# Agricultural Value Added and Economic Growth in the European

## Comments

## Transcription

Agricultural Value Added and Economic Growth in the European

NEW MEDIT N. 4/2013 Agricultural Value Added and Economic Growth in the European Union Accession Process RAHMIYE FIGEN CEYLAN*, BURHAN ÖZKAN* Jel Classification: E25, F43, O47, Q18 As a result, the factors underlying economic deDevelopment has been velopment and the effects an important concept since of development on welthe definition of ‘econofare have been broadly inmy’ was provided for the vestigated, specifically sfirst time. Yet, methodince the 1990s. In addiological evaluation of detion to growth models velopment and its relationstemming from producship with growth has action function and capital quired greater emphasis sstock, index studies measince the World War II. In uring total and marginal addition, four periods of productivity and growth development have been accounting models were observed which display brought into use in the insignificant features (Dutt ternational literature (OyKeywords: European Union, integration, Solow growth, agricultural value added, risk. and Ros 2008). eranti 2000, Knowles and – 1945-1950: Low savings, McCombie 2002). Résumé high population growth, L’objectif de cet article est d’analyser l’impact du revenu agricole dans le cadre du Factor productivity sless willingness to invest, processus d’intégration dans l’Union européenne, en utilisant une extension du mo- tudies focusing on the intensive production in a- dèle de Solow et des outils d’analyse des données de panel. Les effets de la valeur Cobb-Douglas production agricole par habitant sur le revenu par habitant ont été examinés en s’apgriculture and limited in- ajoutée function developed by puyant sur deux échantillons de 25 et 30 États membres et Pays candidats, respecdustry orientation. tivement pour les périodes 1995-2007 et 2002-2007. Une variable nominale repré- Solow (1957) and K– 1950 - Late 1960s: mass sentant le statut de membre de l’UE et une variable de risque composite, calculée endrick (1961), which protectionism and import par PRS Group, fournissant des informations sur les propriétés structurelles des takes employment as a substitution, as no focus économies des pays concernés, ont été utilisées comme variables indépendantes. reference production facSelon les résultats du modèle à effets aléatoires bidirectionnel, l’élasticité de la vaon export was needed. leur ajoutée agricole du revenu par habitant est égale à 0,025 pour la période 1995- tor, were widely observed – Late 1960s-1980s: Re- 2007 et à 0,22 pour la période 2002-2007. Il a été estimé que le revenu moyen par in the literature. These sbirth of the neo-classical habitant est 5,6% plus élevé dans les pays membres de l’UE. Avec un changement tudies were also referapproach. Extended focus représentant 1 point de hausse de la valeur du risque composite, ce qui signifie une enced as growth accountréduction du risque auquel le pays fait face, il a été démontré que le revenu par haon integration of coun- bitant a augmenté de 1% dans les deux périodes. Les résultats indiquent également ing studies, and included tries, emphasis on the que l’agriculture garde son importance économique et que le revenu moyen par ha- output growth against inbitant dans les pays de l’UE est plus élevé par rapport aux pays non-membres en put growth. Generally, overdependence school. growth refers to technical – 1980 - Present: Measure- raison de certains facteurs exogènes. ment of economic devel- Mots-clés: Union européenne, intégration, modèle de Solow, valeur ajoutée, risque. change when it depends on all production factors, opment by mathematical while it is called partial economics. Interpretation of the effects of economic development on distribu- factor productivity when it depends on the efficiency of a stion, income and preferences started with integration of ingle factor. The main convergence studies focussed on the welfare effect of increased commercial activities (Balassa micro-theory with macro-theory. 1971, Chenery 1962, Chenery and Eckstein 1970, Wilhelm 2008) and on changes in total and partial factor productivity levels (Balassa and Bertrand 1970, Hayami and Ruttan *Akdeniz University, Faculty of Agriculture Department of Agricul1970, Komlos 1988, Dowrick and Nyguen 1989). tural Economics, Antalya - TURKEY. 1. Introduction Abstract The aim of this paper is to analyse the impact of agricultural income within the framework of the European Union integration process using an extended Solow growth model and panel data analysis tools. The effects of per capita agricultural value added on per capita income were examined using two samples of 25 and 30 EU member and candidate states respectively for the periods 1995-2007 and 20022007. A dummy variable representing EU membership status and a composite risk variable computed by the PRS Group, which provided information on structural properties of the economies of relevant countries, were used as independent variables. According to the two-way random effects estimation results, the agricultural value added elasticity of per capita income was 0.025 for the 1995-2007 period, and 0.22 for the 2002-2007 period. It was estimated that average per capita income is 5.6 % higher among EU members. With a change representing a 1 point rise in composite risk, which means a reduction of the risk faced by the country concerned, it was demonstrated that per capita income rose 1% during both periods. The results also showed that agriculture retains its economic importance, and that average per capita income among EU members is higher than among nonmembers due to exogenous factors. 62 NEW MEDIT N. 4/2013 However, the convergence of countries has been tested based on the utilization of production functions. By way of example, Pittau and Zelli (2006) measured regional convergence within the EU between 1977 and 1996 and found that convergence from poor regions to rich regions was relatively low. Mayawala (2007) analysed the effects of agriculture on economic growth for 71 countries in three groups (developed – developing – less developed) between 1984 and 2004, with the inclusion of economic and social structural variables. The study revealed that both the capital structure and the rise in exports affected growth positively in all countries, while investment risk affected inversely medium income level countries. Government instability affected growth positively in the top two groups. Indeed, structural convergence has been an important research topic in recent years. In their push studies focusing on commercial gains across the EU member states, Lejour et al., (2009) found that EU membership increased commercial relationships by 33 % when considered alone and by 