= — — : = = — — -— — — —. — — — — LIBRARY Michigan State University This is to certify that the thesis entitled TRADE ANALYSIS OF SPECIFIC AGRl-FOOD COMMODITIES USING A GRAVITY MODEL presented by Cristébal A. Aguilar has been accepted towards fulfillment of the requirements for the Master degree in Agriculture Economics Jwfl ' / Major Professor’s Signatfire 1/ ’- /o , 0 (a Date MSU is an Affirmative Action/Equal Opportunity Institution PLACE IN RETURN BOX to remove this checkout from your record. TO AVOID FINES return on or before date due. MAY BE RECALLED with earlier due date if requested. DATE DUE DATE DUE DATE DUE 2/05 p:/C|RC/DaIeDue.indd-p.1 TRADE ANALYSIS OF SPECIFIC AGRI-FOOD COMMODITIES USING A GRAVITY MODEL By Cristobal A. Aguilar A THESIS Submitted to Michigan State University in partial fulfillment of the requirements for the degree of MASTER OF SCIENCE Department of Agricultural Economics 2006 TRADE ANALYSIS OF SPECIFIC AGRI-FOOD COMMODITIES USING A GRAVITY MODEL ABSTRACT By Cristobal Aguilar This thesis explores the use of a gravity model (GM) to evaluate the determinants of trade for specific products defined at the 8-10 digit SITC-level. Frozen tart cherries (FTC) are used as an empirical example, allowing for estimation of the determinants of international trade and evaluation of trade potential among the largest exporters and importers of this specific good. Given statistical characteristics of the data, a Tobit model with fixed effects using panel data estimated by MLE is used as the estimation procedure. Results indicate both income and output effects are positive with a significant influence on trade values. A marginal analysis of the model shows FTC trade value is income elastic. According to results of the country-pair fixed effects analysis, a common border, trade agreement and historical trade relationships all contribute to establishing trade partners. The results of this study can generate insights for policy-makers who address needs in specific local markets which are impacted through strong global linkages. This thesis is dedicated to my beautiful children Anika and Joaquin and to my unconditional partner, Valeria iii ACKNOWLEDGEMENTS I’ (1 like to thank the Department of Agricultural Economics for giving me the opportunity to be part of the Ag Econ family. I would like to especially thank Dr. Suzanne Thomsbury for her belief in me, for her constant support and her friendship. Thanks Suzanne!!! I also want to thank: — Dr. Robert Myers for being part of this thesis since it was just an idea and for his invaluable insights. — Dr. Steve Matusz for joining to the team in the middle of the journey and for bringing to the table provocative thoughts. — Dr. Jeffrey Wooldridge for always having his door open, and for his guidance. — Dr. William Green for taking the time to respond to every single question I had. I would like to thank every one of my fellow students for being so supportive and encouraging, and a special thanks to “Frutopia”. A very special thanks to Dr. Crawford for being always The Advisor. And of course, thanks to my family for being my base. iv TABLE OF CONTENTS ACKNOWLEDGEMENTS iv TABLE OF CONTENTS v LIST OF FIGURES vii LIST OF TABLES viii ACRONYMS ix CHAPTER I l I. Introduction ............................................................................................................................................ l 1.1 Background and Motivation ............................................................................................................ I 1.2 Objectives ........................................................................................................................................ 4 1.3 Procedures ....................................................................................................................................... 5 1.4 Implications ..................................................................................................................................... 8 CHAPTER II 10 2. Literature review .................................................................................................................................. 10 2.1 Gravity Model and lntemational Trade ......................................................................................... 10 2.2 Gravity Model and Selected Empirical Applications .................................................................... 15 2.3 Gravity Model and Econometric Tools .......................................................................................... 19 CHAPTER III 27 3. Model Derivation and Empirical Analysis ........................................................................................... 27 3.1 The Frozen Tart Cherry Sector ...................................................................................................... 27 3.2 Development of a Commodity Specific Gravity Model ................................................................ 31 3.3 Empirical Methods for Frozen Tart Cherry Trade Analysis .......................................................... 36 CHAPTER IV 50 4. Results .................................................................................................................................................. 50 4.1 Model Estimation Results .............................................................................................................. 50 4.2 Marginal Effects ............................................................................................................................ 54 4.3 Frozen Tart Cherry Export Potential .............................................................................................. 64 CHAPTER V 67 5. Conclusions .......................................................................................................................................... 67 5.1 Frozen Tart Cherry Trade Analysis ............................................................................................... 67 5.2 The Model ...................................................................................................................................... 72 5.3 FTC Sector and Policy Makers ...................................................................................................... 73 5.4 Future Research ............................................................................................................................. 75 ANNEXES 76 REFERENCES 80 vi LIST OF FIGURES Figure 1. World Tart Cherry Production, annual average of selected countries .............. 28 Figure 2. Import and Export Shares, annual average from 2000 to 2003 of selected countries. ............................................................................................................ 3O vii LIST OF TABLES Table 1. Selected authors and their approach to the GM considering censored dependent variable, unobserved and time invariant variables and panel data. ...................... 8 Table 2. Average yearly value of frozen tart cherry exports of selected countries, 1993 -— 2003. ................................................................................................................... 29 Table 3. Value of continuous independent variables for 2003 and average rate of growth between 1993 and 2003. ..................................................................................... 39 Table 4. Log-likelihood ratio test and AIC estimates for the alternative models ............. 47 Table 5. Tobit estimates with fixed effects for trade value ............................................... 51 Table 6. OLS estimation individual effects ...................................................................... 54 Table 7. Marginal effects of Tobit estimates conditional on trade value greater that zer058 Table 8 Marginal effects of Tobit estimates not conditional on trade value greater that zero ..................................................................................................................... 59 Table 9. Export potential for selected countries expressed as percent of the annual export average. ............................................................................................................... 65 viii AIC AUS AUT BEL CAN CES DEN EU FRA FTA FTC GATT GDP GDP-PPP GER GM GRE GSP H-O HUN ISIC ITA JAP LR-Test MEX MLE NED OLS POL SITC SME TUR UK US. WTO ACRONYMS Akaike’s Information Criterion Austria Australia Belgium Canada Constant Elasticity of Substitution Denmark European Union France Free Trade Agreement Frozen Tart Cherries General Agreement on Tariffs and Trade Gross Domestic Product Gross Domestic Product Expressed in Terms of Purchasing Pan'ty Power Germany Gravity Model Greece General System of Preferences Heckscher-Ohlin Model Hungary lntemational Standard Industrial Classification Italy Japan Log-likelihood ratio test Mexico Maximum Likelihood Estimator Netherlands Ordinary Least Squares Poland Standard lntemational Trade Classification Small and Medium Enterprises Turkey United Kingdom United States World Trade Organization ix CHAPTER I 1. Introduction 1.1 Background and Motivation lntemational trade is a field that has been studied broadly and one in which researchers have covered a wide range of topics. Yet, most of the empirical studies deal with aggregated variables and address issues from the perspective of a country-level or a sectorial level within a country. Empirical studies of international trade that address particular issues of a specific subsector or a particular commodity are less common. Empirical analysis of international trade for specific products is an angle of international trade which has received little attention. The lack of attention to this trade topic may be explained, in part, by the shortage of readily available information prior to the 19905, although efforts are still limited. For instance, Harrigan (1994) analyzed intra-industry trade using 3-digit lntemational Standard Industrial Classification (ISIC) data. This level of aggregation classifies agriculture (excluding forestry) into five categories such as livestock, crop production and fishing or hunting. Lee and Swagel (1997) used a lower level of aggregation studying the effects of trade barriers across countries and industries using 4-digit ISIC data. At the 4-digit level, the food manufacturing industry is classified into categories such as slaughtering, dairy products, canned goods and fruit preserves, and grain mill products. Compared with aggregate trade volumes, these 3- and 4-digit international trade studies provide richer information to policy makers who want to address a particular sector, and to producer groups of high volume trade commodities such as tomatoes, oranges, and apples. Still, when it comes to a very specific product (e. g. cabbage, marmalades, mushrooms, pears, tart cherries, 8 — 10 digit Standard lntemational Trade Classification, SITC), previous studies are still too general and results cannot be applied to decision making in specific subsectors. Even though contributions of a specific agri-food subsector may seem marginal at a national level, their importance to the economy at a local level cannot be neglected. Subsectors like these play a key role in creating up-stream and down- stream linkages that generate economic opportunities at the local level. This study explores the use of a gravity model (GM) to evaluate the determinants of trade for specific products (8-1 0 digit SITC). The GM, analogous to the gravity law, is based on the fundamental premise that trade between two countries is directly proportional to the size of their economy and inversely proportional to the distance between them. The application of the GM in international trade was initially developed by Tinbergen (1962) and independently by Poyhonen (1963). The GM has also been used in areas such as migration flow analysis (Karemera, et al, 2000); foreign direct investment (Brenton, et al, 1999; Frankel and Cavallo, 2004); and market area analysis (Baker, 2000). Use of the GM in international trade at an aggregate level has been extensive: examination of bilateral trade, (Prentice et al, 1998); block trade analysis, (Carrillo and Li, 2002; Martinez-Zarzoso & Nowak-Lehmann, 2003); multilateral trade agreements, (Rose, 2002); trade policy analysis, (Wall, 2000); estimation of trade potential, (Pravorne et al, 2003); hypothesis testing such as the Linder Hypothesis (McPherson, 2001) and border effect (McCallum, 1995; Wall, 2000). But, despite its multiple uses, there is a lack of information about the econometric specification of a GM to analyze international trade when it comes to a specific disaggregate commodity.l An example of a disaggregated traded commodity sector with important local economic impacts is the tart cherry industry in Michigan. Approximately 70 percent of all US. production takes place in Michigan. Total farm-gate cash receipts from tart cherries approached $US 58 million during 2003 (MASS, 2004). In addition, it is estimated that a majority of the processing capacity is also in Michigan. Using an output multiplier of 1.87 annual contribution to the state economy would approach $US 108 million.2 In this study, trade of frozen tart cherries (FTC) among the seventeen biggest country traders of this product is evaluated using a GM. Frozen tart cherries have been chosen as an empirical case because it is a product that clearly exemplifies the kind of subsector addressed by this study. The motivation for selecting FTC was based on changes that developed in this subsector after a severe shortage in US. production during 2002.3 This production shortage caused international players, such as Poland (the world’s largest exporter) to move quickly to fill ' The work done by Koo and Karemera (1991) is an exception; their work is discussed later. 