Transcript I(1)

The Effect of Government Expenditure
on the Environment
Prof. George Halkos & Mr Epameinondas Paizanos
Lab of Operations Research, Dept of Economics University of Thessaly
6th International Scientific Conference on Energy and Climate Change
Athens, 10 October 2013
This research has been co-financed by the European Union (European Social Fund – ESF) and Greek national funds through the
Operational Program "Education and Lifelong Learning" of the National Strategic Reference Framework (NSRF) - Research Funding
Program: Heracleitus II. Investing in knowledge society through the European Social Fund.
Prof. George E. Halkos & Epameinondas A. Paizanos
1
Introduction
• A large fraction of national GDP is spent by governments
affecting a variety of economic variables and wealth in
general.
• Despite the important influence that public spending may
have on the environment, this relationship has not been
studied extensively in the literature.
• The effect of government spending on the environment
may be distinguished between direct and indirect effects.
Prof. George E. Halkos & Epameinondas A. Paizanos
2
Introduction (cont’d)
• Four mechanisms by which level and composition of
fiscal spending may affect pollution levels:
• Scale effect (increased environmental pressures
due to economic growth),
• Composition effect (increased human capital
intensive activities instead of physical capital
intensive industries that harm environment more),
• Technique effect (due to higher labor efficiency)
• Income effect (increased income raises demand for
improved environmental quality).
Introduction (cont’d.)
• The direct effect of government spending on pollution has
been studied in Frederik and Lundstrom (2001), Bernauer and
Koubi (2006), Lopez and Palacios (2010), Lopez et al. (2011),
with contradicting empirical results.
• In addition, government size has been found to influence
prosperity (Bajo-Rubio, 2000; Folster and Henrekson, 2001;
Bergh and Karlsson, 2010) which in turn may impose an
indirect effect on pollution, depending on the shape of the
Environmental Kuznets Curve (Grossman and Krueger, 1995).
• Our paper is the first to distinguish and empirically estimate
both direct and indirect effects of fiscal spending on the
environment.
Prof. George E. Halkos & Epameinondas A. Paizanos
4
Data
• Panel data for 77 countries with full set of SO2, CO2,
share of GOVEXP, GDP/c and other explanatory
variables for 1980-2000. 1617 observations per variable.
• The two pollutants vary in geographical range of impact.
Since SO2 moves away from the atmosphere within 10
days after its emission, impact is mainly local or regional,
whereas CO2 emissions, whose atmospheric life is
between 50 to 200 years, have a more global impact.
• Based on sources of pollution, SO2 pollution characterized
as production-generated, while CO2 emissions are mix
between
production
and
consumption–generated
pollution.
Prof. George E. Halkos & Epameinondas A. Paizanos
5
• Model
Methodology
• Control variables in Eq. (1) include GDP/c (scale effect),
household final consumption (income effect) and democracy
level (proxy for environmental regulation).
• In Eq. (1), coefficient of GOVSHARE (i.e. direct effect) mainly
captures composition effect and part of technique effect.
• Equation 2 is augmented Solow model: is a production function
based formulation and expresses income as function of share
of GOVEXP in GDP and other explanatory factors (investment,
education, population growth, trade, inflation rate)
Prof. George E. Halkos & Epameinondas A. Paizanos
6
Methodology (cont’d)
• To consider correlation between cross-section specific errorcomponent (e.g. country specific climate and geography) and
explanatory variables, FE are used instead of RE.
• Since our panel consists of large N and T dimensions, dynamics and
non-stationarity are taken into account by employing the Dynamic
Fixed Effects (DFE) estimator proposed by Pesaran and Smith.
• To address bias occurring from potential endogeneity between
GOVSPEN with pollution and income respectively, we apply A-B
GMM which treats predetermined and exogenous variables as
instruments in systematic way, while taking into account dynamics.
• For equation (1) we set-up an initial general ADL. Nonlinearity in
parameters requires models are estimated using ML.
Prof. George E. Halkos & Epameinondas A. Paizanos
7
Methodology (cont’d.)
Prof. George E. Halkos & Epameinondas A. Paizanos
8
Panel data methods
• First method employed imposes same intercept and slope parameter for
all countries (equivalent to OLS estimation).
• Second method is the FE allowing each individual country to have a
different intercept treating the δi and ζt as regression parameters. This
practically means that the means of each variable for each country are
subtracted from the data for that country and the mean for all countries
in the sample in each individual time period is also deducted from the
observations from that period. Then OLS is used to estimate the
regression with the transformed data.
• Third model is the RE in which individual effects are treated as random,
with δi and ζt treated as components of the random disturbances.
Residuals from an OLS estimate of the model with a single intercept are
used to construct variances utilized in a GLS estimates.
• To control for non-observable specific effects 2SLS was applied but the
results were insignificant.
GMM Arellano - Bond is adopted using the orthogonality conditions between the lags of the dependent
and the disturbances as additional instruments. This is the GMM non-linear instrumental variable estimator
q
Yit  t  i   k Yi (t k )   ( L) X it  vit
k 1
t = q+1,…,Ti i = 1,2,...,N and where λt and ηi are specific and individual effects respectively,
Xit is a vector of explanatory variables, β(L) is a vector of associated polynomials in the lag operator and q is
the maximum lag length.
Identification of the model requires restrictions on serial correlation of the error term vit and on the properties of
the independent variables Xit allowing only for MA or white noise errors.
If the error term was originally autoregressive, the model is transformed
The (Ti –q) equations for individual unit i can be written as
Yi  wi  dii  vi
where δ is a parameter vector including the ακ's, the β's and the λ' s and wi is a data matrix containing the
time series of the lagged endogenous variables, the x' s, and the time dummies. di is a (Tiq) x1 vector of
ones.
Following Arellano and Bond (1998), linear GMM estimators of δ are computed by

