Do Stringent Environmental Policies Deter FDI? M&A versus Greenfield

This study examines how environmental stringency affects the location decision of foreign direct investments. We analyze a firm-level data set on German outbound FDI and innovate on previous studies by controlling for the mode of entry and applying the mixed-logit analysis. The results show that Greenfield projects react to environmental regulation in a strongly different way than M&As. We find robust support for pollution haven hypothesis for polluting Greenfields. M&A investments in low polluting industries, on the other hand, seem to be attracted by stricter environmental regulation. We introduce a new instrumental variable for environmental stringency and apply it to verify the results.


Introduction
The pollution haven hypothesis (PHH) posits that di¤erences in environmental regulation cause the production of dirty goods to relocate from jurisdictions with stringent standards to more lenient locations. This relocation,if present, should be re ‡ected in changes in the international trade patterns as well in foreign direct investment (FDI) ‡ows from …rms that ‡ee highly regulated locations.
Theoretical foundations for the PHH were laid with the help of Heckscher-Ohlin model with pollution as a factor of production. Accordingly, in jurisdictions that provide a low cost of pollution the producers should make intensive use of this factor (Siebert [46], Pethig [41], Markusen et al. [38]).
The models built up at the outset of the PHH literature made rather strong and unequivocal statements about whether the production relocates in response to regulation. Some papers suggested that, given the free mobility across frontiers, even marginal di¤erences in environmental stringencies may induce polluting industries to relocate entirely from high to less regulated economies (McGuire [39]). More recent studies point to the factors that could weaken the pollution haven e¤ect, among others corruption and endogeneity of environmental policy as in Fredriksson et al. [19], small market sizes as in Dong et al. [15] and endogenous market structure as in Elliott and Zhou [16].
If true, the predictions of the PHH have important implications for environmental and trade policies. The carbon leakage and other e¤ects implied by the PHH would render unilateral regulations futile. However, at least when it comes to the FDI channel, the theoretical predictions concerning …rm location have gained only mixed empirical support on the macro and micro level. 2 Notably, early studies like Bartik [6] and Levinson [29], which due to the dearth of international data worked mostly with U.S. new plant locations and performed a cross-section analysis of the aggregated data, were inconclusive about the e¤ect of regulation on FDI ‡ows. The latest studies have brought mixed results. While many studies …nd support for PHH (e.g. Hanna [21], Wagner and Timminis [50], Kellenberg [27], Xing and Kolstad [51], Chung [10]), a considerable share of the publications indicates a lacking or small impact of the environmental regulation on the investment patterns (Javorcik and Wei [26], Dean [13], and Manderson and Kneller [37]). Some studies point at heterogeneous e¤ects for di¤erent types of countries (developed vs. developing) as found by Kheder and Zugravu [28] or di¤erent types of investments (vertical vs. horizontal) as shown by Rezza [45]. Finally, Poelhekke and van der Ploeg [43] put forward the idea that in some industries a reputation for sustainable management and corporate social responsibility (CSR) may be more important than avoiding stringent environmental policy ("green haven e¤ect"). Indeed, they were able to corroborate the hypothesis with the empirical …nding that highly regulated countries attract FDI in machines, electrical and automotive sectors.
Brunnermeier and Levinson [8] and Dong et al. [15] provide reviews of the empirical publications and comment on the mixed results. A possible explanation for the inconclusive results could be the failure of the literature to su¢ ciently account for the heterogeneity of the investment which may dilute the e¤ect. Indeed, heterogeneity has been identi…ed as an important factor that makes …nding evidence for the PHH di¢ cult in trade studies (Levinson,Taylor [32]). Moreover, Hanna [21] points out that the studies which test on the macro level whether the e¤ect of environmental stringency varies across industries with di¤erent pollution intensity could confuse industry speci…c trends in FDI like recessions or changes in consumers taste with regulation.
We avoid some of the potential problems by analyzing individual location choices. An additional advantage of microdata is that, unlike aggregate FDI ‡ow model, it enables one to focus on individual …rms, thus better representing location choices as an individual …rms'pro…t maximization decision. It also allow analyses that are otherwise not possible, such as computing cross elasticities of choosing among alternatives.
Simply looking at individual observations may not be enough without controlling for sources of heterogeneity. It seems to us that distinguishing between the two modes of FDI -Green…eld investments and mergers and acquisitions may be crucial for the sensitivity of FDI towards environmental regulation.
The intuition is that Green…eld projects usually need to obey all the latest environmental requirements whereas M&As involve local …rms that usually, due to grandfathering policies, remain una¤ected by the latest rules and need to adhere to the older regulations only.
The so-called grandfathering of existing sources of pollution is a quintessential feature of the environmental regulation. Objects in operation at the time of the enactment of new regulatory requirements are usually exempted from these requirements or granted a long time for transformation due to the high cost of adjusting their operations and the need for minimizing the general investment uncertainty. One of the examples comes from the U.S. where under the 1970 amendments to the Clean Air Act, Congress decided to subject new sources of air pollution to stringent pollution control standards while grandfathering preexisting sources, leaving them free of federal regulation. In the course of time, some rules were implemented to regulate which expansions of a grandfathered plant subject it to the newest regulation. Nevertheless, the legislation favoured the object existing before 1970 (Nash and Revesz [40]).