2 Doherty and McKissick (2000) 3 In 2002 tart cherries yield in Michigan were 10 percent of the average yield from 2000 to 2004, excluding 2002 (NASS 2004). the void and gain access to new markets. So, after 2001 , the vulnerability of the sector and close connection to international was more evident. This is a case that exemplifies the reality of many agri-food subsectors, mostly small and medium enterprises (SME) based, that are still thinking locally without acknowledging the opportunities and threats that the new international scenario entails. This study also has important implications for SMEs. In 2001, between one-quarter and two-fifths of worldwide manufactured exports were goods from SME (OEDC, 2002). In the United States (US), 97 percent of exporters are classified as SME, and two-thirds of all companies that exported in 1999 were firms with fewer than 20 employees. SME exports represented one-third of US exports in 1999 and both, the number of SME exporters and export volume from SME, show an increasing trend (U .8. Small Business Administration, 1999). In the US, for example, 56 percent of firms in the manufacturing industry of frozen fruit, juice, and vegetable manufacturing have 100 or less employees (US. Census Bureau, 2004a) and 50 percent of fruit and vegetable canners have less than 20 employees (US. Census Bureau, 2004b). Thus, an empirical tool to analyze trade determinants and trends of a specific commodity could be used by researchers to address issues important to SMEs. 1.2 Objectives The general objective of this study is to identify trade flow determinants for a specific agri-food commodity (frozen tart cherries) at an 8-10 digit SITC using a gravity model. This study presents three more specific objectives, the first two are focused on relating the statistical characteristic of trade data for a specific agri-food subsector to the theoretical construction of a GM, and the third one is centered on an empirical case. 1. Determine adequate basic independent variables and an appropriate specification of the time-invariant variables in a gravity model for a disaggregated agri-food product. 2. Determine minimum data requirements of the specific product to be analyzed using a gravity model. 3. Estimate the determinants of international trade in frozen tart cherries and evaluate the trade potential among the largest exporters and importers of this product. 1.3 Procedures The primary objective of this study is to identify the trade flow patterns of frozen tart cherries among the world’s largest trading countries using a GM. This model is seen as an empirical tool to provide information about the determinants of international trade flows for specific agri-food goods and specifically FTC. The selection of a GM is based on its success in other empirical studies, although criticisms regarding weak linkages to a theoretical basis are acknowledged. In recent literature the theoretical foundation of the GM has been addressed, and focus has turned towards econometric specification of the model, albeit at an aggregate level. Here efforts are concentrated on analyzing relevant theoretical studies to reconcile trade of a disaggregated product with constructs of the GM.4 4 The most relevant studies in this field are presented in a special section in chapter 11. Once the model is specified, attention will turn to the data. K00 and Karemera (1991) argue that the use of panel data is desirable in international trade analyses, particularly if there is high volatility in trade flows, which often is the case for agri-food products. A particular year may not provide accurate information to evaluate trade flows of a commodity with cross-sectional data only. The availability of a panel data set is not a constraint. Data used in this empirical analysis is based mainly on the EUROSTAT data base; it covers exports and imports of FTC among 17 selected countries from 1993 to 2003. After a descriptive analysis of the data, two statistical issues, common to many agri-food products, must be addressed. The first issue is the reduced number of countries engaged in the trade of frozen tart cherries. The 17 largest exporters and importers of frozen tart chem’es account for more than 85 percent of world trade. Given the limited number of countries, a random selection is illogical. Therefore sample is predetermined and the 17 countries will be included in the analysis. Such a constraint on the number of countries is likely to be more acute when the product analyzed is mainly produced by SME. For instance, according to the US Small Business Administration (1999) 63 percent of the SMEs export to just one country. With a limited sample, time-invariant factors that affect trade between two countries (distance, cultural similarities, border effect, etc.) cannot be considered random and it is important to determine the proper way to include these effects in the analysis. The fixed effects specification of a GM, particularly evaluating it against random effects, has been studied to some extent.5 Nonetheless, it is still not clear how to manage the fixed effects in order to account for all the factors that may determine trade among countries that are time invariant. A second statistical issue is the large number of observations where the value of trade between two countries is zero. This issue is especially important for a disaggregate commodity like FTC. For example in 2004, even in a broad category such as fruit juices and vegetable juices (SITC - 059), the U.S. reported zero export values for almost 20 percent of the trading countries and zero import values for more than 30 percent of the trading countries. As food sectors are defined more specifically the number of zero observations is expected to increase. A review of previous literature shows that some authors have dealt with a subset of these statistical issues (limited sample, censored dependent variable and panel data) in their studies, with differing degrees of depth, but none have incorporated all three, particularly not in the context of a specific product. Table 1 summarizes some approaches of researchers that have touched this topic. Specific contributions of these authors are discussed in greater detail in chapter 11. 5 See Matyas ( 1997), Egger (2000) and Martinez-Zarzoso & Nowak-Lehmann (2003) Table 1. Selected authors and their approach to the GM considering censored dependent variable, unobserved and time invariant variables and panel data. Author Panel Data Unobserved . a variables Limited dependent variable Time-invariant . a vanables Koo & Karamera (1991) ¢ Matyas (1997) Soloaga and Winters (1999) Egger (2000) McPherson (2000) Martinez-Zarzoso (2003) ¢ Carrillo & Li (2002) Cheng & Wall (2004) This study \RRRRS RS «a «as «& R R R R t These two characteristics are closely related to the limited number of countnes included In the study. A GM that has been modified to account for limited sample size, censored dependent variable, and panel data is estimated to evaluate the determinants of international trade in FTC and evaluate the trade potential among the largest exporters and importers of this product. The statistical program STATA (version 8) was used to estimate the model. 1.4 Implications The results of this study will address shortcomings of empirical tools available to analyze trade flow patterns of specific agri-foods and specifically of FTC. Such an analysis can generate insights for policy-makers who address particular agricultural subsectors impacted through strong linkages with international markets. This study will also be relevant for agri-food industries that are in the process of expanding their focus to include global competition. In addition, the procedures and techniques proposed in this study may be the base for future studies, particularly now that researchers have started to focus attention on the econometric issues of GM. The empirical results of this study address just the FTC subsector, nevertheless, the methodology presented can be widely used with other specific products. CHAPTER II 2. Literature review 2.1 Gravity Model and International Trade This section introduces the gravity model and discusses linkages between international trade theory and the gravity model. 2.1.1 Gravity Model Newtonian Physics were first incorporated into human behavior studies by H. Carey in the 18605. Since then, the “gravity equation” has been adopted for use in several social science fields including labor migration, commuting, consumer services, medical services and international trade. Gravity models were first applied to international trade by Tinbergen (1962) and Poyhonen (1963). Tinbergen developed the model to determine the normal or standard pattern of international trade that would prevail among 42 countries in the absence of trade barriers. He justified his model selection by claiming that this model, regardless of its simplicity, captures aspects relevant to the aim of his study. Besides the standard GM, Tinbergen also estimated other models including dummy variables for trade agreements and the presence of a common border among trading countries. Poyhonen (1963) presented, what he called, a tentative trade volume model to analyze trade volumes of ten European countries in a single year. A close connection between the 10 “estimator function” and the “gravitational problem” was highlighted using the analogy that body mass is replaced by the square root of country GDP and transportation cost is replaced by the distance between countries. After the seminal contributions of Tinbergen and Poyhonen additional authors have increasingly used the model in an international trade context, adjusting the model to their specific needs. Linnemann (1966) included population as an additional measure of country size. Other authors used per capita income to express the level of economic development (K00 and Karamera,199l; Carrillo and Li, 2002). Models that include population are often referred to as augmented gravity models (Cheng and Wall, 2004). Other variables that are commonly used in a GM are dummy variables to control for cultural similarity among trade partners, such as language or historical relationships such as colonialism. The basic form of the gravity equation is: Y-Y- _ '1 Til—AD” (I) 1} Where T ,j is the value of exports (imports) from country i to country j, A is a constant, Y is a measure of the size of the economy in each country (i and j), and Dy‘ is the distance between countries. In empirical estimation other variables can also be incorporated into this basic form such as: exchange rates, specialization indices, and dummy variables for the presence of a common border among the trading countries, a common language, or trade agreements. (Deardorff, 1995) 11 Prentice, et al (1998) derived the GM from a one-commodity, two-region theoretical trade model, which they represent in its reduced form as: QE = QE(F: SI, Sx) (2)- Assuming equilibrium between exporting and importing regions, the trade volume (Q5) is determined by transportation costs (F) and supply and demand shifters in both importing and exporting regions (8;, Sx). Given this specification, the authors conclude that “this is a gravity model in terms of explicitly relating trade flows to factors derived from interregional trade models and it is a reduced form of the interregional model” (Prentice, et a1. 