55 % when changes in the structure of economic institutions were also taken into consideration. Moreover, income levels of candidate countries were expected to rise between 31% and 43 % in the long run after they had became EU members. Another study investigating the single sector neo-classical growth model also incorporated an exogenous agriculture sector, and aimed to clarify the reasons behind the late industrial development of some countries regardless of resource abundance (Gollin et al., 2002). Research on 62 developing countries in the period 1960-1990 demonstrated a significant rise in agricultural efficiency when interpreting the uprising trend of GDP per capita. Later, the same study found that agriculture cannot be substituted for the relevant economies. Another important question addressed has been whether agriculture, having inelastic production factors, affects economic growth and per capita income. Therefore, a survey on the effects that agriculture has on economic growth can be important to assess the productivity of the sector and the supportive financial transfers to the sector. The European Union (EU) integration process involves specific tools and produces certain outcomes for economies, and one of the main impacts of the integration process has been recorded in agriculture. The aim of this paper is to analyse the effects of agriculture on economic growth within a sample of EU member and candidate countries. By analysing the impact of agriculture on the national economies, it seems possible to provide information on the productivity of the sector and on efficiency change according to the time considered and the countries involved. Therefore the impact of agricultural value added on per capita income has been estimated using a version of the Solow growth model extended with structural coherence variables to cover the European Union member and candidate countries in two reference periods, 1995-2007 and 2002-2007, taking into account the accession process. Samples were constructed according to data availability and the number of member states. The analysis was based onto two samples to include more candidate and new member states specifically for the 2002–2007 sampling period. Therefore, the main objective of the study was to understand the effect of the relationship between agriculture and growth levels through differences in per capita GDP according to the membership status of the countries with respect to the European Union. 2. Materials Description of the model and variables The Extended Solow Growth Model was applied using panel data analysis for two periods between 1995 and 2007. By the year 2013 there will be 28 EU members and 4 candidates. While it was possible to acquire data for Iceland, Turkey, no data was available for Macedonia and Montenegro. Therefore, 30 countries were included in the study. Within the scope of the study, an analysis was made to measure the impacts of agricultural value added, specific macroeconomic variables and indices representing social and institutional structure over per capita income as growth indicator. Most of the fundamental economic indicators used were retrieved from World Development Indicators and Global Development Finance (Anonymous, 2011a). In addition, structural coherence indices retrieved from the International Country Risk Guide (ICRG) were used in the analyses in order to represent the institutional and social structures of the countries. Political, economic and financial risk The composite risk score is the arithmetic mean of political (100 points), economic (50 points) and financial (50 points) risk scores. The composite risk score is divided into the following rating categories: too low (80-100) and too high (0-49.9) according to ICRG methodology (Hoti, 2003). A new basket composed of government stability, socio-economic conditions, investment profile, openness for democratic control and bureaucratic quality was established, and the scores were indexed to 100 points for political risk. The components of economic risk are per capita GDP risk (ratio of the Dollar-based national income to the average income of all countries), real GDP growth risk (growth rate calculated relating to 1990 prices), annual inflation risk (non-weighted Consumer Price Index), budget balance risk (budget balance of the country with domestic currency to the overall GDP) and current account risk (balance of payments in Dollars to the GDP of the country). Financial risk components are external debt risk (debt to GDP ratio), external debt service risk (current account total to Dollar value of exported goods and services) net international liquidity risk (Dollar value of official annual reserves to monthly Dollar value of import cost of goods) and exchange rate stability risk (depreciation or appreciation of 63 NEW MEDIT N. 4/2013 the national currency). For this study, the economic and financial risk components indexed for 50 points were adjusted to a 100 points scale. The average of the three risk scores is treated as the composite risk variable in the growth model implemented. al observation points, other proposed tests should also be considered (Hoyos and Sarafidis, 2006). The Pesaran CD test, which also depends on the Lagrange Multiplier methodology, is appropriate for situations where T is smaller than N, and when there is a balanced panel (Hoyos and Sarafidis, 2006; Pesaran, 2004). In addition, the non-parametric statistic of Friedman (1937), which relies on the Spearman rank correlation coefficient and Frees statistic (1995, 2004) depending on the square of the correlation coefficient, can be used for the measurement of cross-sectional dependence (Hoyos and Sarafidis, 2006). All three tests assume cross-sectional independence. When independence is inapplicable, it is clear that the model fits best the random effects estimation. In this situation, it is necessary to implement the Hausman specification test to determine whether to use random effects or fixed effects estimation methodology (Baltagi, 2005). In this test, the h coefficient retrieved after taking differences of the regressors with respect to cross-sections and time series can be used to test whether the random effects model is preferred to fixed effects or not (Arellano, 2003). This test focuses on correlations of independent variables with random group effect variables. If this correlation is equal to 0, the random effects model is preferred. At the opposite end of the scale, the fixed effects model is preferred (Lloyd et al., 2001, Washington et al., 2003; Kunst, 2009). 