The environmental regulation may also enhance pro…ts of existing producers by restricting access to common property and thus creating a scarcity rent. In general, grandfathering regimes give a competitive advantage to the industries, …rms, and regions where the preexisting plants are located. Some quanti…cation of the 'new source bias'is provided by Levinson [30] based on state variation in toxic air pollution regulations in the U.S. and by Ackerman et al. [1] for coal-burning power plants.
Moreover, in the case of an M&A project, the acquisition price may already be a function of the regulation faced by the company as the purchaser of the existing plant is only willing to pay the present discounted value of future pro…ts. This is in analogy to the taxation literature which states that in a high tax country a portion of the tax burden may be capitalized, reducing the acquisition price (Hebous, Ruf and Weichenrieder [23]). Huizinga et al. [25] …nd similarly that additional international taxation in form of non-resident dividend withholding taxes and home -country corporate income taxation is fully capitalized into takeover premiums implying that the incidence of this taxation is primarily on target-…rm shareholders.
Particular regulations may be easier to comply when starting a …rm from the scratch. However, we believe that the above listed reasons may overweight those costs and so we expect Green…eld projects to have a significantly higher sensitivity with respect to environmental requirements than M&A investments.
To the best of our knowledge, the distinction between the two modes of entry has not been taken care of in the literature on the e¤ects of environmental regulation on FDI location. List and Co [34] could be seen as an exception here as they explicitly acknowledge the possible problems connected to the grandfathering rules. However, they do not compare the two investment modes but instead con…ne themselves to Green…eld investments 4 for their estimations. They …nd evidence that environmental policies matter for multinational corporations'new plant location decisions.
The other problem that we see in the hitherto existing literature using the …rm-level data is the potentially inappropriate econometric modeling. The conditional logit approach is by far the most popular one (with some notable exceptions of Poisson regressions, probit and nested logit estimation, e.g. by Dean et. al. [13], propensity score matching estimator model by List et. al. [35] and Millimet et. al. [36], panel data approach [12], di¤erence-in-di¤erence estimation by Chung [10] as well as nonparametric estimation by Henderson and Millimet [24]). The logit, conditional logit and independent probit approaches depend on the underlying independence of irrelevant alternatives (IIA) assumption. The IIA implies that if one alternative became unavailable, the probability of all other alternatives to be chosen would increase proportionally, which strongly restricts the extent to which countries are di¤erent substitutes from the point of view of a foreign investor. The nested logit approach overcomes to some extent the problem of rigid substitution patterns, it requires, however, the researcher to identify the nests which are open to subjectivity.
In general, the IIA assumption may be too restrictive, especially in situations where the number of alternatives in the choice set is large, such as in the model of country destination choice for the FDI. A Hausman-McFadden test that we conducted showed inconsistency of the German data with the IIA and made us turn to estimating a mixed logit model in addition to a usual conditional logit model.
Mixed logit (also known as random-parameters logit) generalizes standard logit by allowing taste variations among individuals. It enables one to control for the fact that companies may attach di¤erent weights to the location factors which in terms of the model involves replacing the coe¢ cients in the regression by i where the i index refers to the parent company-speci…c sensitivity towards the covariate. The econometric approach involves estimation of the so-called deep parameters that describe the moments of the distribution of parameters in the population (the number of those parameters depends on the functional form of the distribution function assumed).
Variance in the unobserved customer-speci…c parameters induces correlation over alternatives in the stochastic portion of utility. Consequently, mixed logit does not exhibit the restrictive substitution and forecasting patterns of standard conditional logit. Additionally, it allows e¢ cient estimation when there are repeated choices by the same decision makers, as it is the case in our application (Revelt,Train [44]).
Our study analyzes additionally the economic signi…cance of the …ndings by looking at the magnitudes of the marginal e¤ects of environmental regulation.
The observation used are all FDIs that were undertaken from Germany in years 2005-2009. The data was obtained from Microdatabase Direct Investment (MiDi) gathered by the Deutsche Bundesbank in accordance with the provisions of the Foreign Trade and Payments Regulation. MiDi keeps a comprehensive account of all the FDIs where the balance sheet total of the foreign direct investment exceeds 3 million Euro and the obtained voting rights are 10% or more. The data contain industry characteristics of both the investing and the target company. Due to the reliability of the data we can exclude any measurement errors for the FDI choice variable. What is rare among FDI data sets, the German data di¤erentiates between the modes of new entries allowing us to account for the investment heterogeneity discussed above.
Since MiDi contains con…dential individual data reports, the use of the database is subject to restrictions; notably, the data may be used only at the premises of the Deutsche Bundesbank.
