Ecological Economics 68 (2009) 2818–2824

Contents lists available at ScienceDirect

Ecological Economics

j ourna l homepage: www.e lsev ie r.com/ locate /eco lecon

ANALYSIS

Analysis of the major drivers of the ecological footprint using the STIRPAT model andthe PLS method—A case study in Henan Province, China

Junsong Jia a, Hongbing Deng a, Jing Duan a, Jingzhu Zhao a,b,⁎a State Key Laboratory of Urban and Regional Ecology, Research Center for Eco-Environmental Sciences, Chinese Academy of Sciences, Beijing 100085, Chinab Institute of Urban Environment, Chinese Academy of Sciences, Xiamen 361021, China

⁎ Corresponding author. State Key Laboratory ofResearch Center for Eco-Environmental Sciences, Chines100085, China. Tel.: +86 10 62849112; fax: +86 10 628

E-mail address: jzhao@rcees.ac.cn (J. Zhao).URL: http://www.rcees.ac.cn/ (J. Zhao).

0921-8009/$ – see front matter © 2009 Elsevier B.V. Adoi:10.1016/j.ecolecon.2009.05.012

a b s t r a c t

a r t i c l e i n f o

Article history:Received 26 November 2008Received in revised form 20 May 2009Accepted 20 May 2009Available online 18 June 2009

Keywords:EFDriversSTIRPATPLSHenan Province

Taking Henan Province of China as an example, we computed and analyzed the ecological footprint (EF) in1983–2006. The results showed that the EF in Henan Province quadrupled in the 23 years, but its ecologicalcarrying capacity (EC) was rather low and was in a state of slow decline, indicating that Henan's ecologicaldeficit (ED) had become a remarkable social problem. Therefore, the major drivers of the EF's change wereanalyzed. According to the simulations with STIRPAT model, the major drivers of Henan's EF were humanpopulation (P), GDP per capita (A1), quadratic term of GDP per capita (A2), percent of economy excluded inthe service sector (Ta1) and percent of urban population (Tb1). However, these drivers themselves had strongcollinearity, which might produce some uncertain impact to the final results. In order to avoid the impact ofcollinearity, the method of partial least squares (PLS) was used. The results showed that the major drivers ofEF were P, A1, A2 and Tb1. Ta1 was excluded. Compared with the results by the STIRPAT model, which showedthat P is the most dominant driver and the effect of the other drivers could almost be ignored, the results byPLS method were considered as more reasonable and acceptable because the impacts of the A (Affluence)and T (Technology) conditions to the regional EF were still too important to be ignored. In addition, theresults acquired by both methods showed that the curvilinear relationship between economic developmentand ecological impact (EF) or the classical EKC hypothesis didn't exist in Henan Province.

© 2009 Elsevier B.V. All rights reserved.

1. Introduction

Sustainable development has become an important goal around theworld since the Rio Earth Summit (Wackernagel et al., 2004; Chen et al.,2006). So far, a variety ofmodels, methods and indicators have beenputforward to evaluate sustainable development quantitatively at regionalscales. As one of approaches, the ecological footprint (EF) indicator,developed by Wackernagel and Rees (1996), has received considerableattention and been broadly used (Erb, 2004; Holmberg et al., 1999).Although the EF has some inherent drawbacks and thus are criticized bysome scholars (Ayres, 2000; Van den Bergh and Verbruggen, 1999), wecannot deny the fact that it is an effective evaluation tool to estimate theoverall ecological impact caused by human activities (Rosa et al., 2004;Xu et al., 2005; Dietz et al., 2007) and provide some reasonable clues tohelp achieve regional sustainable development.

The STIRPAT model, which was originated by Dietz and Rosa (1994,1997), was converted from the widely recognized IPAT that has beenused to analyze environmental impacts (Chertow, 2001; York et al.,

Urban and Regional Ecology,e Academy of Sciences, Beijing49510.

ll rights reserved.

