Which Boats are lifted by a Foreign Tide? Direct and Indirect Wage Effects of Foreign Ownership Sourafel Girma University of Nottingham Holger Görg Kiel Institute for the World Economy and University of Kiel Erasmus Kersting Villanova University February 2016 Abstract This paper examines the effects of foreign ownership on wages both theoretically and empirically. The focus is on direct as well as spillover effects: Foreign ownership directly raises a firm s wage level, and this effect is shown to be stronger the more firms are already foreign owned in a given cluster. In contrast, an increase in the share of foreign owned firms lowers the average wage paid by domestic firms. The strength of this effect also depends on the share of foreign owned firms in the cluster. These theoretical effects are derived from a model with heterogeneous firms and labor market frictions, and we find broad empirical support using data from over 140,000 manufacturing firms in China. Keywords: Foreign Direct Investment; Wage effects; Spillovers. JEL Classification Numbers: F12, F16, F23, F61.
1 Introduction Policy makers aiming to improve economic development in their countries frequently look to inward foreign direct investment. Such investments by multinational companies are expected to bring in new capital and technology as well as create new jobs in the economy. But are these new jobs good jobs? This is a frequently debated question, with critics usually arguing that multinationals exploit local labor, while proponents point to new employment opportunities, higher wages and more advanced skills for the local economy (OECD, 2008). While studying the impact of multinational engagement on working conditions and labor standards is important, we focus on the wage effects. Do multinationals pay higher wages than domestic firms? And do they have any implications for wages paid in domestic firms in the host country? These are the questions we aim to address in this paper. A number of recent studies have focused on the first question in particular and examined whether foreign firms pay higher wages than domestic firms (e.g., Hijzen et al., 2013, Heyman et al., 2007, Huttunen, 2007, Girma and Görg, 2007). To be more precise, the research has aimed to establish whether foreign ownership leads firms to increase their wages i.e., whether there is a causal relationship between foreign ownership and wages that is driven not by observable or unobservable firm characteristics, but by ownership per se. A much smaller literature has attempted to look at the implications of the presence of foreign multinationals on wages paid by domestic firms (e.g., Aitken et al., 1996), so-called wage spillovers. However, these two questions are generally studied in isolation. Doing so is problematic, however. Take the first issue, whether foreign firms pay higher wages. The econometric analyses generally define this as a treatment effect (i.e., the treatment is receiving foreign ownership as opposed to staying locally owned) and use estimation strategies from the treatment effect literature in many cases propensity score matching. However, these approaches implicitly, if not explicitly, assume that there are no spillovers, i.e. no indirect effect of the foreign ownership treatment on domestic firms. 1 This is clearly at odds with the literature on spillovers. Another econometric issue is that there is selection when analysing both types of questions. For the first, it is the case that foreign multinationals tend to select better domestic firms (which also pay high wages) as targets for the treatment. This is why treatment effects techniques are used. There is also selection in the spillovers literature, where the question is whether the presence of multinationals in a sector affects wages in domestic firms: Multinationals will not randomly disperse across sectors in the economy but are likely to locate in specific sectors. And the characteristics of the sector that attract multinationals are likely to be correlated with wages. Apart from controlling for unobserved time invariant sector specific effects (via dummies / fixed effects), this concern is generally not dealt with. We implement a unified framework to estimate direct and indirect (spillover) effects from 1 This is the so-called Stable Unit Treatment Value Assumption (SUTVA). 2
foreign ownership on wages which allows for interaction between foreign and domestic firms. 2 Our approach enables us to estimate differing direct or indirect (spillover) effects depending (possibly non-linearly) on the level of foreign presence in a well-defined cluster. This is done within a well-defined potential outcomes framework that lends itself to a counterfactual analysis. Furthermore, we recognize and deal with the problem that there are two levels of selection as pointed out above. Another advantage of our estimation technique is that we can let the estimated direct or indirect effect vary with the strength of the presence of foreign owned firms. In other words, we can estimate different direct or spillover effects in clusters in which there is a high level of foreign presence compared to clusters with low presence. This approach is novel and of high importance for policy as it enables us to determine how effects depend on the presence of foreign owned firms already located in a cluster. 3 In order to motivate our empirical investigation we present a theoretical framework that defines the direct and indirect effects we seek to estimate. While there are a small number of theoretical approaches towards explaining why foreign firms may pay higher wages (e.g., Egger and Kreickemeier, 2013, Malchow-Møller et al., 2013), these papers do not consider whether this premium may be dependent on the presence of other foreign firms in a cluster (i.e., on interference in econometric parlance). Also, there is little theory to motivate why domestic firms wages may be affected by the presence of foreign firms. We provide a simple extension of the model introduced by Helpman et al. (2010) featuring firms that are potentially heterogeneous in terms of productivity and search frictions in the labor market. Foreign firms are assumed to have lower search costs than domestic firms. Domestic firms can choose to pay an additional fixed cost to make themselves attractive as targets to multinationals and consequently be acquired, which lowers the firm s search costs. In this setup, foreign-owned firms pay higher wages than domestic firms. More importantly, the model predicts that this direct wage effect of foreign ownership is higher the larger is the share of foreign ownership in a cluster. The model also predicts negative indirect effects on the wages paid in domestic firms. Again, this effect is stronger the larger is the presence of foreign firms in a cluster. We implement our econometric approach using firm level data for Chinese manufacturing. China is an interesting case to look at, given the importance of inward FDI in the economy. Since its accession to the WTO in 2001, China has liberalized policy towards foreign direct investment inflows substantially in order to attract foreign investors (e.g., Chen, 2011, Long, 2005). In 2014, the Chinese economy was host to about 10 percent of the global stock of foreign direct investment and received 19 percent of world-wide FDI inflows (United Nations, 2015). 2 The approach is adopted from Girma et al. (2015), who look at productivity effects of FDI. We focus on wage effects in this paper, relying on a different theoretical set-up that identifies potential mechanisms at work. 3 By way of contrast to our methodology we also experiment with a simpler, albeit naïve, approach of estimating the direct and indirect effects of foreign ownership within a single framework. The details are provided in Appendix C. 3