German investment behavior should be particularly relevant in the PHH context as Germany is one of the largest economies with 10% of the total world exports (Francis [18]) and a share of 5-8% in the world FDIs in the years considered according to UNCTAD data. Chung [10] criticizes utilization of data coming from developed countries in PHH studies on the ground that …rms employing clean technologies in response to the domestic environmental regulations, which is usually the case in the industrialized countries, would have less incentive for outward migration. However, we believe that such claims in general do not invalidate the information coming from data on highly developed countries. Looking from Chung's perspective and knowing that Germany belongs to one of the environmentally most regulated economies, we could formulate our question as follows: given that they have green technologies at their disposal, do the …rms nevertheless want to use the dirty technologies? Additionally, the approach we take -conditioning the results on …rm's decision to go abroad without investigating who and why wants to perform FDI in the …rst place makes the critique even less germane to our study.
There exist already one study (Wagner and Timminis [50]) that explores German (manufacturing sector) investment decisions on the macro level and …nds robust evidence for pollution haven e¤ect for the chemical industry. We believe we can considerably enhance those results by working with individual projects, including more sectors in our analysis and di¤erentiating between modes of entry.
Our main …nding is that investors' preferences for environmental laxity depend strongly on both the mode of FDI and the pollution intensity of the sector. Environmental stringency is shown to reduce the probability of investing in a given country for projects in dirty industries with the deterrence e¤ect being much more pronounced in the case of Green…eld projects. For clean investments of M&A type we …nd evidence that increased environmental requirements may boost the attractiveness of a given location.
The reminder of our paper is split into four parts. Section 2 describes our empirical approach and the data. Section 3 compares the estimation results in di¤erent setups to …ndings in the literature. We also provide some robustness checks, among others we instrument the environmental stringency. Section 4 analyses the economic importance of the …ndings while Section 5 provides concluding remarks.

Methodology and the Data
The theoretical framework of our model is derived from the standard location model for …rms establishing a new a¢ liate in a host country. Like in Head and Mayer [22], we use the partial-equilibrium framework for the equation explaining the determinants of FDI decisions. The parent …rms, after having made up their mind concerning the mode of the investment (Green-…eld vs. M&A) as well as industry they want to invest in (decisions that we take as given) selects the country for the location of its investment. The only criterion applicable for the decision-making is the expected pro…t associated with di¤erent countries -the …rm settles its a¢ liate there where it expects the pro…ts to be the highest possible. 3 For each investment they select one of the 70 countries. 4 The set of potential locations has been determined by the availability of the environmental stringency index and other covariates. 5 The pro…t associated with a given location i = 1; 2 : : : 70 for a company making an investment j is a function of country and investment characteristics and gives rise to following probability function of FDI in a given location: (1) where f is some function appropriate for the chosen econometric model, envI i measures the degree of environmental stringency of the host economy, Greenf ield j is a dummy variable that captures the mode of investment, industry j describes the qualities of the industry in which the …rm is investing and x ij is a vector containing the remaining control variables. The variable of key interest for our study is the degree of environmental stringency of the host economy. It was a contentious issue in the literature how to compare and capture the level of regulation. Lately, though, more and more publications have used the Stringency of Environmental Regulation Index from the Executive Opinion Survey published annually by the World Economic Forum (WEF). The index re ‡ects the perception of Partner Institutes of the Forum (recognized research institutes, universities, business organizations, and in some cases survey consultancies) of the environmental policy run by di¤erent countries. 6 The values of the index range between 1 and 7, where 1 marks lax regulatory standard and 7 indicates a country among the world's most stringent. We employ the policy stringency index in our regressions and interact it with the Environmental Policy Enforcement Index (also from the World Economic Forum) as it was done in Kellenberg [27] and rescale it down by factor 10. A thus-created environmental index may take on values between 0.1 and 4.9. We multiply the policy stringency and enforcement indicators as they are highly correlated and a separate inclusion of both may lead to multicollinearity. Second, we expect a strong 4 For years 2005-2006 they choose between 69 countries due to the unavailability of data for Saudi Arabia. 5 With the resulting collection of countries we cover 94,5% of the investments undertaken by German investors. Some 400 entries had to be dropped, investment in the Cayman Islands constituted there a major group (86 investments).
In the robustness analysis we work additionally with the FDI data for 2009-20011. For those years we manage to gather the covariates for around 120 countries, thus covering around 98,5% of the investments performed. 6 The merits of using the WEF data are discussed thoroughly by Kellenberg [27]. 8 complementarity between the stringency of rules and the intensity of enforcement that should best be captured by interacting the indexes. Figure  1 plots the number of conducted FDIs against the values of environmental index. Intuitively, highly polluting sectors are more likely to be a¤ected by the regulation than clean ones. Additionally, as we have argued, we expect the investment pro…ts of companies entering the market in form of M&A to be less in ‡uenced by the environmental requirements. To account for those effects, we include in our model the interaction of environmental stringency variable with the dummy for the Green…eld investment (Greenf ield j ) and a variable describing pollution intensity of the sector (industry j ). Consequently, the coe¢ cients associated with environmental stringency re ‡ect the e¤ect of the regulation per se as well as its in ‡uence on the composition of the FDIs ‡owing into a country.
