Environmental regulations and plant exit

Environmental and Resource Economics 11: 35–59, 1998. c 1998 Kluwer Academic Publishers. Printed in the Netherlands. 35 Environmental Regulations and Plant Exit A Logit Analysis Based on Establishment Panel Data ERIK BIØRN, ROLF GOLOMBEK and ARVID RAKNERUD 1 University of Oslo, Department of Economics and Foundation for Research in Economics and Business Administration, Oslo, Norway ( corresponding author, email: rolf.golombek@snf.uio.no) Accepted 11 March 1997 Abstract. A qualitative response model is applied to study the relationship between environmental regulations and plant exit. The data set is Norwegian panel data for establishments in three manufacturing sectors that have high shares of units which have been under strict environmental regulations. In two of the sectors, the estimated exit probability of regulated establishments is, ceteris paribus, only one third of the exit probability of non-regulated establishments. We also find that the probability to change regulation status from being non-regulated to become regulated depends significantly on economic factors. In particular, establishments with weak profitability had the highest probability to come under environmental regulation. Key words: environmental economics, environmental standards, firm exit, logit model, panel data JEL classification: C25, D24, I18, Q38 1. Introduction At present, there is a debate in several countries on measures to be used in order to achieve a sustainable industrial development. In particular, attention has been focused on whether it is possible to achieve both better environmental quality and higher employment levels by means of a tax package whereby increased government revenues from higher environmental taxes are used to cut the payroll tax rate. However, because ‘command and control’ instruments have been the predominant environmental policy measures, we believe it is also of interest to examine effects following from imposing quantitative requirements, i.e., environmental standards. This is the topic addressed by the present article: Focusing on three manufacturing sectors with high shares of units under strict environmental regulations, we show that imposing environmental standards may not raise the rate of exit. To our surprise, we find that for two of these sectors, establishments under strict environmental regulations had a lower tendency to exit than did establishments under weak or no environmental regulation. For the third sector, environmental regulations had no significant impact on the rate of exit. Although we are not in a position to explain the results, they could be related to Porter (1990, 1991) who argues that environmental regulations may trigger off internal changes in firms, 36 ERIK BIØRN ET AL. leading to more cost effectiveness and development of resource-efficient products through e.g. transformation of former residuals into marketable products. In Norway, as in several other countries, increased concern for the environment has materialized as extensive regulations of harmful emissions. The national regulator imposes environmental standards on either the amount of emissions or the concentration of emissions of polluting establishments. In general, these requirements are related to local receptors and the technology of the establishments. Imposed environmental standards can be honoured by  changing the output mix,  modifying the present technology by installing anti-pollution devices, e.g. a scrubber for the smokestack, or  installing a new, and environmentally more friendly, technology. Through changes in revenues and costs (due e.g. to acquisition of more capital, and requirements to change inputs per unit of output), profits of the regulated establishments shift. Two interesting questions are then: (a) Have environmental regulations an impact on the exit decision of establishments when other factors affecting the exit decision, e.g. profitability, have been ‘controlled for’? (b) If such regulations have a ‘net’ effect on the exit decision, does the probability to exit increase or decrease as a consequence of imposed environmental standards? This paper presents an empirical investigation attempting to answer these questions. The framework is a qualitative choice model, and our data base is a set of panel data for Norwegian manufacturing establishments. We distinguish, for each year, between establishments that are operating and establishments that are closing down. Furthermore, operating units are divided into two categories, environmentally regulated and environmentally non-regulated establishments. We aim at explaining, on the one hand the probability of closing down, and on the other hand the probability of being regulated, by means of a set of covariates, including, inter alia, an indicator of profitability and the number of employees. The primary issues we wish to throw light on are:  Given that a firm is operative in a certain year, does its exit probability in the next year depend on whether it is regulated or not?  Can differences in the exit probabilities between firms be ascribed to other factors than environmental regulations?  Is there evidence that the profitability of a firm in a certain year affects its probability of being regulated in the next year? Our model has the main characteristics of a multinomial qualitative response model of the logit type, but it also has elements of a Markov chain model. It is estimated, by maximum likelihood, for three Norwegian manufacturing sectors which have a large proportion of establishments that have been under strict environmental regulations in the data period 1976–1991. We find that in Manufacture of pulp, paper, and paperboard and in the Manufacture of iron, steel, and ferroalloys, the probability to exit (in an arbitrary year) differs significantly between regulated and non-regulated establishments. Contrary to our expectations, we find that, ceteris ENVIRONMENTAL REGULATIONS AND PLANT EXIT 37 paribus, the probability to exit is highest for non-regulated units. The estimated exit probability of regulated establishments is one third of the exit probability of regulated establishments. In the third sector, Basic industrial chemicals, mainly due to the very few exits during the data period, we are unable to make precise inference. Our main findings are in line with Golombek and Raknerud (1997), who use the same data set in combination with a Markov model to compare the transition probabilities to increase employment, to decrease employment, or to exit between regulated and non-regulated establishments. They find that for two of three sectors, regulated establishments had a significantly lower probability to exit than the non-regulated units. Brännlund et al. (1995) analyse the impact of environmental regulations on firm profits in the Swedish pulp and paper industry. According to their study, most pulp and paper mills were unaffected (in terms of profits) by the environmental regulation in 1990. Deily and Gray (1991) take a different approach, by focusing on the enforcement activity of the environmental regulator. They claim that the U.S. regulator directed fewer enforcement actions towards U.S. steel plants with a high predicted probability of closing down and/or plants that were major local employers. Finally, there are several studies examining the impact on productivity of imposed environmental standards, e.g., Barbera and McConnell (1986), Conrad and Morrison (1989), Jorgenson and Wilcoxen (1990), and Gray and Shadbegian (1993). As these studies find a negative impact on productivity following from environmental regulations, they indicate that imposition of environmental standards leads to a larger not a smaller number of exits as our study suggests. A survey of recent studies of post-entry and exit decisions of firms, without specific attention to the effect of environmental regulations, is given in Audretsch and Mata (1995); examples of empirical studies based on plant data are Audretsch (1995) and Doms et al. (1995), using logit and probit analysis, respectively. The rest of the paper is organized as follows. Section 2 presents the econometric, multinomial choice, model (focusing on the responses operative/closing down and regulated/non-regulated). We