Smooth transition pollution-income paths

Transcription

1 Smooth transion pollution-income paths Nektarios Aslanidis and Anastasios Xepapadeas Universy of Crete Department of Economics Universy Campus 74 00, Rethymno Greece September, 004 Abstract We explore the idea of regime swching as a new methodological approach in the analysis of the emission-income relationship. A static smooth transion regression model wh fixed-effects is developed. The basic idea is that when some threshold is passed, the economy could move smoothly to another regime, wh the emission-income relationship being different between the old and the new regime. We justify our methodology by proving that the quadratic or cubic polynomial model used in the lerature derives from the smooth transion regression specification. The methodology is applied to a panel dataset of US state-level sulfur dioxide and nrogen oxide emissions covering 48 states over the period We find robust smooth N-shaped and smooth inverse-vshaped pollution-income paths for the sulfur dioxide. For the nrogen oxide emissions, environmental pressure tends to rise wh economic growth in the early stages of economic development, then slows down but does not decline wh further income growth. Keywords: Environmental Kuznets Curve, smooth transion regression, regime swching, thresholds. JEL Classification: C, O, Q.

2 . Introduction In the analysis of the emission-income relationship, there exists a set of theoretical models, which derives inverted V-shaped curves by having pollution increasing wh income until some threshold point is passed, after which pollution is reduced. John and Pecchenino (994) consider an overlapping generations model where economies wh low income or high environmental qualy are not engaged in environmental investment, that is, pollution abatement. When environmental qualy deteriorates wh growth, the economy moves to posive abatement, then the environment improves wh growth and the relationship is an inverted V-shaped. Stokey (998) generates an inverted V-shaped curve by considering a static optimization model where below a threshold income level only the dirtiest technologies are used. As economic activy and pollution increase, the threshold level is passed and cleaner technologies are used. Jaeger (998) derives the inverted V- shaped curve by considering a threshold in consumer preferences. Below the threshold the marginal benef from improving environmental qualy is small, whereas when pollution increases and the threshold is passed, qualy may be improved. The main purpose of this paper is to empirically explore the inverted V model by introducing the idea of regime-swching as a new methodological approach in the analysis of the emission-income relationship. More specifically, econometric techniques appropriate for a static smooth transion regression (STR) wh panel data are developed. In this model regression functions are not identical across all observations in a sample but fall into classes. The basic idea is that when some threshold is passed, then the economy could move smoothly to another regime, wh the emission-income relationship being different between the old and the new regime. The low-income regime might correspond to an increasing emission-income relationship, while in the regime after the threshold the emission-income relationship might be decreasing. The abrupt regime swch is a special case of the smooth regime swch implying that the discrete inverse V-shaped emissionincome paths are a special case of the more general smooth inverse V-shaped ones. Furthermore, we justify our methodology by proving that the quadratic or cubic polynomial model used to examine the emission-income relationship derives from the STR model, and therefore, the latter can be seen as the underlying structural specification For a recent lerature review, see, for example, Levinson (00).

3 in this lerature. Finally, we also develop N-shaped emission-income paths in a smooth transion regression framework where pollution can be increasing at low levels of income, decreasing at high levels, and then increasing again at even higher levels of national income. The methodology is applied to a panel dataset of US state-level sulfur dioxide ( SO ) and nrogen oxide ( NO ) emissions covering 48 states over the period from 99 to x 994. We find robust smooth N-shaped and smooth inverse-v-shaped pollution-income paths for SO. In particular, emissions of SO are found to peak at a relatively early stage of economic development and then decrease at middle-to-high levels of income. What is more interesting, however, and perhaps more policy relevant is that while in high-income states environmental qualy continues to improve, in low income states pollution increases wh the later increases in economic activy. As for NO x emissions, environmental pressure tends to rise wh economic growth in the early stages of economic development, then slows down but does not decline wh further income growth.. Methodology. Model By regime-swching behaviour we mean that the regression functions are not identical across all observations in the sample or that they fall into discrete classes. One of the most prominent among the regime-swching models in the macroeconomics area has been the threshold class of models and s smooth transion generalization (STAR models) promoted by Teräsvirta and his co-authors (Teräsvirta, 994, Teräsvirta and Anderson, 99, Granger and Teräsvirta, 993). Regime-swching models are flexible enough to allow several different types of effects that could be observed in the relationship between pollution and income. The equation of interest is the static singlethreshold smooth transion regression (STR) model wh state-specific fixed effects given by 3