2002). STIRPATwasfirst applied to analyze EF, CO2 emissionsandenergyfootprint by York, Rosa, and Dietz (2003a,b). The indicators above wereused to describe ecological impact of human activities and STIRPATmodel was used to analyze the driving forces of the impacts. Later on,Rosa et al. (2004) utilized EF to denote ecological impact and STIRPATmodel to analyze the impact's driving forces. Dietz et al. (2007) still usedEF and STIRPAT model to analyze anthropogenic ecological impact andits drivers. Taking the EFs of most provinces as an example, Xu andCheng (2005) conducted some similar studies in China. Meaningfulconclusions were found from these studies and some significantsuggestions were provided for local governments. Our study is differentin several ways from the previous studies (Table 1). Particularly, thecross-section datawere used at a global scale in the previous studies butthe time-series data in a certain region have hardly been used. Moreimportantly, STIRPAT model can essentially be transformed into arandom form of the Ordinary Least Squares (OLS) regression model,which allows for determining variables subjectively before selectingindependent variables (X) and has uncertainties that cannot totallyexclude the impact of collinearity among the independent variables.

The phenomenon of the existence of prefect correlation or a highdegree of correlation between the explanatory (independent) variablesin the linear regression model can be considered as collinearity (Wang,1999, 2006). When there are more than two covariates that are highlycorrelated, this is called multicollinearity (Wang, 1999, 2006). O'Brien

Table 1A comparison between the previous studies and this paper.

Items York et al. (2003a,b) Rosa et al. (2004) Xu et al. (2005) Dietz et al. (2007) This paper

Study scales World World China World A province in ChinaData features Cross-section Cross-section Cross-section Cross-section Time-seriesMethods ImPACT&STIRPAT STIRPAT STIRPAT STIRPAT STIRPAT & PLSEcological impact's indicators CO2 emissions and energy footprint, EF O3 and CO2 emissions and EF EF EF EF

Fig. 1. The location of Henan Province in China.

2819J. Jia et al. / Ecological Economics 68 (2009) 2818–2824

(2007)believes that the values of theVariance Inflation Factor (VIF, usedin measuring the degree of multicollinearity) of 10, 20, 40, or evenhigher donot, by themselves, discount the results of regression analyses,and suggests to eliminate one or more independent variables from theanalysis by using the ridge regression. From this point of view,collinearity (multicollinearity) may not be a serious problem. However,other scientists (Wang, 1999, 2006; Næs and Mevik, 2001; Tu et al.,2004; Shacham and Brauner, 1997) consider that collinearity is aproblem that should receive more attention. For example, if there ismulticollinearity in the regressionmodel, regression coefficients will nolonger have the meaning for general interpretation (Wang, 1999);further, collinearity can sometimes lead to serious instability of thevariables' coefficients (Næs and Mevik, 2001).

Partial least squares is a class of regression estimation methodsinitially developed by Wold (1966, 1973), which has recently beenincreasingly popular among scientists as a technique for treatinghighly collinear data. The method can exclude the impact of theindependent variables' collinearity problem successfully and inte-grates the multiple regression (MR), principal component analysis(PCA) and canonical correlation analysis (CCA) into a model at thesame time. It can choose accurately (not stochastic) dominant latent(integrated) component (t) from independent variables (X) anddetermine quantitatively the size of the contribution from X variablesto Y variables (dependent variables) (Wang, 1999, 2006) and, thus,appears to be more appropriate in analyzing EF's drivers.

Therefore, we tried to utilize this innovative PLS method inanalyzing EF's drivers. Firstly, we took Henan Province of China as astudy site and used Henan's historical (time-series) data to calculateits EF and EC and analyzed the condition of natural resourcesutilization; Secondly, we used STIRPAT model to analyze EF's majordrivers and identified the main influencing factors of changes inHenan's EF; Thirdly, we introduced the PLS method to further analyzethe major drivers of Henan's EF and further confirmed theseinfluencing factors. Lastly, a comparative analysis between the twomethods as well as between the two results was also made.

2. Study area

Henan Province, geographically ranging from 31°23′ to 36°22′N andfrom 110°21′ to 116°39′E, is located in central China (Fig. 1). Its total landarea is 1,670,000 km2, accounting for 1.73% of entire land area in China.Henan Province has the largest population in China and its regionalresourcesper capita arebecoming less as thepopulationkeeps increasingcontinuously. For example, the arable land area per capita is 0.087 ha in2003, which is much lower than the national average (0.112 ha, Xing etal., 2005). At the same time, its economic or technical development lagsbehind relatively developed regions in China. Henan's ecologicalconditions are getting worse (Li and Fan, 2005). In particular, in recentyears, a lot of forests are destroyed, which leads to more serious regionalsoil erosion and environmental degradation.