3. Method Measurement of economic development using panel data For the empirical analysis of economic development, time series analysis can be preferred as a methodology, with the utilization of measures retrieved from a unique data source for different time periods. However, panel data studies are preferred when measuring time and place variations of growth, and when comparing the growth performances of different countries (Ederveen et al., 2006, Islam 2003, Chenery and Taylor 1968). Panel data analyses are also preferred in cross-country macro-economic studies due to the short time series (Olofin et al., 2009, Lloyd et al., 2001). Panel data may be estimated with the Least Squares (OLS) or Maximum Likelihood (ML) methods. Yet, what is more important in panel data analysis is whether the estimation could be bound to time series and/or cross sectional points. Due to the decision made on this extended model, panel estimation, fixed effects or random effects estimation methodology is selected (Arellano, 2003). In order to select the proper estimation method, it is important to test the stationarity of a series using the specification tests. After determining whether the series analysed are stationary initially or after being differentiated, it is also important to determine whether the series could be estimated with a co-integration procedure or not. After co-integration was defined as a specific field of study in 1995, many different tests were developed. In our study, Pedroni’s (1999) four-statistic test is used. This test was developed by investigation of ρ and t statistics in the panel and in the group of series. Panel estimation incorporates OLS estimation which disregards cross-sectional and time effects for the valuation of the model. In case of models for which panel estimation is not suitable, the one-way error components model, which refers to the inclusion of unobserved cross-sectional effects, or the two-way error components model involving unobserved time effects as well, are considered to be the appropriate estimation model. In order to understand whether the set could be estimated in the form of a panel, it is important to first implement a cross-sectional dependency test. If we reach a concrete result according to which cross-sectional variables are independent of neglected fixed effects and random effects, then utilization of unified panel regression produces more accurate estimation results. The first testing methodology proposed for a panel series was the Breusch-Pagan Lagrange Multiplier (LM) test (Korkmaz et al., 2010; Berke, 2009). The LM test gives accurate results when the number of time related data points (T) is larger than the number of cross-section related data points (N). By contrast, when there are more cross-section- Panel data analysis methods - Two-way error component regression model When the objective is to measure the impact of an unobserved time series variable, as well as the unobserved crosssectional variable in panel data analysis, it is necessary to implement a two-way error components regression model. Accordingly, the error term takes the following form (Baltagi 2005). ݑ௧ ൌ ߤ ߣ௧ ݒ௧ i=1,……,N; t=1,…..,T [1] Here, while µi is the unobserved cross-sectional effect, λτ is the unobserved time series effect. In addition, λτ is independent of the cross-section and provides information on all time-related data excluded from the regression. In a two-way fixed effects analysis, when µι and λτ are interpreted as constant parameters to be estimated, νit is the two-way error term [1]. Statistical inference depends on the values that N observation points take during definite time periods (Baltagi, 2005). Then the function takes the following form in this situation [2]. ݕ௧ ൌ ߙ ߚݔ௧ ߤ ߣ௧ ݒ௧ [2] The main assumption for random effects estimation is that µi, λτ and νit components have independent distributions with 0 mean and constant variance (Wooldridge, 2002). Inference with random effects analysis gives general information on the population from which the sample is 64 NEW MEDIT N. 4/2013 constructed (Baltagi, 2005). In this analysis, it is assumed that the error terms have constant variance for all i and t as var ሺݑ௧ ሻ ൌ ߪఓଶ ߪఒଶ ߪ௩ଶ . By contrast, covariance of error terms is σµ2 for the same cross-sections (i=j) and for different time periods (t≠s) and is σλ2 for the same time period (t=s) and different observation periods (i≠j) and 0 otherwise. In addition, random effects estimation allows inference on three different error components for the population from which that sample is collected. Accordingly, random effects estimation is referred as the error components analysis in the literature as well. The two-way error components model is preferred when the data set and degrees of freedom are available. Yet, when no cross-sectional dependence is detected, it is more appropriate to estimate the data set in a simplified way. GDP (Gross Domestic Product) is used as the explained variable of the model as a proxy for growth measurement. All quantitative variables in the model were logarithmically transformed. As the composite risk variable includes information on the openness of the economy, the debt situation of the country, and the payment schedule of internal and external debts and inflation, direct variables were not used to prevent potential multi-collinearity problems. Accordingly, the financial performance of the relevant country is assessed using the value and significance of the composite risk coefficient. Model [4] is estimated for two different time periods and the cross-section structure while taking the data set limitations into consideration. The relevant countries and time periods are demonstrated in Table 1. Table 1 - Time series and cross section definition for modeling function. Model Structure – The effect of agriculture and European Union Membership on economic growth 1995-2007 This study implies a specifically modified Solow Growth Model. The model is extended with sector productivity variables and social and economic indicators. The Solow Growth Model stemming from the production function is turned into the following general form [3] after including agriculture, exports and inflation (Hwa, 1988). N = 25 Germany Austria Belgium Bulgaria Czech Rep. Denmark Finland France Netherlands U.K. Ireland Spain Sweden [3] ܻሶ ൌ ܽ ߙܭሶ ߚܮሶ ߛܣሶ ߠܺሶ ߜܲሶ ߝ . . where: Y = average annual GDP .growth rate; K = average annual capital. stock growth rate; L = average annual labour. growth rate; A = average annual agricultural growth rate; X . = average annual export growth rate; P = average annual inflation growth rate. Humphries and Knowles (1998) included non-farm labour, education and health expenditures variables in this system, while Barro (1991) extended the model with economic and structural coherence variables. Since the aim of this study was to measure the impacts of integration to the Union’s social and economic structures on economic growth in the EU member and