4 P = µ + β c; Y ) + u, i =,..., N ; t =,..., T () i Y + ( β Y ) F( γ, where P is a measure of per capa air pollution in state i in year t, Y is per capa GDP in state i in year t, β ( β, β ) is the parameter vector, µ i represents statespecific effects and u is an IID error term. The function F γ, c; Y ) is the transion ( function, continuous and bounded by zero and uny, γ and c are s parameters, whereas Y is assumed to act as the transion variable. By wring () as P = µ + β + β F ) Y + u i ( we can see that the model is locally linear in Y and that the combined parameter ( β + β F) is a function of the transion variable Y. As F is bounded between zero and one, the combined parameter fluctuates between β and β + β. Values of zero of the transion function identify regime one and values of uny identify the alternative regime. In the analysis of emission-income relationship, this property makes possible, for example, to derive smooth inverted V-shaped curves by having pollution increasing wh income ( β > 0 ) until income passes some threshold, after which pollution is reduced ( β +β 0). It is also reasonable to assume that in utily terms the disutily 3 < of pollution is related to the flow of new pollutants and is thus adequate to use a static model. An alternative way of wring () is µ i+ β Y + u P = µ i+ ( β + ) β Y + u F F = 0 = We also included a time-trend to account for aggregate time effects such as exogenous technical progress. The time trend is insignificant and the results did not change. We argue that both the composion and (endogenous) technology effects can be captured by per capa income, so we focus solely on the relationship between pollution and income. 3 Since the flow of pollution affects utily in a negative way. 4

5 Obviously, a weighted mixture of these two models applies if 0 < F <. The practical applicabily of the above specification depends on how F is defined. One form of transion function used in the lerature is the logistic function ( + exp( γ ( Y ))) F( γ, c; Y ) = c, γ > 0 () where the parameter c is the threshold between the two regimes or the location of the transion function, and the parameter γ determines the smoothness of the change in the value of the logistic function and thus the speed of the transion from one regime to the other. When γ, F becomes a step function ( F = 0 if Y c and F = if Y > c ), and the transion between the regimes is abrupt. In that case, the model approaches a threshold model (Hansen, 999). A smooth transion between the two extremes may be an attractive parameterization because, from a theoretical point of view, the assumption of two regimes may sometimes be too restrictive compared to the STR alternative. For instance, instead of assuming that in the emission-income relationship there are two discrete regimes, degradation and improvement, may be more realistic to assume a continuum of states between these two extremes. Another argument is that economic agents may not all act promptly and uniformly at the same time probably due to heterogeneous beliefs. Nevertheless, the two viewpoints are not in conflict since the abrupt swch is a special case of the STR model and can therefore be treated whin that framework. Model () has a single threshold. An obvious extension would be to perm multiple thresholds. For example, the double threshold or three-regime STR model takes the form P = µ + β ; Y ) + u i Y + ( β Y ) F ( γ, c; Y ) + ( β 3Y ) F ( γ, c where β β, β, β ) is the parameter vector and γ, c and γ, c are the ( 3 parameters of F and F, respectively. It is assumed that income Y determines both transions, while the second transion function is defined analogously to (). If is 5

