗∗ [0.000]
Joint linearity test for all three functions
f1 , f2 and f3 are all linear. 43.735∗∗∗ [0.000]

The asterisks in the superscript denote the signifi-

cance level, †p < 0.10; ∗ p < 0.05; ∗∗ p < 0.01;
∗∗∗ p < 0.001.

Table 7
Linearity tests of nonlinear functions in Model PL4.
Null hypothesis Wald statistic p-value

Linearity tests for single functions

f1 (lwage) is linear. 3.437 [0.179]
f2 (lbr_qlty) is linear. 4.806† [0.090]
f3 (lwage, lbr_qlty) is linear. 46.86∗∗ ∗ [0.000]
Joint linearity test for all three functions
f1 , f2 and f3 are all linear. 61.28∗∗ ∗ [0.000] Fig. 6. Estimated marginal effect of log wage on log FDI as a function of log
𝜕̂g
The asterisks in the superscript denote the signifi- wage and labor quality: 𝜕lwage (lwage, lbr_qlty )
cance level, †p < 0.10; ∗ p < 0.05; ∗∗ p < 0.01;
∗∗∗ p < 0.001.

tures the marginal effect of wage on FDI more effectively than the linear
FE models does.
Using the estimation results of Model PL4, we illustrate intuitively
the nonlinear effects of wage and labor quality on FDI in Figs. 6 and
7. Fig. 6 shows the estimated marginal effects of log wage on log FDI
inflow as a function of lwage and lbr_qlty; similarly, Fig. 7 shows the
estimated marginal effects of labor quality on log FDI as a function of
log wage (lwage) and labor quality (lbr_qlty), i.e.
𝜕̂
g(lwage, lbr_qlty ) 𝜕̂
g (lwage, lbr_qlty )
and .
𝜕 lwage 𝜕 lbr_qlty
The two 3-dimensional surface plots imply that the marginal effects
of log wage and labor quality also depend on each other. In the 3D
plots, the arrow alongside the axis indicates the direction in which the
value of the corresponding variable increases; the color of the surface
indicates the value of the marginal effect. In Fig. 6, the marginal effect
of log wage on log FDI is generally negative and decreases as labor
quality increases. Specifically, when labor quality is low, the marginal
effect of log wage on log FDI inflow could be positive. In Fig. 7, the
marginal effect of labor quality on log FDI inflow generally diminishes
as log wage increases. Moreover, when log wage is low, the marginal Fig. 7. Estimated marginal effect of labor quality on log FDI as a function of log
effect of labor quality is positive. 𝜕̂g
wage and labor quality: 𝜕lbr_qlty (lwage, lbr_qlty )
Furthermore, the estimated marginal effects of log wage and labor
quality shown in Figs. 6 and 7 are statistically significant when log wage
and labor quality are within certain ranges. Fig. 8 plots the estimated
value of 4.550‰, i.e.
marginal effect of log wage on log FDI as a function of labor quality
with lwage fixed at its 60% percentile value of 10.130.20 Fig. 9 plots the 𝜕̂
g
(l wage60%pctl = 10.130, lbr_qlty) and
estimated marginal effect of labor quality on log FDI as a function of 𝜕 lwage
log wage (lwage) with labor quality (lbr_qlty) fixed at its sample median 𝜕̂
g
(lwage, l br_ql ty50%pctl = 4.550).
𝜕 lbr_qlty
Generally, both estimated curves are nonlinear and the marginal effects
are statistically significant on certain sub-intervals of labor quality and
20 log wage respectively. In Fig. 8, the estimated curve is downward slop-
This confirms the key result of the empirical analysis that the estimated
marginal-effect curve is downward sloping. That is to say, the marginal effects
ing and the marginal effects of log wage at the 60% percentile are
of wage on FDI are decreasing with respect to labor quality. While it is true that marginally significant at the 10% level when labor quality is smaller
the marginal effects of log wage on log FDI are not significant at larger values than 3.5‰. In Fig. 9, the estimated curve is hump-shaped, which shows
of labor quality, it does not affect the downward sloping shape of the estimated that the marginal effect of labor quality is statistically significant with
curve. an upward slope when log wage is smaller than 9.6 and with a down-