1998, p. 8). Thus the authors express the reduced form of a gravity model as: Qijt = Q (Du, Sits Sjt) (3) where: Qijt = Trade volume from region i to region j in year t Dii = Distance from region i to region j (Proxy for transportation cost, substituting for F from equation 3) Sit, Sjt = Vectors of supply and demand shifters (income) in exporting region i and importing region j, respectively, at time t. Using income (Y1, , Yj,) as supply and demand shifters, expressing equation 3 in its multiplicative form, and taking natural logarithms this equation gives the expanded form of the gravity model: 12 1“ Qijt=flo+flt 1n Dijt’l'fizln th+133 111 Yit+8ijt (4) 2.1.2 lntemational Trade Theory and Gravity Model The GM has been praised as probably the most successful trade device of the last twenty five years (Anderson, 1979). More recently, Eichengreen and Irwin (1998, p.33) described the GM as the primary methodology (“workhorse”) for empirical studies of regional integration. Despite success in the empirical analysis of trade patterns, the GM has been a target of critics for its lack of theoretical foundation. Several authors have worked on reconciling international trade theories with a gravity model specification. Here the most relevant theoretical supports developed for GM are presented. Linnemann (1966), one of the first researchers trying to provide some theoretical foundation for the GM, justified its use by claiming that the GM is a reduced form of a four-equation partial equilibrium model of export supply and import demand, where prices are excluded. According to Linnemann, this initial attempt was found to be inconsistent with the multiplicative form of the partial equilibrium model. The first formal theoretical explanations for GM were provided by Anderson (1979). He derived a GM using the properties of the expenditure function in countries where the preference structure is similar for traded goods. An important assumption was the hypothesis of homothetic preferences across regions, i.e. consumers with different 13 income levels in different regions facing the same prices will demand goods in the same proportions. First he assumed Cobb-Douglas and then constant elasticity of substitution (CES) preferences. In both cases he also assumed that goods are differentiated by country of origin. Anderson concluded this study stating that the gravity equation can be derived from the properties of the expenditure system. Bergstrand (1985) built on Linnemann’s work and derived the GM from a general equilibrium framework. He demonstrated that a GM is a reduced form of a partial equilibrium sub-system of a general equilibrium trade model. A GM can be explicitly derived from a general equilibrium model of world trade. The reduced form of this system defines trade flow as a function of available resources, transport cost, and barriers to trade. The main assumptions underling his derivation are perfect substitutability, single factor of production in each country, and products that are differentiated by nations. Deardorff (1995) established the link between the GM and the Heckscher—Ohlin Model (H-O) by deriving the GM from two separate cases of H-O. The first case assumes frictionless trade, where consumers and producers of homogeneous products are indifferent among trading partners. In the second case he considers countries that produce different goods and derived a GM with Cobb-Douglas preferences and then with CBS preferences. Evenett & Keller (1998) analyzed the accuracy of perfect specialization versions of the H-0 model and the increasing returns to scale model and concluded that both supported 14 the GM. In addition, they indicate these models are a limiting case of a model with imperfect specialization. More recently, Anderson and Wincoop (2001) developed a CBS expenditure system to derive an operational gravity model. This derivation was made under the assumptions that all goods are differentiated by place of origin, that each region is specialized in the production of only one good, and that consumers have homothetic preferences. These contributions have provided a solid theoretical foundation for the GM, refuting most of its earlier criticisms. In fact, Jeffrey Frankel (1998, p.2 ) pointed out that the gravity equation passed from a poverty of theoretical foundation to an overwhelming richness. 2.2 Gravity Model and Selected Empirical Applications Use of the GM within international trade covers a wide range of subjects. This section presents empirical cases relevant to this study. The discussion is organized in two parts, estimation procedures and variable selection. 2.2.1 Gravity Model Estimation Procedures Applying a GM, Rose (2002) estimated the effect on international trade of multilateral trade agreements - World Trade Organization (WTO), General Agreement on Trade and Tariffs (GATT) and General System of Preferences (GSP) - among a panel data set of 175 countries over fifty years. Besides the standard variables of a GM, common language, landlocked situation and historical colony links were also included as explanatory variables. The basic GM was estimated using ordinary least squares (OLS), 15 computing standard errors that are robust to clustering by country-pairs. In order to control for global business cycles and other time factors, year-specific fixed effects were included. This setup of country-pair and time panel data was estimated using random effects (GLS) as well as fixed effects (OLS) estimators as robustness checks. Rose filtered his data in several ways; in one he included in the analysis non-members (both trading partners) and the case where just one of the countries is a GATT/WTO member. To estimate this particular case he used a maximum likelihood Tobit estimator. Wall (2000) demonstrated that a gravity model allows for heterogeneity across countries, i.e. fixed effects, and provides unbiased estimates of the volume of trade. He performed his analysis using data from trade among Canadian provinces and U.S. individual states. The dependent variable, ln(l+x,-j,), allowed inclusion of observations where measured trade was zero, but the empirical estimation did not account for the limited dependent variable.6 The model in this study was estimated by general least squares (GLS). Since the model includes fixed effects, time invariant variables can’t be estimated directly. However, a second stage OLS regression on the fixed effects on the time-invariant variables was applied to estimate the impact of time-invariant variables. Carrillo and Li (2002) applied the GM to examine the effects of the Andean Community and Mercosur on both intra-regional and intra-industrial trade in the period 1980-1997. They solved their model using random-effects Tobit left-censoring estimation to account for country-pairs with zero trade. To account for intra-industrial movement, data was 6In an earlier study, McCalum (1995) analyzed the border effect using trade data from U.S. and Canada where all observations with a zero dependent variable were omitted. 16 disaggregated at the three-digit SITC, covering eleven countries. The study does not present a discussion about the author’s choice of random effects over fixed effects. Results indicate that Andean Community preferential trade agreements had a significant effect on capital intensive and differentiated goods, whereas Mercosur preferential trade agreements had a positive effect on capital intensive goods only. 2.2.2 Variable selection While the basic variables of gravity model are well known (i.e. purchasing capacity, output capacity, and transport cost) there is still some discussion on the best proxy variables to capture those effects. In the following section, some relevant studies that addressed this issue are presented. K00 and Karamera (1991) performed a non-nested J A7 test where GDP expressed in current U.S. dollars was evaluated against GDP-PPP8 as an alternative variable to control for income effects in the importer country, and GDP, also expressed in current dollars, was evaluated against agricultural GDP as a proxy variable to capture the output capacity of the exporter country. In both cases, the second alternative was preferred. Nine wheat exporting countries and 34 importing countries between 1981 and 1987 were included in the analysis. Observations where the dependent variable was zero or small were excluded from the analysis. After performing a Hausmann test, the authors decided to use fixed effects. They presented two model specifications, first a covariance model estimated 7 The JA test is a hypothesis testing procedure. 8 GDP expressed in purchasing power parity. 17 with OLS using variables expressed in deviation form, and second an error component model following the Fuller and Battese (1974) procedure. Prentice et al (1998) specified a GM for interregional trade of Canadian pork exports to the United States to identify potential markets for each of the five major Canadian exporting provinces. Dummy variables are included to distinguish high-volume from low-volume market regions. Although this is an implicit recognition of the censored nature of the dependent variable, the estimation procedure does not explicitly consider this issue. Instead the models (one for each of the five provinces) are estimated using a pooling technique that combines the cross-section and time-series data. Regional trade analysis is another trade field where the use of GM is common. Paas (2003) explored international trade flows of countries involved in EU eastward enlargement processes. Included is a discussion about the most accurate selection of GDP as independent variable, i.e. GDP expressed in current currency or expressed in terms of purchasing parity power (GDP-PPP) which captures the income effect. Results indicate that statistical estimations are the best in equations using GDP-PPP. The model (cross sectional) was estimated by OLS using white heteroskedasticity-consistent covariance estimators. Previous results of modeling bilateral trade flows between the countries of the Baltic Sea Region using 1998 data show that the statistical estimations have a better fit in the equations with GDP-PPP (Pass, 2003). Growing empirical literature finds that historical linkages are important determinants of 18 international trade flows (Frankel, Stein and Wei, 1995; Frankel, 1997; Eichengreen and Inrwin, 1998). Lagged bilateral trade flows are significant in determining current trade in a large cross-section of countries, after controlling for income and distance. This indicates that past trade linkages adjust slowly to new conditions. 2.3 Gravity Model and Econometric Tools In the following section, the most relevant econometric topics for a specific good are discussed, emphasizing censored panel data, fixed effects vs. random effects, and discussing the implications of these effects on a Tobit model. 2.3.1 Censored Panel Data Wooldridge (2002) defines panel data as a set of data formed by repeated observation of the same cross-sectional units over time. Panel data generally require extra time and cost, and also creates some econometric challenges. Nerlove (2002) indicates that the problem of latent individual heterogeneity is the central problem in panel data econometrics. Kennedy (2003) highlights the following advantages of panel data: 0 Allows control of heterogeneity of the cross-sectional units. It is to be expected that each cross-sectional unit has some intrinsic and immeasurable characteristics distinguishing it from others. 0 The combination of cross-sectional and time elements generates more variability, and at the same time reduces multicollinearity problems. 19 0 Panel data can be used to identify the effect of time-varying variables (e. g. technology) and cross-sectional variables (e. g. economies of scale) simultaneously. 0 Panel data allow better analysis of dynamic adjustments. Some dependent variables are limited in their range. The econometric analysis of data sets with limited dependent variable requires special treatment. Maddala (1983) distinguishes two types of limited dependent variables, truncated and censored. A truncated regression is one where there is not any observation for either the dependent or the independent variables when the value of the dependent variable is above (or below) a certain threshold. (e. g. negative-income tax experiment). In a censored regression the data of explanatory variables exists for all observations, but there is group of observations where the value of the dependent variable is missing (non- observable). Examples of this type of sample are reservation wage or demand for durable goods. It is not difficult to imagine an economic agent maximizing utility at a comer solution. (i.e. y = 0). Wooldridge (2002) calls these kinds of models “comer solution models”. Estimation of censored regression using OLS will generate biased and inconsistent parameter estimates. A preferred estimation procedure is maximum-likelihood (Pindyck and Rubinfeld, 1998) where the likelihood estimation is the product of expressions for the probability of obtaining each observation. A common model to estimate censored 20 regression is the Tobit model, named after James Tobin, who was the first to analyze issues raised by censored data. The Tobit model is defined as follows: yi = ,6x i + gi if the dependent variable is > 0 (5) y, = 0 otherwise where y is the censored dependent variable, x is a vector of explanatory variables, ,8 is a set of parameter to estimate, and a is the error term. In the case of a Tobit model, maximum likelihood will be as defined above for dependent variable observations greater than zero. The likelihood for observations with censored dependent variables equals the probability of getting an observation below a threshold (Kennedy, 2003). Censored panel data include a set of observations for the same cross-sectional units over a period of time in which the dependent variable of a subset is censored. Wooldridge (2002) presents the basic form of censored panel data as: yit = max(0 aflxit +811), t: 1,2,... T (6) a), | xi, ~ Normal (0, oz) where the variables and parameters are defined as in equation 5. In this case the cross- sectional unit i is observed over time t. Since the model does not maintain strict exogeneity of x", lagged variables can be incorporated in the model. In addition the error terms, a), are allowed to be serially dependent. In practical terms, this means that xi, can contain any conditional variables such as interaction of time dummies with time-constant or time-varying variables. Both characteristics are relevant to this study. Maximum 21 likelihood can be used to estimate the model depicted in equation 6; however, to control for serial correlation across time a robust variance matrix is needed. 