6 assumed that c < c, the parameters of this model change smoothly from β via β to β 3 for increasing values of Y.. Estimation A tradional method of eliminating the individual effect µ i is to remove individualspecific means. While straightforward in linear models, the non-linear specification in () calls for a more careful treatment. Once we have removed individual-specific means to estimate the STR model, is computationally convenient to first concentrate on the transion function parameters. Note that assigning fixed values to the parameters in the transion function makes the STR model linear in parameters. That is, condional on the transion function, the parameters of the STR can be estimated by OLS. We first carried out a two-dimensional grid search procedure using 50 values of γ ( to 50) and at least 00 equally spaced values of c whin the observed range of the transion variable. Essentially, Y is ordered by value, extremes are ignored by omting the most extreme 5 values at each end and the 00 values are specified over the range of the remaining values. This procedure guarantees that the values of the transion function contain enough sample variation given the choices of γ and c. The model wh the minimum RSS value from the grid search is used to provide γ and c. Following Teräsvirta (994) the exponent of the transion function is standardised by the sample standard deviation of the transion variable. This makes γ scale-free and helps in determining a useful set of grid values for this parameter. Specification of the double threshold model involves a modeling procedure analogous to that in the single transion case. Here a four dimensional grid search is performed over γ,γ =,,50 and 8 values of c,c over the range of the transion variable 4. We have described an algorhm to estimate a STR static model wh fixed effects. As for the consistency of the estimator vector β, in linear static models this estimator is consistent. If we assume that the dependence on γ and c is not of first-order asymptotic 4 Essentially, the first threshold is considered to be over the left part of the observed range of income series whereas the second threshold is over the right part. 6

7 importance, then inference on β can be proceeded as if the estimates γˆ and ĉ were the true values. Hence, β is asymptotically normal and conventional standard errors can be reported..3 Inference Before estimating the STR model is important to determine whether the threshold effect is statistically significant. The null hypothesis of no regime-swching effect in () is H : 0 against H : 0. The null hypothesis can be equally well expressed as H γ 0 = β 0 : β = γ > 0. This is an indication of an identification problem in (); the model is identified under the alternative but not under the null hypothesis, so classical tests have non-standard distribution. This is typically called the Davies Problem (see Davies, 987); for later contributions in the econometrics lerature see Shively (988), Granger and Teräsvirta (993), King and Shively (993), Andrews and Ploberger (994) and Hansen (996). The fixed-effects equation () falls into the class of models considered by Granger and Teräsvirta (993), who find a way of solving the identification problem by circumventing. To discuss this idea, take the logistic transion function in () and s first-order Taylor series approximation wh the null hypothesis γ = 0 as the expansion point. This can be wrten as T ( = δ 0+ δy + R γ, c; Y ) where R is the remainder and 0 = γ = 0 F in () yields δ F, δ = F γ = 0 are constants. Substuting T for P = µ + θ Y + θ Y + u (3) i * * where θ ( β + δ 0β ), θ δ β and u = u + ( β Y ) R( γ, c; Y ). Use of this approximation amounts to giving up information about the structure of the alternative 7

8 hypothesis in order to circumvent the identification problem and obtain a simple test for the null hypothesis. Thus the null hypothesis in () H : γ 0 implies H θ 0 and 0 = 0 : = H θ 0 in (3). Standard asymptotic inference is used to test the null hypothesis since : (3) is linear in the parameters and therefore a t-type test is performed. Next, consider a second-order Taylor series approximation to F. This can be wrten as T ( = δ 0+ δy + δ Y + R γ, c; Y ) where R is the remainder and 0 = γ = 0 Substuting T for F in () yields δ F, δ = F γ = 0, δ = 0. 5F γ = 0 are constants. P = µ + θ Y + θ Y + θ Y + u (4) i 3 * 3 where θ ( β + δ 0β ), δ β * θ, θ 3 δ β and u u + ( β Y ) R ( γ, c; Y ) =. The null hypothesis of lineary is a straightforward F-test of H θ = θ 0 in (4). 0 : 3= As turns out equations (3) and (4) are the benchmark econometric specifications used in environmental Kuznets curve (EKC) studies 5. Therefore, we empirically justify the idea of regime-swching in the analysis of the emission-income relationship by proving that in fact the above auxiliary regressions derive from a STR specification, and so the STR model can be seen as the underlying structural specification in this lerature. 3. Empirical results We use a long-term panel dataset on US state-level sulfur dioxide ( SO ) and nrogen oxide ( NO ) emissions previously used by List and Gallet (999). The data source is the x 5 In many studies the full specification also includes the average GDP per capa over the prior three years and other covariates, which can be recovered in our model if included in (). 8