Hence, an investigation of the wage effects of additional incoming FDI and their dependence on how many firms in a sector are already foreign owned ought to be of high interest to policymakers. Our empirical exercise shows that there are positive direct treatment effects of foreign ownership. While this result is frequently found in the literature, we show that the direct effect strengthens with increasing presence of foreign firms in a cluster (measured by employment). This is in line with our theoretical prediction. We also show that the indirect (spillover) effect varies with the strength of the externality, as well. For low levels of foreign presence in a cluster, the estimated effect is low but positive, while it turns negative (and increasingly so) once the presence of foreign firms is higher than about 20 percent. Combining the estimated direct and indirect effects we find that the total effect of foreign ownership on wages is positive and depends on the level of foreign presence. It is highest when the share of employment by foreign owned firms is around 25 percent. The remainder of the paper is structured as follows. In Section 2 we define the direct, indirect and total treatment effects of foreign ownership on wages which we seek to estimate in this paper. Section 3 develops a theoretical framework which motivates our empirical exercise. Section 4 presents details on our empirical strategy to estimate the treatment effects. In Section 5 we describe our data set and provide some preliminary analysis. Section 6 then presents the results of our estimation of treatment effects. Section 7 concludes. 2 Defining Treatment Effects Recent work in the statistical and econometric literature has sought to estimate treatment effects in the presence of interactions (e.g., Hudgens and Halloran, 2008, Manski, 2013). Accordingly, treatment externalities (or spillovers) occur when an individual s potential outcome is affected by other individuals treatment status within a group or cluster. The approach we take, following Hudgens and Halloran (2008), is to use the employment-weighted proportion of foreign firms within a well-defined cluster as a measure of interaction between individual firms. This is similar to what is usually used in the literature on wage spillovers from FDI (e.g., Aitken et al., 1996). The potential outcomes are thus expressed as a function of the firm s treatment status (i.e., foreign or domestic-owned) and the employment share of foreign-owned firms in a particular cluster. We define r = 1... R sufficiently heterogeneous economic clusters. 4 A key identifying assumption is that we only allow for interactions between firms in the same cluster. This definition is based on the empirical findings in the literature that spillovers are strongest within particular regions or broad set of industries (e.g., Driffield and Girma, 2003). For ease of exposition we assume that there are i = 1... N firms in each cluster and N r of these receive the treatment, i.e., 4 As detailed below, our study defines clusters in terms of province two digit SIC industry combinations. See also Table A1 in the appendix. 4
in our case, are foreign-owned. Defining a binary treatment variable d ir = 1 if firm i in cluster r receives foreign ownership and d ir = 0 if not, the (employment-weighted) proportion of treated firms in the cluster is s r = Lr = N i d ir emp ir L L emp ir employment in firm i situated in cluster r., with L being total cluster level employment and There are two potential outcomes, w 0 and w 1, corresponding to the wages for treated (foreign) and untreated (domestic) firms. Only one of those can be observed for any one firm. The potential outcomes under the two treatment states are defined in terms of the individual s treatment status and s r. That is wis d = wis(s d r ); d = 0, 1 and s r [0, 1]. (1) One can then define the average (across independent clusters) potential outcomes corresponding to the two treatment states as w d s = E r [ w d (s r )]; d = 0, 1 and s r [0, 1]. (2) Once the average potential outcome w d s by cluster are constructed for d = 0, 1 and all relevant values of s r [0, 1], we can define the direct and indirect causal treatment effect parameters as differences between the two average potential outcomes. To be specific, for a given level of foreign presence s, the direct treatment effect parameter is given as γ 10 ss = w 1 s w 0 s (3) This is the difference in wages between treated (foreign) and non-treated (domestic) firms for a given level of foreign presence in a cluster. The second parameter of interest is the indirect treatment effect on the non-treated, again keeping the cluster-specific proportion of treated firms at s. This is defined as γ 00 s0 = w 0 s w 0 0 (4) This, hence, gives the difference in wages between domestic firms in a cluster with foreign presence level s and in a cluster without any foreign presence. In other words, the above parameter is defined as the change in the potential outcome of non-treated firms resulting from increasing the cluster-specific proportion of foreign firms from 0 to s. 5 treatment effects, we can also calculate a so-called total treatment effect as Based on these two γ 10 s0 = w 1 s w 0 0, (5) which captures the change in the potential outcome of foreign firms when the proportion of 5 Note that the benchmark treatment levels need not be 0; comparisons can be made between any two treatment levels. 5