2.3.2 Fixed vs. Random Effects An inherent issue raised by the use of panel data is the decision of how to treat unobservable variables, i.e. fixed or random effects approach. Wooldridge (2002) presents the following basic unobserved effects model yitzflxit+ci+£it t=l,2,...T (7) where x,-, is 1x K and contains explanatory variables that for propose of this analysis can be classified in three categories: time-invariant variables (e. g. distance between countries), variables that change over time but not among cross-section units (e. g. international exchange rates), and variables that change over time as well as from one cross-sectional unit to another. (e. g. GDP, crop yield). The variable c1, captures the unobservable effects; in the case of the gravity model cross-sectional heterogeneity such as variation among pairs of trading countries (e. g. UK — U.S. and Canada-Germany). Finally, a”, is the idiosyncratic error. An important consideration in this unobservable effects model is whether to treat c,-, as a random effect or as a fixed effect. Random effects implies that c. is considered a random variable and it becomes part of the error term. The error term is a composite formed by two parts: one, in the case of a gravity model, is the “random intercept” of a county-pair, and the second is the normal error term. Kennedy (2003) shows that in a random effects 22 model not all of the off-diagonal elements of the variance-covariance matrix of this composite error term are zero (i.e. non-spherical), which could indicate a certain degree of autocorrelation. On the other hand, considering c. as fixed implies that it is a parameter that has to be estimated for each cross-sectional unit, generally by adding a series of dummy variables. Fixed effects estimation has two major drawbacks: a) the use of a dummy variable for each cross-sectional unit generates a loss in degrees of freedom; b) time-invariant variables are not identified and must be discarded from the equation. Wooldridge (2002) states that the key issue that determines if c,~ should be considered random or fixed is whether or not it is correlated with the observed variables (i.e. x”). If c,~ is uncorrelated, random effects estimation is more appropriate. McPherson (2000), using a GM, argued for random effects assuming that the differences between cross-sectional units of his study were randomly distributed. These cross-section units, the pair of potential trading partners, were postulated to be different in terms of economic, social, political and geographic conditions. In his model he used a random- effects Tobit (weighted maximum likelihood) estimation procedure. Later the same author (2001) estimated a gravity model to test Linder’s hypothesis in developing countries by using fixed effects. In this case he argues that fixed effects can be used if the sample represents a relatively large proportion of the population and if unobservable factors that differentiate each country-pair are best characterized as parametric shifts of the regression function. 23 2.3.3 Panel Data, Tobit Model and Fixed Effects Fixed effects in non-linear models, such as Tobit, present difficult statistical problems, particularly if the period of time covered by the sample is small. One of the most critical issues with fixed effects and a limited dependent variable is the “incidental parameters” problem. Heckman and Macurdy (1980) argue that if the number of periods observed for each cross-sectional unit is small (eight in their case) it is not possible to consistently estimate fixed effects. In these cases the structural parameters are themselves a function of the fixed effects, thus inconsistency in the fixed effects is transmitted to the estimated structural parameters. Maddala (1983) expands Heckman and Macurdy’s arguments using the following non-linear model (Tobit): y; =01 +51%: +51? 8” ~1N(0’0) (83) y. =y2; if y; >0 y,-, = 0 otherwise In this case it is not possible to devise estimators of ,6 and a that are not independent of a (fixed effects), particularly if T (number of observations per i) is small. Heckman and Macurdy concluded that the inconsistency of the estimates might not be a serious problem if there are no lagged dependent variables. This model was estimated for a Probit model. Results indicate a similar procedure could be applied to a Tobit model, which is expected to have even better behavior, since it combines the linear regression model with the Probit model. 24 Maddala presents an alternative to the “incidental parameter” problem originally presented by Heckman and Macurdy (1980), he derives model estimators by iterative methods. Let d1: =1 if y; >0 ; al,, =0 otherwise. Then the likelihood function is defined: LogL = 20 _d,.,)log¢(:“_‘/’fl)+zd,{-llogaz —L, 5 and the degree of censoring is over 40 percent. The variables included in a basic GM are likely to have one of the three distributions studied by Greene (GPD, AR(l); Normal). Thus, a GM using panel data for a specific commodity can be estimated by MLE, with some confidence if data covers a period of time longer than five years, more than ten countries are included in the analysis, the proportion of zeros in the dependent variable is greater than 40 percent, and the explanatory variables included in the model follow a Normal, Chi-squared or AR(l) distribution. 3.3 Empirical Methods for Frozen Tart Cherry Trade Analysis Prior to explaining the empirical methods used in this study, a brief discussion of the data utilized follows showing basic statistics and trends of the values of the independent variables. 9 In the Monte Carlo simulation used by Greene, he set N=1000. So, N has to be at least equal to 1000 to have consistent results. 36 3.3.1 Data The data set for this analysis has 2992 observations [i X (i-l) XT], over a period of 11 years (1993 — 2003). It is unbalanced panel data, which is common in a GM setup, since no country trades with itself. Almost 63 percent of the observations are zero (non-trade among the countries). Thus the data set fits into the minimum requirements to estimate a GM model with fixed effects. The countries included in the analysis are Australia, Austria, Belgium, Canada, Denmark, France, Germany, Greece, Hungary, Italy, Netherlands, Japan, Mexico, Poland, Turkey, United Kingdom and United States. These seventeen countries are the largest exporters and/or importers of FTC. The volume traded among these countries in 2003 accounted for more than 85 percent of world trade volume. The specific trade name of the item analyzed in this study is “Sour Cherries Prunus cerasus, Whether or not Boiled or Steamed, Frozen, not Containing Sugar or Other Sweetening Matter” The value of FTC trade (dependent variable) is measured in thousands of current U.S. dollars. This information was gathered from different sources. The most important source was Data Service and Information (DSI-Eurostat) for the European countries and their trade partners. It accounts for almost 80 percent of the observations. Other data sources are STAT-USA (U.S. and trade partners), Canadian Statistic Service (Canada and trade partners), and OECD- lntemational Trade Data Base (Australia, Hungary, Poland, Mexico and Turkey). 37 With one exception, data was available at a ten-digit Standard International Trade Classification (SITC) level of disaggregation. The data from OECD was at a different level of aggregation (six-digit SITC), so it was necessary to estimate the portion of this item that corresponded to FTC. This estimation was based on the data reported from other sources. Less than two percent of the observations were estimated using this procedure, since most of the observations were zero. In the cases where the data reported from a country did not coincide with that reported by the trading partner, a simple average was taken from the two observations. Explanatory variables included in the model are current gross domestic product per capita at purchasing power parity (GDP-PPP),lo agricultural value-added,ll population, and lagged total yearly export/import. Data for the two first variables were obtained from the World Bank Development Database. Distance, measured as kilometers between capitals cities, was obtained from the Great Circle Distances Between Capital Cities web page. Lagged yearly export/import data is the aggregation of total exports or imports of FTC in the previous year. Table 3 shows values for the year 2003 of the three time-variant variables: total agriculture value-added, GDP-PPP and population. As an indicator of trends, the average rate of growth between 1993 and 2003 for each variable is also included in the table. The '0 Purchasing power parity (PPP), an international dollar has the same purchasing power over GDP as a U.S. dollar has in the United States. For example, GDP per capita in China is about $US 1,500, while on a PPP basis, it is about $US 6,200. ” Agricultural value-added corresponds to the lntemational Standard Industrial Classification (ISIC), divisions l to 5, and includes forestry, hunting, and fishing, as well as cultivation of crops and livestock production. Value-added is the net output of a sector after adding up all outputs and subtracting intermediate inputs. It is calculated without making deductions for depreciation of fabricated assets or depletion and degradation of natural resources. (World Bank, Development Indicators) 38 agricultural output of the U.S. is three-times larger than the second largest agricultural output from the selected countries. Belgium and Netherlands have a relatively small agricultural output, however they are important players in the FTC market. Japan, on the other hand, has the second largest agricultural value-added but its role in FTC markets is limited as buyer and it has a low share of the worldwide demand. The negative rate of growth in the agriculture sectors of Mexico, Poland and Turkey, can be explained in part by the dynamism of other economic sectors in those countries that compete for resources with the agricultural sector. As countries develop agriculture normally becomes a smaller part of the economy. Table 3. Value of continuous independent variables for 2003 and average rate of growth between 1993 and 2003. Agriculture value-added GDP per capita PPP Population Countries _ Current Avg. rate of International Avg. rate of Thousands Avg. rate of Mrllron USS growth dollars growth growth AUT 12,076 0.5% 29,632 4.3% 19,881 1.2% AUS 5,327 0.7% 30,094 3.6% 8,090 0.2% BEL 3,565 -1 . 1% 28,335 3.5% 10,376 0.3% CAN 15,496 0.4% 30,677 4.1% 31,630 0.9% DEN 3,875 -1.7% 31,465 3.7% 5,387 0.4% FRA 42,635 1.0% 27,677 3.4% 59,762 0.4% GER 24,707 0.1% 27,756 3.1% 82,541 0.2% GRE 10,442 1.5% 19,954 4.8% 1 1,033 0.5% HUN 2,367 0.0% 14,584 5.4% 10,128 -0.2% ITA 34,856 0.8% 27,1 19 3.3% 57,646 0.1% JAP 46,053 -7.2% 27,967 2.7% 127,573 0.2% MEX 22,981 -1.5% 9,168 2.7% 102,291 1.5% NED 9,156 -2.0% 29,371 3.9% 16,222 0.6% POL 5,280 -1.6% 11,379 6.0% 38,196 -0. 1% TUR 26,931 -3 .6% 6,772 2.7% 70,712 1.7% UK 15,487 -0.6% 27,147 4.1% 59,329 0.2% US 158.947 3.1% 37,562 3.6% 290,810 1.1% Source: World Bank Development Indicators 39 The difference in agricultural output between U.S. and the rest of the countries is less evident when it comes to GDP — PPP per capita. Net exporting countries such as Turkey, Poland and Hungary, have a GDP — PPP per capita below the average of the countries included in the analysis, but Poland and Hungary have the highest average rate of growth of GDP - PPP per capita. The values for this variable for FTC net importing countries (Germany, US, UK, Japan) are above the average, and have large rates of grth for this variable. The negative rate of population growth in Hungary and Poland could be attributed to the migration rates from those countries to Western Europe. The relatively high and positive population rate of growth of Australia, Mexico and U.S. can be attributed to migration trends and/or high rate of births. Distance (measured in kilometers between capitals) presents a large range of variation. The greatest distance was found between Australia and the United Kingdom (16,973 Km. or 27,309 miles ). On the other hand, the distance between the capital of Belgium and Netherlands is just 173 Km. (278 miles). Annex 2 shows the distance between all 17 capitals. This table also presents information about the countries that share borders as well as those that have signed a free trade agreement. 