9 US Environmental Protection Agency (EPA) in their National Air Pollution Emission Trends. For both emissions, there are 368 annual observations from 48 6 states over the period from 99 to The estimated single-threshold log-str 8 model for SO and NO x are presented in Table. To determine whether the threshold effect is statistically significant, we also report results for the test of no threshold effect by estimating the auxiliary regression (4). We find that the test is highly significant wh a p-value of in both emissions, implying strong regime-swching behaviour. In the first panel, the model for SO gives a threshold at per capa GDP of $3,00, which is a relatively small value in the distribution of the GDP transion variable. What is interesting, however, is that there are two classes of states, those wh low-income (associated wh F(GDP) = 0) where pollution increases wh economic growth, and those wh middle-to-high income (associated wh F(GDP) = ) where pollution begins to decline. Also seems that there is a continuum of states between the two extremes, where the transion from one regime to the other is smooth. In the second panel, the model for NO x estimates a threshold at per capa GDP of $5,84 implying two classes of states, those wh low-to-middle income and those wh high-income. The effect of income on pollution is posive throughout the sample, though smaller in magnude for high-income states. According to this specification the transion from one class to the other is abrupt (γˆ = 50), and therefore the model behaves similarly to a threshold model (Hansen, 999). To check the robustness of these finding we also included a time-trend to account for aggregate time effects such as exogenous technical progress. The results showed that the time trend is insignificant for both emissions, and that the estimated slopes and transion function parameters didn t qualative change 9. Thus, consequently, we argue that both the composion and endogenous technology effects can be captured by per capa income, so we focus solely on the relationship between pollution and income. To ease the interpretation of the models, Graph plots the shape of the estimated transion functions versus the transion variable, that is, per capa GDP (in logarhms). 6 Data are not available for Alaska, Hawaii and Washington DC. 7 See List and Gallet (999) for more details on the data. 8 We consider logarhmic transformations of Eq. (). 9 We do not report the results here but they are available from the authors upon request. 9

10 Every point indicates an observation so that one can readily see which values the transion function has obtained and how frequently. In the first panel, we have identified regimes of low-income states (associated wh F(GDP) = 0) and middle-to-high income states (associated wh F(GDP) = ). On the other hand, in the second panel, low-tomiddle income states and high-income are identified. The location parameters (thresholds) are not distributed equally between the left hand-side and right hand-side tails of the functions. In particular, in the NO x graph (second panel) the high- income states class applies to only 8.3% of the sample. The double-threshold models presented in Table are more intuive. In the first panel, the model for SO which estimates thresholds at $3,368 and $7,9 implies three classes of states. It is interesting to note that the coefficient estimates generate a smooth N-shaped pollution-income path, wh trough and peak pollution levels somewhere in the range of low-income and high-income states, respectively. In other words, we observe a pollution-income path increasing at low-income states, decreasing at middle-to-high income states, and then increasing again at very-high income states. However, the first panel of Graph shows that the very-high income class contains sparse data (associated wh F(GDP) = and F(GDP) close to ). Grossman and Krueger (995) dismiss the upper tail of this pattern as an artificial construct of the fact that they use a cubic functional form. Millimet and Stengos (999) find a pattern similar to ours wh a semiparametric specification. As before, NO x emissions seem to increase wh income at decreasing rates. A consistent result from all the above models is that the thresholds occur at reasonable values and consequently spl the states into groups. In order to check the robustness of our findings and provide a more elaborate presentation of these groups, we spl the sample into two groups: we employ a dummy variable, which indicates whether a state s per capa GDP is above or below the average per capa GDP of all states in every year from 99 to 994. In this way we try to develop regime-swching specifications for low-income and high-income states separately. By doing so we also tried to better identified the composion and technology effects 0. The final double- 0 For a detailed explanation of these effects, see Grossman and Krueger (995). 0