3. Methods and data

3.1. EF computation

The EF is the total area of productive land and water requiredcontinuously to produce all the resources consumed and to assimilate all

the wastes produced, by a defined population, wherever on Earth thatland is located (Wackernagel et al., 1997). Eq. (1) was used to calculateEF in Henan Province in this paper (Wackernagel et al., 1997, 1999):

EF = N EFið Þ = NX

rj aaið Þ = NX

rj Ci = Yið Þh i

ð1Þ

where, EF is the total regional EF (gha); N is the population; EFiis the per capita EF (gha/cap); i is the consumption item number(i=1, 2, …, n); rj is the equivalent factor of different types of land,obtained from the literatures (Wackernagel et al., 2004); aai is theecologically productive land converted from per capita consumptionof i item (gha/cap); Ci is the per capita consumption of i item (kg/cap) which is decided by the productivity plus the trade balanceamount (import minus export); Yi is the average productivity ofproducing i item in a certain biological productive area in a certainyear (kg/gha). The per capita EF is classified into six types,corresponding to arable land, pastureland, forestland, fossil energyland, built-up land and water land, respectively.

The EC reflects the ability of regional available land resources.Eq. (2) was used to calculate EC (Wackernagel et al., 1997, 1999):

EC = N ECið Þ = NX

aj × rj × yj� �

ð2Þ

where, EC is the total regional EC (gha); ECi is the per capita EC (gha/cap); aj is the per capita biological productive area of j type land in a

Fig. 3. Total Henan's EF, EC and ED in 1983–2006.

2820 J. Jia et al. / Ecological Economics 68 (2009) 2818–2824

region; and yj is the yield factor of different types of land, got from theliteratures (Wackernagel et al., 1997).

The data sources of consumption were mainly collected fromHenan Province Yearbooks in 1984–2007 and Henan Statistics in1949–1998 published by China Statistics Press. The data sources oftrade balancewere obtained fromChina Statistical Yearbooks in 1984–2007, New China's 55-Year Statistical Data in 1949–2004 and the FAOdatabase provided by the Food and Agriculture Organization of theUnited Nations.

3.2. Driver analysis methods

3.2.1. STIRPATThe STIRPAT is calculated with a non-linear regression formula and

the coefficients in it represent the hypotheses (Eq. (3)) which need tobe tested:

I = aPbAcT de ð3Þ

where, the constant a scales the model, b, c and d are the exponents ofP, A and T, respectively and e is the error item.

In order to estimate the parameters, Eq. (3) is converted to thelinear logarithmic form as follows:

ln Ið Þ = ln a + b ln Pð Þ + c ln Að Þ + d ln Tð Þ + ln e ð4Þ

3.2.2. PLSPartial least squares (PLS) is particularly useful when independent

variables have a strong collinearity (Wang, 1999, 2006).The method will extract the latent variables (integrated variables)

ti(ui) from X(Y) variables and make the ti(ui) represent originalinformation in X(Y) as much as possible. In this paper, the dependentvariable (Y) has the only one (EF), thus the analytical process has beengreatly simplified. Firstly, PLS calculation process requires to standar-dize the raw data X(Y) and then the first component t1 is extractedfrom standardized data (X). Secondly, the regression of X to t1 and theregression of Y to t1 are implemented in PLS. If the accuracy of result ofregression is satisfactory, the computation terminates; if unsatisfac-tory, the second component (t2) need to be extracted from residual

Fig. 2. Computation results of different types of EF in Henan Province in 1983–2006.

information of X and the second regression analysis is implemented.The regression is implemented repeatedly until the satisfactoryaccuracy is achieved and the latent variables t1, t2…… tm are obtained(Wang, 1999, 2006). In this study, we adopted directly the softwareSimca-P11.5, developed by the Umetrics company, to perform PLScalculations. The input data were the same as those in STIRPAT model.