derive the likelihood function, imposing a logit parameterization of the choice probabilities. The raw data set, covering 278 Norwegian manufacturing establishments, is described in Section 3. It is a panel data set, but due to exits and entries of establishments it is unbalanced. The selection of explanatory variables (covariates) for the choice probabilities is also discussed. As no formal theory of optimization (on the part of neither the firm nor the regulator) have been used in formulating our econometric model, the choice of covariates is based mainly on a priori (common sense) considerations. Section 4 reports the empirical results and points out some problems of interpretation. First, we estimate the model and test whether the parameters of regulated and non-regulated establishments are equal. Second, we calculate (and construct a confidence interval for) the log-odds ratio, which shows (at the sample means of the covariates) whether regulated establishments have a lower or a higher exit 38 ERIK BIØRN ET AL. probability than non-regulated establishments. Third, we estimate the conditional probabilities of being regulated in an arbitrary year, given the regulatory status in the previous year. This is the Markov chain element of our model. Finally, in Section 5, we summarize our findings and point at some remaining interpretation problems and topics for further research. 2. Econometric Model Let T denote the number of years covered by the data set, and let N be the number of firms which are observed in at least one of these T years. A firm i is observed only in the year(s) during which it is operative. Assuming that the firms are numbered consecutively from i = 1 to i = N , we let It  [1; 2; : : : ; N ]; t = 1; 2; : : : ; T ; represent the numbers of those firms which are observed in year t according to the above criterion. We assume that an exit decision is irreversible, which means that a firm is not restarted in one of the years t + 1; t + 2; : : : if a closing down decision has been made in year t. On the other hand, the pollution control authorities may, in principle, impose or abolish an environmental regulation on any firm in any of the T years under observation, regardless of its previous ‘regulation history’. According to Norwegian practice, however, the authorities take a decision of whether to impose or not impose environmental standards to be valid for several successive years. In any year t, a subset of the N firms, given by the index set It , is observed, so that the combinations of indices (i; t) in the data set can be symbolized as i 2 It  [1; 2; : : : ; N ]; t = 1; : : : ; T : Since the data set shows simultaneous variation across firms and years, it is a panel data set. But because It 6= Is and It \ Is 6= ; for at least one s 6= t, it is a set of unbalanced panel data. For firm i in year t we define the two binary state variables (C and R symbolizing ‘closed down’ and ‘regulated’, respectively)  Cit =  Rit = c o r n if firm i has closed down in year t, if firm i is operative in year t, (1) if firm i is regulated in year t, if firm i is non-regulated in year t. (2) Let Xit denote the covariates to be used in explaining the choice probabilities regarding Cit and Rit for firm i in year t. As explained above, the coverage of the data set is described by i 2 It ; t = 1; : : : ; T . However, for simplicity, we proceed for the moment as if the index set for (Cit ; Rit ; Xit ) is i = 1; : : : ; N ; t = 1; : : : ; T , i.e. as if all firms are observed in each year. Let Hit denote the history of Cit , Rit , and Xit up to, and including, year t, which implies Hit = (Cit ; Rit ; Xit ; Hi;t 1 ); i = 1; : : : ; N ; t = 1; : : : ; T : 39 ENVIRONMENTAL REGULATIONS AND PLANT EXIT j H 0 ) can be factorized recursively as follows The probability P (HiT i iT Y P C ;R ;X T P (H j H 0 ) = i ( =1 it it it jH i;t ; 1) (3) t where P (Cit ; Rit ; Xit j Hi;t P (C ; R ; X j H it it it 1) 1) = i;t P (C j R ; X ; H it it can be factorized as it i;t 1) P (R j X ; H it it i;t 1) P (X j H it i;t : 1) Hence, P (H j H 0 ) iT where L1 i = i YP C T = ( =1 it L 1 L 2 L 3; i i (4) i jR ;X ;H it it i;t ; 1) (5) t L2 i YP R T = ( =1 it jX ;H it i;t ; 1) (6) t L3 i YP X T = ( =1 it jH i;t : 1) (7) t Equation (4), with Equations (5)–(7) inserted, defines the factor of the joint likelihood function of (Cit ; Rit ; Xit ); t = 1; : : : ; T , which would relate to firm i if this firm was observed in all the years 1; : : : ; T . Recalling that the observations available are described by i 2 It and imposing the standard assumption of independent observations from the different firms, we can write the likelihood function for the complete sample, conditional on Hi0 for all i, as L = YL i 2t i 1 L 2 L 3: i i (8) I The econometric model we use to analyse the establishments’ and the regulator’s choices is a qualitative choice model with likelihood function given by Equations (5)–(8). Let and be the coefficient (column) vectors characterizing the conditional choice probabilities P (Cit j Rit ; Xit ; Hi;t 1 ) and P (Rit j Xit ; Hi;t 1 ), respectively. We assume that the covariate vector can be written as Xit = (xit ; zit ), where the subvector xit characterizes the first probability and the subvector zit characterizes the second probability. The covariate (row) vectors xit and zit , however, may have common elements, and are described in more detail below. Since there could be ‘structural differences’ between regulated and non-regulated firms, 40 ERIK BIØRN ET AL. not only with respect to the values of the covariates, but also with respect to their 0 ; 0 ) and 0 = ( 0 ; 0 ), coefficient values, we partition 0 and 0 as 0 = ( R N R N where R and N refer to regulated and non-regulated firms, respectively, and consider R = N and R = N as interesting testable hypotheses. We assume, furthermore, that , , and the parameters characterizing P (Xit j Hi;t 1 ) are variation free.1 We parametrize P (Cit jRit ; Xit ; Hi;t 1 ) and P (Rit jXit ; Hi;t 1 ) as logit probabilities as follows exit R for Rit = r; xit R P (Cit j Rit ; Xit ; Hi;t 1 ) = 1 +xeit N (9) e for R = n; it 1 + exit N 8 >< >: i 2 It ; t = 2; : : : ; T; and P (Rit j Xit ; Hi;t 8 ezit R >< 1 + ezit R 1) = >: ezitzitN N 1+e for Ri;t 1 = r; for Ri;t 1 = n; (10) i 2 It ; t = 2; : : : ; T: The latter expression implies that R and N are the coefficient vectors describing the probability of regulation for firms which were regulated and non-regulated, respectively, in the previous year.2 Maximum likelihood estimation of R , N , R , and N implies maximization of L, defined by Equations (5)–(8), subject to Equations (9)–(10), with respect to these coefficients. This is equivalent to maximization of the partial likelihood function LC = LC ( R ; N) = T Y Y t=1 i2It P (Cit j Rit ; Xit ; Hi;t ; 1) (11) with respect to R and N , subject to Equation (9), and maximization of the partial likelihood function T LR = LR ( R ; N ) = (12) P (Rit j Xit ; Hi;t 1 ); t=2 i2It YY with respect to R and N , subject to Equation (10). As R , N , R , and N are assumed to be variation free, this can be interpreted as follows: We perform (i) one binomial logit analysis of the closing down/remaining operative decision confined to regulated firms/years, (ii) one binomial logit analysis of the closing down/remaining operative decision confined to non-regulated firms/years, (iii) one ENVIRONMENTAL REGULATIONS AND PLANT EXIT 41 binomial logit analysis of the regulation/non-regulation decision confined to firms that were regulated in the previous year, and (iv) one binomial logit analysis of the regulation/non-regulation decision confined to firms that were non-regulated in the previous year. 3. Data and Choice of Covariates This study relies on two data sources. The first consists of raw data from Statistics Norway’s Manufacturing Statistics, which provide annual observations on gross value of production, intermediates, employment, man-hours, wage costs, investments, and fire insurance for all Norwegian manufacturing establishments (i.e., ‘functional units which, at a single location, are engaged mainly in activities within a specific activity group’). These data, although containing no information on pollution compliance costs, facilitate the identification of exits. In this