12
L. Hou et al. Economic Modelling 102 (2021) 105557

over time from 1993 to 2018 but also across different regions of China.
This finding further suggests that these effects may be nonlinear. Hence,
we propose a partially linear panel data model, allowing for a nonlin-
ear relationship between FDI inflow and the two variables of interest,
and apply it to FDI location choices in China. Our main finding is that
the marginal effect of wage is generally a decreasing function of labor
quality. This implies that when faced with low labor quality, FDI firms
would prefer locations with high wages and thus higher labor quality.
When facing high labor quality, FDI firms would prioritize locations
with low wages in order to reduce labor costs. Moreover, we find that
the marginal effect of labor quality on FDI inflows generally diminishes
as the log wage increases. This suggests that when the wage level is
moderate, labor quality will attract more FDI than when the wage is
high.
FDI is one of the driving forces behind the growth of transitional
economies. Hence, attracting FDI inflows is the policy priority of vari-
Fig. 8. Estimated marginal effect of log wage on log FDI as a function of labor ous transitional economies. Our results help the transitional economies
𝜕̂g
quality: 𝜕lwage (l wage60%pctl , lbr_qlty ) where lwage is fixed as its 60% percentile better understand the wage effect on FDI, i.e., low labor costs do not
value. necessarily mean an effective comparative advantage in attracting FDI
inflows. Understanding that local labor quality is vital for the low labor
cost countries, the transitional economies could dedicate themselves to
promoting the human capital formation and enhancing the labor quality
level. Hence, the transitional economies could activate their low labor
cost advantages and successfully attract FDI.
There are limitations in our current study. Although the province-
level data enables us to explore heterogeneity across regions and over
time, and to confirm the nonlinear relationship among wages, labor
quality, and FDI location choices, it does not include the FDI indus-
trial and bilateral information. Hence, we cannot investigate how wage
effects vary across industries with different production technologies and
across different FDI sourcing countries. We expect that more detailed
data with the industrial and bilateral information can help us generate
new results, which is the direction of our future research.

Acknowledgements

Fig. 9. Estimated marginal effect of labor quality on log FDI as a function of log We are very grateful to the editor, Angus Chu, an associate edi-
𝜕̂g
wage: 𝜕lbr_qlty (lwage, l br_ql ty 50%pctl ) where lbr_qlty is fixed as its sample median tor, and two anonymous referees for their constructive comments and
value. suggestions that have greatly improved the paper. Lei Hou grate-
fully thanks the financial support from the Humanities and Social Sci-
ence Research Planning Foundation from the Ministry of Education of
ward slope when log wage is larger than 9.6. China (Grant No. 19YJA790073). Yanfei Wang acknowledges the finan-
In sum, the results shown in the two 3D plots and the two 2D plots cial support from the National Natural Science Foundation of China
suggest that when labor quality is high, a higher wage would tend to (Grant No. 72073022). Xintong Yang acknowledges the financial sup-
impede FDI inflows; when the wage is moderate, higher labor quality port from the National Natural Science Foundation of China (Grant
would attract more FDI than when the wage is high. No. 72003136). We extend our gratitude to the participants at China
Econometrics Forum 2017 at Xiamen University, the Economic Mod-
6. Conclusion elling Special Issue Conference on Applied Finance, Macroeconomic
Performance, and Economic Growth 2019 at Southwestern University
Although in theory low-wage locations do attract FDI, the empirical of Finance and Economics, the Asian Economic Symposium 2019 at
evidence for this is rather mixed. Using linear panel data models with Chonnam National University of Korea, and the International Sympo-
province fixed effects and a linear time trend, we find that the effects sium on Regional Integration and Financial Development in Northeast
of wage and labor quality on FDI inflow are heterogeneous, not only Asia 2020 at Liaoning University. All errors are our own.