foreign firms in the cluster is s > 0 compared to the outcome for non-treated firms that would occur if s = 0 (i.e., no foreign firms in the cluster). In other words, it gives an estimate of the wage premium for a treated (i.e., foreign) firm in a specific cluster compared to a non-treated (domestic) firm in a cluster with no foreign presence. The total effect can be calculated as the sum of direct and indirect effects. 3 Theoretical Motivation Before we move on to the details of how we estimate the direct and indirect treatment effects we present a simple theoretical framework which illustrates a potential mechanism by which foreign ownership may affect wages in domestic and foreign firms. This motivates the empirical analysis in the subsequent sections. Our model builds on Helpman et al. (2010) and assumes that firms are heterogeneous with respect to productivity and that there are search frictions in the labor market. Workers are ex-ante homogeneous, but there is a match-specific productivity draw whenever a worker is paired with a firm. Firms have the ability to screen workers in order to avoid low-productivity matches. Expanding on Helpman et al. (2010) we assume that a firm s screening cost depends on its ownership structure. In particular, following a vast literature on the difference between locally owned and foreign-owned firms, we model foreign firms (multinationals) to have a cost advantage when screening the labor force. The idea behind this assumption is that multinationals have on average better management practices (Bloom and Van Reenen, 2010) which are also reflected in better hiring practices. They may have centralized human resources departments with specialists that have plenty of experience hiring on international markets. This may put them at an advantage vis-à-vis an average firm that only operates domestically. 3.1 Basic Set-Up To illustrate the mechanisms at work we focus on one country with a continuum of risk neutral workers. The standard setup of CES-preferences with monopolistic competition in the final good market yields the following expression for equilibrium revenue of a firm producing variety j. r(j) = Aq(j) β with 0 < β < 1 (6) where A = E 1 β P β (7) is a demand factor exogenous to the firm and 1/(1 β) is the elasticity of substitution between varieties. Total expenditures on differentiated goods are given by E and the price index for differentiated goods is given by P. Furthermore, demand for good j from (6) is q(j) = A 1 1 β p(j) 1 1 β (8) 6
There is a fixed entry cost that firms have to pay in order to remain in the market given by f e. Firm productivity is determined by a draw of the parameter θ, and the distribution is Pareto. A second set of fixed costs determines the mode of market entry. Here, the firms face the following choice: either they choose a lower profile which results only in the usual market entry costs of f d but also implies that the firm will not be able to partner up with a multinational. Or the firm pays f d plus an additional fixed cost f m and, as a result, becomes attractive to multinational corporations and is consequently acquired. 6 This additional cost can be interpreted as reflecting investments by local firms undertaken with the specific purpose of being able to cooperate with multinationals. In other words, the local firm attempts to make itself attractive to a multinational buyer, and this is a costly process. Many local business support agencies provide advice for businesses as to how to make themselves attractive to investors, which entails the kind of preparatory actions and expenditures we have in mind here. 7 A firm s output depends on its productivity, a measure of workers hired and the average ability of those workers: y = θh γ ā (9) Firms begin the hiring process by posting vacancies. The workers that are subsequently matched with the firm are subject to screening. Upon being interviewed, each worker draws a matchspecific ability parameter a, which is also assumed to be distributed Pareto with shape parameter k. 8 The firms choose a level of screening effort by setting a threshold a c. Workers with a match-specific productivity below a c will be recognized and not hired. Screening costs increase in a c and also differ depending on whether the firm is part of a multinational or not. Screening costs are given by c La δ c and c M a δ c for locally owned and foreign-owned firms, respectively, and δ δ c L > c M. One can show that a firm sampling n workers and choosing screening intensity a c ends up with a measure of workers given by h = n ( amin a c ) k (10) 6 Note that we model the difference between foreign- owned and domestically owned firms only insofar as it relates to the labor market. The firm s decision whether to incur the additional fixed cost f m will hence solely depend on the benefits it gets from lower screening costs. 7 See, for example, http://www.finance.scotland.gov.uk/prepare/ready/ready/how-to-make-your-business. Javorcik and Spatareanu (2009) show in survey evidence that local firms invest in obtaining ISO certification in order to become suppliers to multinationals. While this is a vertical relationship that is not our focus here, it shows that our assumption does not appear to be a priori implausible. 8 Throughout we assume that 0 < γk < 1 in order to ensure that firms have an incentive to screen (see Helpman et al. (2010) for discussion). Intuitively, firms may find it disadvantageous to screen if i) the importance of number of workers relative to their (average) productivity level in the production function is high (high γ) and ii) there is low dispersion of match-specific productivity levels (high k). 7
and average ability of the hired workers is given by a c ā = k k 1 (11) Note that a min is the minimum of the support for the Pareto distribution of the match-specific productivity draws by workers. The firm maximizes profit given by π(θ) = Here κ y = max n 0,a c a min,i m {0,1} k k 1 aγk min 1 1 + γβ A(κ yθn γ ac 1 γk ) β bn (1 I m ) c La δ δ I c M a δ m f d I m f m (12) δ is a constant and b is the cost of sampling one worker. The firm chooses the number of workers to sample n, the screening threshold a c and whether to invest in being acquired by a multinational (and thus lower its screening costs) or not. Note that the factor 1 1+βγ is a result of bargaining with the workers: it represents the share of revenue that the firm obtains. For now we continue with the case of the locally owned firm (I m = 0). The first-order conditions are given by and β(1 γk) 1 + βγ r(θ) = c La c (θ) δ (13) βγ r(θ) = bn(θ). (14) 1 + βγ Here we used the fact that the firm s revenue is given by r(θ) = A(κ y θn γ a 1 γk c ) β. The total wage bill w(θ)h(θ) is a constant fraction of firm revenue, so we derive w(θ) = βγ ( ) k r(θ) 1 + βγ h(θ) = bn(θ)/h(θ) = b ac (θ), (15) using the second first-order condition as well as (10). In addition, we can solve for the screening threshold and the number of sampled workers: a c (θ) = n(θ) = ( Aβ ) 1 1 + βγ ( Aβ ) 1 1 + βγ a min 1 γβ (1 γk) γ γβ (κ y θ) β (16) c L b β(1 γk) (1 γk) βγ+ γ (κ y θ) β. (17) c L b Here (1 βγ) β(1 γk) > 0. Firms with higher revenue (due to a better productivity δ draw) choose a higher screening threshold and sample more workers. This, in turn, raises their wage, because they hire a higher quality workforce. Note that a drop in the screening cost from c L to c M indicates a jump upwards in both. All of the firm s decisions depend on its initial productivity draw. 8
