The majority of these studies have utilized data compiled after World War II, although a study by Panayotou, Peterson and sachs used data from 1870 to1994. Nearly all studies involved a panel of countries.We utilized CO2 data provided by the World Bank originating from 15 Latin American countries over a 21 year period . For those countries for which CO2 data were not available from the World Bank, data from the Oak Ridge National Laboratory data was used instead. In turn, the CO2 emissions were calculated by measuring the total fossil fuel consumed. Per capita income was measured in dollars, which were obtained from the World Bank economic indicators for Latin America. Per capita income data were adjusted by purchasing power parity in order to construct comparable values across countries. Population density was measured by the number of people per square kilometer . The illiteracy rate was calculated as a percentage of the population aged 15 years and who were not able to read. Forestry data were collected from different sources. Those sources are listed in the Appendix A. The availability of forestry data are in deed the asset of this study as this set of data are not presently readily available. Weight variables, such as income weight and CO2 weight, are calculated using a queen contiguity matrix. In essence, the weight variables represent the average of income or CO2 emissions in the neighboring countries. We discussed weight variables in detail when presenting results from sensitivity analysis tests.Descriptive statistics of the data used in this paper are provided in Table 1.
Using quadratic specification in model ,stackable planters we conducted an F-test for the joint significance of fixed effects model and rejected the hypothesis that all the coefficients are zero at a significance level of 1%. The calculated F-statistic was 530.98, while the critical value was 2.16.The absence of the time effects given the individual effects was also tested. In effect, the study was unable to reject the null hypothesis at a significance level of 5%, as the statistic was -0.84, while the critical value was 1.1310 . Since several authors have used cubic models to estimate EKC, a cubic model was then tested against the quadratic model. To do this, an F test was then run to determine whether the quadratic model should be rejected in favor of the cubic model11 .The calculated F-value was 1.10. The critical F-value at 5% level was 3.90. This indicated that the quadratic model could not be rejected. We also conducted a Hausman specification test for the systematic difference between fixed and random effects. The m-statistic for Hausman test was 2.02, and the critical value at 0.05 significance level was 0.72, which means fixed effects model was more appropriate. In the panel data setting in developing countries, the fixed effect has proven effective by other studies as well . The following section presents the specification test in an attempt to compare the fixed effect quadratic model against the fixed effect semiparametric model 12. The results from these tests are listed in table 2. Our preliminary visual observation of data points reveals some sort of inverted ìuî curve for a limited number of countries, . For example, Brazil,Colombia and Peru seem to indicate an increasing tendency to pollute as income rises. However,Argentina, Ecuador, Guatemala, and Bolivia seem to reveal some kind of concavity in their income pollution curve.
It should be obvious that even if a country shows a rising or increasing pollution level to coincide with income, the country may still not contradict the inverted U hypothesis, since it may simply be on the rising part of the curve . Next the focus turns to revealing the importance of including forestry as a covariate, rather than including only income as a covariate and running a non-parametric model to estimate the EKC. The study is similar to the method used by Duncan and Blundell to justify the use of a semiparametric model in their study of estimation of Engel Curve. The countries in the sample were divided into three different groups. The first group contained countries with significantly low forestry to population ratios . Countries in this group have less than one hectare of forest per thousand people. The second group included countries such as Argentina, Chile,Ecuador, Colombia, Honduras and Nicaragua which have an intermediate level of forestry to population ratios. The countries in this second group possess more than 1 hectare but less than 2 hectares of forest per thousand people. The third group included countries with the highest forestry to population ratios. The pollution elasticity of income curve for these three groups are markedly different as presented in figure 2. This difference justifies the use of country specific heterogeneity in the forestry to population ratio. . Following Duncan and Blundell’s suggestion, the forestry variable was entered linearly into semiparametric specification. The curve for EKC from the semiparametric specification is given in figure 3. This indicates the presence of ìNî shaped curve. The N-shaped curve indicates that CO2 initially increases with an increase in income, then decreases, and eventually increases yet again. The ìturning pointî was about $3500, but the per capita consumption of carbon dioxide rises again at about $4500.