11 threshold specifications are illustrated in Table 3, and the transion functions in Graph 3. Our major finding concerns the model for SO which suggests that low income states follow a smooth N-shaped pollution-income path, whereas high-income states seem to follow a smooth inverse-v-shaped path. One possible explanation for this finding could be that low-income states, which try to catch up wh high-income states, may have received less attention from policymakers in the late years where most of the right handse observations cluster. Thus, the scale of economic activies appears to deteriorate environmental qualy. On the other hand, in high-income states structural economic changes and abatement activies seem to have offset this effect even during the later stage of economic growth and thus environmental qualy improved. Once again the model for NO x shows that in both low and high-income states, environmental pressure tends to increase wh economic growth in the early stages of economic development, then slows down but does not decline wh further income growth. These results can be compared to those in the seminal papers in the lerature. Selden and Song (994) confirm the EKC hypothesis for suspended particulate matter (SPM), SO and NO x, but not for CO emissions. Our finding concerning NO x is not in line wh the above work, however Selden and Song employ a different dataset (a panel of 30 countries). Grossman and Krueger (995) find a robust inverse-u-shaped pollutionincome path for SO, smoke and for most of the indicators of water qualy. Shafik and Bandyopadhyay (99) find that only two out of ten types of environmental pollutants, namely SPM and SO, follow an EKC. These studies have allowed for intercept heterogeney, but have ignored the possibily of slope heterogeney running a higher risk of producing inconsistent and biased parameter estimates. Only List and Gallet (999) address this issue by allowing US states to have heterogeneous slopes and provide evidence of the quadratic and cubic polynomial model for most of the states for both SO and NO x. We generalize their approach by allowing for two and three slopes We also considered single-threshold models before settling on the proposed double-threshold specifications. The results are similar to the first and second regime obtained from the double-threshold specifications and are available from the authors upon request. However, this effect gets weaker: compare the coefficients of to

12 (regimes 3 ) and a further continuum of slopes between the extremes. US states differ in a number of ways (in terms of resource endowments, infrastructure, technological developments, public pressures, etc.) but is perhaps too strict to assume that every state has also undergone a distinct pollution-income path. List and Gallet actually employ an F statistic where the null hypothesis tests for identical slopes across all states. The null hypothesis specified like this might be too restrictive since is not surprising for at least one state to have a different slope. 4. Concluding remarks The main purpose of this paper is to re-addresses the pollution-income path from a different perspective. First, we justify the idea of regime-swching and develop the smooth transion model as a more general specification than the quadratic or cubic polynomial model used in the lerature. Second, we empirically estimate a robust smooth N-shaped and smooth inverse-v-shaped pollution-income paths for SO. The thresholds, which can be viewed as turning points, occur at reasonable values 4. Emissions of SO are found to peak at a relatively early stage of economic development (before a state reaches a per capa income of $3,500), and then decrease at middle to high levels of income. What is more interesting from a policy point of view is that while for highincome states environmental qualy continues to improve, for low-income states pollution appears to still increase during the later increases in economic activy. As to NO x emissions, pollution tends to rise wh economic growth in the early stages of economic development, then slows down but does not decline wh further income growth. 3 In fact, the final specification allows for three slopes for low-income states plus three slopes for highincome states. 4 Many studies in the lerature find very high turning points that are not achievable for the majory of the world population.