3.3.2 Empirical Model The value of trade is estimated using a Tobit model with fixed effects by MLE (based on equation 12) which takes the form: In ya, = maxlo, p0 +111ow + wmy + {ya + m. + am) (13) z= 1,2,... T; 8,, l xw, r, t, ~ Normal ( 0, oz) 40 where: yij, = trade value expressed in current U.S. dollars from region i to region j in the period t expressed in current dollars. '2 xij, is a vector that contains the following variables expressed in logarithms: ag,~, = agricultural value-added expressed in thousands of current $U.S. in exporting country i in period t px,-, = population expressed in thousands in exporting country i in the period t ng, = per capita GDP at purchasing power parity expressed in thousands of international $U.S. of importing country j in period t pmj, = population in thousands in importing country j in period t lxj,- 1 = value of total FTC exports expressed in thousands of current $U.S. of country i in period t-l 1m”- 1 = value of total FTC imports expressed in thousands of current $U.S. of country j in period t-l f0- : vector (1x (i X j-1)/2 ) that contains a dummy variable for each pair of countries regardless of whether the country is importer or exporter. The number of variables contained in fij is] 36. t, = vector (1 x T) of dummy variables for each year, where 1993 is the base year. '2 A unit was added to ya, in order to be able to estimate the its logarithm when ya, = 0. This monotonic transformation does not affect the estimates results, but requires special consideration at the moment of estimating the expected value of y, (Wooldridge, 2005). 41 The classic GM uses total GPD as a proxy for output capacity in the exporting country, which is appropriate for studies using aggregated total export data. In the case of a single commodity, total GDP would overestimate output capacity of the country for that particular commodity. A better proxy to capture the output capacity for a specific commodity is the portion of GDP for the sector that is related to the good under consideration. In the case of this study FTC is part of agricultural GDP, and so agricultural GDP is used as a proxy for the output capacity of the exporting country. The parameter for this variable is expected to be positive. A larger agricultural sector is expected to have a significant positive influence (synergy effect) on a subsector such as FTC. Population of the exporting country was included in the model to account for the capacity of the country to consume the good domestically. It is assumed that the larger the population the less the country will export. Thus the significance for this parameter is expected to be negative. Income effect for the importing country in a GM is generally controlled by including total GDP. Given the countries under analysis have important differences in terms of economy size, living cost and income per capita, a variable with stronger explanatory power for income effect in the importing country is per capita GDP expressed in purchasing power parity (PPP). This analysis is consistent with the findings of Paas (2003) who formally tested the adequacy of the variables concluding that GDP expressed in term of PPP is more suitable. A positive sign for this parameter is expected, meaning that an increase in 42 income in the importing country would lead to an increase of FTC imports. Since FTC is a processed good it is also expected that the parameter for this variable would be one or greater. Population in the exporting country is included to complement per capita GPD-PPP, since the later controls for the income effect of one individual in the importer country but does not consider the size of that economy. By including population, total purchasing capacity of the importing country is captured. Once initially processed, FTCs can be stored for long periods of time. Therefore, the volume imported or exported in previous years can have a direct impact on trade in the current year. On the other hand, it is logical to believe that once a trade relationship has been established between two countries it is likely to continue over time. The lag variables, Ix,” and 1m”- 1, are included in the analysis to account for these occurrences. Signs for these parameters are ambiguous since storage could have a negative influence, while trade relationships could have a positive effect. Equation 13 will be estimated by maximizing the likelihood function: I N T logL = Z Zlogf(yij,,flo + lnxw8+fija,0) t=l t=l (14) 43 Where f(. . .) is the density firnction of the Tobit model containing the parameters to be estimated, 0 is a vector of disturbance standard deviation. Equation 14 usually does not have an explicit solution but must be solved iteratively. Inclusion of a set of dummy variables (1) may make the maximization viable, but proliferation of dummy variables may create a serious constraint in the procedure due to “nuisance parameters” (Greene, 2004b). In the case of FTC, it would be necessary to create 272 parameters, but since fixed effects are considered to be the same regardless of whether a country exports or imports, the number of parameters added to the model was reduced to 136. This characteristic of GM drastically reduces the “nuisance parameters problem”, i.e. the proliferation of dummy variables for each fixed effects to be controlled. Additionally, this characteristic makes the maximization of the function feasible by “brute force”, namely by adding a dummy variable for each pair of countries. The inclusion of fixed effects creates another empirical issue: all time-invariant variables (distance and the dummies the border and free trade agreement) effects will be aggregated and captured by the fixed effects dummy variables. The individual effect of these variables on trade can not be observed from equation 13. Following the procedure of Martinez-Zarzoso and Nowak-Lehmann (2002) a two-step procedure is used to estimate the effect of time-invariant variables. Country-pair effects (a) are regressed on independent time-invariant variables. Hence, based on empirical results from equation 13 the following equation is estimated using OLS with robust standard errors. at] =50+511ndij +62b01j +63flit+gfi (15) where: 44 al,-j = Distance from country i to country j bog- = presence common border between country i and j ftij, = existence of free trade agreement between country i and j in the period t. Distance is a proxy for transport cost and is expected to have a negative effect on trade value. In this case the effect of distance will be reflected on the fixed effects, which have a direct and positive effect on trade value. The presence of a common border and free trade agreement are two dummy variables commonly included in GM. The presence of either of these two variables is expected to generate an increase in trade, i.e. parameter signs would be positive. Free trade agreements relevant to this study are NAFTA (U .S., Canada, Mexico), EU (Belgium, Denmark, France, Germany, Greece, Italy, Netherlands, UK), Eastern Europe Free Trade Agreements (Hungary, Poland, Turkey) and EU and Mexico. All the agreements were established prior to 1993 with the exception of EU — Mexico that was signed in 2000. Variables included in equation 13 were selected based on previous studies that used log- likelihood ratio (LR) tests and Akaike’s Information Criterion (AIC). The properties of LR tests and AIC for specification of non-linear models used in this study have not been completely specified. Therefore, results of LR test are included as one method to support the selection of variables and not the only criterion. 45 The LR test is applied to test the null hypothesis that restrictions imposed on the original model do not apply. Five different alternatives to Equation 13 were tested. Alternatives include models where either variables were omitted or alternative proxies were substituted for those variables in Equation 13. The six alternative models that were evaluated are: 1) Including total GDP measured in current U.S. dollars of importing and exporting countries substituting for per capita income at purchasing power parity and population. 2) Including total GDP at purchasing power parity of importing and exporting countries substituting for per capita income at purchasing power parity and population. 3) Excluding the lagged value of total imports and exports. 4) Excluding country-pair fixed effects and including distance and dummy variables for border and PTA (random effects). 5) Excluding time effects. 6) Alternative specification of fixed effects by export or import country. Models 1 and 2 used alternative proxies to capture income effects while models 3-5 test more parsimonious specifications of the original model and model 6 tests a different set of fixed effects variables. The interest of the LR test is to compare explanatory power of the variables included in the alternative models with the final set of variables selected for the study. Likelihood ratio tests are used to compare two models where the simpler model (alternative model) is 46 a special case of the more complex model (original model), i.e., “nested". The test statistic is distributed as a chi-squared random variable. The Akaike Information Criterion (AIC) of each model is also included and compared; here the model that presents the smaller estimates is likely to have a stronger explanatory power. Table 4. Log-likelihood ratio test and AIC estimates for the alternative models Alternative Model variables a Degrees of No. LR chi2 AIC Included} Excluded C freedom I gdpi, gdpj gpi, agj, pm), pxj 121 6204*" 8406.77 2 gdp_p,-,gdp_pj gp;,agj;pm,-, pxj 121 6919*" 8413.91 3 1mm, Ix,“ 121 383.01*** 8727.73 4 fij 8 959.32**"‘ 9078.04 5 tij 113 21.26 " 8349.99 6 fofmJ- in 49 632.1?” 8832.85 Unrestricted model 123 n/a 8348.72 a The number of observation for all the models: 2310 b gdpi, gdpj, gdp_p,-, gdp_pj: GPD in current dollars and GDP in PPP of exporting and importing countries respectively. fxi , fnrj: fixed effects for exporting and importing countries. C gpi, agj-n pmi, pxj , (mu- 1, Ix,“ : GPD in current dollars, population and one year export/import lagged, respectively. "*, ** denotes a significant test statistic at the 0.01 and 0.05 level respectively. Significant indicates that the alternative model does not have a better fit than the original. From Table 4 we can see that neither of the two alternatives for capturing the income effects of the exporting country and the output capacity (model 1 and 2) had a better performance than the original model. Model 1 includes total GDP in current dollars for both trade countries as alternative proxy variables. Total GDP in the importing country is 47 substituted for per capita GDP-PPP and population, and total GPD in the exporting country is substituted for agricultural-GDP and population. The results of the LR-test and AIC indicate that the original model has a stronger explanatory power; therefore, the hypothesis that the alternative model has a stronger explanatory power is rejected. The second alternative model presents total GPD-PPP for both trader countries as a proxy for agricultural GDP and population in the case of exporting country and GDP-PPP per capita and population in the case of importing country. According to the LR—test the unrestricted model has a better performance and it is preferred to the alternative. The AIC also presents a smaller number for the original model in both cases. Models 3, 4 and 5 exclude some variables from the original model. The LR test indicates whether coefficients of the excluded variables are equal to zero or not. Results do not show coefficient value of the excluded variables equal to zero for any of the cases, therefore, the null hypothesis that restrictions do not apply is rejected. The total value of exports/imports from a previous year contributes significantly to explain the trade flow of FTC. To omit fixed effects from the model would reduce the explanatory power of the model, as well as leave out business cycle dummy variables. Model 6 presents an alternative set of fixed effects variables. Here it is assumed that fixed effects are more related to individual characteristics of the exporter and importer rather than to the specific trade characteristics between two countries. These individual characteristics could be factors such as trade facilities (roads, harbors, airports, etc.) or trade services (telecommunication, market information services, etc.). According to the 48 LR test it was possible to establish that the original model is preferred to the alternatives. The AIC also indicates the unrestricted model has a better performance. 