study, an establishment is defined to have exited in year t if year t 1 is the latest period in which Statistics Norway was able to gather (economic) information about the unit.3 Statistics Norway also collects information about inputs and outputs (both in physical units and in Norwegian kroner) for establishments with at least 20 employees. The second data source is the data bank of the Norwegian environmental regulator – the State Pollution Control Authority (the SFT – Statens forurensningstilsyn). In Norway, comprehensive regulation of harmful emissions was gradually introduced in the early 1970s. Establishments are regulated by annual emission quantities and/or by maximum concentrations (quantity per unit of volume). Due to the small size of the Norwegian economy and the close interrelationships between the SFT and other public bodies, the regulator is aware of all non-marginal emissions from manufacturing industries. As pollution is unlawful unless a manufacturing establishment has been granted an emission permit from the SFT, all big pollutants have (after 1975) permits. Emission permits are usually granted for a 10-year period. The SFT operates with four control classes, according to amount of emissions and concentrations (measured without any purification), type of emissions and condition of receptors (in the area where production takes place). Class 1 generally comprises permits for ‘large emissions into weak or medium receptors’, cf. SFT (1987), whereas class 4 encompasses permits for ‘medium emissions into good receptors and small emissions into medium receptors’. Inspections and enforcement vary significantly across control classes. Around 15% of all Norwegian manufacturing establishments have emission permits, and establishments in control classes 1, 2, and 3 comprise more than 90% of total emissions from manufacturing industries. Emission permits are related to local receptors (e.g., whether the plant is located close to wetlands) and the substance being emitted. Because there are several types of emissions and numerous conditions of receptors, each permit is – even within a sector – unique. This makes it virtually impossible to compare emission 42 ERIK BIØRN ET AL. permits within a simple formalized model. In order to discriminate between permits, we shall rely on the rule of thumb that class 1 permits require larger drops in amount/concentration of emissions (i.e., implying higher pollution abatement costs) than class 2 permits, etc. In particular, according to the regulator, class 4 permits may often imply only marginal compliance costs. The ranking of permits (in terms of pollution compliance costs) motivates us to group establishments according to the control class of their permits. In this study, establishments which have at least one class 1 or one class 2 permit are termed ‘regulated’, whereas all other establishments are termed ‘non-regulated’. According to the SFT classification, all ‘regulated’ establishments have either large emissions or weak/medium receptors. The establishments denoted as ‘non-regulated’ have been under either weak or no environmental regulation. This definition may, of course, be criticized: Because an establishment may have several permits, both the number of permits and the classification of each permit may be of significance. To this we can argue that because it is mainly ‘regulated’ establishments that have several permits, there will be an overall significant difference between the two groups of establishments. Moreover, when confronted with our definition of a regulated establishment, the SFT found it sound. Finally, our definition has the convenient property that units are divided into groups that are not too small to make testing of hypotheses on regulated establishments possible. Needless to say, as the severity of environmental regulations varies over time and across firms, it would have been better to apply a regulatory intensity variable which accounts for the legally mandated reduction in emissions as well as the effective enforcement imposed on establishments. One possibility could be to follow Gollop and Roberts (1983) who construct a measure of regulatory intensity for the US fossil-fuel electric power industry in which the intensity is a function of the legal standard, actual emission and unconstrained emission (all factors measured as pounds of SO2 per million Btu). In the present study, such an approach is not feasible. First, we are not able to estimate unconstrained emission rates. It should also be noticed that Gollop and Roberts estimate the unconstrained emission rates in a rather ad hoc way. In particular, they do not fully take account of market forces and the fact that all observed fuel shares reflect regulatory constraints. Second, while the regulatory intensity variable in Gollop and Roberts (1983) is uni-dimensional (representing emission of SO2 only), we are faced with several types of emissions, and it is far from obvious how different types of emissions should be weighted. In particular, the fact that types of emissions vary across units complicates matters severely. A standard answer to these difficulties is to use (appropriate) economic data. However, such data are not available. Finally, the intensity measure of Gollop and Roberts has the following unwarranted implication: If the difference between the unconstrained and the legal emission rate is constant from year s to year t, the intensity measure is independent of the legal environmental standard from year s + 1 to year t. ENVIRONMENTAL REGULATIONS AND PLANT EXIT 43 The two data sources were merged by Statistics Norway for the years 1976–1991 16). In 1991, the SFT regulated around 150 establishments. An examination of the data reveals that a significant share of these establishments was found in four sectors (according to the four-digit ISIC code): Pulp, paper, and paperboard, Basic industrial chemicals except fertilizers, Iron, steel, and ferroalloys, and Nonferrous metals. In 1991, these sectors comprised around 90% of all establishments with a class 1 emission permit. However, we removed establishments in the sector Non-ferrous metals from the data set as in this sector, (i) large establishments are regulated whereas small establishments are not, and (ii) there were only a few exits in the data period. That leaves us with three manufacturing sectors that can be analysed. A further description of the data set and the institutional setting is given in Golombek and Raknerud (1995). Figure 1 illustrates the distribution of employment in regulated and non-regulated establishments, showing, for each category, the smoothed relative frequency, calculated as an average over the years 1976–1991. As seen from this figure, regulated establishments have the highest fraction of large units. This probably reflects, inter alia, the criteria used by the SFT when granting emission permits. %beginfigure[tbp] Figure 2 shows employment in 1976 and 1991 in the sector Iron, steel, and ferroalloys, shaded and non-shaded triangles representing regulated and non-regulated establishments, respectively. To make the figure lucid, it only comprises establishments that (i) had at least 50 employees in 1976 and (ii) did not change their regulatory status after 1976 (i.e., did not change from being non-regulated to become regulated, or vice versa). If an establishment is located below the line running from the origin to the north-east corner, the level of employment was lower in 1991 than in 1976. In particular, establishments that exited after 1976 are found on the horizontal axes. Figure 2 gives a clear indication that in terms of employment, regulated establishments have done better than the non-regulated units: From 1976 to 1991, the number of establishments decreased from 36 to 22, whereas the proportion of regulated establishments increased from 47% to 73%. A closer examination reveals that the exit rate for regulated establishments is significantly lower than the exit rate for non-regulated establishments (6% versus 68%). An examination of Iron, steel and ferroalloys reveals the same type of development. On the other hand, the development in Basic industrial chemicals is ambiguous. We next present and comment on the choice of covariates, i.e., the vector xit