13
L. Hou et al. Economic Modelling 102 (2021) 105557

A. Numerical example

Suppose the production function take the Cobb-Douglas form of Q(L(z), x, z) = zL(z)𝛼 · x1−𝛼 . We assume L(z) = Lz , where L is a constant. It
means that the higher
( ) the labor quality is, the smaller amount the labor demand for production is. Hence, the production function can be written
𝛼
L
as Q(L(z), x, z) = z z
· x1−𝛼 and we have:
( )𝛼−1
L
wj (zj ) = QL (L(zj ), xj , zj ) = 𝛼 zj xj1−𝛼 = 𝛼 z2j −𝛼 L𝛼−1 xj1−𝛼
zj
( )𝛼−2
L
QLL (L(zj ), xj , zj ) = 𝛼(𝛼 − 1)zj xj1−𝛼
zj
( )𝛼−1
L L
QLz (L(zj ), xj , zj ) = 𝛼 xj1−𝛼 , L′ (zj ) = −
zj z2j
( )𝛼
L
Qz (L(zj ), xj , zj ) = z · x1−𝛼 , Qzz = 0
zj

w′j (zj ) = QLL (L(zj ), xj , zj )L′ (zj ) + QLz (L(zj ), xj , zj ) = 𝛼(2 − 𝛼)z1j −𝛼 L𝛼−1 xj1−𝛼

w″j (zj ) = 𝛼(1 − 𝛼)(2 − 𝛼)z−𝛼

j
L𝛼−1 xj1−𝛼

According to w′j (zj )L(zj ) = Qz (L(zj ), xj , zj ), we have:

z1j −𝛼 L𝛼 · xj1−𝛼 = 𝛼(2 − 𝛼)z1j −𝛼 L𝛼−1 xj1−𝛼 Lz−

j
1
= 𝛼(2 − 𝛼)z1j −𝛼 L𝛼 xj1−𝛼 z−
j
1

So we can solve for zj = 𝛼(2 − 𝛼).

QLL (L(zj ), xj , zj )(L′ (zj ))2 + w′ ′(zj )L(zj )

= 𝛼(𝛼 − 1)z3j −𝛼 L𝛼−2 xj1−𝛼 (−Lz−

j
) + 𝛼(1 − 𝛼)(2 − 𝛼)z−𝛼
2 2
j
L𝛼−1 xj1−𝛼 Lz−
j
1

= 𝛼(𝛼 − 1)zj−𝛼−1 L𝛼 xj1−𝛼 + 𝛼(1 − 𝛼)(2 − 𝛼)zj−𝛼−1 L𝛼 xj1−𝛼

= 𝛼(𝛼 − 1)2 zj−𝛼−1 L𝛼 xj1−𝛼 > 0 = Qzz (L(zj ), xj , zj )

Furthermore,

QLL (L(zj ), xj , zj )L′ (zj ) + QLz (L(zj ), xj , zj )

1 𝛼−2 1 𝛼−1
= 𝛼(𝛼 − 1)zj (Lz−
j
) · xj1−𝛼 · (−Lz−
j
2
) + 𝛼(Lz−
j
) · xj1−𝛼

= 𝛼(1 − 𝛼)L𝛼−1 z1j −𝛼 xj1−𝛼 + 𝛼 L𝛼−1 z1j −𝛼 xj1−𝛼

= 𝛼(2 − 𝛼)L𝛼−1 z1j −𝛼 xj1−𝛼 > 0

Hence, we have the two assumptions Qzz (L(zj ), xj , zj ) < QLL (L(zj ), xj , zj ) · (L′ (zj ))2 + w′′ (zj )L(zj ) and QLL (L(zj ), xj , zj )L′ (zj ) + QLz (L(zj ), xj , zj ) > 0
hold in this numerical example.

