The impact of internet use on air pollution: Evidence from emerging countries

Abstract

The goal of this study is to analyze the impact of Internet use, employed as a proxy for information and communications technologies (ICTs), on CO₂ emissions. Using a panel of 20 emerging economies spanning the period 1990 to 2015, this paper finds that increased Internet access results in lower levels of air pollution. Moreover, panel causality test results highlight a unidirectional causality running from Internet use to CO₂ emissions. This result also has crucial policy implications for the governments in emerging markets. For instance, increased investment in the ICT sector could be a plausible channel to reduce air pollution level.

Appendix

Methodology for the MG estimator of Pesaran and Smith (1995)

MG estimator of Pesaran and Smith (1995) allows for all coefficients to differ across countries. It estimates separate regressions for each country and calculates the coefficients as unweighted means of the estimated coefficients for the individual countries. If we define a model in Eq. (1),

$$ \Delta {y}_{i,t}=\sum \limits_{j=1}^p{\alpha}_{ij}{y}_{i,t-j}+\sum \limits_{j=0}^q{\beta}_{ij}^{\prime }{X}_{i,t-j}^{\prime }+{\mu}_i+{\tau}_t+{\varepsilon}_{i,t} $$

(1)

where the number of countries i = 1, 2, …, N and the number of time periods t = 1, 2, . . …, T. y _{i, t} is the log of per capita CO₂ emissions, and X _{i, t} is a vector of explanatory variables that includes ICT and other determinants of CO₂ emissions. μ _i is the time-invariant fixed effect for country i, τ _t is the country-invariant time effect for time t, and ε _{i, t} is the idiosyncratic error term. If y _{i, t − 1} is included into each side of Eq. (1), we obtain Eq. (2):

$$ {y}_{it}=\sum \limits_{j=1}^p{\lambda}_{ij}{y}_{i,t-j}+\sum \limits_{j=0}^q{\beta}_{ij}^{\prime }{X}_{i,t-j}+{\mu}_i+{\tau}_t+{\varepsilon}_{i,t} $$

(2)

where ${\lambda}_{i,j}={\alpha}_{i,j}{\forall}_j\ne 1$ and λ _{i, 1} = α _{i, 1} + 1.

When we reparametrized Eq. (2) into the error-correction equation, Eq. (3) is obtained:

$$ \Delta {y}_{i,t}={\phi}_i\left[{y}_{i,t-1}-{\theta}_i^{\prime }{X}_{i,t}\right]+\sum \limits_{j=1}^{p-1}{\lambda}_{ij}\Delta {y}_{i,t-j}+\sum \limits_{j=0}^{q-1}{\beta}_{ij}^{\prime}\Delta {X}_{i,t-j}+{\mu}_i+{\tau}_t+{\varepsilon}_{i,t} $$

(3)

where

$$ {\phi}_i=-\left(1-\sum \limits_{j=1}^p{\delta}_i\right),{\theta}_i=\sum \limits_{j=0}^q\frac{\beta_{ij}}{1-\sum \limits_k{\lambda}_{ik}},{\lambda}_{ij}=-\sum \limits_{m=j+1}^p{\lambda}_{im},{\beta}_{ij}=-\sum \limits_{m=j+1}^q{\beta}_{im,} $$

Equation (3) could be estimated for each country separately, and then an average of the coefficients might be calculated. This yields the results of the MG estimator proposed by Pesaran and Smith (1995). For instance, in the case of the MG estimator, the error-correction speed of adjustment term is defined as ${\widehat{\phi}}^{MG}={N}^{-1}\sum \limits_{i=1}^N{\widehat{\phi}}_i$ along with the variance ${\widehat{\Delta}}_{\widehat{\phi} MG}=\frac{1}{N\Big(N-1}\sum \limits_{i=1}^N{\left({\widehat{\phi}}_i-\widehat{\phi}\right)}^2$.