explaining, for firm i in year t, the exit probabilities, and the vector zit explaining the regulation probabilities. As is evident from the above discussion, no formal optimization model, on the part of the firm and/or the regulator, has been used in constructing the multinomial choice model applied in this paper.4 Hence, in choosing covariates, we have to draw on our general understanding on what determines exits and what determines the regulation decision. (T = 44 ERIK BIØRN ET AL. Figure 1. Frequency of employment. Average 1976–1991. According to standard textbook theory, it is optimal to exit at time t if the scrapvalue of the plant exceeds the expected discounted profit of not exiting at time t. Using for instance a putty-clay theory of producer behaviour, in which capital is considered as malleable before investment, whereas the factor proportions are treated as fixed thereafter [cf. Johansen (1959)], we may conclude that a decrease in both present and expected allowed emissions may reduce discounted profits to such an extent that an instantaneous exit becomes optimal. In general, for both regulated and non-regulated establishments, expected changes in future profits due to, e.g., shifts in market prices, technological changes, and allowed emissions may affect the optimal time to exit. ENVIRONMENTAL REGULATIONS AND PLANT EXIT 45 Figure 2. Employment in 1976 and 1991. As already noted, present profitability may have an impact on exit behaviour. In this study, it is measured gross of capital cost and defined as the difference between value added and wage costs (measured at factor prices) divided by total man-hours. We have not adjusted for capital costs because our data set contains no direct information on this variable. Although capital costs can be derived from transformations of reported fire insurance values (which could be taken as proxies for the value of the capital stocks), being aware of the uneven quality of the fire insurance data, we found that an extensive use of these data is not warranted. In this study, fire insurance values are only included in one covariate. Profitability is normalized against the number of man-hours because we want differences in scale of operation to be picked up by another covariate, cf. the discussion below. Present profitability may also have an impact on future profitability through the formation of expectations. In general, with adaptive expectations, future profits depend on the 46 ERIK BIØRN ET AL. whole profitability history of the establishment, i.e. lagged profits may be relevant covariates, and the inclusion of such covariates is possible because we have a panel data set. In this study, due to data limitations, we only use the profit (per man-hour) in the last year. Furthermore, we use the rate of increase of a composite relative price variable (from year t 1 to year t) as a covariate, to be referred to as a net price index. For each establishment and year, this net price index is calculated as the difference between a weighted annual rate of increase of the output prices minus a weighted annual rate of increase of the input prices, cf. Golombek and Raknerud (1995, Appendix A). A positive net price index indicates that the output prices have, on average, increased by more than the input prices. This variable is, of course, related to the profitability variable, but being a rate of increase, it has a different ‘dimension’ and therefore has its place as a potential covariate. Finally, the stock of finished goods (also measured per man-hour) is also included; a large stock may reflect limited sales, at least in the near future. In general, environmental regulations require real investments at the plant level. If these investments are primarily ‘unproductive’, profitability drops, i.e., imposed environmental standards may lead to investments that increase the probability to exit. Assuming that differences in capital stock (measured by fire insurance values) across establishments reflect scale differences, we included the ratio between gross investment and capital as a covariate when testing the impact of investment on the exit decision. The exit decision could also depend on the size of the establishment. First, a large establishment may have more output activities than a small establishment. If some activities of a large establishment run into profitability problems, the establishment may simply close down these activities and continue production of its other outputs. On the other hand, a small establishment with only one output has to close down sooner or later once its product becomes nonprofitable. Second, a large establishment may have larger financial resources than a small establishment. Suppose, e.g., that present production is nonprofitable, whereas future production is expected to be profitable, and that, due to high hiring and firing costs, total longrun profit is expected to be higher if production takes place in all future periods as compared with a short-term close-down followed by an expected reentry. While a large establishment may be in a position to avoid the short-term close-down, the small firm could be forced to exit. Turning to the decision of the regulator, we want to investigate whether economic factors as well as institutional arrangements have impact on the regulatory status of establishments. A decision lag of one year is assumed, i.e., the variables in zit are lagged one year. As emission permits are, in general, granted for 10 years, it is highly likely that the regulatory status of a firm in period t depends on its regulatory status in period t 1 [cf. Equation (10)]. 47 ENVIRONMENTAL REGULATIONS AND PLANT EXIT 4. Empirical Results: Testing and Estimation In this section, we analyse the empirical models presented in Section 3. Recall that the probability models are logistic regression models. The first is given by Equation (9), where xit R and xit N are the log-odds of closing down, given that the firm is regulated and non-regulated, respectively, in the current year.5 The second model is given by Equation (10), where zit R and zit N are the log-odds of being regulated, given that the firm is regulated and non-regulated, respectively, in the previous year. We consider the following covariate vectors (cf. Section 3):6 xit = zit = (changes in output and input prices from year t number of employees in year t; profitability in year t; investment in new capital in year t; stock of finished goods in year t; profitability in year t 1); (number of employees in year t profitability in year t 1): 1 to year t; 1; Investment is scaled by the size of the capital stock, whereas profits at constant prices and stock of finished goods are measured per man hour. This choice of scaling may seem arbitrary, but is motivated mainly on pragmatic grounds; some scaling is necessary in order to compare establishments of different size. Furthermore, data on employment are, in general, much more reliable than capital data.7 We found that the R estimates were poorly determined in all sectors (the Hessian of the log-likelihood function had several eigenvalues close to zero). This indicates that the proposed model has too many parameters relative to the amount of information in the data. To obtain a more parsimonious parameterization, we 0 0 apply a two step test procedure: First, we make the partition R = ( eR ; cR ), where e contains the slope coefficients and c is the intercept of the log-odds for the R R 0 0 = ( eN ; cN ) is partitioned similarly for the non-regulated regulated firms, and N firms. In the first step, we test the hypothesis HA : eR = eN , i.e., all structural differences, if any, between regulated and non-regulated establishments can be represented by a shift in the intercept. Since eR has 6 elements, the likelihood ratio statistic is asymptotically distributed as 26 (i.e. chi-square with 6 degrees of freedom) under HA . If HA is not rejected at the 5% level, we proceed to test for equality of the constant terms, i.e., HB : cR = cN , given that HA is satisfied, which implies that the exit probability is independent of regulatory status. The likelihood ratio statistic of the latter test is asymptotically distributed as 21 . The estimation results for each of the three sectors are presented below. Due to the absence of price information for establishments with less than 20 employees, 48 ERIK BIØRN ET AL. Table I. Pulp, paper, and paperboard. Estimated parameters for the probability to exit. Standard deviations in parentheses. Parameter of Net price index No. of employees Profit Investment Stock Lagged profit cR cN Regulated 0.52 3.46 0.85 2.43 1.09 1.27 1.50 Non-regulated (0.73) (2.62) (1.05) (6.26) (1.31) (1.70) (1.47) 0.03 0.99 2.47 0.10 0.32 0.08 (0.17) (0.41) (0.53) (0.12) (0.11) (0.55) 2.25 (0.31) Table II. Pulp, paper, and paperboard. Estimated parameters for the probability to exit when eR eN . Standard deviations in parentheses. = Parameter of Net price index No. of employees Profit Investment Stock Lagged profit cR cN Estimate 0.00 1.12 2.01 0.10 0.31 0.11 3.30 2.15 (0.16) (0.38) (0.43) (0.12) (0.11) (0.49) (0.59) (0.29) these units are removed from the sample. Furthermore, a small fraction of the observations (less than 1%) was omitted due to missing information on some of the covariates. 