candidate countries, the following extended form [5] of the Solow Growth Model was applied. i= 1,2,……., N; t = 1,2,……., K; ݈ܻ݊௧ ൌ ߚ ߚଵ ݈݊ܣܸܣ௧ ߚଶ ݈݊ܵܥ௧ ݈݊ߚସ ܧ௧ ߚହ ܷܧ௧ ߚ ܴܥ௧ ߝ௧ 2002-2007 Italy Iceland Cyprus Luxemburg Hungary Malta Poland Portugal Romania Slovakia Turkey Greece K = 13 N = 30 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 Germany Austria Belgium Bulgaria Czech Rep. Denmark Estonia Finland France Croatia Netherlands U.K. Ireland Spain Sweden K=6 Italy Iceland Cyprus Latvia Lithuania Luxemburg Hungary Malta Poland Portugal Romania Slovakia Slovenia Turkey Greece 2002 2003 2004 2005 2006 2007 The cross-section set is longer than the time series set in both periods. Accordingly, for the use and interpretation of descriptive and diagnostic statistics in analytical processes, the N > T situation is considered. Models are estimated by the SAS 9.2 Enterprise 4.3 statistical program. Moreover, E-VIEWS 7.1 is used for unit root tests and co-integration tests, and STATA 11.0 is used for model specification tests. 4. Description and Results Data generation process for 1995-2007 Prior to the estimation of the econometric model, the stationarity of the variables was tested. The results for per capita income, agricultural value added and composite risk variables are provided in Table 2. According to the test results, lny was stationary on level. All tests rejected the existence of a unit root specifically when a time trend was included to the level measure, and agricultural value added series were stationary. The composite risk variable also produced an outcome similar to agricultural value added. The composite risk was clearly stationary on level according to all individual statistics (IPS, ADF and HT) and common unit root statistic (LL). Besides, as the series did not have common unit root, there was no need to test for co-integration across these series. [4] The main variables for the model were the following: Yit : Per capita income in country i and time t AVAit : Per capita agricultural value added in country i and time t CSit : Per capita capital stock in country i and time t Per capita education expenditure amount in counEit try i and time t EUit : EU membership condition in country i and time t (member=1; non-member=0) CRit : Per capita composite institutional quality index value 65 NEW MEDIT N. 4/2013 Table 2 - Stationarity test results for 1995-2007 period. Table 4 - Two-way error component estimation results (1995-2007). lnYit lnYit N: 25 K: 13 Test statistic Hypothesis IPS IPS LL LL – ADF2 Fisher Fisher – ADF2 HT HT it lnAVA Test statistic lnAVA it H0: All panels have unit roots : All panels have unit roots HH 0:0Panel has a unit root : Panel has a unit root HH 0:0All panels have unit roots : Allpanels panels have unit roots HH have unit roots 0:0All H025 : All panels have unit roots N: Hypothesis N: 25 Level + Time !1 1 ! Level + Time Drift trend Drift trend 5,65 1,00 0,73 0,77 1,00 0,77 -2,2425,650,01** -9,90 0,73 0,00** -2,242< 0,01** 0,01** 0,00** 3,2948 3,29-9,90 < 0,01** 3,29481,00< 0,01** 0,7510 3,29 1,01 1,00 < 0,01** 0,7510 1,00 K: 13 1,01 1,00 Level Time ! K: +13 ! Drift trend Level + Time 1,75 0,96 -4,44 Drift trend 0,00** -0,791,750,210,96 -7,06-4,44 0,00** 0,00** 6,57 2,51-7,06 < 0,01** -0,790,00** 0,21 0,00** 0,81 0,68 0,28 < 0,01** 6,57 0,00** 2,51 < 0,01** K: 13 0,81 0,68 0,28 < 0,01** Level + Time ! ! K: 13 Drift trend Level Time -2,66+ < 0,01** -1,18 0,12 Drift trend -10,71 0,00** -9,97 0,00** -2,660,00** < 0,01** 0,12 12,16 2,31-1,18 0,00** -10,71< 0,01** 0,00** 0,67 0,51-9,97 0,56 0,00** Test statistic Test statistic N: 25 K: 13 Hypothesis Hypothesis IPS LL IPS Fisher LL – ADF HT Fisher – ADF CRit HT Test statistic H0: All panels have unit roots HH has a unit root 0: Panel 0: All panels have unit roots HH panels have unit roots 0: All 0: Panel has a unit root H0: All panels have unit roots H0: All panels have unit roots N: 25 H0: All panels have unit roots Hypothesis LL IPS – ADF Fisher LL HT H0: Panel has a unit root : Allpanels panels have unit roots HH have unit roots 0:0All : Panel hashave a unit root HH panels unit roots 0:0All Fisher – ADF HT H0: All panels have unit roots H0: All panels have unit roots CRit Test statistic IPS N: 25 HHypothesis 0: All panels have unit roots 12,16 0,67 0,00** < 0,01** 2,31 0,51 0,00** 0,56 Null hypothesis is rejected on* 0,05 and on** 0,01 significance level. 2 Statistic computation for Fisher ADF with time trend does not involve drift. 1 Table 3 - One-way error component estimation results (1995-2007). lnAVAit CRit EUit Constant N K*N R2 Wald Parameter Test("2(3)) B-P CSD 3 Test ("2(1)) FRE CSD Test 4 RAVE CSD Test Peseran CD Test Hausman S. T. ("2(3)) Wooldridge Test (F(1,24)) Wald Test ("2(3)) 4 0,5493 (0,00) 0,0658 (0,00) 0,5285 (0,00) 0,8140 (0,02) 25 325 0,77 F Test 3 One Way Random Effect One Way Fixed Effect -0,2857 (0,00) 0,0139 (0,00) 0,1720 (0,00) 10,013 (0,00) 25 325 In-group: 0,65 Between groups: 0,04 Total: 0,05 F (3,321): 362,45 (0,00) -0,3418 (0,00) 0,0122 (0,00) 0,1647 (0,00) 10,48 (0,00) 25 325 In-group: 0,66 Between groups: 0,003 Total: 0,0099 F (3,297): 188,06 (0,00) 409,63 (0,00) 80,21 (0,00) 5,471 (0,00) 108,253 (0,00) 18,227 (0,00) CSD = Cross Sectional Dependence. Q values 0,10: 0,1984; 0,05: 0,2620; 0,01: 0,3901. R2 F Test Wald Parameter Test ("2(15)) B-P CSD Test("2(1)) FRE CSD Test 5 RAVE CSD Test Peseran CD Hausman S. T. ("2(15)) Wooldridge Test (F(1,24)) Wald Test ("2(12)) OLS Two-Way Random Effect 0,5493 (0,00) 0,0658 (0,00) 0,5285 (0,00) 0,8140 (0,02) 25 325 0,77 Two-Way Fixed Effect 0,0245 (0,41) 0,0081 (0,00) 0,0561 (0,00) 8,5247 (0,00) 25 325 In group: 0,86 Between groups: 0,83 Total: 0,37 F (3,321): 362,45 (0,00) -0,0019 (0,94) 0,0075 (0,00) 0,0549 (0,00) 8,7348 (0,00) 25 325 In group: 0,86 Between groups: 0,73 Total: 0,29 F (15,285): 112,57 (0,00) 1279,94 (0,00) 831,25 (0,00) 7,056 (0,00) 1,899 (1,00) -1,966 (1,95) Q values 0,10: 0,1984; 0,05: 0,2620; 0,01: 0,3901. 