There are two important tables or plots used to explain theapplicability of the PLS Method: the t1/t2 scatter plot (also called to beT2 oval plot) and t1/u1 scatter plot. In the T2 oval plot, t1 and t2 are theintegrated or latent variables extracted from the X variables. They canrepresent the key information of X variables as much as possible andhave strong capacity to explain the Y variables as far as possible(Wang, 1999, 2006). If the t1/t2 relationship of the sample data is allincluded in the oval (Fig. 6), these sample data are homogeneous andgood and can be accepted perfectly (Wang, 1999, 2006). In the t1/u1scatter plot, u1 is the integrated or latent variables extracted from theY variables. If the t1/u1 relationship of the sample data is near linear,the PLS linear regression model built is appropriate to the studyproblem (Wang, 1999, 2006).

Similarly, there are three important tables or plots used to explainor determine the fitting effect of PLSmodel: the Observed vs. PredictedPlot, Variable Importance Plot (VIP) and the Coefficients Table. TheObserved vs. Predicted Plot, which presented the relationship betweenthe observed variables and the predicted variables, can expressintuitively the fitting effect of the model. An almost linear relationshippresented in the Observed vs. Predicted Plot meant the explanation orfitting effect of the results made by the PLSmethod is excellent (Wang,1999, 2006). The VIP plot reveal quantitatively the statisticalimportance of independent variables (X) to all dependent variables(Y) in the model. The sum of squares of all VIP value is equal to thenumber of X items in themodel hence the average VIP should be equalto 1 (Wang, 1999). VIP values more than 1 indicate X variables are“important”, and values less than 0.5 indicate X variables are“unimportant”. The interval between 1 and 0.5 is a gray zone, wherethe importance level depends on the values of VIP (Wang, 2006). TheCoefficients Table showed the specific correlation coefficient betweenX and Y variables and the cumulative explanatory capacity from Xvariables to Y variables after the t1, t2 or t3 were extracted.

4. Results and discussion

4.1. Analysis of ecological conditions in Henan Province

Fig. 2 shows the computation results of different types of EF during1983–2006. It can be found that all types of EF had generally increasedin the whole period. Fig. 2a shows that although the EFs of built-upland, forest land and water land increased, they took small shares inthe total EF and thus could almost be ignored (0.03bb3 ha/cap, Fig. 2).However, it should be noted that the increase in EF for built-up landchanged significantly. It could explain the fact that the area used forconstruction such as housing and road, was increased rapidly in the

Fig. 4. Computation results of EC for different land types in Henan Province in 1983–2006.

Table 2Analyzing results of EF's drivers by STIRPAT model.

Independent variables Model 1 Model 2 Model 3 Model 4

P—population 2.368⁎⁎⁎ 3.675⁎⁎⁎ 3.253⁎⁎⁎ 3.583⁎⁎⁎A1—GDP per capita 0.210⁎⁎⁎ 0.142⁎⁎ 0.228⁎⁎⁎ 0.180⁎A2—quadratic term of GDP per capita 0.092⁎⁎⁎ 0.098⁎⁎⁎ 0.137⁎⁎⁎ 0.118⁎⁎Ta1—percent of GDP not in services sector 0.751⁎ 0.435Tb1—percent of urban population −0.333⁎ −0.172F 1615 1437 1428 1112Adjusted R2 0.995 0.996 0.996 0.996Constant (intercept) −42.91 −66.69 −58.16 −64.58Durbin–Watson 1.636 2.222 1.997 2.146Sample size 24 24 24 24

Note: ⁎Sig.b0.05; ⁎⁎ Sig.b0.01; ⁎⁎⁎ Sig.b0.001.

2821J. Jia et al. / Ecological Economics 68 (2009) 2818–2824

study period. Forest land EF and water land EF also increased but theirtotal growth was less than that of built-up land EF.

Similarly, Fig. 2b shows that fossil energy EF and pasture land EFexperienced sharp increases, which reflect people's consumption ofenergy and meat resources or the level of industrial technology andliving standard have been improved in Henan. Especially, per capitafossil energy EF has exceeded per capita arable land EF and pastureland EF since 2001, meaning that energy has become one of the mostimportant issues in daily life in recent years. In other words, EF inHenan Province has increasingly come from the energy-relatedproduction's consumption from then on.