4.1. PULP, PAPER, AND PAPERBOARD The sample from this sector (ISIC 3411) consists of 95 establishments and 40 exits. The estimates of R and N in the unrestricted model are presented in Table I. Each component of the parameter vector is referred to by the corresponding variable (except for the intercepts cR and cN ). The estimates of standard deviations are obtained from the observed information matrix. The P -value obtained from testing HA : eR = eN is 0:12. Hence, there is no substantial evidence against assuming equality of the slope coefficients for regulated and non-regulated firms and representing the differences in the logit probabilities (9) as differences in the intercept term only. The estimates of the restricted model are shown in Table II. 49 ENVIRONMENTAL REGULATIONS AND PLANT EXIT Table III. Iron, steel, and ferroalloys. Estimated parameters for the probability to exit. Standard deviations in parentheses. Parameter of Net price index No. of employees Profit Investment Stock Lagged profit cR cN Regulated 0.00 0.53 0.02 4.81 0.39 0.75 2.44 (0.42) (0.91) (0.41) (2.86) (0.31) (0.43) (0.91) Non-regulated 0.46 1.49 1.09 0.02 0.02 0.08 (0.35) (1.05) (0.57) (0.22) (0.38) (0.63) 2.33 (0.45) As expected, the impact on the exit probability of the number of employees and profitability is negative: The probability of exit decreases significantly with increased number of employees and improved profitability. It follows from Equation (9) that when eR = eN = e, the conditional log-odds ratio between a regulated establishment i with covariate vector xit and a non-regulated establishment j with covariate vector xjt is:  ln P (Cit j r; Xit ; Hi;t 1 ) 1 P (Cit j r; Xit ; Hi;t 1 )   ln P (Cjt j n; Xjt ; Hj;t 1 ) 1 P (Cjt j n; Xjt ; Hj;t 1 )  = (xit xjt ) e+ cR cN : If the covariates are equal (xit = xjt ), this log-odds ratio is simply cR cN . In Pulp, paper, and paperboard, the estimated log-odds ratio is 1:15 when xit = xjt . The negative sign reflects that regulated establishments have a lower conditional exit probability than non-regulated establishments. Let PbC jr (xit ) and PbC jn (xit ) be short hand expressions for the estimated exit probabilities of a regulated and a nonregulated establishment with covariate vector xit , respectively. At the sample mean, x, their ratio, which can be denoted as their relative risk, is PbC jr ( x)=PbC jn ( x) = 0:3, i.e., the estimated exit probability of regulated establishments is, ceteris paribus, only one third of the exit probability of non-regulated establishments. An estimate of the standard deviation of cR cN , i.e. the log-odds ratio between a regulated and a non-regulated establishment with equal covariates, can be obtained from the observed information matrix by means of: var(cbR cbN ) = var(cbR ) + var(cbN ) -2 cov(cbR ; cbN ). We find that the standard deviation of this log-odds ratio is 0.55. Hence, ( 2:25; 0:05) is an approximate 95% confidence interval for the log-odds ratio. This is easily transformed into a confidence interval for the odds ratio: (0:11; 0:95). Testing the hypothesis HB : cR = cN , conditional on HA , by means of a likelihood ratio test, we found the P -value 0.03. Hence, there is evidence that the exit probability of regulated establishments is significantly lower than the exit probability of non-regulated establishments in Pulp, paper, and paperboard. 50 ERIK BIØRN ET AL. Table IV. Iron, steel, and ferroalloys. Estimated parameters for the probability to exit when eR eN . Standard deviations in parentheses. = Parameter of Net price index No. of employees Profit Investment Stock Lagged profit cR cN Estimate 0.18 1.02 0.30 0.17 0.08 0.60 3.05 2.32 (0.24) (0.73) (0.29) (0.22) (0.22) (0.30) (0.63) (0.35) 4.2. IRON, STEEL, AND FERROALLOYS The sample from this sector (ISIC 3710) consists of 87 establishments and 22 exits. The estimates of R and N in the unrestricted model are presented in Table III. The hypothesis HA : eR = eN is not rejected at the 5% level; the P -value is 0.08. The estimates of the restricted model are displayed in Table IV. As seen from the table, the estimated impact on the exit probability of number of employees and profitability has the expected signs. However, the estimates are not significant at the 5% level. The estimated log-odds ratio between regulated and non-regulated establishments with equal covariate values is cbR cbN = 0:73 (st.dev. = 0.54). Evaluated at the sample mean of the covariate vector x x) = 0:01 and PbC jn ( x) = 0:03, , PbC jr ( i.e., also in this sector the exit probability is lowest for regulated establishments, with an estimated relative risk, PbC jr ( x)=PbC jn ( x), around 0.3. The estimated odds ratio is 0.48, with an approximate 95% confidence interval (0:16; 1:41), i.e., it is not significantly different from 1. Testing the hypothesis HB : cR = cN by means of a likelihood ratio test gave the P -value 0.19. We conclude, then, that there are indications that regulated establishments have the lowest exit probability in Iron, steel, and ferroalloys, although the difference is not significant. 4.3. BASIC INDUSTRIAL CHEMICALS In this sector (ISIC 3511), there are 76 establishments but only 4 exits. Hence, it is impossible to draw firm conclusions about the impact of regulations on the exit probabilities. We therefore only report the estimated log-odds ratio between regulated and non-regulated establishments with equal covariate values: 0.79 (st.dev. = 1.14). 51 ENVIRONMENTAL REGULATIONS AND PLANT EXIT 4.4. THE CONDITIONAL PROBABILITY OF BEING REGULATED We now present estimation results characterizing the conditional probability of being regulated in year t, given the regulatory status (non-regulated or regulated) in the previous year. Table V presents the estimated N coefficients for the three sectors. Note that all observations, not only those representing transitions from non-regulated to regulated (4–6 establishments in each sector), contribute to the result. The number of employees is a significant variable in all the three sectors: Increased employment is accompanied by a higher probability to become under environmental regulation. An interpretation of this may be that large establishments are potentially large polluters and may therefore receive more attention from the regulators than small ones. More interestingly and contrary to our expectations, in all sectors, the estimated conditional probability to change regulatory status from non-regulated to regulated is a decreasing function of lagged profitability, and the effect is significant at the 5% level in Iron, steel, and ferroalloys and in Basic industrial chemicals. How can this be explained? Table V. Estimated parameters for the probability to become regulated, given nonregulation in previous year. Standard deviations in parentheses. Parameter of Lagged no. of employees Lagged profit Intercept Pulp, paper, and paperboard Iron, steel, and ferroalloys Basic industrial chemicals 0.50 (0.17) 0.43 (0.43) 3.61 (0.33) 0.21 (0.10) 1.25 (0.41) 2.30 (0.21) 1.08 (0.36) 3.98 (1.73) 2.97 (0.51) According to the regulator, imposed environmental standards are independent of profitability, though the allowed time period to adjust to these standards are negotiable and may to some extent reflect economic conditions. Moreover, the