The industry j variable was assigned one of the three values: H (high polluting), M (medium polluting) or L (low polluting) depending the sector in which the FDI took place. The assignment of sectors to pollution levels was conducted for manufacturing industries using the German data on the relative pollution abatements costs and the relative green investments completed in 2009 (data taken from Statistisches Bundesamt [47]). To make sure that the data do not re ‡ect some preference of German authorities for particular sectors or lobbying e¤orts but rather the di¤erences in pollution intensity we cross-checked the resulting classi…cation against computations obtained from the analog U.S. data and found no important di¤erences (U.S. Census Bureau, [49]). For services, the classi…cation relies on the data gathered by Levinson [31]. The classi…cation of the industries is presented in table 9 (in the appendix).
We created dummy variables for each of the manifestations of the industry variable (lowP; medP; highP ) and interacted them with dummies for the entry mode (Greenf; M &A) so that the complete regression estimated reads: Such a setup allows for nonlinear changes in sensitivity to environmental regulation when altering the pollution footprint of the investment. Individual components of the interaction -industry and entry mode as well as the interactions between them are excluded from the regression as they are unidenti…ed in that framework. Consequently, the 1 coe¢ cient is to be interpreted as the e¤ect of environmental regulation on the probability of investment for the M&A projects in clean sectors. As the mixed logit choice probability does not have closed form formulation, simulations need to be performed for the estimations. To assure reasonably low simulation error in the estimated parameters 300 Halton draws were used. Train [48] discusses the e¢ ciency of Halton draws compared to random draws and, among others, concludes that with random draws, the simulation variance decreases at a rate of approximately 1=R, where R is the number of draws whereas with the Halton draws, the rate of decrease is faster: doubling the number of draws decreases the simulation variance by a factor of about three.
The resource dependent industries, transportation, mining and agriculture were excluded from the study as we expect the locational characteristics to dominate strongly in these areas and lead to a very di¤erent hierarchy among the drivers of FDI decisions. We analyzed some 6500 new cross-border projects, out of which 37.5% took the form of Green…eld investments. We disregarded expansions of already existing investments. Most of the analyzed projects were conducted in the low polluting industries (73%), geographically they concentrated in Europe (63%) and the Americas (20%). Expanding the data for it to …t the logit and mixed-logit structure gives in total around 459 000 observations.
The observed location decisions are made by 1892 di¤erent companies. On average, a …rm in our sample performs 3,5 di¤erent investments, with …ve …rms performing over 100 investments. 1071 …rms were observed only once in their choice. The distribution of di¤erent types of investments between the most important host countries is shown in …gure 2. The visual analysis of the …gure already seems to reveal some interesting patterns. For instance, while China receives only a small share of clean projects, it is a major host for dirty investments, especially of Green…eld type.
The remaining explanatory variables employed in the main model are typical for the location decision literature: logarithm of GDP per capita (gdp), logarithm of population (population), logarithm of the distance to Germany (distance), The Heritage Foundation index of corruption freedom (corruption f r) and labor freedom (labor f r), the statutory corporate tax rate (ctax) and openness (openness) measured as ratio of summed imports and exports over the country's GDP. F DIstock is measuring the value of the stock of the inward FDIs for a given country (data taken from UNCTAD) and its purpose is to proxy the factor endowments and the agglomeration e¤ects as in Wagner and Timminis [50]. In some speci…cations we control also for the country …xed e¤ects.
The correlation between major variables is shown in …gure 3. The descriptive statistics for the variables are given in table 1 together with information about the data sources.
Importantly, we run the regressions with various other explanatory variables as well. The variables that are included in our preferred regressions presented here were selected on the basis of signi…cance across empirical models or their prevalence in the literature. It should be emphasized that the coe¢ cients of interest (i.e. those re ‡ecting the e¤ects of environmental stringency) proved robust in terms of sign and signi…cance when including (or excluding) additional variables.
For the mixed logit estimations, we assume the coe¢ cients to be inde-

Estimations
The results are presented in table 2. Since we assume that the coe¢cients are normally distributed in the population of …rms we …rst report the mean of the coe¢ cients. Stars attached to these coe¢ cients imply whether …rms'sensitivity towards a covariate is, on average, di¤erent from zero. Secondly, we report the variances of the sensitivity in the population (given in brackets). The variances may be starred as well, indicating whether there is heterogeneity in the population. Insigni…cant variance implies that every …rm reacts in the same way towards a given covariate.