B. Estimation procedure of the partially linear model

We assume that both Xit and Zit are distributed on compact intervals. Let Lx , Lz ≥ 1 be the numbers of interior knots for the compact supports
[ax , bx ] and [az , bz ] associated with Xit and Zit respectively. We divide the closed intervals [ax , bx ] and [az , bz ] into Lx + 1 and Lz + 1 subintervals.
L
For the interval [ax , bx ], we have Ij,x = [tj,x , tj+1,x ), j = 0, … , Lx − 1, ILx = [tLx ,x , bx ], where {tj,x }j=x 1 is a sequence of equally spaced interior knots,
satisfying

t−2,x = t−1,x = t0,x = ax < t1,x < · · · < tLx ,x < bx = tLx +1,x = tLx +2,x = tLx +3,x .

L
Similarly, for the interval [az , bz ], we have Ij,z = [tj,z , tj+1,z ), j = 0, … , Lz − 1, ILz = [tLz ,z , bz ], where {tj,z }j=z 1 is a sequence of equally spaced
interior knots, satisfying

t−2,z = t−1,z = t0,z = az < t1,z < · · · < tLz ,z < bz = tLz +1,z = tLz +2,z = tLz +3,z .

Based on de Boor (2001), we define the univariate cubic B-spline basis as Mx = {Bj,x (x) ∶ 1 ≤ j ≤ Mx } and Mz = {Bj,z (z) ∶ 1 ≤ j ≤ Mz } with the
number of spline basis functions Mx ≡ Lx + 3 and Mz ≡ Lz + 3. Then we define the tensor product cubic B-spline basis by Mxz = {Bj,x (x)Bk,z (z) ∶
1 ≤ j ≤ Mx , 1 ≤ j ≤ Mz } with the number of spline basis functions Mxz = Mx × Mz for simplicity of notation. In the empirical application, we

14
L. Hou et al. Economic Modelling 102 (2021) 105557

∑Mx ∑Mz
use Akaike Information Criterion to select M. The unknown smooth function g(x, z) can be approximated by j =1 k=1
Bj,x (x)Bk,z (z)𝛽jk . Using this
approximation, the partially linear model (5.1) with ANOVA decomposition (5.2) can be rewritten as

Yit = f1 (Xit −1 ) + f2 (Zit −1 ) + f3 (Xit −1 , Zit −1 ) + Wit′ −1 𝜸 + 𝜇i + 𝜆t + 𝜀it

Mx Mz Mx Mz
∑ ∑ ∑ ∑
= Bx,j (Xit −1 )𝛽x,j + Bz,k (Zit −1 )𝛽z,k + Bx,j (Xit −1 )Bz,k (Zit −1 )𝛽jk + Wit′ −1 𝜸 + 𝜇i + 𝜆t + 𝜀it
j =1 k=1 j=1 k=1

= Bx (Xit−1 ) 𝜷 x + Bz (Zit−1 ) 𝜷 z + Bxz (Xit−1 , Zit−1 )𝜷 xz + W ′it−1 𝜸 + 𝝁i + 𝝀t + 𝜺it

′ ′ ′

= B′ (Xit−1 , Zit−1 )𝜷 + Wit′ −1 𝜸 + 𝜇i + 𝜆t + 𝜀it , (B.1)

where
( )
B′x (Xit −1 ) = Bx,1 (Xit −1 ) ··· Bx,j (Xit −1 ) ··· Bx,Mx (Xit −1 )
( )
B′z (Zit −1 ) = Bz,1 (Zit −1 ) ··· Bz,k (Zit −1 ) ··· Bz,Mz (Zit −1 )

B′xz (Xit −1 , Zit −1 ) = B′x (Xit −1 ) ⊗ B′z (Zit −1 )

( )
B′ (Xit −1 , Zit −1 ) = B′x (Xit −1 ) B′z (Zit −1 ) B′xz (Xit −1 , Zit −1 )
( )
𝜷 ′x = 𝛃x,1 , … , 𝛃x,Mx
( )
𝜷 ′z = 𝛃z,1 , … , 𝛃z,Mz
( )
𝜷 ′xz = 𝛃11 , … , 𝛃1Mz , … , 𝛃Mx 1 , … , 𝛃Mx Mz
( )
𝜷 ′ = 𝛃′x , 𝛃′z , 𝛃′xz
∑∑ ∑∑
and by construction we restrict that i t Bx,j (xit −1 ) = 0 and i t Bz,k (zit −1 ) = 0 for all j and k so that the interaction terms
f3 (xit −1 , zit −1 ) = Bxz (xit −1 , zit −1 )𝜷 xz are orthogonal to the main effects f1 (xit ) = Bx (xit −1 )′ 𝜷 x and f2 (zit ) = Bz (zit −1 )′ 𝜷 z .
′