In equilibrium, we define the key productivity cut-offs θ d and θ : Only firms with a productivity draw of θ d or higher will enter production and only firms with a productivity draw of θ or higher will enter production and pay the fixed cost to attract multinational investment and lower their screening costs. Furthermore, we use θ dm to denote the level of productivity in between the two where revenue net of fixed costs is exactly sufficient to cover the sum of both fixed costs, f d + f m (but a firm would still prefer to remain locally owned due to higher profit). We derive solutions for n, a and w for local (L) and multinationals (M) firms which only depend on the ratio of firms productivity levels and the minimum possible levels that need to be reached to enter either mode of production, θ d for locally owned firms and θ dm for multinational firms. Details are in the appendix. 3.2 Wage Effect of Foreign Ownership 3.2.1 Direct Wage Premium We established above that a firm s wage depends on its screening threshold: Once a firm pays the additional fixed cost, becomes foreign owned and benefits via a more sophisticated screening process, it sets a higher screening threshold and ends up paying higher wages to its (on average) more productive work force. Figure 1 depicts the relationship between initial productivity draw, choice of ownership and subsequently chosen wage. [ Figure 1 about here ] Firms with productivity draws between θ d and θ will enter the market but not pay the fixed cost required to attract foreign investors. Firms that draw a productivity level of θ or higher instead do pay the additional cost, attract foreign investors and gain a screening cost advantage. They consequently pay higher wages to their workforce. As we saw above, more productive firms will have foreign ownership while less productive firms remain domestically owned. Helpman, Melitz, Yeaple 2004). This pattern is well documented in the literature (e.g., We want to go beyond the claim that a more productive, foreign owned firm will pay a higher wage than a less productive, domestically owned one. We are particularly interested in the effects of a stronger or weaker presence of multinationals in a particular sector and location (in order to make predictions concerning the two treatment effects derived in Section 2). Within our context, this relates to the level of the productivity threshold θ relative to the entry threshold θ d. In what follows, we denote θ d = ρ < 1, as in Helpman et al. (2010). Sectors with strong θ international presence are sectors with relatively high values of ρ, since this implies that a larger number of firms have invested to become an attractive target for multinationals (the two productivity cut-offs are closer together). Sectors with low ρ will have a relatively small share of foreign-owned firms (here the two productivity cut-offs are further apart, leaving a larger mass of firms that enters without investing in attracting foreign multinationals). The question 9
thus becomes: How does the wage premium of foreign ownership depend on the presence of other foreign firms? Or, within the model, what is the link between this wage premium and ρ? Since ρ is endogenous we have to deepen the analysis. Differences between sectors that would result in different values for ρ stem from one of the two parameters: the degree of screening cost saving reflected by c M cl or the cost of the fixed investment required to be acquired by a multinational and gain access to better screening technology f M. First, we depict the wage premium using the ratio of the wage paid by the average foreignowned firm w M and the wage paid by the average domestically owned firm w L. w M w L = wm w d ( θd θ = wm w ρ βk d ) βk 1 ( θd ) βk z 1 θ 1 ρ βk z 1 = wm 1 w d ρ z ρ βk We assume z > βk, ensuring a strictly positive value for the wage ratio. Here wm is the wage paid by a firm with θ = θ that chooses to lower its screening costs. Note that the wage premium depends only on two ratios. The first is the ratio of the wages paid by the least productive firm in each group, that is firms with draws of θ d or θ, respectively. The second is the direct ratio of these two productivity cutoffs given by ρ. The wage premium of foreign owned firms depends positively on ρ: The closer are the two cutoff productively levels, the smaller the support [θ d, θ ] of productivity draws that lead to domestically owned producers. But it is obvious that a decrease in θ for any reason increases ρ, so we focus on the first factor ( wm ) in what follows. w d One can show that ( ) w M k fd + f m c ( ) δ L θ kβ =, (18) w d f d c M θ dm which indicates that the difference in wages paid by the two kinds of marginal firms fundamentally depend on the additional fixed cost and the relative screening costs. Let us look at the latter first. Intuitively, a lower relative screening cost for multinationals (i.e., a decrease in c M /c L ) leads to a higher screening threshold and wages paid by foreign-owned firms. It also moves the productivity cut-off, since firms that previously were exactly indifferent between local and foreign ownership now find it beneficial to make the investment in order to obtain the lower screening cost. As a result, the average wage of domestically owned firms falls unambiguously and the average wage of foreign owned firms rises. To show the second part of that sentence 10
more formally, we recall that the average wage paid by foreign owned firms is given by ( θ w M = w M = w M [ θ z θ z βk ) βk dgθ (θ) (19) ( ) θ min z ]. (20) The decrease in c M /c L lowers θ, so it increases the term in brackets. For the first factor (the wage paid by a firm with θ = θ that invests in the lower screening costs) we need to examine the cases of an increase in c L or a drop in c M separately. All other things equal, a drop in c M will always raise the wage paid by any foreign owned firm due to the resulting higher average quality of the work force. An increase in c L will lower w M, however one can show that this effect on average foreign owned wage w M second factor (details are in the appendix). θ will always be outweighed by the increase in the To summarize, the average wage paid by foreign owned firms w M increases as a result of a drop in c M and/or an increase in c L. This is one of two theoretical reasons why we would see a stronger foreign presence in a given sector provided by this model (because it moves the cut-off productivity threshold θ ). The other theoretical reason could simply be differences in f M : consider the case of local firms facing lower costs of attracting foreign investors. For example, there may be particularly low barriers to entry for foreign firms in certain sectors. As it turns out, the model s predictions regarding the wage premium wm are ambiguous in this w d case, though we can find conditions under which the wage premium rises (both the average wage paid by locally owned and foreign owned firms declines, so the question becomes which declines by more). Interestingly, a downward movement in the cut-off level θ leads to a bigger wage premium especially in those cases where the initial share of foreign-owned firms is already high. This is set out in the appendix, in order to save space. 3.2.2 Effect of Foreign Presence on Wages of Locally-Owned Firms We can also use the model to investigate the effect of an increase in the number of foreign-owned firms on the wages paid by the average locally-owned firms w L : θ ( θ w L = w d θ d = w d z z βk θ d ) βk dgθ (θ). [( ) θ min z ( ) θ min z ] Since the change in θ does not impact the required productivity level for entering production θ d, the expression shows that there is no need to separately analyse the two different channels θ d θ 11