To visualize how using mixed logit a¤ects our prediction compared to the most popular models used in the literature, we report their results in table 3. The structure of that table is analogue to 2 with b representing the linear probability model (LPM), c logit and d conditional logit. 7 This does not restrict the ‡exibility of the model in the sense of departure from IIA. 13 The basic setup (Ia) uses the environmental index (envI), but ignores any possible interactions between the environmental regulation and investment characteristics assuming the same sensitivity pattern for all …rms. In this simple framework the investors seem to have very mixed attitude towards environmental regime but are on average negligent of it with most of the other coe¢ cients staying in accordance with standard predictions. The insigni…cance of the corruption freedom (corruption f r) variable is in contrast to some literature, among others to …ndings of Fredriksson et. al. [19] Kheder, Zugravu [28] and Kneller, Manderson [37]. Presumably this is due to the fact that our environmental index, part of which derives from the enforcement of the regulation, is correlated with the corruption level of the country just as log of FDI stock is (F DIstock). This can be seen in the correlation pattern between the variables shown in …gure 3.
In a next step we interacted the environmental stringency index with pollution intensity of the sector (setup IIa, terms medP #envI and highP #envI) which is the typical speci…cation in the FDI-PHH literature. We notice that sharpened environmental requirements increase the probability of attracting FDI in low polluting industries and reduces the probability for middle and high polluting industries. The coe¢ cients of interest still exhibit signi…cant heterogeneity. The envI coe¢ cient has signi…cantly increased compared to Ia speci…cation, nevertheless, for some 33% of low polluting investments negative weight is placed on the regulatory stringency.    .0028 .0027 .0018 .0017 .0017 .0205 .0205 .0207 .1349 .1351 .1363 .1642 .1675 .1696 Log-likelihood Note: ***, ** and * denote signi…cance at the 1%, 5% and 10% level respectively. Clustered standard errors at the investing company level were used. Number of observations: 459267.
Interestingly, the positive and signi…cant coe¢ cient of envI in IIa suggests that more stringent environmental regulation may increase the attractiveness of a given location in case of low polluting …rms. This may be due to the "green e¤ect" reported by Poelhekke and van der Ploeg [43]: some …rms that put much weight on the sustainable management image and on corporate social responsibility may want to avoid settling in low regulated regions to prevent potential reputation losses. As their expenses for obeying the regulation are probably low (they are in the low polluting sector) this image boosting does not come at a high cost. There could also be competition for input factors between various sectors. High regulatory standards put the polluters at a competitive disadvantage and may potentially deter them from the market. That, in turn, means for low polluters less competition for inputs, such as land and labour force.
The …nal speci…cation (IIIa) uses interactions between mode of entry and pollution intensity to gauge the e¤ect of environmental stringency. The results suggest that investors'perception of the environmental policy is strongly dependent not only on how polluting the investment sector is, but also on the mode of investment, and the di¤erence is signi…cant. Just by looking at the gap in magnitudes of coe¢ cients of interest for Green…eld investments as compared to M&A (e.g. Greenf #highP #envI vs. M &A#highP #envI) one may assume that the two investment types exhibit structurally di¤erent sensitivity towards the environmental regulation. For the medium polluting and high polluting industries the interaction coe¢ cients for the Green…eld investments were 2-3 times higher in absolute terms but pointing in the same direction (negative). At the same time the respective sensitivity of the clean M&A projects was positive.
Introducing the interaction terms leads to the standard deviation of the environmental index coe¢ cient envI becoming insigni…cant. It means that we are thereby able to capture an important part of the heterogeneity in the tastes. Likewise, the standard deviation of the Greenf #highP #envI coe¢ cient is statistically insigni…cant meaning that all pollution intensive Green…eld investments respond in a similar, negative manner to environmental regulation.
Considerable variability in tastes can still be observed for the Green…eld projects in clean industries. As for mergers in highly polluting sectors, around 30% of the …rms exhibit positive value of the coe¢ cient, which corroborates our intuition that M&A are very di¤erent from Green…elds when it comes to pollution regulation. There are probably other …rm and sector characteristics that drive up the variability of sensitivity towards environmental regulation, like the extent of grandfathering, technology used and R&D spending but that could not be controlled for as our information on parent characteristics is limited.
The magnitudes of the estimated standard deviations relative to the estimated means are important for some other variables as well. For example, the very pronounced heterogeneity in the responses towards taxation (37.5% of the companies having a positive ctax coe¢ cient) seems to mirror how differently the pro…ts of various companies are a¤ected by the corporate tax rates and point towards individual issues like tax holidays or possibilities of transfer pricing which are not observed by the researcher. Note that the standard logit model conceals this e¤ect: its slightly negative coe¢ cient for the corporate tax variable would be interpreted as companies shunning hightaxation countries whereas in reality the picture may be more complicated.
The results from the "traditional" regressions reported in the table 3 are congruent to a great extent with the …ndings from the mixed logit model 8 . When comparing the coe¢ cients from conditional logit and randomparameters logit, we notice that most of them are even of similar magnitudes. The major di¤erences appear for coe¢ cients with relatively high estimated standard deviations. Furthermore, openness loses its signi…cance in the mixed logit model. However, based on the likelihood-ratio tests, the likelihood values of all the mixed logit models are statistically di¤erent from their conditional logit counterparts. In other words, allowing the parameters to vary across individual decision makers signi…cantly improves the …t of the model.