Methodology for the Emirmahmutoglu and Kose (2011) panel causality test

The LA-VAR model (k _i  + dmax _i ) in heterogeneous mixed panels is specified as

$$ {x}_{i,t}={\mu}_i^x+\sum \limits_{j=1}^{k_i+d{\max}_i}{A}_{11,\mathrm{ij}}{x}_{i,t-j}+\sum \limits_{j=1}^{k_i+d{\max}_i}{A}_{12,\mathrm{ij}}{y}_{i,t-j}+{u}_{i,t}^x $$

(1)

and

$$ {y}_{i,t}={\mu}_i^y+\sum \limits_{j=1}^{k_i+d{\max}_i}{A}_{21,\mathrm{ij}}{x}_{i,t-j}+\sum \limits_{j=1}^{k_i+d{\max}_i}{A}_{22,\mathrm{ij}}{y}_{i,t-j}+{u}_{i,t}^y $$

(2)

where i = 1, 2, …, N and t = 1, 2, . . …, T denote the number of countries in the panel and the time dimension, respectively. ${\mu}_i^x$ and ${\mu}_i^y$ are two vectors of fixed effects, and ${u}_{i,t}^x$ and ${u}_{i,t}^y$ are column vectors of error terms. The lag structure (k _i ) of the process is assumed to be known and may differ across cross-sectional units. dmax _i denotes maximal order of integration suspected to occur in the system for each i. Searching for the causality from x to y is described with the following steps (see Emirmahmutoglu and Kose 2011):

• 1. Determine the maximum order (dmax _i ) of integration of variables in the system for each cross-sectional unit based on the ADF unit root test, and select the lag orders k _i s via some information criteria by estimating Eq. (2) by OLS for each individual.

• 2. By utilizing dmax _i and k _i from step 1, re-estimate Eq. (2) by OLS under the non-causality null hypothesis and obtain the residuals for each individual.

$$ {\widehat{u}}_{i,t}^y={y}_{i,\mathrm{t}}-{\widehat{\mu}}_i^y+\sum \limits_{j={k}_i+1}^{k_i+d{\max}_i}{\widehat{A}}_{21,\mathrm{ij}}{x}_{i,t-j}-\sum \limits_{j=1}^{k_i+d{\max}_i}{\widehat{A}}_{22,\mathrm{ij}}{y}_{i,t-j} $$

(3)

• 3. Then, the residuals are centered as

$$ \tilde{u}={\widehat{u}}_t-{\left(T-k-l-2\right)}^{-1}\sum \limits_{t=k+l+2}^T{\widehat{u}}_t $$

(4)

where ${\widehat{u}}_{t=}\left({\widehat{\mathrm{u}}}_{1t},{\widehat{u}}_{2t},\dots, {\widehat{u}}_{Nt}\right)$, k = max(k _i ), and l = max(dmax _i ). After that, ${\left[{\tilde{u}}_{i,t}\right]}_{NxT}$ is developed from the centered residuals. We choose randomly a full column with replacement from the matrix at a time to preserve the cross covariance structure of the errors. The bootstrap residuals is denoted as ${\tilde{u}}_t^{\ast }$.

• 4. Under the null hypothesis, a bootstrap sample of y is generated as in Eq. (5).

$$ {y}_{i,t}^{\ast }={\widehat{\mu}}_i^y+\sum \limits_{j=1}^{k_i+d{\max}_i}{\widehat{A}}_{21,\mathrm{ij}}{x}_{i,t-j}+\sum \limits_{j=1}^{k_i+d{\max}_i}{\widehat{A}}_{22, ij}{y}_{i,t-j}^{\ast }+{\tilde{u}}_{i,t}^{\ast } $$

(5)

where ${\widehat{\mu}}_i^y$, ${\widehat{A}}_{21,\mathrm{ij}}$, and ${\widehat{A}}_{22,\mathrm{ij}}$ are estimated from step 2.

• 5. Wald statistics are calculated to test the non-causality null hypothesis for each individual by substituting ${y}_{i,t}^{\ast }$ for y _{i, t} and estimating Eq. (2) without imposing any parameter restrictions.

Finally, the Fisher test statistic $\left(\lambda =-2\sum \limits_{i=1}^N\ln \left({p}_i\right),\kern0.5em i=1,2,..\dots, N\right)$ is derived by using the individual p values that correspond to the Wald statistic of the ith individual cross section. Additionally, the bootstrap empirical distributions of the Fisher test statistics are generated by repeating steps 3–5 10,000 times and specifying the bootstrap critical values by selecting the appropriate percentiles of these sampling distributions.