regulator claims that its information is primarily limited to sectorial output prices and enterprise accounts. As an enterprise may comprise more than one productive unit (establishment), the regulator has limited information on the establishments. If this is correct, it is hard to see why there should be a negative statistical relationship between profitability and the probability to change regulatory status. A different view could be that the regulator has more information on the establishments and uses it all to identify establishments which have weak profitability due to old (or even obsolete) technologies. As these establishments have to undertake heavy investments to survive, the additional costs of installing a clean technology may be moderate. This could be the reason why the regulator imposes emission standards on (some) establishments with weak profitability. An alternative explanation of why establishments with weak profitability had the highest probability to become under environmental regulation could run as follows: The majority of the establishments that were regulated in the years 1976– 1991 became under environmental regulation before 1976. Assume (hypothesis 1) 52 ERIK BIØRN ET AL. that prior to the regulation, these establishments had a low probability to exit (e.g., due to strong profitability). If this hypothesis is valid, we should find that between 1976 and 1991 establishments with weak profitability had the highest probability to become under environmental regulation, simply because polluting units with strong profitability had already been regulated. Although this hypothesis contrasts with the view of the regulator, it would be desirable to have it tested: there may be omitted variables, relating to the pre-sample period, which explain both the regulatory status at the start of the sample period and the probability to exit during the sample period. Unfortunately, we do not have data which permit the testing of this interesting hypothesis. Hypothesis 1 can also explain why regulated establishments had the lowest exit probability in the data period (high profits at the start of the data period). An alternative explanation could be related to a sample-selection hypothesis: Assume (hypothesis 2) that the partial effect of regulation is to temporarily increase the probability to exit. Then, regulated establishments had, ceteris paribus, a higher exit rate than non-regulated establishments before 1976. On the other hand, at the start of the the data period (i.e., when the negative effect following from regulation had vanished) regulated establishments may have had the lowest exit probability as the ‘weakest’ regulated units had already closed down. However, an examination of the data does not support this sample-selection hypothesis: We found no exits among regulated establishments prior to 1976. Regarding the probability of being regulated in year t conditional on being regulated in year t 1, in two sectors, the conditional probability is estimated to 1. Hence, we are unable to identify any effects of the covariates: All combinations of parameter values that yield a regulation probability of one explain the data equally well. In the third sector – Pulp, paper, and paperboard – for an establishment with average employment and average profitability, the probability to remain regulated is estimated to 0.996. In all three sectors, present regulatory status is entirely explained by lagged regulatory status only. These results mainly reflect institutional facts: An emission permit is (usually) granted for a period of 10 years. Regulation is cancelled only if the emission source disappears (e.g., due to a close down of the polluting activity or installation of a clean technology). 4.5. ROBUSTNESS ANALYSIS AND GOODNESS OF FIT In order to assess the robustness of the results, we performed some supplementary analyses. The first examines whether observations on establishments within the same sector fail to be independent, conditional on the covariates. There are several reasons why dependence may occur. One possibility is the existence of latent variables at the subsector level, as defined by the 5 digit ISIC code (recall that our sector definition follows the 4 digit ISIC code), which cause establishments belonging to the same subsector to have, ceteris paribus, a significantly different risk of exit than establishments in other subsectors. This hypothesis can be exam- ENVIRONMENTAL REGULATIONS AND PLANT EXIT 53 ined by including subsectoral dummies and testing whether their coefficients are zero. For Pulp, paper and paperboard, the hypothesis of zero dummy coefficients was rejected at the 5% level. However, the coefficient estimates of the covariates are not significantly altered, and the hypothesis HB : cR = cN is rejected in the model with subsectoral dummies as well (P -value = 0:01). For Iron, steel and ferroalloys, the hypothesis of zero coefficients of all the subsectoral dummies was not rejected at the 5% level, although the effect of regulation is weakened (it was not significant in the original model). In this sector, differences between regulated and non-regulated establishments could be attributed solely to subsectoral differences. Due to random effects, establishments belonging to the same subsector may differ in their probability to exit, conditional on the covariates. Unfortunately, to test the hypothesis of establishment specific random effects requires advanced computer algorithms that are still not part of standard packages and (in generalized linear models) typically requires Markov Chain Monte Carlo methods in combination with the Expectation Maximization (EM) algorithm [cf. e.g., Fahrmeir and Tutz (1994)]. Because of these substantial difficulties, we did not test this random effects hypothesis. In the original model, the effect of employment on the log-odds is linear. However, from a priori reasoning one might anticipate that the marginal effect of increased employment is decreasing, i.e. that the correctly specified employment variable is concave. Can such a misspecification explain the estimated negative relation between the regulation variable and the exit probability? Our answer is no, since (i) we have identified a negative employment effect, and (ii) the regulation variable is positively correlated with employment, so that regulation can act as a proxy for ‘high employment’. If there were no regulation effect, only a misspecified linearity instead of concavity, the estimated regulation parameter should have been positive to modify the negative effect of high employment levels on the exit probability. Third, in order to sharpen the distinction between regulated and non-regulated establishments, we removed class 2 establishments from the data set and reestimated the model. The estimates turned out to be very similar to the original ones, although the degree of discrimination between regulated and non-regulated establishments became lower (as we would expect, since the data set is diminished). In particular, in Pulp, paper, and paperboard, the P -value of the test of HB : cR = cN increases from 0.01 to 0.05. Finally, we conducted an ‘informal’ goodness of fit test, by calculating a ‘misclassification’ table. An observation was termed ‘misclassified’ if the estimated probability of exit when the unit did exit was below 50%, or if the estimated probability of exit when the unit did not exit was above 50%. It turned out that in all sectors the proportion of correct ‘exit predictions’, according to this criterion, was around 20%, whereas the proportion of correct ‘no-exit predictions’ was around 70%. This reflects that in general it is difficult to predict exit. It should be recalled 54 ERIK BIØRN ET AL. that such a test, although common in the analysis of binary response data, is not a test supported by statistical theory [cf. e.g. Cramer (1991, chapter 5)]. 