13 Acknowledgements It is acknowledged that the GAUSS programs used here derive from programs made available by Timo Teräsvirta. Also, we thank John List for providing us the data. Any remaining errors are exclusively ours. 3

14 Bibliography Andrews D.W.K. and W. Ploberger (994), Optimal tests when a nuisance parameter is present only under the alternative, Econometrica 6, Davies R.B. (987), Hypothesis testing when a nuisance parameter is present only under the alternative, Biometrika 74, Granger C.W.J. and T. Teräsvirta (993), Modelling nonlinear economic relationships, Oxford: Oxford Universy Press. Grossman G.M. and A.B. Krueger (995), Economic growth and the environment, The Quarterly Journal of Economics 0, Hansen B.E (996), Inference when a nuisance parameter is not identified under the null hypothesis, Econometrica 64, Hansen B.E (999), Threshold effects in non-dynamic panels: Estimation, testing and inference, Journal of Econometrics 93, Jaeger W. (998), Growth and environmental resources: A theoretical basis for the U- shaped environmental path, mimeo, Williams College. John A. and R. Pecchenino (994), An overlapping generations model of growth and the environment, The Economic Journal 04, King M.L. and T.S. Shively (993), Locally optimal testing when a nuisance parameter is present only under the alternative, Review of Economics and Statistics 75, -7. Levinson A. (00), The ups and downs of the environmental Kuznets curve, in J. List and A. de Zeeuw (Eds.), Recent Advances in Environmental Economics (Edward Elgar: Cheltenham). List J.A and C.A. Gallet (999), The environmental Kuznets curve: does one size f all? Ecological Economics 3, Millimet D.L. and T. Stengos (999), A semiparametric approach to modeling the environmental Kuznets curve across US states, Mimeo, Southern Methodist Universy. Selden T.M. and D.S. Song (994), Environmental qualy and development: is there a Kuznets curve for air pollution emissions?, J. Environ. Econ. Manage. 7, Shafik N. and S. Bandyopadhyay (99), Economic growth and environmental qualy: time-series and cross-country evidence, World Bank Working Papers, WPS 904, Washington, 5 pp. Shively T.S. (988), An analysis of tests for regression coefficient stabily, Journal of Econometrics 39,

15 Stokey N.L. (998), Are there lims to growth?, International Economic Review 39, - 3. Summers R. and A. Heston (99), The Penn World Table (mark 5): An expanded set of international comparisons , Quarterly Journal of Economics 06, Teräsvirta T. (994), Specification, estimation and evaluation of smooth transion autoregressive models, Journal of American Statistical Association 89, Teräsvirta T. and H.M. Anderson (99), Characterizing nonlinearies in business cycles using smooth transion autoregressive models, Journal of Applied Econometrics 7, S9- S36. 5

16 Table : Single-threshold STR models Fixed-state effects STR model for SO SO = 0.56*GDP + (-0.648*GDP)*F(GDP) (3.98) (-3.39) Classification of regimes SO = 0.56*GDP, when F(GDP) = 0 SO = -0.49*GDP, when F(GDP) = ĉ = $3,00(antilog of 8.07), γˆ = 6 R-sq = 0.47 Threshold effect (p-value) Fixed-state effects STR model for NOx NOx = 0.579*GDP + (-0.63*GDP)*F(GDP) (50.6) (-6.065) Classification of regimes NOx = 0.579*GDP, when F(GDP) = 0 NOx = 0.46*GDP, when F(GDP) = ĉ = $5,84(antilog of 9.68), γˆ = 50 R-sq = Threshold effect (p-value) Notes: Models are estimated in logarhmic levels; threshold effect tests for the null of no regimeswching behaviour; values in parentheses are t-ratios. 6