As a whole (Fig. 3), the total EF in Henan Province increased from0.73 ha/cap in 1983 to 2.66 ha/cap in 2006, which grew nearly threetimes. At the same time, the total EC in Henan Province was 0.53 ha/cap in 1983, but it decreased to 0.43 ha/cap in 2006 (Fig. 3), whichmeans that the ability of regional natural resources to support humanactivities is limited and has been in a state of slow decline. Based onthe change of the total EF and EC, Ecological deficit (ED, per capita ECminus per capita EF) of the Henan Province was improved from0.20 ha/cap in 1983 to 2.23 ha/cap in 2006, indicating that theregional resources' supply capacity increases inadequately to supportregional human activities.

For the individual types of EC in Henan Province in 1983–2006, thedetails could be seen in Fig. 4. The change in forest EC (Fig. 4a) over

Fig. 5. Differences in EF, EC and ED between Henan and the average level in China in2003.

time is not obvious, but the EC of pastureland, built-up land (Fig. 4a)and the arable land (Fig. 4b) decrease obviously. Water EC (Fig. 4b)dropped down during the whole period, except a rapid increase from1995 to 1996 that could be resulted from the construction of regionalwater conservancy projects in 1995.

Fig. 5 shows that Henan's EF is higher than China's average (WorldWildlife Fund, 2006). Furthermore, Henan's EC is only 0.44 ha/cap,which is less than the national average level (0.80 ha/cap). Therefore,the ED of Henan Province (1.43 ha/cap) is much more than thenational average level (0.80 ha/cap), whichmeans that “human–land”relationship in Henan Province is tenser than China's average andHenan's anthropogenic ecological impact has become a moreimportant social problem in China.

4.2. Major drivers

The potential drivers included the total population (P), affluence(denoted by GDP per capita, A1), quadratic term of affluence (A2),technology (denoted by the percent of the economy not in servicesector or percent of the urban population, Ta1 or Tb1), quadratic termof technology (Ta2 or Tb2), fixed investment and percentage of alltypes of student, etc. The data-processing method adopted thecommonly used procedure based on works by York et al. (2003b),Rosa et al. (2004) and Xu and Cheng (2005). In these hypothesizeddrivers, Ta2, Tb2, fixed investment and percentage of all types ofstudent, etc. are ultimately excluded through computations due to thelow correlations with the dependent variable (EF).

The regression results of STIRPAT models are showed in Table 2.Overall, it can be seen that P had an elasticity of about 2.3–3.7, whichmeans that a 1% change in population may induce about 2.3–3.7percentages change in EF. Compared with the P drivers, the otherdrivers may have almost no effects on Henan's EF because theirelasticities were so low (bb2.3–3.7).

Model 1 is the basic equation including variables P, A1 and A2, andthey can explain almost all ecological impacts (Adjusted R2=0.995,Durbin–Watson=1.636≈2, Sig.b0.001). Compared with the Pdrivers, A1, A2 may have relatively small effects to Henan's EF becausetheir elasticities were low (0.092–0.210bb2.3). Then, the impacts

Table 3Collinearity correlation of independent variables.

P A1 A2 Ta1 Ta2 Tb1 Tb2

P 1.000 0.851⁎⁎⁎ 0.760⁎⁎⁎ −0.873⁎⁎⁎ −0.873⁎⁎⁎ 0.848⁎⁎⁎ 0.326A1 1.000 0.982⁎⁎⁎ −0.592⁎⁎ −0.592⁎⁎ 0.983⁎⁎⁎ 0.630⁎⁎⁎A2 1.000 −0.488⁎⁎ −0.488⁎⁎ 0.953⁎⁎⁎ 0.745⁎⁎⁎Ta1 1.000 1.000⁎⁎⁎ −0.645⁎⁎⁎ −0.088Ta2 1.000 −0.645⁎⁎⁎ −0.088Tb1 1.000 0.552⁎⁎Tb2 1.000

Notes: ⁎⁎ Sig.b0.01; ⁎⁎⁎ Sig.b0.001.

Fig. 6. T2 oval plot. Fig. 8. Observed vs. Predicted Plot.

2822 J. Jia et al. / Ecological Economics 68 (2009) 2818–2824

from Ta1 and Tb1 are added in Model 2 and Model 3, respectively.The explanation capacity of the two model are improved slightly(Adjusted R2=0.996N0.995), which means the two drivers haveabout the same net effects on EF. The effect from Ta1 is positive but theeffect from Tb1 is negative (−0.333b0). We also can get the similarresult from Model 4, where the cumulative effect of the two drivers(Ta1 and Tb1) are considered but all of their contributions to the EF'schange are not important (Sig.N0.05).