4.6. DISCUSSION Our results agree well with the estimates in Golombek and Raknerud (1997), who analysed the same three sectors by a multistate Markov model and estimated the probability for regulated and non-regulated establishments to increase employment, decrease employment, and exit. In the present study, the estimated relative exit risk between regulated and non-regulated establishments with equal covariates is 0:3 (P -value = 0:03) in Pulp, paper, and paperboard, and also 0:3 (P -value = 0:19) in Iron, steel, and ferroalloys. The corresponding figures in the Golombek – Raknerud study are 0:18 (0:005) and 0:28 (0:08). Hence, according to both studies, regulated establishments have the lowest exit probabilities. However, the evidence of discrimination between regulated and non-regulated establishments is less clear in the present study. One difference between the two modelling approaches is that Golombek and Raknerud (1997) treat the number of employees as an endogenous variable, whereas we condition on this variable. They find that in both Pulp, paper, and paperboard and in Iron, steel, and ferroalloys, the probability to increase employment is highest for regulated establishments. This is one reason why we find less evidence of discrimination between regulated and non-regulated establishments than they do: Recall that we have identified a negative relationship between employment and the exit probability. Hence, if regulations lead to a rise in the probability to increase employment, regulated establishments will, ceteris paripub, have the lowest exit probability. If all variables in xit and zit could be treated as exogenous with respect to the parameter vectors and ,8 and provided that none of the omitted variables were correlated with xit or zit , there would be no problems with biased estimators or test statistics. However, it is not difficult to raise arguments why such an exogeneity assumption may be questionable. First, when an establishment becomes regulated, due to e.g. pollution abatement costs, its profit usually shifts. A change in present profitability may (partly through the formation of expectations) change the optimal time to exit, i.e. the operation decision is affected. Then there may be simultaneity between the exit decision, the employment, and the profitability. Second, if the regulator tends to use, for instance, profitability and establishment size as criteria of whether to impose environmental standards or not, then there may be simultaneity between the regulation decision and these covariates. The possibility that regulated establishments, ceteris paribus, may need more employees than non-regulated ones to comply with the environmental standards may be another source of simultaneity. Ideally, we should condition only on variables that are not affected by environmental regulations. This implies, in the notation of Section 2, that the conditional distribution of the covari- ENVIRONMENTAL REGULATIONS AND PLANT EXIT 55 ates given history, P (Xit j Hi;t 1 ), should contain no information on the effects of environmental regulations. However, environmental regulations may influence several variables, inter alia, some of our covariates. If e.g. regulations change profitability, it may be misleading to compare and base policy recommendations on estimated exit probabilities of regulated and non-regulated establishments, conditional on profitability. If there is a systematic difference in the distribution of the covariates between the regulated and non-regulated establishments, regulation could have an impact on the estimated exit probabilities P (Cit j Rit ; Xit ; Hi;t 1 ) even if R = N [cf. Equation (9)]. This would be the case if there is causality from regulation to some of the covariates. Then, attempts to make inference, and base policy recommendations, on the estimated ’s in the partial likelihood function (11) could be misleading. Contrary to the approach in this paper, Golombek and Raknerud (1997) estimate conditional exit probabilities where all covariates in the present model, except prices, are ‘integrated out’. Hence, their coefficient estimates represent to a larger extent ‘reduced form coefficients’ than those in the present study and do not rely upon the strong assumption of non-causality from regulatory status to profitability. On the other hand, their results are based on the assumption of non-causality in the other direction, i.e. from profitability to regulatory status. It is not easy to decide which of the non-causality assumptions are most plausible.9 Probably, the bona fide solution to the exogeneity problem is to specify a simultaneous equation model in which one estimates the joint distribution of all non-exogenous variables, including, for instance, the regulation indicator Rit , the exit indicator Cit , the scale of operation, the profitability, and the employment. We are, however, badly in need of econometrically applicable theories in this field. 5. Concluding Remarks The purpose of this study has been to compare the exit probability between environmentally regulated and non-regulated establishments in Norwegian manufacturing industries. Focusing on three sectors with high shares of units which have been under environmental regulations between 1976 and 1991, we find that in Pulp, paper and paperboard and in Iron, steel and ferroalloys, non-regulated establishments had, ceteris paribus, the highest exit probability. In both of the latter sectors, the estimated exit probability of regulated establishments is about one third of the exit probability of non-regulated establishments (evaluated at the sample mean of the covariates). At the conventional 5% level, however, the exit probability of regulated establishments is significantly lower in Pulp, paper, and paperboard only. Finally, while non-regulated establishments have the lowest estimated exit probability in Basic industrial chemicals, due to very few exits it is not possible to make precise inference for this sector. If the environmental regulator has extensive information about the establishments, we may give the following explanation of why regulated units (in two 56 ERIK BIØRN ET AL. sectors) had the lowest estimated exit probability in our data period: Our results may reflect that the regulator imposes strict standards mainly on firms with a low predicted probability of exit. To provide an argument against this hypothesis, assume that the regulator (i) believes that a logit model including profits and employment as covariates provides an adequate representation of the mechanism determining exit, (ii) has perfect information on these covariates (in 1976), and (iii) knows the parameter values that would maximize the likelihood function for all exits between 1976 and 1986. Then, although the P -values related to the covariates are less than 0.1%, it turns out that as much as 27% of the predictions (operative vs. closed down) would be incorrect. If the regulator uses historic information from 1976–1980 (unweighted average values) only, the share of incorrect predictions increases from 27% to 35%. To sum up, based on historic information it is difficult to predict which firms that will exit during the next 10 years. As the regulator has less information than assumed above (claiming that its information is primarily limited to sectorial output prices and enterprise accounts), and certainly does not know the Maximum Likelihood estimates of the coefficients, we find it hard to believe that the regulator is able to pick out establishments with a low 10-year exit probability. Our findings cannot be explained by standard economic variables like profits, stock of goods and scale of operation as the estimated exit probabilities are functions of these factors. Hence, the results must reflect other factors, e.g. internal changes that may take place in the establishments due to the imposed environmental regulations. This hypothesis is in line with Porter (1990, 1991) who argues that if the regulator focuses on pollution prevention rather than solely on cleanup, and if there are no constraints on the technology used to achieve the environmental goals, then strict environmental regulations may work in favour of domestic industry. According to Porter, regulations may encourage firms to re-engineer so that they pollute less and perhaps even become more cost effective. Moreover, an environmental regulation may encourage firms to develop more resource-efficient products. In a subsequent study, we therefore plan to examine internal effects following from environmental regulations by undertaking in-depth interviews of managers, engineers, and unions in regulated