17 Table : Double-threshold STR models Fixed-state effects STR model for SO SO = 0.8*GDP + (-.07*GDP)*F(GDP) + (.50*GDP)*F(GDP) (6.945) (-8.45) (3.84) Classification of regimes SO = 0.8*GDP, when F(GDP) = 0 & F(GDP) = 0 SO = -0.86*GDP, when F(GDP) = & F(GDP) = 0 SO = 0.684*GDP, when F(GDP) = & F(GDP) ĉ = $3,368(antilog of 8.), ˆ γ = 3 ĉ = $7,9(antilog of 9.758), ˆ γ = 5 R-sq = Fixed-state effects STR model for NOx NOx = 0.579*GDP + (.5*GDP)*F(GDP) + (-.496*GDP)*F(GDP) (4.8) (8.7) (-8.863) Classification of regimes NOx = 0.579*GDP, when F(GDP) = 0 & F(GDP) = 0 NOx =.830*GDP, when F(GDP) = & F(GDP) = 0 NOx = 0.334*GDP, when F(GDP) = & F(GDP) = ĉ = $6,608(antilog of 8.796), ˆ γ = 5 ĉ = $9,84(antilog of 9.36), ˆ γ = R-sq = 0.56 Notes: Models are estimated in logarhmic levels; values in parentheses are t-ratios. 7

18 Table 3: Double-threshold STR models wh dummy for high-income states Fixed-state effects STR model for SO SO = 0.6*GDP +0.6*GDP*Dh + (-0.85 *GDP *GDP*Dh)*F(GDP) + (0.885*GDP *GDP*Dh)*F(GDP) (5.098) (.949) (-.45) (-3.940) (0.87) (-.436) Classification of regimes SO = 0.6*GDP, when SO = 0.477*GDP, when SO = *GDP, when SO = -.008*GDP, when SO = 0.86*GDP, when SO = -0.47*GDP, when ĉ = $3,368 (antilog of 8.), ˆ γ = 3 ĉ = $5,063 (antilog of 9.60), ˆ γ = R-sq = F(GDP) = 0 & F(GDP) = 0 low income states F(GDP) = 0 & F(GDP) = 0 high income states F(GDP) = & F(GDP) = 0 low income states F(GDP) = & F(GDP) = 0 high income states F(GDP) = & F(GDP) = low income states F(GDP) = & F(GDP) = high income states Fixed-state effects STR model for NOx NOx = 0.535*GDP -0.83*GDP*Dh + (0.73*GDP -0.55*GDP*Dh)*F(GDP) + (-0.869*GDP *GDP*Dh)*F(GDP) (30.5) (-4.756) (.30) (-5.8) (-.8) (3.668) Classification of regimes NOx = 0.535*GDP, when NOx = 0.35*GDP, when NOx =.66*GDP, when NOx = 0.568*GDP, when NOx = 0.397*GDP, when NOx = 0.9*GDP, when ĉ = $5,464 (antilog of 8.606), ˆ γ = ĉ = $3,508 (antilog of 9.5), ˆ γ = 4 R-sq = F(GDP) = 0 & F(GDP) = 0 low income states F(GDP) = 0 & F(GDP) = 0 high income states F(GDP) = & F(GDP) = 0 low income states F(GDP) = & F(GDP) = 0 high income states F(GDP) = & F(GDP) = low income states F(GDP) = & F(GDP) = high income states Notes: Models are estimated in logarhmic levels; the dummy Dh indicates whether a state s per capa GDP is above or below the average per capa GDP of all states in each year from 99 to 994; the values in parentheses are t-ratios. 8

19 Graph : Transion function F of single-threshold SO (upper panel) and NOx (lower panel) models versus GDP (in logarhms). 9

20 Graph : Transion functions F and F of double-threshold SO (upper panel) and NOx (lower panel) models versus GDP (in logarhms). 0

21 Graph 3: Transion functions F and F of double-threshold (wh high- income states dummy) SO (upper panel) and NOx (lower panel) models versus GDP (in logarhms).