6,06 (0,9788) 271,63 (0,00) 5,42*e16 (0,00) seran statistics, the test result was used to reject cross-sectional independence. In any case, the difference between in group (0.86) and total (0.37) goodness of fit (R2) values retrieved from the two way random effects estimation also indicated that the data could not be estimated using joint panel (Kunst, 2009). What is more, the Hausman statistic, which compares the fixed effects model with the random effects model, was not significant with a 0.98 p-value. Accordingly, it was found that the random effects model would lead to more significant and unbiased results when compared with the fixed effects model. While all parameter estimates were significant in the panel solution, the per capita agricultural value added parameter estimate was not significant in the random and fixed effects solutions. Even so, the Wald parameter significance test1 indicated that the parameter estimates of random effects had common significance. The parameters were jointly significant due to the 0 p-value of the Wald statistic. Furthermore, the 37% total goodness of fit (R2) indicated that explanatory variables could explain the 37% variation in the dependent variable. Prior to interpretation of the findings, diagnostic tests were implemented to understand whether there was some autocorrelation in error terms, and whether the error variance was homoscedastic. As a result of the Wooldridge autocorrelation test results, the null hypothesis of ‘no autocorrelation’ was rejected. In any case, the likelihood ratio based Wald heteroscedasticity test rejected ‘no heteroscedasticity’ hypothesis. Accordingly, the model was reestimated using the White error correction methodology. The test findings and consistent parameter estimates are provided in Table 5. Due to one-way error components estimation outputs shown in Table 3, Frees, Peseran CD and Friedman RAVE tests indicated that the model had cross-sectional dependence, and that the data could not be estimated in panel form. However, the Hausman specification test results showed that one-way random effects estimation would produce biased results. However, as the parameter estimate of per capita agricultural value added did not comply with economic expectations, and as the length of the series was long enough, it was decided to execute comparative estimation for two-way error components. The outputs of the two-way error components estimation are provided in Table 4. Initially, considering the FRE and B-P statistics it was clear that the model had cross-sectional dependence. As the Frees test was developed by evaluation of Friedman and PeOLS lnAVAit CRit EUit Constant N K*N 5 Estimation results for 1995-2007 Independent Variable Independent Variable 16,53 (< 0,01) 271,63 (0,00) 2.045,24 (0,00) 1 Degrees of freedom for the statistic (k-1): 13 years + 3 explanatory variables - 1. 66 NEW MEDIT N. 4/2013 The EU dummy variable parameter estimate of 0.0561 indicates, after taking the anti-log, that average per capita income is around 1 Dollar higher in EU member countries. Like the composite risk interpretation, the EU membership semi-elasticity of per capita income is 0.057. The semi-elasticity for categorical variables is calculated by taking the anti-log of the parameter estimate and deducting 1 in semilogarithmic models as demonstrated by Halvorsen and Palmquist (1980) (Gujarati, 2003). Following this calculation, almost the same value is reached using the parameter estimate. Accordingly, being a member of the European Union leads to a 0.057 % rise in per capita income. Also, time has a significant effect on per capita income rise. Average per capita income rises with time. Table 5. Two-way random effects estimates with consistent standard errors for 1995-2007 period. Independent Variable Two-Way Random Effects 0,0245 (0,79) 0,0081 (0,00) 0,0561 (0,07) 8,5247 (0,00) 0,01 (0,16) 0,02 (0,10) 0,05 (0,00) 0,09 (0,00) 0,13 (0,00) 0,15 (0,00) 0,15 (0,00) 0,17 (0,00) 0,18 (0,00) 0,21 (0,00) 0,25 (0,00) 0,29 (0,00) 25 325 In group: 0,86 Between groups: 0,83 Total: 0,37 807,47 (0,00) lnAVAit CRit EUit Constant 1996 1997 1998 1999 2001 2001 2002 2003 2004 2005 2006 2007 N K*N R2 Wald Parameter Test ("2(15)) Data generation process for 2002-2007 The same unit root tests were applied as in the previous period in order to test whether the variables were stationary or not. Stationary test results are indicated in Table 6 for per capita national income, agricultural value added and composite risk variables. According to results with time trend, the Fisher ADF test indicated that per capita income variable in logarithmic form was stationary. Accordingly, lny was stationary with trend on level. The per capita agricultural value added series was stationary for LL and Fisher ADF statistics. However, the same series was not stationary according to the IPS and HT statistics. Even so, when interpreted with trend on level, the per capita agricultural value added series was found to be stationary using the Fisher ADF and HT statistics. In contrast, the composite risk variable was stationary on level due to the LL and Fisher ADF statistics and it was found to be non-stationary with trend. Accordingly, the series were found to be stationary on level, and investigating co-integrating relationship among the series was considered to be unnecessary. Consequently, the agricultural value added elasticity of per capita national income proves to be 0.0245. Per capita income rises by almost 1% (0.08 %) due to a 1 point increase in the composite risk variable, which means an appreciation of the risk position of the country. Consequently, the composite risk semi-elasticity of per capita income is 1% as semi-elasticity refers to the parameter estimates directly in variables taking numerical values (Gujarati, 2003). Table 6 - Stationarity test results for 2002-2007 period. lnYit N: 25 K: 6 Test statistic Hypothesis Level + Drift !6 IPS LL Fisher – ADF7 HT lnAVAit H0: All panels have unit roots H0: Panel has a unit root H0: All panels have unit roots H0: All panels have unit roots N: 25 13,82 3,57 -2,63 1,05 K: 6 1,00 0,99 0,99 1,00 Test statistic Hypothesis Level + Drift !8 IPS LL Fisher – ADF HT CRit H0: All panels have unit roots H0: Panel has a unit root H0: All panels have unit roots H0: All panels have unit roots N: 25 Test statistic Hypothesis IPS H0: All panels have unit roots LL H0: Panel has a unit root Fisher – ADF H0: All panels have unit roots HT H0: All panels have unit roots 6 * Null hypothesis is rejected on 95 %, ** 0,84 -4,09 6,02 0,47 Estimation results for 2002-2007 Time trend 21,08 0,24 ! As for the 1995-2007 period, an initial comparative analysis was made between the one-way error component and the panel LS estimation outputs. 0,00** 0,98 Table 7 - One way error component estimation findings (2002-2007). Independent Variable 0,79 0,00** 0,00** 0,09 Time trend 14,94 -0,31 OLS One Way Random Effect ! lnAVAit CRit-1 EUit Constant N K*N 0,00** 0,01** 0,6068 (0,00) 0,0856 (0,00) 0,3908 (0,00) -1,0854 (0,00) 30 180 K: 6 Level + Drift 1,12 -24,26 5,27 0,51 !9 0,87 0,00** 0,00** 0,22 Time trend -1,42 0,30 R2 ! 0,83 F Test Wald Parameter Test ("2(3)) B-P CSD Test ("2(1)) FRE CSD Test 10 RAVE CSD Test Peseran CD Test Hausman S.T. ("2(3)) Wooldridge Test (F(1,29)) Wald Test ("2(3)) 0,92 0,99 null hypothesis is rejected on 99%. Fisher ADF statistic with time trend does not include drift. 8 * Null hypothesis is rejected on 95 %, ** null hypothesis is rejected on 99%. 9 * Null hypothesis is rejected on 95 %, ** null hypothesis is rejected on 99%. 