through which the foreign-ownership cut-off may be affected. We find and ( w L θ = z ) 2 θ min z wd z βk (θ z+1 > 0 (21) ) ( 2 w L θ = ( z z ) 2 θ min z 2 1)wd z βk (θ z+2 < 0. (22) ) So an increase in the share of foreign ownership in a sector will always decrease the wage paid by the average locally owned firm. This effect is stronger for small θ, that is when the proportion of foreign-owned firms in the sector is already high. Note that the intuition for this effect does not relate to technology transfer or competition effects, as usually assumed in the interpretation of some empirical results on wage spillovers (e.g., Aitken et al., 1996). Rather, in our model, an increase in the number of foreign-owned firms implies that the remaining domestic firms have on average lower productivity and hence pay lower wages. And this effect increases in the number of foreign firms due to the curvature of the Pareto distribution. 3.3 Summary of Empirical Predictions The model generates predictions regarding the wages paid by foreign-owned versus locallyowned firms. The first variable of interest is the ratio (or difference) of the wage paid by the average foreign owned firm (i.e. firms with draws of θ or higher) to the wage paid by the average locally owned firm (firms with draws of θ or lower). treatment effect parameter in Section 2. This is similar to the direct We are especially interested in how this premium differs depending on the size of the presence of foreign-owned firms in the sector. Prediction 1: The wage premium paid by foreign- versus locally-owned firms is larger if the share of foreign-owned firms in a sector is relatively large. In the econometric parlance used in Section 2, this says that the direct treatment effect should be higher the higher is the level of foreign presence in a cluster. This can be either due to an improvement in the screening cost (in this case, a larger number of foreign owned firms decreases the wage paid by the average domestically owned firm and increases that paid by the average foreign-owned firm) or low costs of attracting foreign investment, f M. (in this case, a larger number of foreign owned firms decreases the wage paid by the average domestically owned firm as well as the wage paid by the average foreign-owned firm, though the net result is still an increase in the wage premium if the productivity dispersion across firms is small.) Unfortunately, we cannot distinguish these two theoretical channels with the data at hand in our empirical analysis. Prediction 2: The second prediction relates to the wages paid by locally owned firms, i.e., the indirect treatment effect parameter in Section 2: The wage paid by the average locally-owned firm decreases as a result of an increase in the share of foreign-owned firms. This effect is 12
stronger the larger is the share of firms that are already foreign-owned. In other words, the negative indirect effect on the non-treated should further decrease the larger is the presence of foreign firms in a cluster. 4 Data Description Our empirical analysis draws on firm level data from the Chinese manufacturing industry. The dataset is based on the Annual Reports of Industrial Enterprise Statistics, compiled by the China National Bureau of Statistics. The dataset covers all firms in China with an annual turnover of more than 5 million Renminbi (about $800,000). These companies account for an estimated 85 90 percent of total output in most industries. For the purpose of this analysis, we have information on more than 147,000 firms over the period 2003-2006. Our treatment is defined as a firm having received foreign investment in 2005; the outcome variable which is average wage is measured in 2006 relative to their pre-treatment (i.e. 2004) values. This helps remove bias due to time invariant firm-specific unobservable effects, and strengthen the conditional independence assumption that underpins much of the microeconometric evaluation literature. The main novelty of our paper is to allow for externalities from foreign ownership within a well-defined cluster. The construction of clusters is not an exact science. Ideally clusters should be constructed in such a way as to maximize the potential of intra-cluster spillovers, while at the same time minimize possible inter-cluster externalities. In doing so, one would need to strike a balance between having a large enough number of clusters and sufficient observations per cluster. In this paper we classify firms into clusters based on the intersection of 11 geographic areas and 13 industry groupings. We impose the condition that at least 4% of firms in a cluster should be foreign-invested and the total number of firms should not be less than 100. For this reason we have to leave out 16 clusters from the analysis. 9 Table 1 provides the definition and summary statistics of pre-treatment covariates by foreign ownership status. Perhaps not surprisingly, the raw data suggest that foreign firms tend to be larger and more productive, and pay higher average wages compared to their domesticallyowned counterparts. Hence, it is important to control for these firm level characteristics in our estimation which deals with selection at the firm level. [Table 1 here] Table 2 provides some summary statistics of the cluster level variables. Average employment share of foreign firms in 2005 stood at 26.6% with an interquartile range value of 24% suggesting quite substantial inter-cluster heterogeneity. Table 3 shows that the proportion of foreign firms 9 Table A1 in the appendix lists the geographic and industry groups which underpin our cluster formation; while Table A2 gives the employment share of foreign firms and the share of foreign firms s in the top and bottom 10 FDI attracting clusters. 13
and their employment share have both significant (unconditional) correlations with all of the cluster level variables. This indicates the importance of controlling for cluster level differences in the empirical analysis. 5 Empirical Implementation [Tables 2 3 here] We now turn to the details of the empirical implementation in which we attempt to calculate the treatment effects defined in the section 2. To do so, we need to estimate the potential outcomes defined in equation (2). This estimation is complicated by the fact that we are likely to have selection on both the firm and the cluster level as indicated by the summary statistics above. This needs to be dealt with in order to obtain unbiased estimates. Firstly, the decision about firm level treatment d is unlikely to be random. Rather, it is the case that foreign investors pick particular types of domestic firms as targets for takeovers. The literature generally argues that there is cherry picking, i.e., foreign investors prefer well performing domestic firms (e.g., Girma et al., 2015a for China). The second source of selection concerns the cluster-level employment share of foreign owned firms, s. It is unlikely that this is randomly distributed, as foreign investors may prefer certain industry-region pairs for their investments. 