49 CHAPTER IV 4. Results This chapter is divided into two parts. The first part presents the result of the GM estimation, including its marginal effects analysis. The second part assesses the FTC export potential of the countries included in this study. 4.1 Model Estimation Results Estimation results for equation 13 are reported in Table 5. The model was run with 2,309 observations, 31 country pairs were excluded from the estimation since they did not report trade during any year of analysis, and consequently it was not possible to estimate a fixed effects parameter for these pairs. For example, Japan and Poland did not report FTC trade during any year included in the analysis. The degree of censoring of the data set after dropping those observations was 53 percent. The coefficients presented in Table 5 cannot be read directly as elasticities; however, the sign and significance of the coefficients indicate the direction of impacts. The overall significance of the model is high, since the probability of non-significance of the Chi-square statistic of the model is very low. The explanatory power of the model is moderate but relevant. The pseudo R2 0.192 is near a range that is considered relevant for non-linear models (0.205 - 0.259). 50 Table 5. Tobit estimates with fixed effects for trade value Variable 8P} WW 08:" PM [xii-l [mu-I Pair AUT AUT AUT AUT AUT AUT AUT AUT AUT AUT AUT AUT AUT AUS AUS AUS AUS AUS AUS AUS AUS AUS AUS AUS AUS BEL BEL BEL BEL BEL BEL BEL BEL BEL BEL BEL BEL BEL Cosnt __se BEL CAN DEN FRA GER HUN JAP MEX NED POL TUR UK US BEL DEN FRA GER GRE HUN ITA NED TUR UK US CAN DEN F RA GER GRE HUN ITA MEX NED POL TUR UK US Coeff. 6.74 3.31 1.97 1.67 0.87 1.67 Coeff. 12.82 16.73 13.55 5.37 0.34 l 1.84 17.59 21.07 12.43 10.11 4.45 6.32 9.14 22.36 13.57 7.42 20.69 12.46 24.29 17.04 20.84 20.18 -8.64 5.09 4.72 9.69 26.47 20.41 18.87 24.07 24.10 13.90 5.78 23.53 19.54 16.72 20.23 8.94 -147.85 6.28 tilt tit iii *#* a 0’. . .00 fitt it. .3. ti! it! it. 0“ Ct. it! it. tit tit t. fitt 0‘. ti‘ it! it? Number of obs LR ch12(111) Prob > ch12 Pseudo R2 Log likelihood Left-censored observations Model Information Country — pair fixed effects Pair CAN CAN CAN CAN CAN CAN CAN CAN CAN CAN DEN DEN DEN DEN DEN DEN DEN DEN FRA ’FRA FRA FRA F RA FRA F RA FRA FRA FRA GER GER GER GER GER GER GER GER GRE GRE FRA GER HUN ITA GER GRE NED TUR UK US GER GRE HUN ITA JAP NED POL TUR UK US GRE HUN ITA NED TUR UK US HUN ITA 1.68 2.19 10.90 -2.01 7.74 10.19 11.08 4.39 7.57 12.57 13.40 19.49 10.11 21.16 18.96 11.92 18.53 0.13 12.84 9.48 12.19 14.81 9.04 16.38 13.68 10.62 15.01 2.37 19.06 19.03 13.89 16.85 17.15 13.17 12.00 -0.38 6.24 14.17 Coeff. 0.0 .‘i it! ‘tt it it. #0 .1 Oi! ti. ‘0 t.‘ Ct. . a it GRE GRE GRE GRE HUN HUN HUN HUN HUN HUN HUN ITA ITA ITA ITA ITA ITA 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 Pair NED TUR NED TUR UK US NED 2,309 1,926 .000 0.192 -4,066.84 1,23 1 Coeff. 13.97 ‘ 15.02 " 19.39 "" 17.81 ‘” 6.96 1.95 15.86 “ 20.39 ”" 26.14 "‘"‘ 17.35 ‘” -1.88 -1.83 16.17 "”" 14.05 ” 11.40 ‘ 1.65 -2.58 ~0.57 6.09 3.60 17.18 ‘” 15.95 " 17.40 ”"‘ 4.96 14.88 " -5.66 14.20 '” -2.89 -0.64 Business cycle effect -1.12 -0.27 -0.50 -0.02 -1. 17 -0.85 -1.06 -2.24 ”‘ -0.54 -2.33 ‘" Significant: *" at 1% , " at 5%, "' at 10% 51 Per capita GPD-PPP in the importing country has a positive effect on FTC trade (at one percent significance). As per capita incomes rise, trade increases. Likewise, as population increases, trade increases, although this coefficient is significant at just the 10 percent level. Both GDP-PPP and population are measures of demand in the importing country and their effects are as expected. The size of agricultural output in the exporting country shows a positive and significant (at 1 percent) effect on FTC trade. Population in the exporting country could not be established as a source of significant variation in the FTC flow, therefore the absorption capacity (domestic demand) for FTC of the exporter shows the export-oriented nature of the good analyzed. The lagged value of trade was significant for both exporting and importing countries with a positive effect on subsequent valued traded. So, the positive impact of the trade relationship built over time outweighs any negative impacts on FTC trade such as storage ability. Looking at the results for country-pairl3 fixed effects some points can be highlighted. In general the fixed effects (distance, border, PTA and any other characteristics that are not likely to change over time, such as historical links) of the Western European countries present a positive and significant influence over trade. Therefore, Western European countries in general have been able create strong trade relationships among themselves based on their geographic location and trade agreements. For example, more than 60 percent of France’s FTC demand is covered by Western European countries. Net exporting countries such as Hungary, Turkey and especially Poland have also developed solid trade linkages with their main trade partners, despite the lack of formal trade ‘3 None of the countries can be identified as exporters or importers. The analysis is centered on the trade relationship. 52 agreements. Poland’s exports to Germany, Belgium and Netherlands have had a consistent and incremental upward trend from 1993 to 2003. On the other hand the U.S., which plays an important role in the FTC world market, has a positive and significant trade link only with Canada (10 percent), its historical trade partner. This suggests that the U.S. has built its FTC trade relationships on time variant characteristics (such as weather) rather than on fixed factors (such as FTA, been Canada the exception). In order to be able to discuss effects of distance, border and F TA on aggregated FTC trade, a second step estimation was undertaken (See details in Chapter III). Results are presented in Table 6. Even though this second step procedure has its limitations, particularly when assessing the effects of the dependent variable directly on trade, it provides a better understanding of the individual effects of time-invariant variables on trade flow analysis. All variables included in this model were found to be significant at the one percent level. Again, a direct interpretation of the coefficients is not possible but the effect of each variable on the individual fixed effects and consequently on FTC trade is as expected. Distance, as was mentioned previously, is a proxy for transportation costs. The negative sign of this coefficient implies that the greater the distance between two countries the lower the level of trade between them. The positive and significant signs of the coefficients of border and PTA seem to indicate that their presence positively affects the country-pair fixed effects and in turn FTC trade among the pair of countries. Based on relative magnitude of the estimated coefficients the effect of a FTA between two countries will have nearly the same effect as if those countries shared a common border. A common border does not just proxy geographic location, but also could imply cultural 53 similitude, historical linkages, migration flows, etc. These factors, as a set, may have a positive influence on trade. Table 6. OLS estimation individual effects Dependent variable: Individual effects from Eq. 16 Independent variables Coeff. Distance -1.281“‘*"' Border 3 .938" * F TA 3 .584" * * Constant 20.59?" N 105 R-squared 0.22 Root MSE 6.06 "" Significant at 1% 4.2 Marginal Effects The marginal effect of any of the variables on trade value depends on the probability that trade value is greater than zero (see table 5). Tobit estimates are multiplied by a factor (for details see Wooldridge, 2002 p. 522) and thus obtaining the marginal effect of the explanatory variables on the conditional expectation. There are two conditional expectations of interest: 1) Expected trade value given trade existence, (i.e. trade value > 0) and the explanatory variables E(Trade value | Trade value > 0, x), and 2) Expected trade value given only the set of explanatory variables E(Trade value I x). The first estimate provides information on the expected trade value for given values of x in the subset of observations where trade was observed. Expected trade value, given trade 54 occurs, is more adequate for a profit assessment, since it considers the real demand. The second expected value calculated is not conditional to existence of trade and uses the whole set of observations to predict value of trade. The latter case is relevant for demand analysis, where the possibility exists that the dependent variable becomes greater than zero (i.e.: a change demand equal zero to some positive value) without a change in independent variables. The factor used to estimate the marginal effects is a value strictly between zero and one and is based on these two expectations (Wooldridge, 2000, p. 567). Table 7 presents results of the marginal effect conditional on value of trade greater than zero and Table 8 results of the marginal effects conditional only on the values of the explanatory variables, i.e. this estimation include the expected value of trade for those country-pair observations of a particular year where the value of trade observed was zero. A key difference between these two calculations is that the second accounts for the fact that a country pair that has not traded in the past may chose to trade if the conditions change. Significance levels as well as the direction of the effects in both cases are very similar to the original model estimation. The discussion below will focus on values reported in Table 8. According to results, FTC trade value is income elastic: a one percent increase in income could represent an expected 3.61 percent increase in the value of FTC traded (using the estimates not conditional on trade value greater than zero), holding the rest of the variables constant. That is consistent with the literature that finds processed goods generally have income elasticities greater than one. (Ash, 2004; Babula and Corey, 2005). 55 Thus the natural target markets for exporters that are expanding their production should be countries where the income growth is constant and solid. It is important to notice that any income decrease from the trade partner would have serious negative consequences in the volume of exports. The fact that the income elasticity is greater for the 19914 “unconditiona analysis could be because the market is considered larger for the same amount of product. An increase in population of the importing country suggests the generation of increased imports of FTC. A one percent increase in population in a given country would mean an expected increase of 1.78 percent increase FTC demand, ceteris paribus.ls However, it is important to consider that some of the countries included in the study, particularly European countries, have less than a 1 percent rate of growth. In addition, the income effect is likely to be the predominant demand shifter. Size of the agricultural sector in the exporting country has a positive effect on exports. This effect is slightly larger when analysis includes those country-pairs that did not trade. Thus Mexico, a country that exports FTC to a few countries, might conceivably engage in trade with other countries if its agricultural sector grew and generated the proper conditions. The growth of exports will be almost proportional to the rate of grth of the '4 The expected value is still conditional on the values of the explanatory variables. '5 Mathematically population is actually considered twice as an independent variable twice in the model, since per capita GPD-PPP also incorporates population. Performing some mathematical operations with logarithms it is not difficult to show that population could have a negative coefficient (from table 7: -1 .3 1). However, since all the variables, including GDP-PPP are held constant, an increase in population would also imply a reduction in per capita income, and given the income elasticity of FTC, this would lead to a decrease in demand. 