In all the estimated models the coe¢ cients for medium polluting and heavy polluting industries are not statistically di¤erent from each other. This could be due to some underlying threshold of pollution footprint above which the …rms become concerned about the regulation. It could be also that our classi…cation of industries into di¤erent polluting categories had some measurement problems.
To ensure that the above described monotonicity in environmental coe¢cients is not just some artifact of using interaction terms in nonlinear models, we investigated the marginal e¤ects and were able to con…rm our statements. The details of this approach are described in section 3.2.

Marginal E¤ects
The e¤ect the regulation exerts on di¤erent types of investments is easily recognizable with the linear model where the deterrence e¤ect clearly increases with the pollution intensity and rises correspondingly for Green…eld FDIs. However, with the nonlinear probability models the interpretation of the coe¢ cients and their comparison is potentially deceptive. We should also stress the problem of assessing the signi…cance of interaction e¤ects in such models pointed out by Ai and Norton [2]. Therefore, to be able to draw precise conclusions on the impact of the pollution prevention on the investment decision, we simulate the marginal e¤ects of environmental stringency -the change in the probability of choosing a particular country when the environmental stringency increases for that country (and remains unchanged in all the other locations). We calculate marginal e¤ects for individuals, average marginal e¤ects (AME) and average conditional marginal e¤ects (CME) of environmental regulation, where CME is de…ned to be the AME within a certain group of investments (e.g., CME Greenf. medium is the average marginal e¤ect for the Green…eld investments in medium polluting sectors).
The results for all the econometric models used are shown in table 4. All industries but the clean ones are on average negatively a¤ected by the regulation and the di¤erence in responsiveness of medium polluting investments as compared to the dirty ones is a rather tenuous one. However, the policy impact is much more pronounced for the Green…eld investments. The clean M&A projects are allured by the increased stringency, for clean Green…elds we observe no signi…cant e¤ects.
Evidence for the pollution haven e¤ect is found in the form of the CME for Green…eld investments in medium and highly polluting sectors -a unit increase in the environmental index lowers the probability of investment by one percentage point. On the other hand, positive CME for M&A in clean sectors implies that such investments tend to be attracted to highly regulated locations. This may be due to the reasons we discussed before (the reputation for sustainable management and corporate social responsibility, competition for local input resources and deterrence e¤ect of changes in environmental stringency on polluting sectors). However, once we di¤erentiate between the modes of entry, a new dimension of the competitive advantage occurs. In the case when an existing …rm is acquired the investor does not have to fear the instantaneous in ‡uence of increased environmental requirements on his operating processes due to the grandfathering rules. The changed regulation will, however, apply to all the companies freshly entering the market, driving the cost wedge between the new units and existing ones. As Gruenspecht [20] points out, that bias against new sources in regulation reduces investment in new facilities and lengthens the economic lifetime of old ones. This e¤ect sustains the longer, the lower the rates of physical deterioration and technical obsolescence in the industry with grandfathering rules. Buchanan and Tullock [9] argue that whenever grandfathering encompasses some assignment of quotas to existing …rms, excess pro…ts may even result in the short term.
Allurement impact is especially visible for the low polluting industries as, we believe, for more pollution intensive sectors the fact that some adjustments need to be done to comply with the altered regulations in the long run prevails over the advantages.
This positive e¤ect, albeit relatively small in magnitude, points to the fact that environmental policy has a bearing on the composition of in ‡owing FDI.
Imposing the sensitivity of the investments reaction to the environmental regulation to be the same for all the …rms blurs the e¤ect. This is re ‡ected in the fact that the unconditional average marginal e¤ect (AME) in the nonlinear models is insigni…cant. Note: Standard errors of the estimates are given in brackets. In case of logit and conditional logit models, the errors were calculated using the delta method, for the mixed logit they were bootstrapped using 229 repetitions.

21
The heterogeneity of responses in various groups is illustrated by …gure 5 that plots individual marginal e¤ects versus the probability of investment for mixed logit. Particularly salient is the wide spread in marginal efects of regulation for the Green…eld projects in the low polluting sectors. Figure  4 gives an overview on the di¤erent responsiveness of particular investment types in the conditional logit. 9 As we have acknowledged when looking at the coe¢ cients, the e¤ect that the environmental regulation exerts on heavy polluting projects is hardly discernible from that exerted on medium polluting investments.

Robustness checks
Our …ndings, especially the allurement e¤ect for clean M&As, stay in contrast to most of the literature where it has been claimed that, as summarized by Kheder, Zugravu [28], "all industries have interest to avoid additional costs induced by stricter environmental regulation" as there are no totally "clean" industries.
