國
立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
The model to be used for the interpretation of results depends on the kind of spatial autocorrelation in the data as indicated by the Moran’s I and Lagrange Multiplier (LM) statistic: if the OLS model is spatially autocorrelated and shows a pattern of spatial dependence between the dependent variables, the spatial lag model is used. If the error terms show a pattern of spatial dependence between the error terms, the spatial error model is used.
IV. EMPIRICAL RESULTS AND DISCUSSION
The purpose of this study is to determine whether access to jobs has an impact on employment levels and to compare the degrees of access to jobs and employment levels between neighborhoods inside and outside Metro Manila. This chapter discusses the results derived from the regression analysis and hot spot analysis. As an initial step, descriptive statistics such as the mean, median, standard deviation, minimum values, and maximum values are presented in order to provide an overview of the data used in the study. Maps are also provided in the descriptive spatial statistics in order to visualize the spatial distribution of the data, and a hot spot analysis is conducted to determine whether the clusters in the spatial distribution map are statistically significant.
The main econometric procedures used to assess the impact of the number of accessible firms on employment levels are OLS regression and spatial regression. The differences between the models will be illustrated and compared in the chapter’s results section.
Finally, the chapter discusses the obtained results and how these results answer the research questions.
‧
while Table 3 presents the summary measures for employment-to-population ratio.According to the statistics, the average employment-to-population ratio in the study area neighborhood is 61.2%, and when examined according to region, the employment level is higher for villages inside Metro Manila (62%) than for those outside Metro Manila (60%). Table 4 also indicates that there is a large disparity in the average share of firms accessible from a given community when examined according to region, with the average share of accessible firms being drastically higher in Metro Manila than outside Metro Manila.
Table 2.
Descriptive statistics of all variables.
Variable Observations Mean Standard Deviation
Minimum Maximum
Employment level 3,336 61.23415 5.486976 31.01449 86.70757
% Accessible jobs 3,336 0.2308512 0.4757931 0.0002485 6.572666
% HS graduates 3,336 78.28966 11.70944 18.47134 99.13043
% Skills training 3,336 2.888209 1.802689 0 13.50575
% Married female 3,336 23.29799 4.469615 4.587156 90.99361
% Under 25 3,336 29.98598 3.855447 9.523809 69.65429
Economic zone 3,336 0.0227818 0.1492295 0 1
Classification 3,336 0.6669664 0.4713691 0 1
Protected area 3,336 0.0562237 0.2303878 0 1
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Table 3.
Summary measures of employment-to-population ratio per region Region Observations Mean Standard
Deviation
Minimu m
Maximum
Metro Manila 1,663 62.34112 5.535484 31.01449 86.70757 Outside Metro
Manila
1,673 60.13379 5.211827 44.63938 83.33334
Table 4.
Summary measures of % of accessible jobs per region Region Observations Mean Standard
Deviation
Minimum Maximum
Metro Manila 1,663 0.2976637 0.6096973 0.0002485 6.572666 Outside Metro
Manila
1,673 0.1644381 0.270516 0.0002485 3.94524
Figure 4 shows the economic zones in the study area, and the graph shows that Metro Manila has the highest number of economic zones, followed by Laguna, Cavite, and finally Rizal.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 4. Comparison of the number of economic zones in the study area.
4.2. Descriptive spatial statistics
As part of the descriptive statistics, maps of the study area were also generated to illustrate the spatial distribution for each of the model’s variables (Figures 5 to 12) and to gain an initial understanding of how the values of the attributes are spread out within the study area.
Based on the map of the employment-to-population ratio in Figure 5, high community employment levels tend to be concentrated in the Metro Manila area. The eastern parts of Laguna also appear to have high levels of employment, while most of Cavite and Rizal exhibit lower employment levels than the rest of the study area.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 5. Spatial distribution of employment-to-population ratio in the study area, classified according to quantiles.
Although the spatial distribution map appears to show some patterns of clustering, this map alone cannot determine whether these clusters are statistically significant. In order to see more detail on which areas have a higher-than-average employment level, a hot spot analysis was conducted and the Getis-Ord-Gi* statistic was also generated and mapped, with the resulting hot spot map given in Figure 6.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 6. Hot spot map of employment levels.
