3.1. Data and Model
To further investigate the effects of compact urban development on air pollution, multi-dimensional panel data models were employed. The panel data model [
41] is a quantitative analytical method that can be used when time-series and cross-section data are both available. Air pollution in cities is influenced by spatial characteristics (e.g., locational characteristics such as coastal or inland, geological characteristics such as mountains or plains,
etc.) and periodic characteristics (e.g., changes in climatic conditions) and, therefore, spatial and temporal variations need to be taken into consideration at the same time in a panel analysis.
The panel data model handles variables that are important to the model but that are not included as explanatory variables. Another advantage is that it can also regulate estimate errors that arise from time-series processes and regional unit data. The model helps overcome the limitations of insufficient sample size (i.e., data) which were a cross-section of 17 cities and 14 year time series in this study. Considering that the atmospheric dispersion can be dependent on the location of city (e.g., whether cities are located inland or close to the coast), the location of city is an important factor. However, it was not used as an independent variable as city location does not change over time. For these reasons, the panel data model is an ideal analytical method for this study, considering that it can account for an unobservable omitted variable that has a significant effect on interurban air pollutant concentration differences.
To regulate omitted variables, error terms are categorized as variables such as individual (regional)-variant but time-invariant (or time-constant) or time-variant but individual-invariant. It also includes remainder stochastic disturbance term that is both dependent on individual and time.
The estimation equation for the panel data model is given below [
42]:
Yit = α + Xitβ + εit
where εi,t = μi + λt + νi,tεi,t = μi + λt + νi,t, i (region) = 1, 2, ..., N, t (year) = 1, 2, ..., T
μi = unobservable individual effect
λt = unobservable time effect
νi,t = remainder stochastic disturbance term.
The model is divided into either a fixed effects model (FEM) or a random effects model (REM) depending on the form of the error term. In the FEM, it is assumed that each subject has its own specific characteristics due to inherent individual characteristic effects in the error term, thereby allowing differences to be intercepted between subjects. Fixed effects are due to the fact that, although the intercept may differ across subjects, each entity’s intercept does not vary over time—that is, it is time-invariant [
43]. The REM assumes that the individual characteristic effect changes stochastically, and that differences in subjects are not fixed in time and are independent between subjects. Individual differences vary over cross-sections (
i.e., subjects) as well as time [
44].
Air pollution levels, as dependent variables, were obtained by averaging observed measurements from each monitoring station in each of the 17 cities (nine in the case of PM10). There was little concern about using the averaged values because the changes in air pollution in relation to the distance from city centers was not large, as was outlined earlier (see
Table 1,
Figure 1,
Figure 2,
Figure 3,
Figure 4 and
Figure 5). The key explanatory variable among the independent variables was the one representing the compactness of urban development. Urban compactness, in general, was measured by the activity densities within cities. Importantly, this is not the average density of the city as a whole but the relative spatial concentration of the density distribution.
Prior to defining the concept of urban compactness in this study, it examined the proportion of built-up area and green area according to the distance from city center in the Seoul metropolitan area. Seoul City Hall was selected to represent the city center and 10 concentric circles with a 20 km radius were used, starting from a concentric circle with a 2 km radius from the City Hall (
Figure 6). The distribution rates of the built-up area and green area were calculated after each of their total areas was extracted. For example, the proportions of built-up area and green area within the entire 2 km radius concentric circle were estimated individually. This spatial analysis was carried out using ArcGIS 10.0 with a 1:25,000 scale land cover map, as published by the Korean Ministry of the Environment [
38]. The results showed that the further the distance from the city center, the smaller the built-up area ratio becomes, whereas the proportion of green area increases (
Figure 7). This means urban compactness relatively increases when the net density in the built-up area increases under the population-based control variable.
Figure 6.
Distribution of built-up area and green area according to the distance from Seoul’s city center.
Figure 6.
Distribution of built-up area and green area according to the distance from Seoul’s city center.
Figure 7.
Concentric circle of Seoul used to identify the built-up area and green area according to the distance from Seoul’s city center.
Figure 7.
Concentric circle of Seoul used to identify the built-up area and green area according to the distance from Seoul’s city center.
This concept can be explained with the following equation when the activity densities are approximated by population density. It can be seen that, even if the gross density of a city is the same, urban development has been carried out in a more compact way if the net density in the built-up areas is higher.
The higher the degree of city compactness, the greater the ratio of green area surrounding the built-up area. Accordingly, the ratio of green area for a total land area (i.e., the proportion of green areas) and the number of people within the built-up area (i.e., the net density) are employed as two complementary indicators that characterize urban compactness in this study. Total land area implies a separate, distinct administrative district. The type of land use in the built-up area is indicated according to plots used for building and factory construction according to the plot-based land-use classification system used in Korea. Meanwhile, the green areas indicate plots classified as forests, parks, and recreational areas.
As additional explanatory variables, population size as well as the presence of manufacturing industries and vehicle dependency are added. Population size is used to control for the absolute level of pollution emissions. The manufacturing dependency of a city is assessed as the net density of workers engaged in a manufacturing industry hiring five or more employees in the built-up area. Vehicle dependency is expressed as the interaction between vehicle ownership and availability of road infrastructure: Vehicle ownership is calculated as the number of registered motor vehicles per capita and road availability is represented as the proportion of the plots classified as road space and parking lot in the administrative district [
45]. The data for these variables were obtained from the Statistical Yearbook (1997–2010) [
43] and the Report of the Census on Establishments (1997–2010) [
46], which were published by government agencies in each of the different cities.