Such statements were made mostly in publications which investigated the manufacturing sectors only. One could therefore expect our contradictory …nding to be due to inclusion of services in our analysis, especially that the share of investment in services in the clean M&As in our sample is abve 80%. To check whether this is indeed the reason, we reestimated our models using manufacturing projects only. The envI coe¢ cient remained positive and signi…cant (.3317*** and .5088*** in logit and conditional logit correspondingly) and the signi…cance level of the corresponding CME improves in conditional logit from 15.6% to 7% upon the exclusion of investments in services. This seems to reinforce that stricter environmental regulation may be a bait for certain types of FDI and that even if no "totally clean"sectors exist, the pollution environmental costs may be outweighed by the competitive aspects of pollution stringency. 9 The graphs plot the individual marginal e¤ects against the investment probability for the conditional logit speci…cation. The values of the marginal e¤ects are given by: p(1 p)( 1i + di ), where di is the coe¢ cient assigned to the appropriate dummies. For the conditional logit case the coe¢ cients are constant across individual …rms ( ji = ji 8i , j = f1; dg). Additionally, the estimated probabilities of investments are relatively low (the highest being 0.14), so the CMEs in the conditional logit seem to be linear functions of probability.  To assure the robustness of our results we perform several additional estimations.
Firstly, we explicitly explore the panel character of the data with …xed e¤ects logit model. We also investigate spatial e¤ects in conditional logit model by introducing spatially weighted exogenous variables to cap the third country e¤ects like in Baltagi et al. [5] but without allowing for spatial autocorrelation in the error term. 10 Countries used for weighting are the ones belonging to the same region where the region classi…cation is taken from the World Bank.
In di¤erent speci…cations we employ various combinations of additional controls: continent dummies to (explicitly) mimic the nested choice structure, tari¤s and number of documents necessary for imports/exports to proxy for trade costs of the host countries, index of FDI Restrictiveness Index prepared by OECD to proxy the investment costs, share of high-tech exports/ R&D in GDP, HDI/share of population with tertiary education to proxy for quality of labour force, exchange rates, GDP growth, exchange rate and in ‡ation. We also check for "announcement e¤ect" -the impact of the announced future environmental policies that we try to capture by adding variable lagged environmental index (envI t 1 ).
The additional estimation results are available on request. 11 In all the cases we observe some changes in the magnitudes of coe¢cients and their signi…cance level. Nevertheless, the previously discussed economic insights concerning the responsiveness towards environmental policy remained (roughly) robust throughout the di¤erent speci…cations.

Endogeneity issues
Starting with Cole et al. [11] the question of endogeneity of the environmental stringency has been permeating the FDI-PHH literature. Some studies, among others Cole [12] and Kellenberg [27], show that, once the endogeneity is accounted for, the deterrence e¤ect of the environmental policy becomes much more pronounced implying a potential positive bias.
A coarse way to deal with that potential vice of our study could be inclusion of country level …xed e¤ects. Table 5 presents the new coe¢ cients on environmental stringency in the case of logit (IIIc) and conditional logit (IIId). Due to the short time dimension and the relatively small year to year changes in many policy variables, many of the coe¢ cients lose signi…cance.
The results give some (weak) evidence for no omitted variable bias for environmental stringency in our study but tells nothing about the potential reversed causality or measurement errors. Table 5: Coe¢ cient on environmental stringency for (IIIc) and (IIId) speci-…cations when using country-…xed e¤ects. Note: ***, ** and * denote signi…cance at the 1%, 5% and 10% level respectively. Clustered standard errors at the investing company level were used.
In our second attempt to handle the endogeneity problem we employ a control function (CF) approach by using "external pressure on environmental regulation" (ext_pressure) as an instrument. We construct ext_pressure as a weighted average of the regulation level in the countries that import the goods produced by a given country. The weights correspond to the shares of the partner countries in total exports. This re ‡ects the expectation that the partner countries exert pressure on the exporters in case the exporters' environmental regulation is lenient compared to the regulation of importing partner. The pressure could come from consumer groups, importing companies protecting their "responsible" image or from legislation imposing certain requirements on the imported goods. To avoid any connection between our instrument and the FDI decision, we leave out Germany in constructing the variable 12 . .8354 F (15,333) .160.19 Note: ***, ** and * denote signi…cance at the 1%, 5% and 10% level respectively. Clustered standard errors at the country-year level were used. Number of clusters: 715.
The control function approach has some limitations compared to two-stage least squares. In particular, it requires the …rst-stage model to be correctly speci…ed and the exactly right set of instruments to be found for the consistency of the estimators (Lewbel et al., [33]). On the other hand, we believe that what makes the relationships we study particularly interesting, the control function allows us to keep the (potential) non-linearities and the heterogeneity of the tastes of the companies in our study. Conversly, this would be di¢ cult with 2SLS. 13 . Additionally, we are encouraged to the usage of the method by its successful application in many areas, for example in estimation of demand for di¤erentiated products (see, e.g. Ferreira [17]).
Our implementation of control function approach follows Petrin and Train [42] with bootstrapped standard errors of the coe¢ cients of the residual in the second stage.
