Competencias docentes básicas

This study utilises the wealth of atmospheric composition satellite data for air quality applications which are primarily collected by National Aeronautics and Space Administration (NASA), 55 specifically the aerosol index (AI). The index is obtained using the total ozone mapping spectrometer (TOMS) instrument or the ozone monitoring instrument (OMI). The former satellite was used from 1978 to 2005, and the latter has been used since 2004.56 The TOMS / OMI aerosol index is a measure of how much the wavelength dependence of backscattered UV radiation from an atmosphere containing aerosols (Mie scattering, Rayleigh scattering, and absorption) differ from that of a pure molecular atmosphere (pure Rayleigh scattering). Quantitatively, the aerosol index (AI) is calculated from the ratio of measured to calculated 360 nm TOMS / OMI radiances. Under most conditions, the AI is positive for absorbing aerosols and negative for non- absorbing aerosols (pure scattering).

We first selected the years of interest and a predefined area (in this case, Indonesia). Then we customised the output plot by using the scale interval and minimum as well maximum range, colour, plot type, and grid lines inclusion. Next, we obtained the output of spatial map data of the aerosol index containing a set of coordinate data (points) with associated attribute tables. We combined the aerosol index layer with Indonesian spatial map data

55 _{http://disc.sci.gsfc.nasa.gov/aerosols/data-access}

56 _{NASA has a series of TOMS data from Nimbus-7 orbit and Earth Probe instrument. The TOMS Nimbus-}

7 was used from November 1978 to May 1993; TOMS EP was used from July 1997 to December 2005. Both of these satellites include 1.0x1.25 spatial resolution. The OMI has been used since October 2004 up to now, with 1 spatial resolution (http://disc.sci.gsfc.nasa.gov)

105 layer with polygon features. We then overlaid the district data with AI data and calculated the AI for each district using a polygon-based simple average method. To get the AI for a district, we made some polygons filling the area of the district and calculated the (Euclidean) distance of each polygon to the nearest point of AI values. Finally, we calculated a simple average AI of each district based on total AI values of all polygons in a district divided by the total polygons in the district.57 The procedure was conducted using ArcGIS software.58

3.4 Methods

In order to investigate the impact of tariff reform on air pollution, we estimate the following specification:

Δyd,t = α + βΔTariffd,t + ΔX'd,t γ + I'dθ + λr,t + Δεd,t (3.3)

where Yd,t is the air pollution measure in district d, year t. Tariffd,t is the district exposure

to tariff reforms. Xd,t is a set of the average time variant district characteristics

(temperature, precipitation, expenditure per capita, population and household access to electricity). We also include interactive island-year fixed effects, λr,t to control for shocks

over time that affects trade across all districts but may vary across different islands within Indonesia.59 The coefficient of interest, β, captures the average effect of trade reform on regional outcomes related to the air pollution index. We examine the aerosol index effects of tariff reforms, using both output and input tariffs. Then to check the robustness of our tariff measures, we combine output and input tariffs as shown by equations (3.4). Later, we also check whether our results are sensitive to the way we measure tariffs. We thus

57 _{See detailed procedure in appendix.}

58 _{The GIS software computing was done with the help of Diana Minita, a Master of Environment from}

Crawford School of Public Policy, Australian National University.

106 replace the tariffs in equations (3.3) and (3.4) with manufacturing weighted output and input tariffs.

Δyd,t = α + β1ΔOutputTariffd,t + β2ΔInputputTariffd,t + ΔX'd,t γ +

I'dθ + λr,t + Δεd,t (3.4)

The first difference specification controls for unobserved district-level heterogeneity and addresses potential bias of time-invariant unobservables. This method removes district fixed effects and eliminates the unobserved heterogeneity that might be instigated by the initial regional sectoral structure in employment and industry output. Furthermore, it eliminates potential bias due to endogenous national tariffs by controlling for country variation over time and by limiting variation only at the district level.

However, if any unobserved time variant confounders exist, the first difference approach can be biased. The potential confounders may include structural change, economic performance and any policies related to initial district sectoral structure and urban-rural differences. To deal with this problem, we incorporate a vector of initial conditions, Id,

which includes the 1990 sectoral labour shares (aggregated to one-digit sectors), 1990 rural population shares and in some specifications, the initial levels of the dependent variable.

Moreover, if the tariff measures are endogenous to air pollution measurement, or if they seize differential trends in air quality between districts, we would also expect air pollution to be correlated with future changes in district tariff exposure. Following Kis-Katos and Sparrow (2015), we conduct a placebo test by regressing changes in the independent

107 variable on future changes in tariff measures,60 with the null of no confounding patterns rejected if the future tariff coefficient is not statistically significant.

Rey and Janikas (2005) argue that spatial interaction might exist between regions which are traditionally ignored in the multiregional analysis. Similarly, Aklin (2014) argues that the effect of trade on pollution is interdependent since pollution is correlated across countries. Thus, the assumption of cross-sectional units independence is not relevant because the disturbance error may come from observed countries or region.

In the context of our study, the assumption of independence of each district in Indonesia might be argued since Indonesia is an archipelago country which has spatial differences in natural resources, population distribution, and many other economic and geographic variables.61 One could also argue that the air pollution index which is used in this study could not be measured accurately since pollution could flow from and to other regions. Thus the analysis results could be biased. Failing to take spatial dependence into account in the standard techniques will lead to inconsistent standard error estimates (Driscoll & Kraay, 1998).

We deal with this issue by extending model (3.3) to incorporate neighbourhood influence or spatial effects. We take into account the potential spatial spillovers by applying a spatial weight matrix, W. In this study, we used a general spatial weight matrix based on the geographic characteristics of the samples. In this method, following Viton (2010), we define two regions as neighbours if they share any part of a common border. We construct a simple binary contiguity matrix: the element of the spatial weight matrix (W) is one, if location i is adjacent to location j, and zero otherwise. The weight matrix is a binary

60 _{That is, we first regress y}

dton ΔTariffdt+1 and re-run with ΔTariffdt+2 for further checking.

61 _{Several studies on Indonesia apply spatial effects for the same reasons. For example: Magrini (2004);}

108 matrix with n by n dimension, where n is the number of districts in our sample. We then utilise the spatial models introduced by Anselin & Griffith (1988) and Anselin (2013) - the spatial autoregressive lag model (SAR) and the spatial Durbin model (SEM) -to examine the effect of other regions over a particular area.62 The SAR and SDM models are defined as specifications in equation (3.5) and equation (3.6), respectively.

Δ y_d,t = α + β · Δ Tariffd,t + ρWyd,t + Δ X'd,t γ + λd,t + Δ u (3.5)

Δ yd,t = α + β1 · Δ Tariffd,t + ρW ΔTariffd,t + Δ X'd,t γ1 +

W Δ X'd,t γ2 + λd,t + Δ u (3.6)

where yd,t the is the spatial lag of the dependent variable and W is the spatial weight matrix.

The SDM specification basically adds average-neighbour values of the independent variables to the specification. The spatial dependence weight matrix consists of 0 and 1, which belong to t h e neighbourhood: a value of 1 is given for any district n that is adjacent with district d and 0 otherwise. The spatial weight matrix has to be a square matrix [NxN] and symmetric.

3.5 Results and discussions