According to the hot spot map in Figure 6, a large part of Metro Manila is composed of high employment clusters, with some exceptions in the central city and the northern part of the metropolitan area bordering on Rizal. Although some parts outside Metro Manila, particularly the Laguna area, also has multiple statistically significant clusters of high employment, the mapped results reveal that overall, the regions outside Metro Manila have larger clusters of communities with low employment levels, with larger areas of lower-than-average employment levels concentrated in the Cavite and Rizal provinces.
The results of the hot spot analysis confirm the hypothesis that employment levels are not evenly distributed throughout the study area, with provinces experiencing lower employment levels than the rest of Metro Manila.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
On the other hand, the spatial distribution map of the percentage of accessible firms as illustrated in Figure 7 shows that the areas with the highest share of accessible firms are all found in Metro Manila and the areas immediately adjacent to the border such as the cities of Antipolo and Taytay in Rizal, and Dasmariñas and Imus in Cavite. There is a noticeable shift as communities move farther and farther outside Metro Manila:
according to the map, the shares of accessible firms decreases with the distance from these high-access centers.
Figure 7. Spatial distribution of accessible firms in the study area, classified according to quantiles.
Just like with the employment levels spatial distribution map, it is necessary to generate the Getis-Ord-Gi* statistic in order to determine which of the high- or low-access clusters are statistically significant. Figure 8 shows that again, majority of Metro Manila is
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
composed of high-cluster communities, compared to areas outside Metro Manila which are composed mostly of low-access clusters. The only exception to this are portions of Imus and Dasmariñas in Cavite, and Biñan and Santa Rosa City in Laguna, which are high-access clusters due to their proximity to the southern border of Metro Manila.
The results of the hot spot analysis confirms that communities with higher-than-average access to firms are mostly found in Metro Manila, and in contrast, majority of communities with low access to firms are outside Metro Manila, which means that the distribution of jobs and industries are not equitable. The presence of an uneven distribution of both jobs and employment levels supports the hypothesis that spatial mismatch exists in the Philippine setting, specifically in the context of Metro Manila and its adjacent provinces.
Figure 8. Hot spot analysis of access to firms.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 9. Spatial distribution of independent variables in the study area, classified according to quantiles.
Maps illustrating the spatial distribution of the control variables are shown in Figure 9.
The maps indicate that the highest share of high school graduates in the working age population are in Metro Manila and the western side of Laguna. Low shares of high school graduates are mostly found in Rizal, southern Cavite, and eastern Laguna.
Moreover, according to the spatial distribution map, there is also a high share of individuals with skills training in the northern part of Metro Manila and in Calamba and San Pedro, Laguna. Clusters of high percentage of working age population with skills training can also be found in Antipolo, Rizal and the southern cities of Metro Manila such as Paranaque and Muntinlupa.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
As for married females, a high share of married females appears to be concentrated in Laguna and Cavite, while the lowest concentrations can be found in Rizal province and northern Metro Manila. Finally, a large share of individuals under 25 can be found in Rizal province, while the lowest concentrations are located in the eastern side of Metro Manila, particularly Pasig and Makati.
Figure 10. Spatial distribution of economic zones in the study area.
In accordance with the graph as given in Figure 4, the map on Figure 10 confirms that a large number of economic zones are in Metro Manila. Cavite also has some economic zones that are spread out throughout the province, while Laguna’s economic zones are concentrated in Calamba, San Pedro, and Santa Rosa.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 11. Spatial distribution of rural and urban areas in the study area.
The spatial distribution map of rural and urban areas as given in Figure 11 shows that Metro Manila is fully composed of urban communities. Though most of the adjacent provinces are composed of rural communities, urban communities are also noticeably present in parts of the provinces that are in the immediate vicinity of Metro Manila’s borders: examples are Cainta, Antipolo, and Taytay in the province of Rizal; Bacoor, Dasmariñas, Imus, and General Trias in the province of Cavite; and San Pedro, Biñan, and Santa Rosa in the province of Laguna.
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Figure 12. Spatial distribution of protected areas in the study area.
Figure 12 illustrates the spatial distribution of protected areas in the study area.
According to the map, protected areas are mostly concentrated in the province of Laguna, particularly in Cavinti, Nagcarlan, Majayjay, and Liliw. When compared to the hot spot map of access to firms, it is noticeable that these same regions were part of the cold spots in that map. Thus, from these maps, it can be inferred that these regions have low access to firms because these areas are protected and are not allowed to be developed.