The …rst stage is reported in table 6 (t-value of ext_pressure is 3.92), its implementation for logit (IIIc) and conditional logit (IIId) in table 7. The regression results point to the fact that endogeneity may not be a problem in our study in the …rst place. As reported in table 7, when the residual from the …rst stage is used in the second stage it fails to be a signi…cant predictor of the …rms' behaviour at conventional levels. This seems to support our previous …ndings 14 .
This being said, instrumenting the environmental stringency makes envI lose its signi…cance. If interpreted as the the result of some weak endogeneity, this suggest that the endogeneity may conceal some of the negative e¤ects of the regulation, i.e. the true e¤ect of environmental stringency may be more negative than reported in the previous chapters. At the same time, the e¤ect of environmental regulation at clean mergers and acquisitions is never negative. The main object of our interest -di¤erence between M&A and Green…eld is preserved, the same holds true for "monotonicity" of estimated coe¢ cients in the pollution-intensiveness. Note: ***, ** and * denote signi…cance at the 1%, 5% and 10% level respectively. Clustered standard errors at the investing company level were used. Number of observations: 958798.

Economic importance of the results and policy implications
The PHH presumes environmental stringency to be a main location determinant for polluting industries. To make a point here, we compare the marginal e¤ects of environmental stringency to the marginal e¤ects associated with the variables corporate tax, gdp per capita and stock of FDI which have been shown to be important location factors in the FDI literature. We perform this exercise for various countries. The marginal e¤ects are determined by all the independent variables at the same time therefore comparison of several countries allows us to fully explore the relative importance of environmental policy. We concentrated on US, France, China and UK, which are vital hosts for German foreign investments as shown in …gure 2 but are quite diverse in terms of their environmental policy, openness, taxation etc. Table 8 reports some of the results for IIIa, IIIc and IIId speci…cations (logit, conditional logit and mixed logit models with all the interaction terms). For every location, the AMEs of environmental stringency and corporate tax calculated for this particular location are reported on the left side of the table, together with the CMEs. In our search for evidence for PHH we decided to concentrate on dirty Green…eld projects as from the previous analysis we know that dirty M&A investment do not react strongly to environmental regulation. For illustration of the supposed allurement effect of environmental legislation we also provide the results for clean M&A projects. The investment probabilities are shown as a benchmark to enable the reader assessing the economic importance of the marginal changes in control variables. For example, a unit increase in Chinese environmental index would reduce, ceteris paribus, the probability of some German multinational choosing China as a location for its Green…eld project in polluting industry from 8.7% to 3.2% according to the mixed logit model. Analogically, a unit increase in the U.S. corporate tax rate would decline the probability of some multinational choosing U.S. as a location for its clean M&A by 0.2% to 11% according to the conditional logit model.
The fact that calculated marginal e¤ects are much smaller for the corporate tax then environmental policy is partly attributable to the di¤erent scales, on which the variables are measured -environmental stringency ranges from 0.4 to 4.2 in our sample whereas tax rates vary from 10 to 40.7. It needs to be reiterated that a unit increase in environmental index of a country marks a major step in the environmental protection (e.g. moving from the environmental regulation stringency of Benin to that of Chile in 2009). To alleviate the problem of incomparability of the marginal e¤ects, we calculated the e¤ect of a one standard deviation change (St. dev. change) of the discussed explanatory variables. The results are also reported in table 6. While the e¤ects of a one standard deviation change of the FDI Stock and GDP per capita are larger than the e¤ect of a one standard deviation change of environmental stringency, the e¤ect of environmental stringency are in the same order of magnitude as tax e¤ects. Note: "Probability" gives the calculated probability of a new investment project of a given type locating in the investigated country.

Discussion
The policymakers in some of the industrialized countries have been balking at sharpening the environmental requirements for fear of impairing the international competitiveness of the economy and losing workplaces. They tend to support their arguments with predictions of pollution haven hypothesis. However, even though a host of high-quality studies on the PHH has been conducted, its existence is still disputed as the gathered empirical evidence has been mixed.
This paper is an empirical analysis of whether, and if yes, to what extent, the German investment location decisions are sensitive towards the spatial variation of the environmental stringency. Our main contribution has been to distinguish between di¤erent modes of entry. Using the information on the German outgoing FDI in 2005-2009 we have shown that the M&A projects respond structurally di¤erently to pollution requirements and that the governments can in ‡uence the composition of FDI by setting the environmental standards.
The application of the mixed logit model allowed us to make some potentially insightful statements about the heterogeneity of tastes of the …rms that would not be otherwise possible.
Our …ndings reveal that tightened environmental stringency is an important deterrent for the FDI in ‡ow in case of polluting Green…eld projects. Importantly though, increased restrictiveness of regulation has a positive or neutral e¤ect on the decision of clean M&As locating in a given jurisdiction. This could be due to competitiveness e¤ects associated with grandfathering as well as the "green image" that German …rms are trying to keep.
Our result appear to be robust to di¤erent speci…cations.