7 10 67 0,0102 (0,88) 0,0116 (< 0,01) 0,1868 (0,00) 8,3212 (0,00) 30 180 In-group: 0,28 Between groups: 0,70 Total: 0,62 F (3,176): 290,10 (0,00) One Way Fixed Effect -0,1524 (<0,01) 0,0029 (0,32) 0,1691 (0,00) 9,9651 (0,00) 30 180 In-group: 0,65 Between groups: 0,04 Total: 0,05 F (3,147): 27,46 (0,00) 56,48 (0,00) 169,80 (0,00) 7,025 (0,00) 90,819 (0,00) 26,215 (0,00) Q values 0,10: 0,4127; 0,05: 0,5676; 0,01: 0,9027. 34,30 (0,00) 1783,356 (0,00) 2669,03 (0,00) NEW MEDIT N. 4/2013 Table 8 - Two-way error component estimation results (2002-2007). Independent Variable lnAVAit CRit EUit Constant N K*N R2 Two-Way Random Effect 0,6068 (0,00) 0,0856 (0,00) 0,3908 (0,00) -1,0854 (0,00) 30 180 0,83 F Test Wald Parameter Test ("2(8)) B-P C-D ("2(1)) FRE CSD Test 11 RAVE CSD Test Peseran CD Test Hausman S. T. ("2(15)) Wooldridge Test (F(1,29)) Wald Test ("2(8)) 11 OLS 0,2178 (0,00) 0,0067 (< 0,01) 0,0558 (< 0,01) 7,4859 (0,00) 30 180 In-group: 0,78 Between groups: 0,65 Total: 0,56 F (3,176): 290,10 (0,00) Table 10 - Two-way fixed effects estimation results for 2002-2007. Independent Variable Two-Way Fixed Effect lnAVAit CRit EUit Constant 2003 2004 BE BL CRT CZ CY DEN ES F# FR GE GR HUN IC IR IT 0,1671 (0,00) 0,0033 (<0,07) 0,0507 (< 0,01) 8,0572 (0,00) 30 180 In-group: 0,79 Between groups: 0,59 Total: 0,46 F (29,142): 422,86 (0,00) 303,96 (0,00) 200,36 (0,00) 8,395 (0,00) -0,950 (1,66) 2,533 (1,00) Q values 0,10: 0,1984; 0,05: 0,2620; 0,01: 0,3901. 9,44 (0,31) 1783,356 (0,00) 156,11 (0,00) As a result of the findings demonstrated in Table 7, the initial cross-sectional dependency was confirmed for each variable with B-P, FRE, RAVE, and CD values. In addition, it is understood that the model could be estimated at least by using the random effects model when the difference between the in-group and total R2 is considered. In addition, the Hausman m statistic indicated that the fixed effects methodology was the most appropriate methodology in one-way error components estimation. Yet, as the concerned variables reached level stationarity only with trend, and due to the availability of the data, the findings were interpreted with two-way error components analysis. Comparison of the twoway error components analysis results and the panel estimation outputs are provided in Table 8. Initially, cross-sectional independence was rejected as a result of the FRE and B-P statistics. However, the CD and RAVE statistics seemed not to reject cross-sectional independence, but the difference between the in-group (0.78) and total (0.56) goodness of fit values (R2) indicated that the series could not be estimated Table 9 - Two-way random error estimation outputs with joint Panel for 2002-2007 with consistent standard errors. (Kunst, 2009). Independent Two--Way Random In addition, Variable Effects the Hausman m-statistic, lnAVAit 0,2178 (< 0,01) CRit 0,0067 (<0,06) which compaEUit 0,0558 (0,02) res the fixed Sabit 7.4859 (0,00) effects and ran2003 0,03 (0,00) dom effects es2004 0,05 (0,00) timation find2005 0,11 (0,00) ing of two-way 2006 0,16 (0,00) 2007 0,20 (0,00) estimation, N 30 was not signifK*N 180 icant with a In-group 0,78 0.31 p-value. As 2 R Between groups: 0,65 a result, as it Total: 0,56 was the case Wald Parameter 156,11 (0,00) Test ("2(15)) for the 1995- 0,1671 (0,00) 0,0033 (<0,07) 0,0507 (< 0,01) 8,0572 (0,00) 0,03 (0,00) 0,05 (0,00) 0,05 (0,14) -2,25 (0,00) -1,30 (0,00) -1,68 (0,00) -0,52 (0,00) 0,18 (0,00) -1,27 (0,00) -0,05 (0,09) -0,12 (0,00) 0,06 (0,06) -0,64 (0,00) -1,34 (0,00) 0,11 (0,15) 0,13 (0,00) -0,23 (0,00) Two-Way Random Effects 2005 2006 2007 LUX LET L#T MAL NL PL POR ROM SLV SLK SP SW UK TR 0,11 (0,00) 0,16 (0,00) 0,20 (0,00) -1,46 (0,00) -1,51 (0,00) 0,79 (0,00) -0,78 (0,00) -0,04 (0,16) -1,43 (0,00) -0,69 (0,00) -2,24 (0,00) -1,19 (0,00) -0,67 (0,00) -0,49 (0,00) 0,17 (0,00) 0,21 (0,00) -1,62 (0,00) 2007 period, the random effects estimation was preferred to the fixed effects estimation. While the composite risk parameter estimate had lower significance than the fixed effects estimation, all parameter estimates were found to be significant in all models. The Wald test2 mentioned earlier, which combined the parameter significance and 56 % of variation of the dependent variable is explained by independent variables. Finally, autocorrelation and heteroscedasticity were tested in order to assess the consistency of estimation. The Wooldridge autocorrelation test findings rejected existence of an autocorrelation in the model and the Wald statistic findings retrieved by Generalized Least Squares rejected constant variance hypotheses. Therefore, the model was reestimated according to the White error correction mechanism, and the outputs are reported in Table 9. Estimation with consistent errors did not lead to loss of significance except for the composite risk variable. However, joint significance was secured as indicated using the Wald statistic. Based on the 2002-2007 estimation outputs, the agricultural value added elasticity of per capita income was 0.22. Per capita income rose by around 1% (0.7 %) with a 1 point rise in the composite risk indicating an appreciation in the risk condition of the country. In accordance with this finding, the composite risk semi-elasticity of per capita income was 1%. Any reduction in the composite risk or appreciation of the countries’ economic, political and financial risk condition affects countries in a similar way to that observed in the previous period. 2 Degrees of freedom for the statistic (k-1): 6 years + 3 independent variables - 1. 68 NEW MEDIT N. 4/2013 As regards the two-way random effects estimation results for 25 countries between 1995 and 2007, the agricultural value added elasticity of per capita income was found to be 0.025, and it was clear that 1 point appreciation of composite risk leads to a 1% rise in per capita income. Also, average per capita income proved to be 5.6 % higher in EU member countries. Following the two-way random effects estimation results for 30 countries between 2002 and 2007, the agricultural value added elasticity of per capita income was found to be 0.22, and it was again clear that a 1 point appreciation of composite risk leads to a 1% rise in per capita income. However, the average per capita rise due to EU membership was the same as in the previous period i.e. 5.6%. Hence, the agricultural value added and the rise in agricultural value added contributed to the average per capita income in the two sub-periods. The reason why the quantitative effect of agricultural value added was higher in the second period seems to be related with the accession of Central and Eastern European countries, which proceeded from low equilibrium to equilibrium during the 2002-2007 period. This explains why the above countries display a higher population density