The maps of the variables show that places with high employment also tend to have higher accessibility to firms. The hot spot maps of employment levels and accessible firms show that both high employment and high access could be found in Metro Manila, suggesting a direct relationship between the two variables. The variable maps also show that employment levels also appear to be higher in places with a lower share of
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
individuals under 25, as well as in areas with economic zones and protected areas. A regression analysis was run in order to confirm these relationships.
4.3. Regression model results
The hot spot analysis in the previous section was complemented by a regression analysis on the study area in order to determine which attributes contribute to being a high- or low-employment region and, more specifically, whether access to firms does affect employment levels. OLS regression was used first in order to determine the global, or average relationship between the variables. Test statistics were generated to ensure that the model does not violate any of the assumptions of linear regression, and spatial regression was used to correct for spatial autocorrelation.
Overall, the results show that the presence of accessible jobs around a village increases the employment-to-population ratio, and this result is consistent in both the classical OLS and spatial regression models. Besides the number of accessible jobs, the characteristics of a village’s working age population and location are also significant contributors to employment levels. Nevertheless, the effects of the working age population attributes are smaller compared to the effects of the accessible jobs variable. Table 5 shows that the presence of accessible firms has a positive impact on a village’s employment-to-population ratio.
‧
Empirical results for alternative specifications - OLS and spatial regression models Dependent variable: OLS Spatial lag model Spatial error model Employment levels
β Robust SE β Robust SE β Robust SE
Constant 76.7107** 2.140798 40.1957** 1.75203 74.95** 1.32761
% Accessible firms 1.41375** 0.2040959 0.56150** 0.175277 1.30719** 0.283116
% of HS graduates 0.01618 0.0098725 -0.002366 0.0075998 -0.003465 0.0096151
% With skills training
-0.1217** 0.0499517 -0.029207 0.0423496 0.0367531 0.0455519
Married Female -0.2173** 0.0401808 -0.1474** 0.0207156 -0.16698** 0.0234115 Under 25 -0.4285** 0.0420077 -0.3181** 0.0225283 -0.35443** 0.0249354 Economic zone 3.56099** 0.7707971 2.94331** 0.531797 2.53048** 0.539318 Classification 1.40509** 0.2642325 0.67358** 0.203747 0.834475** 0.258411 Protected area 3.23235** 0.4637508 1.58474** 0.362545 2.20275** 0.699662
W_Employment 0.54783** 0.0188492
LAMBDA 0.570887** 0.0193405
Adj. R-squared 0.159415 0.361572 0.360048
Breusch-Pagan test (prob.)
217.9659 (0.0000) 181.5671 (0.0000) 220.2250 (0.0000)
Moran’s I (prob.) 30.5808 (0.0000) LM - lag (prob.) 978.0896 (0.0000) LM- error (prob.) 919.7485 (0.0000) Robust LM - lag
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
4.3.1. OLS model results
The OLS results show that a 1% increase in the number of accessible firms from a given village also increases its employment-to-population ratio by an average of 1.41%, holding other variables constant. On the other hand, working age population characteristics, specifically the share of those who are likely to be disadvantaged or drop out of the labor force are also significant: with every 1% increase in the share of individuals under 25, the average employment-to-population ratio decreases by about 0.42%. Meanwhile, a 1% increase in the percentage share of married females in the working age population decreases the average employment-to-population ratio by about 0.21%.
The dummy variables for location characteristics such as economic zone designation, classification, and protected areas are all positive and significant: being designated as an economic zone would increase the average employment-to-population ratio by about 3.56% holding all other variables constant, while being an urban area increases the average employment level by about 1.40%. Being designated as a protected area, on the other hand, is associated with an increase of 3.23% in average employment levels.
Finally, no statistically significant linear relationship between employment levels and share of working age population with high school diploma was detected. The same goes for employment levels and the percentage share of working age population with skills training.
4.3.2. Checking regression assumptions
Test statistics were generated to ensure that the OLS model does not violate the assumptions of regression analysis as stated in the previous chapter. The variance
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
and Moran’s I statistic were used to detect heteroskedasticity and spatial autocorrelation, respectively.