10 In order to deal with these two levels of selection, our empirical approach proceeds in two steps. We firstly estimate the potential outcomes per cluster based on firm level data considering the firm level treatment d, and in a second step take into account the cluster level share s as treatment. 5.1 First step estimation We first estimate the average potential outcomes w 1 s and w 0 s corresponding to the two states (foreign owned and non-foreign owned) for each cluster. We identify the expected individual outcomes for the two treatment states per cluster by regressing firm level wages on the treatment status. In order to take into account selection at the firm level, we estimate the outcome equation using inverse propensity-score weighted regression and controlling for the pre-treatment covariates, a so-called doubly-robust estimator (Bang and Robins, 2005, Hirano et al., 2003). 11 For each cluster, this implies that we firstly generate the firm-specific propensity-scores (p) of being treated via a logistic regression with a rich list of pre-treatment covariates X subject to balancing conditions being satisfied. The list and precise definition of the pre-treatment covariates can be found in Table 1. Results are reported in Table A3 in the appendix. As we have 127 propensity score estimations, we report summary statistics for the estimated 10 See Table A2 for evidence on this. 11 The estimator is doubly robust as it provides two opportunities to adjust for selection on observables by combining inverse probability reweighting with regression covariates adjustment. 14
coefficients. These corroborate the pattern suggested in the summary statistics productivity, wages, size and export activity are positively correlated with foreign ownership. The last column of Table A3 also shows that propensity score conditioning has done a remarkably good job at balancing firm level observable covariates across the two groups of firms. Using the obtained propensity scores we then estimate the following outcome equation by cluster via inverse probability weighted regression. We impose the common support condition to ensure that the propensity score is balanced across domestic and foreign firms: y is = α + βd is + F (X; δ) + error; i = 1...N. (23) where F (.) represents a function of pre-treatment covariates vector X. 12 In the inverse probability weighting, treated firms receive a weight of 1/p and non-treated firms 1/(1 p). From the regression we can then calculate the cluster specific potential outcomes for the average treated (1) and non-treated (0) firm in a cluster with foreign presence s as w 1 s = 1 N N [ˆα + ˆβ + F (X; δ)] and w s 0 = 1 N i=1 N [ˆα + F (X; δ)] (24) i=1 Note that the wage variable is defined as the change relative to the treatment period, akin to using a difference-in-differences strategy combined with propensity score matching. average difference between the two potential outcomes in (24) would be an estimate of the average treatment effect in the absence of externalities, i.e., if the cluster-specific level of foreign presence did not matter (in other words, if SUTVA is assumed to hold). Figure 2 plots this cluster-specific difference ( w 1 s w 0 s). Under SUTVA, these differences should be independent of cluster specific treatment intensities. While there is no clear picture, this assumption is unlikely to be satisfied by the data. [Figure 2 here] In order to investigate whether there is a causal relationship between the cluster specific treatment intensities s and the cluster specific potential outcomes, we implement a second step estimation. This then allows us to generate the direct and indirect treatment effects as described in Section 2. 5.2 Second step estimation In the second step of the analysis we follow Hudgens and Halloran (2008) and treat w 1 s and w 0 s estimated in the first step as the outcome variables. The employment-weighted proportion of foreign firms in the cluster is taken to be the treatment variable which in this case is continuous between 0 and 1. 13 The In order to control for selection at the cluster level, we employ 12 In Table A5 in the appendix we present summary statistics of estimated coefficients from these regressions. 13 See Ferracci et al. (2014) and Girma et al. (2015) for a similar approach. 15
the causal inference approach for continuous treatments (Hirano and Imbens, 2004). A key result from this literature is that causal inference can be conducted by conditioning on the generalised propensity score (GPS), which is essentially the conditional density of the treatment given some pre-treatment balancing covariates. As we have a dosage variable s r that is continuous and bounded between 0 and 1 we generate the GPS conditional on pre-treatment cluster level covariates, say Ĝr, using the fractional logit model (Papke and Wooldridge, 1996). A full list of these cluster level variables can be found in Table 2. Marginal effects from the fractional logit model and the accompanying covariate balancing tests are reported in Table A4 in the appendix. In line with the correlations in Table 3, we find that the share of employment in foreign firms is higher in clusters with higher average wages, lower average age of firms, lower tax rate and a higher number of foreign owned firms. We also report results from covariate balancing test in this second step estimation in the final two columns. 14 We show the average p-values from these tests (null hypothesis: there is covariate balancing), and it is reassuring to see that conditioning on GPS has done a very good job at covariate balancing. What remains then is to calculate cluster level potential outcomes conditional on Ĝ r and s r using polynomial approximation (Hirano and Imbens, 2004). In this paper we use the following quadratic approximation 15 E[wr Ĝr, d s r ] = β 0 + β 1 Ĝ r + β 2 s r + β 3 Ĝ r s r + β 4 Ĝ 2 r + β 5 s 2 r (25) with sample counterpart over R clusters obtained as w d r = 1 R R ( ˆβ 0 + ˆβ 1 Ĝ r + ˆβ 2 s r + ˆβ 3 Ĝ r s r + ˆβ 4 Ĝ 2 r + ˆβ 5 s 2 r) (26) r=1 The so calculated potential outcomes allow us to generate sample-counterparts of the direct and indirect treatment effects parameters described in Section 2. 14 In order to check for covariate balancing conditional on generalized propensity scores (GPS), we adopt the blocking approach proposed by Hirano and Imbens (2004). Accordingly we first divide clusters into 3 discrete groups defined by treatment intensity (i.e., employment share of foreign firms), and then create 5 blocks within each such group based on the estimated GPS quintiles. Within each block, we carry out differencein-means tests between a treatment intensity group and the other two groups combined by conditioning on the GPS. This process is repeated for each of the 3 treatment intensity groups and 9 covariates, entailing a total of 135 balancing tests. 15 Note that individual parameters from such polynomial approximations do not have any behavioural interpretation (Hirano and Imbens, 2004). 16