3.2. Estimation Results
The FEM is selected based on the Hausman specification test [
47], in which the estimated χ
2 value is highly significant. The FEM is further divided into a one-way and a two-way model. The two-way FEM assumes that both the individual effect and the time effect have a constant influence over all observation units. Individual effects are caused by certain unique, unobservable properties of the 17 cities, while time effects are associated with the unique properties of each time series from 1996–2009. The production of air pollutants may be caused by certain unique and unobservable traits of individual cities. At the same time, air pollution control technologies and policies can potentially influence air quality in a mid- to long-term timeframe, and pollution may improve or worsen accordingly. Therefore, the two-way FEM is employed in order to control both individual (regional) effects and time effects. Time-Series Cross-Section Regression in SAS software (ver. 9.2) was used for estimation.
The estimation results summarized in
Table 2,
Table 3,
Table 4,
Table 5 and
Table 6 indicate that urban compactness has both negative and positive effects on air quality; the former is about the spatial concentration of pollutants resulting from high densities in built-up areas, the latter is about the dispersion of pollutants attributed to green areas. On the one hand, NO
2 and CO increased as the net population density increased, implying that compact urban development can result in greater spatial concentration of pollutants. On the other hand, SO
2 and CO decreased as the proportion of green areas increased, implying that green areas secured by compact development can promote the dispersion of pollutants and thereby mitigate air pollution. Consequently, there is no clear impact of compact urban form on air quality. Air pollution resulting from compact urban development may vary according to pollutant-specific factors and emission source.
Table 2.
Panel data model estimates for urban characteristics and SO2.
Table 2.
Panel data model estimates for urban characteristics and SO2.
SO2 | Estimate | Std. Err | t-Value | Pr > |t| |
---|
Net density | 7.596 × 10−3 | 7.551 × 10−3 | 1.01 | 0.32 |
Proportion of green area | −0.150 | 0.069 | −2.17 ** | 0.03 |
Population | 4.930 × 10−6 | 0.000 | | |
Manufacturing | 0.498 | 0.431 | 1.15 | 0.25 |
Vehicle dependency | −31.530 | 24.800 | −1.27 | 0.20 |
Intercept | 6.316 | 1.760 | 3.58 | 0.00 |
N = 238, R2 = 0.828 |
Table 3.
Panel data model estimates for urban characteristics and NO2.
Table 3.
Panel data model estimates for urban characteristics and NO2.
NO2 | Estimate | Std. Err | t-value | Pr > |t| |
---|
Net density | 0.190 | 0.109 | 1.74 * | 0.08 |
Proportion of green area | 0.005 | 0.012 | 0.39 | 0.80 |
Population | 7.739 × 10−6 | 0.000 | | |
Manufacturing | 0.982 | 0.681 | 1.44 | 0.15 |
Vehicle dependency | −35.480 | 39.100 | −0.91 | 0.37 |
Intercept | 14.549 | 2.780 | 5.230 | 0.00 |
N = 238, R2 = 0.810 |
Table 4.
Panel data model estimates for urban characteristics and CO.
Table 4.
Panel data model estimates for urban characteristics and CO.
CO | Estimate | Std. Err | t-Value | Pr > |t| |
---|
Net density | 1.302 | 0.580 | 2.25 ** | 0.03 |
Proportion of green area | −13.720 | 5.310 | −2.58 ** | 0.01 |
Population | −2.790 × 10−5 | 0.000 | | |
Manufacturing | 7.137 | 33.100 | 0.22 | 0.83 |
Vehicle dependency | −3484.950 | 1904.200 | −1.83 | 0.17 |
Intercept | 79.970 | 13.540 | 5.91 | 0.00 |
N = 238, R2 = 0.767 |
Table 5.
Panel data model estimates for urban characteristics and O3.
Table 5.
Panel data model estimates for urban characteristics and O3.
O3 | Estimate | Std. Err | t-Value | Pr > |t| |
---|
Net density | −0.040 | 0.074 | −0.59 | 0.55 |
Proportion of green area | 9.10 × 10−3 | 8.13 × 10−3 | 1.12 | 0.26 |
Population | 4.01 × 10−6 | 0.000 | | |
Manufacturing | 0.092 | 0.464 | 0.20 | 0.84 |
Vehicle dependency | −26.110 | 26.700 | −0.98 | 0.33 |
Intercept | 27.700 | 1.900 | 14.59 | 0.00 |
N = 238, R2 = 0.606 |
Table 6.
Panel data model estimates for urban characteristics and PM10.
Table 6.
Panel data model estimates for urban characteristics and PM10.
PM10 | Estimate | Std. Err | t-Value | Pr > |t| |
---|
Net density | 0.228 | 0.725 | 0.31 | 0.75 |
Proportion of green area | −43.387 | 29.927 | −1.45 | 0.15 |
Population | 2.20 × 10−5 | 1.10E–05 | 2.05 ** | 0.04 |
Manufacturing | 0.226 | 0.256 | 0.88 | 0.38 |
Vehicle dependency | −727.525 | 431.900 | −1.68 | 0.11 |
Intercept | 75.638 | 22.945 | 3.30 | 0.00 |
N = 126, R2 = 0.698 |
Meanwhile, urban compactness did not have a significant effect on PM10 and O3. Since O3 is not usually emitted directly into the air and is rather created by the chemical reactions of primary pollutants or previously emitted gases, it may be difficult to identify emission sources related to urban spatial structure. PM10 emissions are mainly caused by combustion of traffic and manufacturing, but it may also be hard to evaluate the influence of PM10 non-exhaust emissions on air quality. PM10 significantly increases with a growing number of people, indicating that, as city size increases, PM10 concentrations increase correspondingly. Manufacturing and vehicle dependency had no significant relationship with air pollution levels, although NO2 and CO are particularly related to automobile exhaust gases.