The VIF of the model’s variables are all less than 10 (Table 6), which indicates that multicollinearity is not a concern. The Breusch-Pagan test for heteroskedasticity also gave a value of 217.97 and had a p-value less than 0.001, so the assumption of homoscedasticity has not been met. Although the presence of heteroskedasticity in the data does not bias the estimates, it does make the estimates inefficient, making the standard errors larger and changing the range of the confidence interval. This is addressed in the model by using robust standard errors to draw more accurate conclusions from the hypothesis testing.
Table 6.
Variance inflation factor for the independent variables.
Variables VIF 1/VIF
% Accessible firms 1.17 0.854317
% of High school graduates 1.37 0.729023
% With skills training 1.01 0.991197
Married Female 1.47 0.682177
Under 25 1.26 0.792243
Economic zone 1.09 0.913427
Classification 1.58 0.633059
Protected area 1.19 0.843518
Mean VIF 1.27
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
Judging from the spatial distribution map as given in Figure 5 and the hot spot map in Figure 6, there appears to be a pattern of clustering of the dependent variable, with high employment levels being concentrated in Metro Manila. The existence of this pattern implies that there is spatial autocorrelation in the data, which means that high or low values tend to be clustered together in a certain area. The Moran’s I statistic was generated for the dependent variable in order to confirm the presence of spatial autocorrelation.
Figure 13. Moran’s I test statistic for the study area.
‧
According to the results of the Moran’s I test for the dependent variable (Figure 13), there is a positive spatial autocorrelation as indicated by the slope of the graph. A positive spatial autocorrelation indicates that there are clusters of high and low employment levels. Because of this, spatial regression must be used in order to correct the original OLS model since spatial autocorrelation could bias the model by underestimating or overestimating the coefficients of the variables. When accounting for spatial autocorrelation, spatial regression models can generate a variable which contains the degree of spatial interactions that are affecting the employment levels.
The decision-making process for spatial model selection as provided in Figure 3 states that the selection for the best spatial regression model is based on the Lagrange Multiplier diagnostics run after OLS regression. If the LM diagnostic - lag is significant, then the spatial lag model is run; otherwise, the spatial error model is run. If both are significant, the robust LM diagnostics are generated and whichever model’s robust LM value is significant would be the appropriate model (Anselin, 2005). Judging from the LM and robust LM values provided in Table 5, both the spatial lag and spatial error model are significant and both must be tried in order to select the best model for interpretation.
4.3.3. Comparison of OLS and spatial regression models
This section compares the results of the OLS and spatial regression models, and draws attention to the effects of running the spatial regression model on the values of the coefficients. Furthermore, the section discusses the steps and criteria for choosing which specification is best for modelling employment levels in the context of the study area and the overall objectives of the research.
The OLS model shows that accessible firms has a higher impact on employment levels compared to working age population characteristics. This result is consistent even with
‧
0.56% rise in the average employment levels. The working age population and location characteristics that were significant in the OLS model are still significant in the spatial regression models. However, these coefficients are noticeably lower than the OLS coefficients. On the other hand, the spatial lag variable W_Employment is positive and significant, which means that spatial dependence between each village’s employment levels exists and that a change in the dependent variable or any of the independent variables in one area would positively affect the other neighboring areas’ employment levels. More specifically, the spatial lag model’s results show that about 54% of the significant for the spatial error model, while accessible firms and location characteristics still have higher effects compared to those for population characteristics. The LAMBDA variable, which indicates the degree of spatial dependence between the neighboring areas’ error terms, is also significant, which means that there might be unobservable geographic characteristics that affect employment levels. Again, the changes in the coefficients could be attributed to the addition of the LAMBDA variable, which accounts for the effects of unobserved spatial characteristics that were originally attributed to the variables in OLS.Overall, both spatial models do not differ drastically from one another; both show almost similar significant variables. Both the lag variable and LAMBDA variable are also significant, which highlights the importance of spatial factors in determining employment
‧
國立 政 治 大 學
‧
N a
tio na
l C h engchi U ni ve rs it y
levels. Not only does the addition of spatial variables correct the estimation of regression coefficients, it also improves the model in terms of the fit diagnostics and R-squared.
Given that both LM-lag and LM error statistics are significant and both spatial regression models appear to improve the existing OLS model, the selection of the best model for interpretation will be determined through the statistics for model fit.
The log-likelihood, Akaike Information Criterion, and Schwarz Criterion are the most
The log-likelihood, Akaike Information Criterion, and Schwarz Criterion are the most