Moreover, there is no curvilinear relationship between economicdevelopment and ecological impact (EF) in Table 2, because thecoefficients of A2 are positive in Model 1 to Model 4, which indicatesthat the classical EKC hypothesis does not exist in the Henan's EF. Thisresult is consistent with the study results of Xu and Cheng (2005) andDietz et al. (2007).

In summary, the results of STIRPAT model show that the majordriving forces of EF are P, A1, A2, Ta1 and Tb1. However, there is a strongcollinearity among these independent variables, thus the accuracy andcredibility of the STIRPAT model are not very high according to Wang(1999). The testing methods of collinearity contain correlation matrixmethod and VIFmethod (Wang,1999). The correlationmatrixmethodis utilized in this paper and the correlation coefficients of independentvariables are showed in Table 3. It is obvious that the collinearityamong these independent variables is strong because there are manycorrelation coefficients that are near 1 (very high).

The PLS method was used in this paper in order to reduce thecollinearity impact of these variables themselves. First of all, theapplicability of PLS Method should be tested by checking the T2 oval

Fig. 7. t1/u1 scatter plot.

plot and t1/u1 scatter plot. The sample data in this paper are acceptablebecause there is only one inhomogeneous point in the T2 oval plot ofsample data (Fig. 6). It is obvious that the t1/u1 relationship of thesample data is nearly linear (Fig. 7), thus PLS linear regression modelbuilt is reasonable to the study problem of this paper.

The Observed vs. Predicted Plot showed a perfect linear relation-ship between the predictive value (YPred) and the actual value (YVar)(Fig. 8). This means that explanation effect of the results made by thePLS method is acceptable. The R2VY (cum) in the Coefficients Table(Table 4) denotes the cumulative fraction of the variation of the Yvariables explained after the selected component. R2VY (cum)reached 0.960 after the t1 component extracted, which means thatthe extracted components t1 from the X variables can explain 96.0%information of the Y variables. When t1 and t2 were extracted, theycan explain 98.1% information of the Y variables. When the threecomponents t1, t2 and t3 were all extracted from the X variables, theycan explain 99.5% information of the Y variables. These R2VY (cum)also showed a good explanation effect of the PLS method.

Moreover, Table 4 also shows the regression coefficients of the PLSmethod. It can be seen that the coefficient of P was 0.177 after the t1component extracted. When t1 and t2 were extracted, the coefficienthad changed into 0.169 and when t1, t2 and t3 were all extracted, thecoefficient had changed into 0.231 again. These showed that a 1%change in population (P) may induce about 0.169–0.231 percentageschange in EF or P had an elasticity of about 0.169–0.231. Similarly, A1,A2 and Tb1 had an elasticity of 0.195–0.301, 0.188–0.256 and 0.192–0.278. This suggests that P, A1, A2 and Tb1 may be the major drivers ofHenan's EF. However, the coefficients of Ta1, Ta2 and Tb2 had changedfrom negative (positive) to positive (negative) number and they wereall small (≤0.131), which means that their impact to EF are notimportant or they are not the major drivers of Henan's EF. About thispoint, the same results could be easily concluded from the VIP plot, inwhich the statistical importance of these independent variables (X)relative to the dependent variables Y (EF) is revealed in a graphic

Table 4Coefficients table.

Components extracted t1 t1 and t2 t1, t2 and t3

Constant 2.443 2.443 2.443P 0.177 0.169 0.231A1 0.195 0.240 0.301A2 0.188 0.235 0.256Ta1 −0.12 −0.04 0.01Ta2 −0.12 −0.04 0.01Tb1 0.192 0.226 0.278Tb2 0.114 0.131 −0.02R2VY(cum) 0.960 0.981 0.995

Fig. 9. The VIP plot.

2823J. Jia et al. / Ecological Economics 68 (2009) 2818–2824

representation according to their VIP values (Fig. 9). Because the VIP'svalues of A1, Tb1, A2 and P are greater than 1, they may be the majordrivers of EF (Y). But, because the VIP's values of Ta1, Ta2 and Tb2stayed in the interval of 0.5–1, their impact to EF(Y) cannot bedetermined or were not important (Wang, 1999, 2006).