6 Wage Effects of Foreign Ownership with Externalities - Main Results We plot the estimated direct and indirect treatment effects along with their 95% confidence intervals in Figure 3. We find that the employment share of foreign-owned firms in a cluster matters significantly, both statistically and economically. [Figure 3 here] Our first finding is that the direct average treatment effect, i.e. the wage premium due to having foreign ownership, is always positive. While a positive wage premium is generally found in the literature, we also find that the premium differs strongly as one varies the employment share of foreign firms in the cluster. The higher is the presence of foreign firms in a cluster, the higher is the difference between the wages paid by foreign compared to domestic firms. This is in line with the prediction from our theoretical model above. In the context of our model, note that for a given distribution of productivity levels in a sector a larger share of foreign-owned firms implies a lower cut-off level. Ceteris paribus, a lower cut-off level always implies lower average productivity (and thus lower average wages) for both kinds of firms via a composition effect (the domestically owned firms lose their most productive members and the foreign-owned firms gain new members who are less productive than the rest). But there are two possible mechanisms that lead to a larger share of foreignowned firms, the first of which may overturn the composition effect for the foreign-owned firms. Consider the case of an increase in the screening cost advantage enjoyed by multinationals. In that case, foreign-owned firms will raise their screening threshold and make more productive matches. This counteracts the composition effect and may lead to overall higher wages, while those paid by domestic firms still unambiguously fall. The relative size of these effects depend on the sector s initial position on the Pareto CDF the more foreign-owned firms there already are, the more the screening effect will dominate the composition effect, resulting in foreign firms hiring more productive workers and paying more while domestic firms become less productive on average and pay less. A second mechanism works through lower fixed costs of becoming foreign owned. As these fall, the productivity cut-offs change and the cut offs for entry of domestic and foreign firms move closer together. As the most productive domestic firms switch to becoming multinationals, this increases the difference in wages between domestic and foreign firms in particular for high shares of foreign presence (again due to the composition effects, which in turn depend on the CDF curvature). Unfortunately, with the data at hand we are not able to distinguish between these two mechanisms in our empirical analysis. In the second panel in the graph we look at the indirect effect, i.e., the impact the presence of foreign firms has on wages in domestic firms so-called wage spillovers (Aitken et al., 1996). 17
Recall that our theoretical model predicts that this should be negative and more so the higher is the presence of foreign firms in a cluster. This prediction is largely borne out by the data. However we find indirect effects that are low but positive, or statistically insignificant, for low levels of foreign presence in a cluster. Only for clusters with above average foreign presence (which stands at about 20 percent in the data) do they turn significantly negative. In line with our model, the negative effect becomes stronger the larger the foreign presence in the cluster. In the context of the literature on FDI, we can interpret this result as indicating positive wage spillovers on domestic firms if the presence of foreign firms is not too high. For clusters with a high share of foreign presence, we find negative wage spillovers. This is not unusual in the literature on spillovers, which in many cases finds negative effects of the presence of FDI on productivity or wages of domestic firms (see Görg and Greenaway, 2004) and is generally regarded as indicating negative competition or market-stealing effects from multinationals. Note however that the mechanism we highlight in our theoretical framework is different, as it does not depend on competition effects but on selection. A larger share of foreign owned firms implies that the remaining domestic firms are the less productive ones, which also pay lower wages. Unfortunately, with the data at hand we cannot distinguish these two theoretical explanations. In line with our theoretical predictions, we have seen that the direct (indirect) average effect of foreign ownership on wages increases (decreases) uniformly with the degree of foreign presence in the cluster. From an aggregate welfare point one may wonder whether there is an optimal presence of foreign firms in the sense that the economy-wide wage effect of foreign ownership is maximized. To answer this question we turn to the total treatment effect estimator defined in Section 2. By summing the direct and indirect treatment effects we can calculate a total treatment effect, which is shown in Figure 4 along with bootstrapped 95% confidence intervals. [Figure 4 here] This total treatment effect is always positive but has an inverted u-shape in relation to the share of foreign presence, reaching its maximum value when the employment share of foreign firms is around 25 percent. Specifically, at this level of foreign presence, the total foreign ownership wage effect stands at about 20 percent. In other words, the wage paid by the average foreign firm in a cluster with a foreign employment share of around 25 percent is about 20 percent higher than that paid by the average domestic firm in a cluster with no foreign presence. As this average effect considers both domestic and foreign firms, we may cautiously interpret this as an optimal level of foreign presence from a welfare point of view. We defined our measure of externalities based on the employment share of foreign owned firms. We chose to do so in order to link our analysis to the existing literature on wage spillovers from foreign firms, which defines spillovers in such a way. In our theoretical model it is, however, the number of foreign firms irrespective of size that defines the strength of 18