 
 
1
    wi* Z i .
.  Z iwi*
 1
 i
Z iH i Z i


 i


 1 
 
1


  Z iYt * 
 .  wi* Z i 
1

  i
Z iH i Z i  i


 i

where wi* and Yi* denote some transformation of wi and Yi such as first differences, orthogonal deviations or
levels. Zi is the matrix of instrumental variables and Hi is an individual specific weighting matrix.
Prof. George E. Halkos & Epameinondas A. Paizanos
11
- Estimated effect of GOVEXP share on GDP negative and
significant, regardless of the method used
- Estimated coefficient of GOVEXP equal to -0.210, while DFE
estimate of GOVSIZE effect on income -0.872 suggesting that
consideration of dynamics increases estimated impact of
GOVSHARE on income/c, even without accounting for
endogeneity.
- To account for autocorrelation and heteroskedasticity,
standard errors reported are robust and for FE estimation we
report Huber-White-Sandwich estimates of var-cov matrix.
• Estimated impact of GOVEXP on GDP even greater, suggesting
that increase of 1% in share of GOVSPEN of GDP, ceteris
paribus, reduces income/c by 1.809%.
• Signs and significance of coefficients associated with other
control variables are plausible and consistent with literature,
apart from human capital proxy which although has the
expected sign, is significant only in OLS estimates.
• Impact of capital stock, represented by share of investment in
GDP positive and significant across all estimation methods.
• Population growth has consistent negative and significant effect,
• Τrade-openness coefficient significant with expected positive
sign.
• Applying AB two-step GMM estimator, dynamics are still taken
into account but GOVSHARE is treated as endogenous. We
use FD and OD GMM to control for fixed country effects.
Significance of lagged dependent variable suggests dynamic
specifications should be preferred.
• Tests for 1st and 2nd order serial correlation related to residuals
from the estimated equation are asymptotically-distributed as
normal variables under Ho: no-serial correlation. Test for AR(1)
rejected as expected, while there is no evidence that the
assumption of serially uncorrelated errors is inappropriate.
• For both equations we test validity of instruments with Hansen
test, which failed to reject Ho: IVs are uncorrelated with
residuals. We also report Difference Hansen test for exogenous
IV subset which does not reject Ho that subset is valid.
• Dynamic Fixed Effects: DFE estimation assumes intercepts
differ across countries but that LR coefficients are equal across
countries.
• Mean Group: Alternative estimation method that fits model for
each country individually and calculates arithmetic average of
the coefficients is the mean-group estimator (MG). This method
is less restrictive than DFE since intercepts, slope coefficients
and error variances are all allowed to differ across countries.
• Pooled Mean Group: PMG estimator combines DFE and MG
methods by allowing the intercept, short-run coefficients and
error variances to differ across groups while assuming equality
of LR coefficients.
• Table 4a: Comparing the MG and PMG estimators, with the use
of a Hausman test, we see that the PMG estimator, efficient
estimator under H0, is preferred and thus, assuming LR
coefficients to be equal across panels, is appropriate in our
panel.
Fisher-type Philips-Peron tests allowing heterogeneity of autoregressive parameters
In all cases variables are I(1)
Prof. George E. Halkos & Epameinondas A. Paizanos
16
• We reject Ho: no-cointegration in 4/7 cases for SO2 and in 5/7 in CO2.
Evidence of cointegration.
• Application of DFE requires variables cointegrated (LR relationship).
DFE is applicable.
Prof. George E. Halkos & Epameinondas A. Paizanos
17
Prof. George E. Halkos & Epameinondas A. Paizanos
18
Prof. George E. Halkos & Epameinondas A. Paizanos
19
• Table 4 provides estimates of pollution/c utilizing GMM estimates of Eq.1.
Based on FE estimates GOVSHARE of GDP has negative and significant
direct effect on SO2/c and insignificant negative relationship on CO2/c.
• Dynamics in columns 2 and 4 of Table 4. Comparing the MG and PMG
estimators, with Hausman test, we see PMG estimator, the efficient estimator
under Ho, is preferred and thus, assuming LR coefficients to be equal across
panels, more appropriate.
• Another application of Hausman test suggests that simultaneous equation
bias between error term and lagged dependent variable minimal and DFE
model is most appropriate.
• DFE estimates suggest that GOVSHARE of income possesses a significant
negative relationship with SO2/c and CO2/c.
• Both pollutants have significant cubic relationship with income/c. Taking into
account endogeneity in A-B GMM estimates produces turning points for CO2
well within sample.
• Household income effect is negative, although insignificant in all
cases except for SO2 in FD GMM.