As indicated by the results of STIRPATmodel, there is no curvilinearrelationship between economic development and ecological impact(EF) in Henan Province because the coefficients of A2 are positive allthe time (Table 4), which mean that the classical EKC hypothesis doesnot stand in Henan's EF.

5. Conclusions

(1) Henan's EF doubled nearly two times during the past 1983–2006 years. But at the same time, the supporting capability ofregional natural resources to human activities (EC) was ratherlow and in a state of slow decline. Thus, Henan's ecologialdeficit (ED) was bigger and bigger and its “man–land”relationship was tenser and tenser over time. In particular,Henan's EF was more than the national average level in recentyears such as 2003, which means that ecological condition ofHenan Province has become a remarkable social problem. Thelocal governments need to pay much more attention to it.

(2) Compared with the STIRPAT model, the analysis results of thePLS method are more reasonable and acceptable because thecollinearity among these drivers themselves cannot be avoidedin the STIRPAT model, which can bring some negative impactsto the ultimate results, but the collinearity can hardly impactthe PLS method. In addition, the analysis results of EF's driversby STIRPAT model showed that the major driving forces ofHenan's EF were P, A1, A2, Ta1 and Tb1. The most importantdriving forces is P (its elasticity is extremely high (over 2)) andthe effect of the other drivers (their elasticities are low) almostcan be ignored. The analysis results of the STIRPATmodel are soextreme that it is inconsistent with the fact. In reality, theimpacts of the A (Affluence) and T (Technology) condition tothe regional EF were still large and important (cannot beignored). However, the analysis results by PLS method showedthat the major drivers were P, A1, A2 and Tb1, and Ta1 wasexcluded as the unimportant drivers. Moreover, the P was notamong the extremely important driving forces and the impactof P to EF was almost the same as the other drivers. Thus, theresults using the PLS method are more reasonable andacceptable.

(3) The results acquired by using the both methods showed thatthe curvilinear relationship between economic developmentand ecological impact (EF) or the classical EKC hypothesis didnot stand in Henan Province.

Acknowledgements

This work was financially supported by 100 Talents Programme ofChinese Academy of Sciences: Comprehensive study on sustainableurban construction. We would like to appreciate Xiaobo Xiong, RencaiDong, Bofu Zheng and Feng Liu for their kind help in data collectionand valuable comments. We are grateful to Professor Guofan Shaofrom Purdue University and Dr. Helen Lee from Taylor and Francispublishing company in English revision for their help in Englishlanguage embellishment. We also express our thanks to twoanonymous reviewers and the editors for their helpful commentsand constructive suggestions.

Appendix A. Supplementary data

Supplementary data associated with this article can be found, inthe online version, at doi:10.1016/j.ecolecon.2009.05.012.

References

Ayres, R.U., 2000. Commentary on the utility of the ecological footprint concept.Ecological Economics 3 (3), 347–349.

Chen, M., Wang, R.S., Zhang, L.J., Xu, C.G., 2006. Temporal and spatial assessment ofnatural resource use in China using ecological footprint method. InternationalJournal of Sustainable Development and World Ecology 13 (4), 255–268.

Chertow, M., 2001. The IPATequation and its variants: changing views of technology andenvironmental impact. Journal of Industrial Ecology 4 (4), 13–29.

Dietz, T., Rosa, E.A., 1994. Rethinking the environmental impacts of population,affluence and technology. Human Ecology Review 1 (1), 277–300.

Dietz, T., Rosa, E.A., 1997. Effects of population and affluence on CO2 emissions.Proceedings of the National Academy of Sciences of the USA 94 (1), 175–179.

Dietz, T., Rosa, E.A., York, R., 2007. Driving the human ecological footprint. Frontiers inEcology and the Environment 5 (1), 13–18.

Erb, K.-H., 2004. Actual land demand of Austria 1926–2000:a variation on ecologicalfootprint assessments. Land Use Policy 21 (3), 247–259.

Holmberg, J., Lundqvist, U., Robèrt, K.-H., Wackernagel, M., 1999. The ecologicalfootprint from a systems perspective of sustainability. International Journal ofSustainable Development and World Ecology 6 (1), 17–33.