56 agricultural sector, i.e. an increase of one percent in the agricultural GDP would potentially generate a 1.05 percent increase in FTC trade. The effect of the total value of FTC exported one lagged year has a larger effect on the value of FTC traded in the current year than the total amount imported in the past year. FTC markets are export driven markets and the capabilities that a country may develop in order to promote exports has a stronger impact on FTC flows than the development of domestic markets in importing countries. Nonetheless, the combination of both will have a most favorable impact. The loss of market share due to a bad tart cherry production year, such as what happened with U.S. in 2002, could also have a negative impact on the next year’s exports. In fact, it is expected that for every percent of market loss in a given country, the following year the demand will decrease by almost half percent. The magnitudes of country-pair fixed effects (again reported in Tables 7 and 8) reveal that the European countries have strengthened their trade relationships based on the time- invariant factors. Most of the European countries will trade amongst themselves between 15 and 24 percent more than the expected value of trade among those countries that have not engaged in trade of FTC.l6 Poland, Hungary and Turkey have consolidated their trade links in Europe rather than with countries in other regions. It is expected that the trade of these three countries with Western Europe would be between 10 and 17 percent greater than the expected value of the countries that have not engaged in trade. ’6 The “unconditional” expectation assigns a expected value of trade for those countries that have not shown trade. The expected value is based on their values of x.)- 57 Table 7. Marginal effects of Tobit estimates conditional on trade value greater that zero Variable dy/dx gpj 2.58 *“ pm} 1.27 “ 0g: 0.75 ""‘ px; 0.64 lxm 0.33 m 1m,“ 0.64 m Country — pair fixed effects Pair dy/dx Pair dy/dx Pair Coeff. AUT BEL 8.39 " CAN FRA 0.70 GRE NED 9.45 " AUT CAN 12.08 "" CAN GER 0.93 GRE POL 10.44 ” AUT DEN 9.06 " CAN HUN 6.71 GRE TUR 14.69 ‘” AUT FRA 2.64 CAN ITA -0.70 GRE UK 13.13 ”" AUT GER 0.13 CAN JAP 4.23 HUN ITA 3.68 AUT HUN 7.52 CAN NED 6.12 HUN JAP 0.82 AUT JAP 12.92 “"‘ CAN POL 6.87 " HUN MEX 11.24 ” AUT MEX 16.35 ”" CAN TUR 2.07 HUN NED 15.67 ”‘ AUT NED 8.04 " CAN UK 4.10 HUN POL 21.37 "" AUT POL 6.06 CAN US 8.17 "‘ HUN TUR 12.68 ‘” AUT TUR 2.10 DEN FRA 8.92 ‘ HUN US -0.66 AUT UK 3.25 DEN GER 14.79 "" ITA JAP -0.64 AUT US 5.27 DEN GRE 6.06 ITA NED 11.54 "W AUS BEL 17.62 *” DEN NED 16.44 "" ITA POL 9.53 " AUS DEN 9.08 DEN POL 14.26 ‘“ ITA TUR 7.14 ‘ AUS FRA 4.00 DEN TUR 7.59 ‘ ITA UK 0.68 AUS GER 15.97 ‘" DEN UK 13.84 "‘ ITA US -0.88 AUS GRE 8.07 DEN US 0.05 JAP NED -0.21 AUS HUN 19.54 ”" FRA GER 8.41 ‘ JAP US 3.10 AUS ITA 12.38 "” FRA GRE 5.54 MEX US 1.64 AUS NED 16.11 "" FRA HUN 7.83 " NED POL 12.52 ‘" AUS POL 15.46 ‘” FRA ITA 10.24 “ NED TUR 11.32 " AUS TUR -2.26 "" FRA JAP 5.20 NED UK 12.73 ”" AUS UK 2.48 FRA NED 11.74 "" NED US 2.40 AUS US 2.26 FRA POL 9.18 “ POL UK 10.31 ” BEL CAN 5.71 F RA TUR 6.48 POL US -1.68 BEL DEN 21.71 “" FRA UK 10.42 ” TUR UK 9.67 " BEL FRA 15.69 "‘ FRA US 1.02 TUR US -0.97 BEL GER 14.17 "" GER GRE 14.31 "' UK US -0.24 BEL GRE 19.32 ”’ GER HUN 14.31 "' . Busrness cycle effect BEL HUN 19.35 ”" GER ITA 9.38 " BEL ITA 9.39 ” GER NED 12.19 "W 1994 -0.41 BEL MEX 2.90 GER POL 12.48 "W 1995 -0.10 BEL NED 18.79 "" GER TUR 8.71 " 1996 -0.19 BEL POL 14.83 "W GER UK 7.67 ‘ 1997 -0.01 BEL TUR 12.07 "" GER US -0.14 1998 -0.43 BEL UK 15.51 ‘" GRE HUN 3.19 1999 -0.32 BEL US 5.12 GRE ITA 9.64 " 2000 -0.39 2001 -0.79 "“ 2002 .020 2003 -0.82 "‘ dy/dx is for discrete change of dummy variable from 0 to 1 Significance: "‘ at 1% , ” at 5%, “ at 10% 58 Table 8 Marginal effects of Tobit estimates not conditional on trade value greater that ZCI‘O Variable dy/dx. 8P} 3.61 "* pm} 1.78 * agr 1.05 **"‘ PM 0.90 1le 0.47 m 1mm 0.89 m Country — pair fixed effects Pair dy/dx Pair dy/dx Pair dy/dx. AUT BEL 10.59 " CAN FRA 0.99 GRE NED 11.72 “ AUT CAN 14.45 ‘” CAN GER 1.32 GRE POL 12.76 ” AUT DEN 11.30 " CAN HUN 8.72 GRE TUR 17.10 ‘" AUT FRA 3.71 CAN ITA -0.95 GRE UK 15.52 ‘" AUT GER 0.19 CAN JAP 5.77 HUN ITA 5.07 AUT HUN 9.63 CAN NED 8.05 HUN JAP 1.16 AUT JAP 15.31 "' CAN POL 8.90 " HUN MEX 13.59 " AUT MEX 18.77 "‘ CAN TUR 2.92 HUN NED 18.09 "" AUT NED 10.21 ‘ CAN UK 5.61 HUN POL 23.81 ”‘ AUT POL 7.97 CAN US 10.35 ‘ HUN TUR 15.06 ”‘ AUT TUR 2.97 DEN FRA 11.16 " HUN US —0.89 AUT UK 4.51 DEN GER 17.20 ”" ITA JAP -0.87 AUT US 7.05 DEN GRE 7.97 ITA NED 13.89 "‘ AUS BEL 20.05 ""‘ DEN NED 18.86 "* ITA POL 11.80 “ AUS DEN 11.33 DEN POL 16.66 “" ITA TUR 9.21 ‘ AUS FRA 5.48 DEN TUR 9.71 ‘ ITA UK 0.97 AUS GER 18.39 ‘" DEN UK 16.24 ‘" ITA US -1.17 AUS GRE 10.24 DEN US 0.07 JAP NED -0.30 AUS HUN 21.97 "" FRA GER 10.61 " JAP US 4.31 AUS ITA 14.76 "" FRA GRE 7.37 MEX US 2.32 AUS NED 18.53 “‘ FRA HUN 9.98 " NED POL 14.90 ”’ AUS POL 17.88 “" FRA ITA 12.55 ” NED TUR 13.68 ” AUS TUR -2.52 ‘" FRA JAP 6.96 NED UK 15.12 ‘”" AUS UK 3.48 FRA NED 14.11 ‘” NED US 3.37 AUS US 3.18 FRA POL 11.43 " POL UK 12.62 " BEL CAN 7.57 FRA TUR 8.46 POL US -2.07 BEL DEN 24.14 "" FRA UK 12.74 ” TUR UK 11.95 " BEL FRA 18.11 ‘" FRA US 1.44 TUR US -l.29 BEL GER 16.57 ‘" GER GRE 16.73 ‘1" UK US -0.33 BEL GRE 21.75 ”" GER HUN 16.72 "" . Busmess cycle effect BEL HUN 21.79 ‘" GER ITA 11.64 " BEL ITA 11.65 ” GER NED 14.57 ”" 1994 -0.57 BEL MEX 4.05 GER POL 14.87 ”“ 1995 -O.14 BEL NED 21.22 "" GER TUR 10.93 " 1996 -0.26 BEL POL 17.25 ”" GER UK 9.79 "‘ 1997 -0.01 BEL TUR 14.44 ""‘ GER US -0.20 1998 -0.59 BEL UK 17.93 "W GRE HUN 4.44 1999 -0.44 BEL US 6.87 GRE ITA 11.92 " 2000 ~0.54 2001 -1.07 ”" 2002 -0.28 2003 -1.11 ”‘ For Dummy variables dy/dx is for discrete change of dummy variable from 0 to 1 Significance: ‘” at 1% , ” at 5%, " at 10% 59 An interesting case is Australia (AUT) which, despite the fact that its share in FTC trade is limited, has developed trade flows facilitated by the fixed conditions of trade. This analysis does not show the direction of trade flow, but we can say that Australian FTC trade with Mexico and Canada, and Japan has been strengthened by fixed effects. The U.S. shows a positive and significant (10 percent) fixed effect with Canada. The FTC trade between these two countries is expected to be almost nine percent higher than the expected value of the control countries. The U.S. and their FTC partners (besides Canada), whether as importers or exporters have not been able to make a difference in the value of trade based of factors such as geographic position, cultural similarity or special trade agreements. A surprising, result is the marginal effect for fixed effects between Austria and Turkey. The trade value among these two countries is expected to be 2.52 percent less that the assigned expected value for those countries that do not trade FTC. This is the only country-pair that showed a significant negative impact from fixed effects. Additional empirical results show the business cycle effects are significant for two years, 2001 and 2003 (using 1993 as the base year). Negative signs on these coefficients seem to contradict general trends in the FTC market, since the total value of FTC traded worldwide has increased substantially between 1993 and 2003. The negative and significant results for 2001 and 2003 deserve special consideration. In 2001 the business cycle marginal effect is -1.07 (significant at 1 percent), indicating that the expected value 60 of trade between any two given countries will be 1.07 percent less than it was in 1993 (Table 8). Since it is known that the total value of FTC trade in 2001 is larger than in 1993, the logical conclusion is that individual countries have diversified their markets and are likely to see, on average, lower volumes of trade between any given pair of countries. Expected Values The estimation of the expected value of exports requires the addition of one unit to the dependent variable in order to calculate the logarithm when ym = 0. To calculate E(ylx), from log(1 + y), w was set equal to log(l + y), and then y = [exp(w) — 1], thus y = {exp[max(0, xii + 8)] — 1} (16) To estimate E(yw) equation 13 was estimated 1500 times for each pair of countries, substituting the MLE estimates from equation 13 for I} and a was drawn (1500 times) from a normal distribution with mean equal to zero and variance equal to the standard deviation of the model squared, 8 ~ Normal(0, $2) . Finally y”, is the result of averaging the above equation across all the draws (Wooldridge, 2005 ). Trade Shares and fixed effects Figure 3 shows a graphical analysis of fixed effects average values, trade share and number of significant fixed effects within the region. As you move from right to left the trade between those specific regions accounts for a greater share of world volume. As you move from bottom to top, the fixed effects factors have a greater (i.e. positive) 61 impact on trade between the specified regions. The values next to the circles represent the number of significant fixed effects that countries within the regions have, for example, between West European and Eastern European countries there are 21 significant fixed effect. 21 g 18 a _ _ _ _ > E-Efl 0 M E OWEN—2 m 4 15 o . . Vile-5321 = we-omg _ W's-WEE _, ._ ,2 E 01141 9 _______ __ _________ ______._..__. -._. 9 '3 Otis-cm 1 5 le- , EE-CAN;1 6 = 2 — % - —1 3 g“ EE~LIS-CAN;1 . f . fl . . . o 40% 0% 10% 20% 30% 40% 50% 60% 70% 30% Region's trade share of the market Figure 3. Selected regions: Graphical analysis of fixed effects average values, trade share and number of significant fixed effects within the region. Source: Study results EE, Eastern Europe; WE, Western Europe; OTH, Others (Australia, Mexico, Japan) FTC trade between Western and Eastern European countries accounts for more than 60 percent of total FTC trade in the model. Fixed effects have a significant impact on FTC trade in 21 (out of 27) country-pairs within these two regions and ,on average, the countries represented are expected to export 13 percent more than the expected value assigned by the marginal effect analysis. Likewise, the trade among Western European countries reveals that these countries account for 26 percent of the market, nine countries from this region show 27 significant fixed effects, and on average, trade among these countries is expected to be 11 percent higher than the expected value assigned by the 62 analysis to those countries that do not trade. These two inter-regional analyses clearly prove the importance of building trade relationships based on time-invariant factors in order to consolidate market participation. The trade among Eastern European countries shows a high expected value of trade (17), however, this potential is not translated into market share. The three countries of this region are net exporters to other regions having insignificant trade flows amongst themselves. The FTC trade among Canada and the U.S. represents 5 percent of the market. This important market share is backed up by the single significant fixed effect that U.S. showed with Canada. The fixed effects value is relatively low (8), and in this case, the importance of the time invariant variable to strengthen a market relationship is also evident. Analyzing the trade among Eastern European countries, and the US and Canada as a block, it is possible to identify just one significant fixed effect (Canada and Poland). The trade between these two regions represents slightly more than two percent of the FTC trade from which a large part is given by the trade between U.S. and Poland. This could mean that the Eastern European countries could have a more relevant role in the US — Canada trade block. More specifically, Poland could gain some space in Canada’s market pushing the U.S. aside from the existing market. 63 The FTC trade among “others”, being Japan, Australia and Mexico, has a relatively high fixed effects value, as well as the Western European countries and “others”. However, the share of the market of this intra- and inter-regional trade is insignificant. This can be explained from two different perspectives. The first being the intra trade among “other” countries is incipient since none of these countries are important FTC producers, therefore, they can not take advantage of the presence of — probably — natural time invariant factors that could promote FTC trade. On the other hand, Western European countries and “others” have not consolidated a significant share of the market probably because Western European countries have not targeted these markets. 4.3 Frozen Tart Cherry Export Potential The export potential of FTC was estimated by subtracting the expected value of exports conditional on the independent variables (xm) from the average of exports from 1993 to 2003 for each pair of countries. Positive values indicate an export potential while negative values (not reported) indicate that the exporting country has exported on average more than the model results would predict. 