compared to former members and their range of economic activity is less diversified. Finally, EU membership, which is measured categorically, also contributed to the rise in average per capita income. Therefore, EU membership affects per capita income externally as an independent factor for the period. In other words, it is estimated that if different countries have the same factor endowments, if they have institutional structures in conformity with each other, and if they are likely to have more open economic structures, they will have comparatively higher per capita income levels. Consequently, EU membership contributed to average per capita income by 5.6% in both the reviewed periods. Appreciation of economic, financial and political risk conditions also leads to positive improvements in average per capita income. The composite risk score is taken as the average of political risk, meaning the institutional quality and administrative capacity of a country, economic risk, meaning the economic stability of an economy, and financial risk, meaning the economic openness and commercial capacity of an economy. A one point rise in this average score, or a one point appreciation of the economic, financial and political situation of a country resulted in an almost 1% rise in per capita income in both periods. The findings of the relevant reviewed periods and samples indicate that EU membership and appreciation of the composite risk score leads to positive improvements in per capita income. Moreover, in order to measure inter-country differences, a two-way fixed effects model was also estimated for the 2002-2007 period. Based on these findings, an average per capita income reduction was observed in 22 countries with reference to the 2002 Austrian data. The reduction measured for Turkey was 0.81 %. As for the 1995-2007 sampling period, the average per capita income was around 1 Dollar (1.0574) higher in EU member countries, and at the same time the EU membership elasticity of per capita income was 0.057. After assessing the estimation results, the objective was to interpret the two-way fixed effects estimation results for 2002-2007 to conform to the economic expectations and quantitative proximity to random effects estimates. The main reason was to evaluate the country parameter estimates. The estimates, in accordance with this aim, are reported in Table 10. Considering the estimation outputs from which the Austrian data for 2002 is taken as a base, the average per capita GDP shows changes according to the countries concerned. In 22 countries from the sample of 29 countries (excluding Belgium, Denmark, Iceland, Ireland, Latvia, Sweden and the United Kingdom), the semi-elasticity of per capita income with reference to Austria and year 2002 is negative. Alternatively said, the average per capita income in these 22 countries is lower than the 2002 average per capita income level of Austria. The reduction rate for Turkey was 0.81 % when the antilog of the parameter estimate (-1.62) was taken. As with Turkey, the average per capita income reduction was higher in Poland, Bulgaria, Romania and the Czech Republic. This finding is compatible with the economic expectations of recent new member countries from Central and Eastern Europe. The analysis of the impacts of agricultural value added, composite risk and EU membership on per capita income in two different samples considering different time dimensions were completed accordingly. The fundamental findings indicated that contribution of agricultural value added to per capita income has increased over the 20022007 analysis period when the 5th enlargement has been completed. In addition, EU membership and appreciation in the risk position has affected per capita income positively. 5. Discussion The objective of this study was to investigate the impact of agriculture on national economies within a sample constructed from member and candidate countries of the European Union. The study focused on panel data estimation of the effect of agricultural value added on per capita income by applying empirical methods to two sub-samples. An extended form of the Solow Growth Model was applied to 25 countries in the period ranging from 1995 to 2007, and to 30 countries between 2002 and 2007. The model used was an extension of the Solow Growth Model, and per capita income was used as a growth indicator. The impacts of per capita agricultural value added, and the composite risk variable determined through economic and social risk indicators pre-calculated by the PRS group were used as explanatory variables. In addition, the effect of EU membership was measured by a dummy variable used as an explanatory variable. 69 NEW MEDIT N. 4/2013 Barro, R. J. (1991). Economic Growth in A Cross Section Of Countries. The Quarterly Journal Of Economics 106, 407-43. Beck, N. and Kazt, J. N. (1995). What To Do (And Not To Do) With Time-Series Cross-Section Data. The American Political Science Review, 89 (3):634-647.Chenery, H. B. 1962. Development Policies for Southern Italy. The Quarterly Journal of Economics, Vol. 76, No. 4 (Nov., 1962): 515-547. Chenery, H. B. And Taylor, L. 1968. Development Patterns: Among Countries and Over Time. The Review of Economics And Statistics, Vol. 50:4, 391-416. Chenery, H. B. and Eckstein, P. 1970. Development Alternatives for Latin America. The Journal of Political Economy, Vol. 78, No. 4, Part 2: Key Problems of Economic Policy in Latin America (Jul. - Aug., 1970): 966-1006. Dowrick, S. and Nyugen, D-T. (1989). OECD Comparative Economic Growth 1950-85: Catch-Up and Convergence. The American Economic Review, Vol. 79, No. 5 (Dec., 1989): 1010-1030. Drukker, D. M. (2003) Testing For Serial Correlation In Linear Panel-Data Models. The Stata Journal (2003): 3-2, 168-177. Dutt, A. K. and Ros J. (2008). (ed) International Handbook of Development Economics. Volumes 1 and 2. Edward Elgar Publishing. Ederveen, S., de Groot, H. L. F., Nahuis, R. (2006). Fertile Soil for Structural Funds? A Panel Data Analysis Of The Conditional Effectiveness Of European Cohesion Policy. Kyklos, 59-1, 17-42. Frees, E.W. (1995). Assessing Cross-Sectional Correlation in Panel Data. Journal of Econometrics 69, 393-414. Frees, E.W. (2004). Longitudinal and Panel Data: Analysis and Applications in the Social Sciences. Cambridge University Press. Friedman, M. (1937). The Use of Ranks To Avoid The Assumption of Normality Implicit in the Analysis of Variance, Journal of The American Statistical Association, 32, 675-701. Gollin, D., Parente, S. and Rogerson, R. (2002). The Role of Agriculture in Development. The American Economic Review, Vol. 92, No. 2, Papers and Proceedings of the One Hundred Fourteenth Annual Meeting of the American Economic Association (May, 2002): 160-164. Gujarati, D. (2003). Basic