Li, J., Fan, Y., 2005. Problems and countermeasures of farmland quality in HenanProvince. Areal Research and Development 24 (2), 88–91 (in Chinese).

Næs, T., Mevik, B.-H., 2001. Understanding the collinearity problem in regression anddiscriminant analysis. Journal of Chemometrics 15 (4), 413–426.

O'Brien, R.M., 2007. A caution regarding rules of thumb for variance inflation factors.Quality& Quantity 41 (5), 673–690.

Rosa, E.A., York, R., Dietz, T., 2004. Tracking the anthropogenic drivers of ecologicalimpacts. Ambio 33 (8), 509–512.

Shacham, M., Brauner, N., 1997. Minimizing the effects of collinearity in polynomialregression. Industrial & Engineering Chemistry Research 36 (10), 4405–4412.

Tu, Y.-K., Clerehugh, V., Gilthorpe, M.S., 2004. Collinearity in linear regression is aserious problem in oral health research. European journal of oral sciences 112 (5),389–397.

Van den Bergh, J., Verbruggen, H., 1999. Spatial sustainability, trade and indicators: anevaluation of the ‘ecological footprint’. Ecological Economics 29 (1), 61–72.

Wackernagel, M., Rees, W.E., 1996. Our Ecological Footprint: Reducing Human Impacton the Earth. New Society Publishers, Gabriola Island, Philadelphia, USA. 66–150 pp.

Wackernagel, M., Monfreda, C., Schulz, N.B., Erb, K.-H., Haberl, H., Krausmann, F., 2004.Calculating national and global ecological footprint time series: resolvingconceptual challenges. Land Use Policy 21 (3), 271–278.

Wackernagel, M., Onisto, L., Bello, P., Linares, A.C., 1999. National natural capitalaccounting with the ecological footprint concept. Ecological Economics 29 (3),375–390.

Wackernagel, M., Onisto, L., Linares, A.C., 1997. Ecological footprint of nations: howmuch nature do they use? How much nature do they have? Commissioned by theEarth Council for the Rio+5 Forum. Toronto: International Council for LocalEnvironmental Initiatives, Canada. 10–21 pp.

Wang, H., 1999. Partial Least Squares Regression: Method and Applications. NationalDefence Industry Press, Beijing, China. 200–207 pp. (in Chinese).

Wang, H., 2006. Linear and Nonlinear Methods of Partial Least Squares Regression.National Defence Industry Press, Beijing, China. 279–308 pp. (in Chinese).

Wold, H.,1966. Nonlinear estimation by iterative least squares procedures. In: David, F.N.(Ed.), Research Papers in Statistics. Festschrift for J. Neyman. Wiley, New York. 411–444 pp.

Wold, H., 1973. Nonlinear iterative partial least squares (NIPALS) modelling: somecurrent developments. In: Krishnaiah, P.R. (Ed.), Multivariate Analysis. AcademicPress, New York. 383–407 pp.

WorldWildlife Fund. 2006. Living planet report 2006. http://www.panda.org/downloads/general/lpr.

Xing, X., Wu, K., Lu, Q., Zhang, J., 2005. Studies on changes and the driving forces ofcultivated land in Henan Province. Journal of Henan Agricultural Sciences 33 (12),56–61 (in Chinese).

2824 J. Jia et al. / Ecological Economics 68 (2009) 2818–2824

Xu, Z., Cheng, G., 2005. Impacts of population and affluence on environment in China.Journal of Glaciology and Geocryology 27 (5), 767–773 (in Chinese).

Xu, Z., Cheng, G., Qiu, G., 2005. ImPACTS identity of sustainability assessment. ActaGeographica Sinica 60 (2), 198–208 (in Chinese).

York, R., Rosa, E.A., Dietz, T., 2002. Bridging environmental science with environmentalpolicy: plasticity of population, affluence and technology. Social Science Quarterly83 (1), 18–34.

York, R., Rosa, E.A., Dietz, T., 2003a. Footprints on the earth: the environmentalconsequences of modernity. American Sociological Review 68 (2), 279–300.

York, R., Rosa, E.A., Dietz, T., 2003b. STIRPAT, IPAT and ImPACT: analytic tools for un-packing the driving forces of environmental impacts. Ecological Economics 46 (3),351–365.