and non-regulated establishments. From these data, we may investigate whether environmental regulations lead to reduced internal slack, changes in administrative routines, more resources spent on R & D, and transformation of former residuals into marketable products. Our multinomial qualitative choice model has examined factors determining not only the exit decision of establishments, but also the regulator’s decision of whether to regulate or not. As emissions permits are, in general, granted for 10 years, the estimated probability to be under environmental regulation in the next period, given that an establishment is already regulated, is (approximately) one. Regarding the probability to change regulatory status from being non-regulated to becoming regulated, we find – as expected – a positive relationship between firm size (number of employees) and the probability to become regulated. However, to ENVIRONMENTAL REGULATIONS AND PLANT EXIT 57 our surprise we have identified that establishments with weak profitability had the highest probability to become under environmental regulation. Needless to say, several unsolved problems remain. Owing to problems of simultaneity and causality pointed out above, one has to be cautious in giving policy recommendations from our results. For instance, a statement that ‘more environmental regulation is favourable since it tends to reduce the number of establishments going out of business’ would be premature. Let us therefore end up by mentioning a few topics for future research. First, the role of the capital input should be analysed in more depth, since regulations often require particular capital investment, which may affect capital (and total factor) productivity as well as profitability. Second, it is desirable to take a larger step towards a joint analysis of the entry and exit decisions of firms and how their respective probabilities are related to environmental regulations, directly as well as indirectly. In such a more ambitious project, it may be necessary to treat, inter alia, scale of operation, employment, capital investment, output price(s), cash-flow, and profitability as determined simultaneously with entry and exit. Acknowledgement Earlier versions of the paper have been presented at the Sixth Biennial International Conference on Panel Data, Amsterdam, June 1996, the Econometric Society European Meeting, Istanbul, August 1996, the E.A.R.I.E. Conference, Vienna, September 1996, and at seminars at the Norwegian Ministry for Industry and Energy and at Statistics Norway. Comments from John Dagsvik, Michael Hoel, seminar participants and two referees are appreciated. We gratefully acknowledge financial support from the European Communities’ Environmental Programme and the Norwegian Ministry for Industry and Energy. Notes 1. That is, the variation of one parameter in the parameter set does not affect the admissible region of other parameters. 2. We have made some attempts with a probit parameterization. Apart from the different ‘scaling’ of the probit and the logit probabilities, only minor effects on the empirical results were found. This is probably explained by the ‘textbook argument’ that if there is no large concentration of observations in the tails of the distributions, the choice between these two parameterizations is for practical purposes immaterial, see e.g. Maddala (1983, p. 23). For a slightly different view, see Greene (1993, p. 638). 3. Ownership changes are not treated as exits. 4. However, a certain class of multinomial choice models can be derived formally by maximizing stochastic utility in cases where utility is assumed to follow an extreme value distribution, cf. McFadden (1984). 5. The odds of a specific response is the ratio between the response probability, say  , and its complementary probability, 1  . The log-odds is the (natural) logarithm of =(1  ). The logit parametrization of a response probability implies that its log-odds is a linear function of the covariates. 58 ERIK BIØRN ET AL. 6. These covariate vectors also include a ‘one’, corresponding to the intercept term of the coefficient vectors. For notational simplicity, it is suppressed here. 7. We also found that using capital stock as a scale variable leads to poorer convergence properties of the maximum likelihood algorithm and much lower likelihood values in optimum. 8. Confer the concept weak exogeneity as defined in Engle et al. (1983, section 2.5), and Davidson and MacKinnon (1993, section 18.2). 9. A general warning should be raised against interpreting conditional models as representing causal relations. The interpretation of the estimated effect of Xit (which includes profitability) on the regulatory probability P (Rit j Xit ; Hi;t 1 ) based on the partial likelihood function (12) essentially reflects correlation, not causality, and hence can be spurious. For a discussion of the causality concept in econometrics, see Davidson and MacKinnon (1993, pp. 629–631). References Audretsch, D. B. (1995), ‘Innovation, Growth and Survival’, International Journal of Industrial Organization 13, 441–457. Audretsch, D. B. and J. Mata (1995), ‘The Post-Entry Performance of Firms: Introduction’, International Journal of Industrial Organization 13, 413–419. Barbera, A. J. and V. D. McConnell (1986), ‘Effects of Pollution Control on Industry Productivity: A Factor Demand Approach’, Journal of Industrial Economics 35, 161–172. Brännlund, R., R. Färe and S. Grosskopf, (1995), ‘Environmental Regulation and Profitability: An Application to Swedish Pulp and Paper Industry Mills’, Environmental and Resource Economics 6, 23–36. Conrad, K. and C. J. Morrison (1989), ‘The Impact of Pollution Abatement Investment on Productivity Change: An Empirical Comparison of the U.S., Germany, and Canada’, Southern Economic Journal 55, 689–698. Cramer, J. S. (1991), The Logit Model. An Introduction for Economists. London: Edward Arnold. Davidson, R. and J. G. MacKinnon (1993), Estimation and Inference in Econometrics. Oxford: Oxford University Press. Deily, M. E. and W. B. Gray (1991), ‘Enforcement of Pollution Regulations in a Declining Industry’, Journal of Environmental Economics and Management 21, 260–274. Doms, M. E., T. Dunne and M. J. Roberts (1995), ‘The Role of Technology Use in the Survival and Growth of Manufacturing Plants’, International Journal of Industrial Organization 13, 523– 542. Engle, R. F., D. F. Hendry and J.-F. Richard (1983), ‘Exogeneity’, Econometrica 51, 277–304. Fahrmeir, L. and G. Tutz (1994), Multivariate Statistical Modelling Based on Generalized Linear Models. New York: Springer. Gollop, F. M. and M. J. Roberts (1983), ‘Environmental Regulations and Productivity Growth: The Case of Fossil-Fueled Electric Power Generation’, Journal of Political Economy 91, 654–674. Golombek, R. and A. Raknerud (1997), ‘Do Environmental Standards Harm Manufacturing Employment?’ The Scandinavian Journal of Economics 91(1), 29–44. Golombek, R. and A. Raknerud (1995), Environmental Regulations and Manufacturing Employment – A Study on Norwegian Data. Memorandum from the Department of Economics, University of Oslo, No. 17, 1995. Gray, W. B. and R. J. Shadbegian (1993), ‘Environmental Regulation and Manufacturing Productivity at the Plant Level’, NBER Working Paper # 4321. Greene, W. H. (1993), Econometric Analysis, Second Edition. New York: Macmillan. Johansen, L. (1959), ‘Substitution versus Fixed Production Coefficients in the Theory of Economic Growth: A Synthesis’, Econometrica 27, 157–176. Jorgenson, D. W. and P. J. Wilcoxen (1990), ‘Environmental Regulation and U.S. Economic Growth’, RAND Journal of Economics 21, 314–340. Maddala, G. S. (1983), Limited-Dependent and Qualitative Variables in Econometrics. Cambridge: Cambridge University Press. ENVIRONMENTAL REGULATIONS AND PLANT EXIT 59 McFadden, D. L. (1984), ‘Econometric Analysis of Qualitative Response Models’, Ch. 24 in Z. Griliches and M. D. Intrilligator, eds., Handbook of Econometrics, vol. II. Amsterdam: NorthHolland. Porter, M. E. (1990), The Competitive Advantage of Nations. New York: The Free Press. Porter, M. E. (1991), ‘America’s Green Strategy’, Scientific American, April 1991, 168. SFT (1987), ‘Control of Manufacturing Emissions’, Mimeo, January 13, 1987, The State Pollution Control Authority.