Econometrics. McGraw Hill, 4th Edition. New York. 1002 pages. Halvorsen, R. And Palmquist, R. (1980). The Interpretation of Dummy Variables in Semilogarithmic Equations. American Economic Review, 70-3, 474-475. Hayami, Y. and Ruttan, V. W. (1970). Agricultural Productivity Differences among Countries. The American Economic Review, Vol. 60 (5 -Dec., 1970): 895-911. Hoti, S. (2003). The International Country Risk Guide: An Empirical Evaluation. Modelling and Simulation Socie- The following inferences are reached regarding the EU accession process, and the impacts of this process on agriculture and economic growth due to the analytical findings. The findings demonstrate that per capita income is higher in EU member countries compared to non-member countries. One of the main factors explaining this condition is the effect of the free movement of goods, services, persons and capital, and the specific implementation provisions developed in accordance with these principles. Specifically, the free movement of persons allowed more opportunities to gain access to income generating activities in the 12 countries which had completed the accession process, without considering Bulgaria, Romania and Cyprus. In addition, the procedures used for cross-border movement of goods provided specific commercial advantages to recently developing new member countries, as well as to developed member countries. From this perspective, the external contribution of European Union membership to per capita income level as an indicator of economic growth can be inferred. Agriculture secures its importance in terms of contribution to the economy of both member and candidate countries of the European Union. Specifically, when new member countries that have a significant rural population rate despite being over-populated and Mediterranean member countries, where agricultural trade continues to be a fundamental economic activity, are considered, agriculture cannot be evaluated as a sector that should be kept out of the integration process. Basically, the positive impact of agricultural value added on per capita income in the two periods under investigation confirms the evidence-based perspective of the European Union, according to which agriculture has to be considered an important sector. References Anonymous. (2011a). World Bank World Development Indicators: data.worldbank.org Arellano, M. (2003). Panel Data Econometrics. Oxford University Press, New York, 231 Ss. Washington, S. P., Karlaftis M. G. And Mannering F. L. (2003). Statistical and Econometric Methods for Transportation Data Analysis. Chapman & Hall/Crc, Boca Raton, Fl. Berke, B. (2009). Public Debt Stock and Inflation Relationship in the European Monetary Union: Panel Data Analysis. Ekonometri ve statistik (Econometrics and Statistics) 9, 30-55. Balassa, B. and Bertand, T.J. (1970). Growth Performance of Eastern European Economies and Comparable Western European Countries. The American Economic Review, Vol. 60, No. 2, Papers and Proceedings of the Eightysecond Annual Meeting of the American Economic Association (May, 1970): 314-320. Balassa, B. (1971). Trade Policies in Developing Countries. The American Economic Review, Vol. 61, No. 2, Papers and Proceedings of the Eighty-Third Annual Meeting of the American Economic Association (May, 1971): 178-187. Baltagi, H. B. (2005). Econometric Analysis of Panel Data. John Wiley & Sons Ltd., 3rd Ed. UK: 302 pages. 70 NEW MEDIT N. 4/2013 ty Of Australia And New Zealand Index. 3/09. http://www.mssanz.org.au/modsim03/volume_03/b09/04 _hoti_international.pdf. (Approach 02.12.2011). Hoyos, de, R. E. and Sarafidis, V. (2006). Testing for Cross–Sectional Dependence In Panel–Data Models. The Stata Journal 6(4): 482-496. Humphries, H. And Knowles, S. (1998). Does Agriculture Contribute to Economic Growth? Some Empirical Evidence. Applied Economics 30, 775-81. Hwa, E-C. (1988). The Contribution of Agriculture to Economic Growth: Some Empirical Evidence. World Development, 16, 1329-39. Islam, N. (2003). Growth Empirics: A Panel Data Approach. The Quarterly Journal of Economics 110 (1995), 1127-70. Kendrick, J. W. (1961). Productivity trends in the United States, Princeton University Press, pp: 630. Knowles, S. and McCombıe, J., (2002). Erkin Bairam: 1958-2001, His contribution to economics. The University of Otago, Economics discussion papers, no: 0212: 22pp. Komlos, J. (1988). Agricultural Productivity in America and Eastern Europe: A Comment. The Journal of Economic History, 48 (3 - Sep., 1988): 655-664. Korkmaz, T. Yildiz, B. and Gökbulut, R.I. (2010). Testing Validity of Asset Pricing Model in Istanbul Stock Exchange 100 Index IMKB with Panel Data Analysis. Istanbul University Journal of Business Faculty 39:1, 95-105. Kunst, R. M. (2009). Econometric Methods for Panel Data – Part II. Unpublished, University of Vienna, http://homepage.univie.ac.at/robert.kunst/panels2e.pdf (Approach 02.12.2011). Lejour, A. M., Solanic, V. and Tang, P. J. G. (2009). EU Accession and Income Growth: An Empirical Approach. Transition Studies Review, 16 (1): 127-144. Lloyd, T., Morrisey, O. and Osei, R. (2001). Problems with Pooling in Panel Data Analysis for Developing Coun- tries: The Case of Aid and Trade Relationships. Centre for Research In Economic Development (Credit), University Of Nottingham Research Paper No. 01/14. Olofin, S. O., Kouassi, E. and Salisu, A. A. (2009). Testing for Heteroscedasticity and Serial Correlation in a TwoWay Error Component Model. Paper Presented in 14 Annual Conference On Econometric Modelling For Africa, 810th July 2009, Abuja, Nigeria. Oyerantı, O. A. 2000. “Concept and Measurement of Productivity” in Productivity and Capacity Building in Nigeria. Proceedings of the Ninth Annual Conference of the Zonal Research Units of the Central Bank of Nigeria, 12th-16th June, Abeokuta: 16-35. Pedroni, P. (1999). Critical Values for Cointegration Tests in Heterogeneous Panels With Multiple Repressors. Oxford Bulletin Of Economics And Statistics, Special Issue, 0305 – 9049. Peseran, M.H. (2004). General Diagnostic Tests for Cross Section Dependence in Panels. Cambridge Working Papers in Economics No. 0435, Faculty of Economics, University Of Cambridge. Solow, R., M. (1957). Technical Change and the Aggregate Production Function. The Review of Economics and Statistics, Vol. 39, No. 3 (Aug., 1957): 312-320. Washington, S. P., Karlaftis M. G. and Mannering F. L. (2003). Statistical and Econometric Methods for Transportation Data Analysis. Chapman & Hall/Crc, Boca Raton, Fl. Wılhelm, S. (2008). Croissance tunisienne et importations agricoles: une analyse longitudinale. New Medit, vol 7, n. 2: 50-56 Wooldridge, J. M. (2003). Econometric Analysis of Cross Section and Panel Data. The MIT Press, Massachusetts, 752 pages. 71