Stacking the equation (B.1), we obtain the following matrix form

Y = X𝜽 + 𝝁 + 𝜺 (B.2)

where X = [B(X, Z), W, R], 𝜽′ = (𝜷 ′ , 𝜸 ′ , 𝜆), and other related terms are defined as follows,

Y = (Y12 , … , Y1T , … , Yn2 , … , YnT )′

B(X, Z) = (B(X11 , Z11 ), … , B(X1T−1 , Z1T−1 ), … , B(Xn1 , Zn1 ), … , B(XnT−1 , ZnT−1 ))′

W = (W11 , … , W1T−1 , … , Wn1 , … , WnT−1 )′

R = 𝛊n ⊗ (2, … , T)′

𝝁 = [𝛍1 , … , 𝛍n ]′ ⊗ 𝜾T−1
𝜺 = [𝛆12 , … , 𝛆1T , … , 𝛆n2 , … , 𝛆nT ]′
and 𝜄d is a d-dimensional column vector of ones.
We propose to use the within-group method to estimate the parameter vector 𝜽 in (B.2). First, by within transformation, we remove the fixed
effects by pre-multiplying the within projection matrix M on both sides of equation (B.2) and obtain

MY = MX 𝜽 + M𝜺. (B.3)

where the within projection matrix M, symmetric and idempotent, is defined by M = In ⊗ QT −1 in which In is a T × T identity matrix,
QT −1 = IT −1 − JT −1 ∕(T − 1) and JT −1 is a (T − 1) × (T − 1) matrix with all entries equal to 1.
Since the variance-covariance structure of the error term 𝜺 is left unspecified, we simply obtain the consistent within-group OLS estimate ̂
𝜽,
̂
𝜽 = (X ′ MX )−1 X ′ MY.
Since both serial and cross-sectional correlation are allowed for the error term, we use with two-way cluster-robust variance matrix estimate
̂ (̂
V ̂ X ′ MX )−1
𝜽) = (X ′ MX )−1 𝛀(
proposed by Cameron et al. (2011) and Thompson (2011). The central matrix is
̂ = X ′ M(̂
𝛀 𝜺̂
𝜺′ ⊙ S)MX
where ̂ 𝜺 is the n(T − 1) × 1 vector of within-group OLS residuals, ⊙ denotes element-by-element (Hadamard) multiplication and S is an
n(T − 1) × n(T − 1) indicator, or selection, matrix with 𝓁m entry equal to one if its corresponding (i𝓁 , t𝓁 ) and (jm , sm ) observations share at
least one of the two clusters in group and time (i𝓁 = jm or t𝓁 = sm or both) and equal to zero otherwise. Mathematically, we have

15
L. Hou et al. Economic Modelling 102 (2021) 105557

S = In ⊗ JT −1 + (Jn − In ) ⊗ IT −1

The smooth function g(x, z) is estimated by ̂ ̂ and its two-way cluster-robust variance estimate is V
g (x, z) ≡ B′ (x, z)𝜷 ̂[̂ ̂ B(x, z).
̂(𝜷)
g (x, z)] = B(x, z)′ V
Under certain smoothing condition on smooth function g(x, z) and regularity conditions on matrices X, B(X, Z), and W, we conjecture that the
estimate ̂
𝜽 is consistent with typical parametric convergence rate and asymptotic normal, and the nonparametric estimator ̂ g (x, z) has typical
nonparametric convergence rate and asymptotic normality; its bias term is asymptotically negligible if certain undersmoothing condition is assumed:
We use a larger number of knots (i.e. base functions) than what is needed for achieving the optimal rate of convergence.