Comparing these values with the turning point estimate of $7954 from the parametric test, it is revealed that these two estimation techniques provide very different predictions . In the semiparametric setting, to test the question of whether certain countries are driving the result, several sensitivity analyses were conducted. Since visual inspection had earlier pointed out Brazil, Colombia and Peru served as major culprits for pollution emission, results were examined when these particular countries were removed from the equation. Figure 4 provides the three graphs from a semiparametric estimation in which each of the three countries is removed, this brings the upper turning point down to $4800, even though the lower turnin gpoint remains unchanged. However, the curvature of the estimated relationship remains essentially the same. The EKC for three different groups was also estimated . Figure 5 reveals the estimated curves. Countries in Group 1 remains on the rising part of the curve; the relationship between income and emission is strictly positive. These countries reveal a similar trend in their income ranges to that of the aggregate data, and interestingly, countries in this group are also primarily poor countries. Countries in Group 2 reveal evidence of an N-shape, and seem likely to reach some level of turning point at about $5000. Countries in group C ,however, behave relatively distinct from other groups. This group’s CO2 emission decreases initially, then increases and then eventually decreases again, with turning point occuring at about $3500. Also, the curvature of overall EKC doesn’t significantly change when three countries are removed .The fact that forest per person is a significant variable in all these estimations means that it is indicating some thing important that income alone. It is likely that variables originating from points other than income drive emissions, and that including income alone will systematically omit the other factors. Li and Hsiao’s serial correlation test of semiparametric model as described in following section indicates that there is some serial correlatedness in this study’s model, which also points towards the omission of some variables. Several authors have argued in favor of adding more variables in the CO2-Income EKC regression . Accordingly, we tested the importance of adding in such variables as population density, the illiteracy rate, and the weighted income variable into the regression models,stacking pots and observe whether these inclusions would affect the results. The justification for including population density in the model is that more dense population will burn more fuel, ceteris paribus. Higher illiteracy levels may mean that the population will resort to inefficient means of energy consumptions, such as burning firewood and so on. The spillover effect of income was also considered. If adjacent countries are wealthy this may also result in increases in their neighboring countries’ pollution levels. Following Paudel, Zapata and Susanto, a weighted income variable was constructed as a representation of the spillover effect of pollution. To account for the spillover effect in the model, the queen contiguity matrix was first calculated. This matrix regards neighboring countries of a country as being in either a vertex or a lateral contiguity of the country. Then average income is then obtained by adding the per capita income of the adjacent countries in that particular year and by dividing it by the number of contiguous countries. The average income, thus obtained, was used as a weighted income variable that measures the spillover effect in the model.
If a spillover effect is present,the coefficient associated with this variable would be positive and significant. Parametric estimation of this full model shows that as population density is introduced, both forestry and population density become insignificant, but the spillover effect remains significant, thus all having expected signs. On the other hand, semiparametric estimation indicates that both population density and illiteracy rates are insignificant, with the spillover effect revealed as minimally significant. The overall curvature of semiparametric EKC remains the same. These results are presented in Table 3. Somewhat surprisingly, the sign of illiteracy in semiparametric model, although insignificant, is positive. Empirical tests of convergence hypothesis have found absolute convergence in productivity only for developed countries . These studies were based on two common assumptions: developing countries are not fundamentally different from industrialized countries and free, world-wide availability of technological knowledge. However, conditional convergence was found in some cases where samples consisted of both developed and developing countries There is a general belief that productivity grows less rapidly in agriculture than in manufacturing sectors . Economists such as Wichman have found that transfer of improved agricultural techniques from the developed countries to developing countries is a lengthy process. It is this notion of slow productivity growth in agriculture that has resulted in developing several theories and policies of economic development that favors the manufacturing sector. For example, Wichmann analyzed technology adoption in agriculture and convergence across economies and found that there exists an optimal technological gap between developed and developing countries, indicating full convergence never takes place between industrialized and developing countries. However, a study performed by Martin and Mitra on productivity growth and convergence in agriculture and manufacturing sectors favored the agriculture sector. The authors found that at all levels of development, technical progress was faster in the agriculture sector than in the manufacturing sector. Moreover, they found strong evidence of a rapid convergence in levels and growth rates of total factor productivity in agriculture, indicating relatively rapid transfer of technological innovations from one country to another. Others who have looked at agricultural productivity convergence issues are Rezitis; Mukherjee and Kuroda and Ball, Hallahan and Nehring. We offer continuity to the existing literature, but interject two new methods to test the convergences a statistical method which improves on the methods used by others, and a redefinition of region in beta convergence. The redefinition of region is not a mere technological quibble in our view. The theory of beta convergence indicates some sort of convergence among a subset of units , but it is not clear what those subsets are. Previous authors have used exogenous criteria such as weather or altitude for region identification. However, it is well known in social sciences that categorization is possible in a large number of ways . Weather, politics, nature of inhabitants etc are just a few of those criteria. It is unlikely that all those criteria will lead to formation of region which can be tested for convergence. By utilizing data based clustering to identify the regions, we believe we provide one answer in that direction. We use a v-fold cross-validation algorithm, normally used in pattern recognition literature for the purpose.In the last decade, two different strands of literature have been prominent in identifying the convergence of total factor productivity or convergence in general. One, started by Kumar and Russell, looks at the distribution of productivity by decomposing it into different components and identifying the components that are converging or diverging to something. The other strand, due to Philips and Sul, provides a new method to test data-based clusters that are converging. One must be careful in applying Philips and Sul’s method in convergence literature since it is related to sigma convergence.