Economic Complexity and the Environment: Evidence from Brazil
13
3.2 Methodology
Our choice of estimation approach is driven by two main considerations: whether
the correlation between time-invariant region-specific effects and all of the explanatory variables is zero; and whether the error term is idiosyncratic. We assume strict
exogeneity, that is, all the explanatory variables to be uncorrelated with the error
term.
Intuitively, the assumption of zero correlation between time-invariant regionspecific effects and all the explanatory variables is unlikely to hold, as industrial
policies within the different Brazilian municipalities, states and metropolitan regions
are likely to be correlated with characteristics and features of these subregions. Omitting them could thus be an important source of bias. Unlike in the random effects
approach, which assumes that these time-constant specific effects are randomly distributed across the subregions, fixed effects and first differences estimation allows
for this correlation to be nonzero.
We compute first differences estimates and conduct a Breusch-Godfrey test for
autocorrelation in order to choose between the first differences and fixed effects
approach. For our analysis on deforestation, forest fires, maximal sulphur dioxide
concentrations, and average ultrafine particle concentrations, first differences estimates are preferred over fixed effects estimates. The Breusch-Godfrey results indicate
that the differenced error term is idiosyncratic. In other words, fixed effects estimation is non-stationary with these datasets. Consequently, results could be biased, as
the error terms have a unit root and are highly persistent in the time-series dimension.
For solid waste generation, and the other air pollution indicators, the error term is
idiosyncratic, allowing for the use of fixed effects. Constant region characteristics
are differenced out or time-demeaned, respectively.
To correct for heteroskedasticity and autocorrelation, which are data properties
found in virtually all indicators after conducting Breusch-Pagan and BreuschGodfrey tests, we use clustered standard errors at the municipal, state, and metropolitan region level.
4 Empirics
4.1 Solid Waste Generation
Table 1 reports the fixed effects estimates for our analysis on the relationship between
solid waste generation and economic complexity in Brazilian municipalities in the
state of São Paulo from 2003 to 2011. In our baseline regression, we follow Eq. (1)
in Sect. 3.1 and regress per capita solid waste generation on the ECI and its square
term. Since the two coefficients are not statistically different from zero, we exclude
13
3.2 Methodology
Our choice of estimation approach is driven by two main considerations: whether
the correlation between time-invariant region-specific effects and all of the explanatory variables is zero; and whether the error term is idiosyncratic. We assume strict
exogeneity, that is, all the explanatory variables to be uncorrelated with the error
term.
Intuitively, the assumption of zero correlation between time-invariant regionspecific effects and all the explanatory variables is unlikely to hold, as industrial
policies within the different Brazilian municipalities, states and metropolitan regions
are likely to be correlated with characteristics and features of these subregions. Omitting them could thus be an important source of bias. Unlike in the random effects
approach, which assumes that these time-constant specific effects are randomly distributed across the subregions, fixed effects and first differences estimation allows
for this correlation to be nonzero.
We compute first differences estimates and conduct a Breusch-Godfrey test for
autocorrelation in order to choose between the first differences and fixed effects
approach. For our analysis on deforestation, forest fires, maximal sulphur dioxide
concentrations, and average ultrafine particle concentrations, first differences estimates are preferred over fixed effects estimates. The Breusch-Godfrey results indicate
that the differenced error term is idiosyncratic. In other words, fixed effects estimation is non-stationary with these datasets. Consequently, results could be biased, as
the error terms have a unit root and are highly persistent in the time-series dimension.
For solid waste generation, and the other air pollution indicators, the error term is
idiosyncratic, allowing for the use of fixed effects. Constant region characteristics
are differenced out or time-demeaned, respectively.
To correct for heteroskedasticity and autocorrelation, which are data properties
found in virtually all indicators after conducting Breusch-Pagan and BreuschGodfrey tests, we use clustered standard errors at the municipal, state, and metropolitan region level.
4 Empirics
4.1 Solid Waste Generation
Table 1 reports the fixed effects estimates for our analysis on the relationship between
solid waste generation and economic complexity in Brazilian municipalities in the
state of São Paulo from 2003 to 2011. In our baseline regression, we follow Eq. (1)
in Sect. 3.1 and regress per capita solid waste generation on the ECI and its square
term. Since the two coefficients are not statistically different from zero, we exclude