References Kang, S.J., Lee, H.S., 2007. The determinants of location choice of South Korean FDI in
China. Jpn. World Econ. 19 (4), 441–460.
Amiti, M., Javorcik, B.S., 2008. Trade costs and location of foreign firms in China. J. Li, Q., Stengos, T., 1996. Semiparametric estimation of partially linear panel data
Dev. Econ. 85 (1–2), 129–149. models. J. Econom. 71, 289–397.
Ascani, A., Crescenzi, R., Iammarino, S., 2016. Economic institutions and the location Liu, X., Lovely, M.E., Ondrich, J., 2010. The location decisions of foreign investors in
strategies of European multinationals in their geographic neighborhood. Econ. China: untangling the effect of wages using a control function approach. Rev. Econ.
Geogr. 92 (4), 401–429. Stat. 92, 160–166.
Baltagi, B.H., Li, D., 2002. Series estimation of partially linear panel data models with Nielsen, B.B., Asmussen, C.G., Weatherall, C.D., 2017. The location choice of foreign
fixed effects. Ann. Econ. Finance 3, 103–116. direct investments: empirical evidence and methodological challenges. J. World Bus.
Blyde, J., Molina, D., 2015. Logistic infrastructure and the international location of 52 (1), 62–82.
fragmented production. J. Int. Econ. 95 (2), 319–332. Noorbakhsh, F., Paloni, A., 2001. Human capital and FDI inflows to developing
Cameron, A.C., Gelbach, J.B., Miller, D.L., 2011. Robust inference with multiway countries: new empirical evidence. World Dev. 29 (9), 1593–1610.
clustering. J. Bus. Econ. Stat. 29 (2), 238–249. Rong, S., Liu, K., Huang, S., Zhang, Q., 2020. FDI, Labor Market Flexibility and
de Boor, C., 2001. A Practical Guide to Splines (Applied Mathematical Sciences, vol. 27. Employment in China. China Economic Review, p. 101449.
Springer-Verlag, New York. Thompson, S.B., 2011. Simple formulas for standard errors that cluster by both firm and
Du, J., Lu, Y., Tao, Z., 2008. Economic institutions and FDI location choice: evidence time. J. Financ. Econ. 99 (1), 1–10.
from US multinationals in China. J. Comp. Econ. 36 (3), 412–429. Tomohara, A., 2017. Does immigration crowd out foreign direct investment inflows?
Gao, T., 2005. Labor quality and the location of foreign direct investment: evidence from Tradeoff between contemporaneous FDI-immigration substitution and ethnic
China. China Econ. Rev. 16 (3), 274–292. network externalities. Econ. Modell. 64, 40–47.
Gao, G., Wang, D., Che, Y., 2018. Impact of historical conflict on FDI location and UNCTAD, 1999. Word Investment Report 1999: Foreign Direct Investment and the
performance: Japanese investment in China. J. Int. Bus. Stud. 49 (8), 1060–1080. Challenge of Development. ISBN: 92-1-112440-9.
Head, K., Mayer, T., 2004. Market potential and the location of Japanese investment in Wei, Y., Liu, X., Parker, D., Vaidya, K., 1999. The regional distribution of foreign direct
the European Union. Rev. Econ. Stat. 86 (4), 959–972. investment in China. Reg. Stud. 33 (9), 857–867.
Huang, J.Z., Wu, C.O., Zhou, L., 2004. Polynomial spline estimation and inference for Wood, S., 2017. Generalized Additive Models. Chapman and Hall/CRC, New York.
varying coefficient models with longitudinal data. Stat. Sin. 14, 763–788. Yeaple, S.R., 2003. The complex integration strategies of multinational and cross country
Iwai, N., Thompson, S.R., 2012. Foreign direct investment and labor quality in dependencies in the structure of foreign direct investment. J. Int. Econ. 60, 293–314.
developing countries. Rev. Dev. Econ. 16 (2), 276–290. Zhang, K., Markusen, J.R., 1999. Vertical multinational and host-country characteristics.
J. Dev. Econ. 59, 233–252.

16