the externality. By way of robustness analysis we estimated the various treatment effects of foreign ownership as a function of the proportion of foreign firms in a cluster (rather than the corresponding employment share). The results are plotted in Figure 5 and they are consistent with results discussed earlier. [Figure 5 here] 7 Conclusion Do multinationals benefit or harm host countries labour markets? This is a difficult question to answer, but an important aspect of any answer would have to say something about the wage effects of foreign ownership. This encompasses at least two issues, which have been mostly treated in isolation in the academic literature. First, what is the direct effect of foreign ownership i.e., do multinationals pay higher wages? And, second, what is the effect of foreign ownership on the wages paid by other domestic firms? In this paper we implement an approach which allows us to estimate such direct and indirect effects of foreign ownership on wages in a unified framework. This approach allows for differences in effects depending on the strength of the presence of foreign firms in a well-defined cluster. It also takes into account selection at both the firm and the cluster level. In order to motive the empirics we propose a simple model extending Helpman et al. (2010). Foreign firms differ in terms of having a better matching methodology with lower screening costs, which explains why they are able to pay higher wages. Foreign presence affects average wages of domestic firms because of selection, whereby only the least productive domestic firms (which pay low wages) remain domestic while other firms get taken over by foreign owners. The model generates a number of predictions concerning the relationship between treatment effects and the presence of foreign firms in a cluster. We use firm level data from China for our empirical exercise. Results show that there are positive direct effects of foreign ownership foreign firms pay higher wages and that the estimated effects increase with increasing presence of foreign firms in a cluster. This is in line with predictions from our theoretical model. We also find that the spillover effect on domestic wages varies with the strength of the externality. For low levels of foreign presence in a cluster, the estimated effect is low but positive, while it turns increasingly negative once the presence of foreign firms is higher than about 20 percent. From a policy point of view it may be important to combine the estimated direct and indirect effects into a total effect. Doing so we find that the total effect of foreign ownership on wages is positive and depends non-linearly on the level of foreign presence. The total effect is maximized at a level of foreign presence around 25 percent. For that level, we find that the total effect, which is the difference between the wage paid by a foreign firm in a cluster with a level of foreign presence at 25 percent is 20 percent higher than the wage paid by an average domestic firm in a cluster with no foreign presence. 19
While this number is of course only an illustration for our Chinese data and specific time period, it shows the potential insights one could gain from implementing this approach for other data sets. This may prove very useful for policy makers who aim at maximizing the potential from inward foreign direct investment for their country. 20
References [1] Aitken, B., Harrison, A., and Lipsey, R. E. Wages and foreign ownership: A comparative study of Mexico, Venezuela, and the United States. Journal of International Economics 40, 3 (1996), 345 371. [2] Bang, H., and Robins, J. M. Doubly robust estimation in missing data and causal inference models. Biometrics 61, 4 (2005), 962 973. [3] Bloom, N., and Van Reenen, J. Why do management practices differ across firms and countries? The Journal of Economic Perspectives 24, 1 (2010), 203 224. [4] Chen, C. The development of china s FDI laws and policies after WTO accession. In Rising China: Global Challenges and Opportunities, J. Golley and L. Song, Eds. ANU E Press, 2011, pp. 85 98. [5] Driffield, N., and Girma, S. Regional foreign direct investment and wage spillovers: Plant level evidence from the UK electronics industry. Oxford Bulletin of Economics and Statistics 65, 4 (2003), 453 474. [6] Egger, H., and Kreickemeier, U. Why foreign ownership may be good for you. International Economic Review 54, 2 (2013), 693 716. [7] Ferracci, M., Jolivet, G., and van den Berg, G. J. Evidence of treatment spillovers within markets. Review of Economics and Statistics 96, 5 (2014), 812 823. [8] Girma, S., Gong, Y., Görg, H., and Lancheros, S. Estimating direct and indirect effects of foreign direct investment on firm productivity in the presence of interactions between firms. Journal of International Economics 95, 1 (2015), 157 169. [9] Girma, S., Gong, Y., Görg, H., and Lancheros, S. Investment liberalisation, technology take-off and export markets entry: Does foreign ownership structure matter? Journal of Economic Behavior & Organization 116 (2015), 254 269. [10] Girma, S., and Görg, H. Evaluating the foreign ownership wage premium using a difference-in-differences matching approach. Journal of International Economics 72, 1 (2007), 97 112. [11] Helpman, E., Itskhoki, O., and Redding, S. Inequality and unemployment in a global economy. Econometrica 78, 4 (2010), 1239 1283. [12] Helpman, E., Melitz, M. J., and Yeaple, S. R. Export versus FDI with heterogeneous firms. The American Economic Review 94, 1 (2004), 300 316. 21
[13] Heyman, F., Sjöholm, F., and Tingvall, P. G. Is there really a foreign ownership wage premium? evidence from matched employer employee data. Journal of International Economics 73, 2 (2007), 355 376. [14] Hijzen, A., Martins, P. S., Schank, T., and Upward, R. Foreign-owned firms around the world: A comparative analysis of wages and employment at the micro-level. European Economic Review 60 (2013), 170 188. [15] Hirano, K., and Imbens, G. W. The propensity score with continuous treatments. In Applied Bayesian Modeling and Causal Inference from Incomplete-Data Perspectives (2004), Wiley. [16] Hudgens, M. G., and Halloran, M. E. Toward causal inference with interference. Journal of the American Statistical Association (2012). [17] Huttunen, K. The effect of foreign acquisition on employment and wages: Evidence from Finnish establishments. The Review of Economics and Statistics 89, 3 (2007), 497 509. [18] Javorcik, B. S., and Spatareanu, M. Tough love: Do Czech suppliers learn from their relationships with multinationals? The Scandinavian Journal of Economics 111, 4 (2009), 811 833. [19] Levinsohn, J., and Petrin, A. Estimating production functions using inputs to control for unobservables. The Review of Economic Studies 70, 2 (2003), 317 341. [20] Long, G. China s policies on FDI: Review and evaluation. In Does Foreign Direct Investment Promote Development?, E. G. Theodore Moran and M. Blomstrom, Eds. Institute for International Economics, 2005. [21] Malchow-Møller, N., Markusen, J. R., and Schjerning, B. Foreign firms, domestic wages. The Scandinavian Journal of Economics 115, 2 (2013), 292 325. [22] Manski, C. F. Identification of endogenous social effects: The reflection problem. The Review of Economic Studies 60, 3 (1993), 531 542. [23] Manski, C. F. Identification of treatment response with social interactions. The Econometrics Journal 16, 1 (2013), S1 S23. [24] OECD. Do multinationals promote better pay and working conditions? In OECD Employment Outlook 2008. OECD, 2008, ch. 5. [25] United Nations. World Investment Report 2015. United Nations, 2015. 22
w(θ) Local Foreign owned w M * w d θ d θ * θ Figure 1: Wage and Firm Productivity 23
Figure 2: Cluster-specific average treatment effects of FDI on wages Figure 3: Causal effects of FDI on wages at foreign and domestic firms 24
Figure 4: Total treatment effects of FDI on wages Figure 5: Treatment effects based on proportion of FDI firms in cluster 25