• Share of investment increases pollution, but the effect is
significant only for CO2.
• Coefficient of trade-openness always negative, but mostly
insignificant.
• Effect of population density robustly positive
• Democracy index insignificant in all specifications.
• Negative direct effect of GOVSHARE of income on pollution is
estimated by all models (Table 4) except in case of GMM (CO2).
• In GMM results, increase of GOVEXP by 1%, results in 0.903%
reduction of SO2/c. Direct effect on CO2 insignificant. Indirect
effects negative at median income, leading to negative total
effect for both pollutants.
• Direct effect is insignificant and considerably smaller for CO2, in
absolute values. This result was expected, considering both
pollutants’ impact on human health and technological capabilities of
reducing their levels in the atmosphere.
• In particular, SO2 emissions externalities are local and immediate
while CO2 emissions externalities are global and occur mostly in the
future. Moreover, local and instant environmental degradation, as in
the case of SO2, increases demand for technological improvements to
diminish that impact.
Prof. George E. Halkos & Epameinondas A. Paizanos
22
• The negative sign of indirect effect occurs from the positive relationship
between income and pollution at the median income level. Explicitly, at
this income level an increase in government spending leads to a
reduction in income and, as a result, to a reduction in emissions.
• The estimated indirect effects are notably larger than the direct effects.
Prof. George E. Halkos & Epameinondas A. Paizanos
23
• The total effect of government share on SO2/c is negative for low
levels of per capita income and then turns to positive (above
$10,809), while the total effect on CO2/c is also negative but becomes
positive only for very high income levels (above $16,438).
• These patterns largely depend on the relationship between pollution
and income levels described by the EKC.
Prof. George E. Halkos & Epameinondas A. Paizanos
24
• Figures 1 and 2 present direct, indirect and total
effects of GOVSHARE of income on emission
levels against income/c.
• For CO2 direct effect insignificant. Indirect effect
(GDP / c)
increases with income/c, since (Govshare) =-1.809
 ( P / c)
 (GDP / c)
• and
falls from 1.27 to –7.17 for SO2/c
and from 0.22 to -1.39 for CO2/c throughout the
sample income range. These patterns depend
on relationship between pollution and income
levels (EKC).
• Total effect of GOVSHARE on SO2/c negative for low
levels of income/c and then positive, while total effect
on CO2/c also negative but becomes positive only for
very high income levels.
• Table 5 reports estimated income level at which total
effect changes from negative to positive. GMM
estimates indicate that this level is $10,809 for SO2/c
and $16,438 for CO2/c, i.e. total effect of GOVSHARE
of income on CO2/c negative through most of sample
income range.
• From figures pattern of total effect is determined by
shape of indirect effect.
Sensitivity analysis
Relative tests indicate that the results are robust to :
• Inclusion of a GOVEXP composition variable.
• Omission of time-variant variables.
• Inclusion of interactive terms like (GOVEXP x GDP/c).
• Use of different model specifications.
• Inclusion/exclusion of extreme observations.
Prof. George E. Halkos & Epameinondas A. Paizanos
27
Policy Implications
• Policy implications, differ according to level of income in a country.
• In general, reducing GOVSIZE enhances economic performance.
• However, cutting GOVEXP should be undertaken with particular care
for some levels of GDP. For SO2 and CO2 pollution, results suggest
that reducing government size in countries with an income level less
than $10,809 and $16,438 respectively, leads to deterioration of
environmental quality.
• Therefore, cutting government expenditure in these countries should
be accompanied by appropriate environmental regulation along with
the establishment of international environmental treaties.
Prof. George E. Halkos & Epameinondas A. Paizanos
28
Policy Implications (cont’d.)
• Cutting GOVEXP in countries with higher income levels, leads to
improvements in both income and environmental quality.
• In particular, countries with income level at the decreasing area of the
EKC are more likely to have already established the environmental
legislation and to have undertaken public expenditures for
improvement of environmental quality, hence they are susceptible to
diminishing returns from a further increase in government size.
• Combining our findings with the results from Lopez et al. (2011),
cutting out public spending items that increase market failure will be
the most beneficial in high income countries.
Prof. George E. Halkos & Epameinondas A. Paizanos
29