64 Table 9. Export potential for selected countries expressed as percent of the annual export average. Export potential (%) AUT AUS BEL CAN DEN FRA GER GRE NED POL TUR U.S. AUT - - - - - - - - . - - - AUS - - - - - - 72 - - - - - BEL - - - - - 58 28 - 82 - - - CAN - - - - - - - - - - - 363 DEN 41 - - - - 270 - - - . - - FRA - - - - - - - - - - - - GER - 51 - - 28 34 - 155 146 1195 377 - GRE - - - - - - - . - - - - HUN - - - - - - - - - 123 - - ITA - - - - - - - - - 1 - - JAP - - - - - - - - - - - 9 MEX - - - - - - - - - - - - NED - - - - - - - - - - 25 - POL - - - - - - - - - . - - TUR - - - - - - - - - - - - UK - - 13 - - 29 - - - 7 - - U.S. - - - 100 - - - - - - - - Countries excluded from the chart did not preset any potential or it was minimal. For the complete chart see Annex 3. Table 9 shows the export potential of FTC for selected countries. The numbers in the chart represent the percentage that an exporting country could potentially increase its 1.17 For exports over historical annual average exports given factors in the gravity mode example, results for Poland indicate a large potential to export to Germany. If German demand increases, Poland would likely be the best candidate to provide this supply. This result is not surprising since Poland and Germany are currently the largest exporter and importer respectively of FTC and the pair of counties have the characteristics that a GM '7 That is to say that these countries are “under-traded” with its respective trade partner. That could be due to factors that have not been included in the model or because the trade among these trade partners has not been sufficiently aggressive. 65 would predict as a high flow of trade. That is a large economy vs. a relative small and a close geographic position. Other good candidates to cover increases in German demand are Turkey, Greece and Netherlands. The U.S. has the potential to triple the value of FTC exports to Canada. No other country shows a potential to cover any extra demand from Canada. On the other hand, Canada may also increase its exports to U.S.; this result is backed up by the presence of time- invariant factors that create a favorable trade scenario. Japan represents a market that U.S. exporters could explore further and where the U.S. has the advantage that it is the only potential supplier for such a market. France has four potential markets (Belgium, Denmark, Germany and United Kingdom) to where it could increase its exports of FTC. It is probable that these potential markets are based in part by France’s geographic position and trade agreements. The fixed effects for these countries showed a positive and significant effect on the trade flow. 66 CHAPTER V 5. Conclusions This chapter presents the conclusions of the study at three levels. In the first section the analysis examines the conclusions of the empirical analysis of FTC using a Tobit model with fixed effects. The second part focuses on the specification of the GM for a specific commodity. Finally, the third part presents some general conclusions and recommendations that are relevant for policy makers, and future research. 5.1 Frozen Tart Cherry Trade Analysis Frozen tart cherry (FTC) international trade is concentrated in a reduced number of countries. The 17 largest exporters and/or importers handle 87 percent of world traded volume of FTC. Poland’s exports supply 55 percent of the world FTC demand and Germany imports 50 percent of the world supply. Netherlands and Belgium are also important players in the FTC market, both countries are important suppliers as well as buyers. The U.S. is a net exporter of FTC, although in recent years, its share in world imports has increased as well. The results of the GM using a Tobit model for FTC are consistent with most of the GM literature. The income effect and the output effect showed a positive and significant influence on trade. The total import and export lagged value of FTC had also a positive effect on trade value, which indicates that trade relationships built over time between two countries are relevant to explain the value of FTC traded. 67 According to the results of country-pair fixed effects, the strong trade relationships among Western European countries also apply for FTC. Western European suppliers and buyers have taken advantage of their geographic position and trade agreements to positively influence FTC trade among themselves. In fact, the FTC trade among the nine Western European countries included in the study represents over 28 percent of the total FTC market. This block showed 27 significant fixed effects. Hungary, Turkey and especially Poland, have also created strong FTC trade relationships based on time-invariant variables with their counterparts in Western Europe. This trade relationship has been established in spite of the lack of trade agreements. Between the countries of these two regions 21 significant fixed were found (out of 27 possible), and as a result of this strong connection the trade between these two regions represents 61 percent of the FTC trade. This case as well as the one presented above, demonstrates the importance of time invariant factors in order to consolidate FTC markets. By contrast, the U.S. has not created FTC trade relationships based on time-invariant factors with any other country besides Canada. This sole trade relationship, where time- invariant variables are significant, represents 5 percent of the total FTC market. The rest of U.S. FTC trade relationships (more than 50 percent of its exports) are likely to be based on market opportunities that change over time. The two-step procedure was useful to clarify the role that the variables distance, presence of common border, and FTA have on fixed effects. As was expected, distance has a 68 negative and significant impact on fixed effects and consequently on FTC trade. Moreover, FTA and presence of border have relatively the same positive effect on trade of FTC. Here it is important to highlight the positive and significant relationship of FTA with FTC trade, despite the fact that FTAs have a broad scope and that a commodity such as FTC is unlikely to be treated specifically in these agreements. From the marginal analysis of the model, it can be observed that the FTC trade value is income elastic (3.61). A one percent increase in income in an importer country would increase imports by 3.61 percent. This result was expected for a processed good such as FTC. The average rate of grth of the per capita GDP-PPP during the last five years has been positive for all FTC importing countries, representing an important opportunity to expand markets for the exporting countries. Australia, Canada, and United Kingdom appear to be more attractive markets in this respect. The output elasticity (i.e. agricultural GDP) for the exporting country is almost unitary, which seems to imply that this particular sector is well integrated in the agricultural sector, and any shift in the aggregate sector would impact the export capability of FTC in the same proportion. Countries like Belgium, Netherlands, Poland and Turkey (all of them net exporters) should take this factor into account, since in the last five years they had an agricultural sector with a negative average rate of growth which in turn impacts export capabilities. 69 The positive and significant values for the marginal effects of the lagged total FTC exported value, as well as the lagged total FTC imported value, illustrate the relevance of building trade relationships over time. Even though FTC can be stored for more than a year, importing countries do not take advantage of this feature so as to not affect next year’s imports. In 2002 Poland gained an important share of the U.S. internal FTC market, this fact could be critical for the U.S. FTC industry if Poland decided to be more aggressive with this market and started building long-terrn trade relationships with the U.S. buyers. The analysis of country-pair fixed effects confirms that the Western European countries have strong trade relationships based on the time-invariant factors. It is expected that FTC trade amongst Western European countries would be between 15 and 24 percent more than the expected value assigned by the model to those countries that have not engaged in trade of FTC. This block of countries accounts for 26 percent of the FTC market. In the same analysis, when Eastern European countries (Poland, Hungary and Turkey) trade with Western European countries, the expected value of trade is between 10 and 17 percent greater than the expected value of the control countries. The Eastern European countries have consolidated their trade links with Western Europe in spite of the lack of a FTA. Out of 27 trade relationships between Western and Eastern European countries, 21 resulted in having significant fixed effects. The trade between these two regions represents 62 percent of the FTC market. It is clear that the Western European countries 70 have taken advantage of their geographic location, borders and other time invariant factors (e.g. cultural similitude). Although none of these countries have created FTC trade links based on fixed factors with countries in other regions besides Poland with Canada. The U.S. is an important player in FTC markets despite the fact that it has not established trade relationships based on FTA or geographic position (other than Canada). The trade between the U.S. and Canada accounts for 5 percent of the total FTC market. This important share could be threatened by the fact that Poland and Canada do have significant fixed effects in their FTC trade relationship. Evaluating the expected value from the model, it was possible to make an analysis of export potential. Results indicate that Germany is the most important potential market. Any German demand expansion could be more likely covered by Poland, although Turkey and Netherlands are also potential candidates to cover that demand. The U.S. could have two market opportunities to expand exports: Canada and Japan. In both cases the U.S. is the only potential supplier. On the other hand, according to the results, if the U.S. expands its FTC demand, Canada is the most likely country to fulfill that extra demand. There is a clear correlation between potential markets and fixed effects significance; this could mean that most of the potential markets that the analysis showed are based on factors such as distance, FTA or the presence of common border. 71 5.2 The Model The analysis of international trade in specific agri-food commodities using a GM needs to consider three issues: 1) the use of panel data 2) censored dependent variable (corner solutions) and 3) fixed effects. Statistical characteristics of the data make a Tobit model with fixed effects using panel data, estimated by MLE, the recommended estimation procedure. This estimation procedure, analyzed by Greene (2004b), requires that the data fulfills some minimum requirements. These requirements can be summarized as follows: the period of time should be greater than five years; at least ten countries should be included in the analysis; the degree of censoring, namely the portion of zeros in the dependent variable (value of trade) should be greater than 40 percent; and lastly, the independent variables included in the model have to follow a Normal, Chi-squared or AR(l) distribution. If these requirements are met, it is expected that results of the estimation will be robust. The way that fixed effects (pair wise) are set in a GM helps reduce the over- parameterization problem of Tobit models with fixed effects. Most of the time-invariant variables (fixed effects) will have the same impact on trade between two countries regardless of the trade position of the countries (as exporter or importer). This particular feature of the GM reduces the fixed effects parameters to be estimated by half, and the model estimation is feasible by “brute force”, .i.e. by adding a dummy variable for each pair of countries and maximizing the likelihood function directly. 72 Aggregated trade analysis using GM generally controls for income and output effects by including total GPD of the importing and exporting countries respectively. When it comes to a specific good, particularly an agri-food commodity, those variables may not be the most appropriate to control for income and output effects. According to the LR- test, Agricultural GDP is preferred over total GPD for the case of the exporting countries and per capita GDP expressed in purchasing parity power is preferred over total GDP for the importer countries. The use of fixed effects to estimate a GM implies that all time invariant variables are excluded from the model. The effects of variables such as distance and border are captured by the fixed effects dummy variables. In order to evaluate the influence that individual time-invariant variables have over trade, a second step procedure is applied. In the second step, time-invariant variables are regressed on the estimates of fixed effects, making it possible to see the relevance of these variables (that are observable) on the set of unobservable time-invariant variables, and consequently also on trade value. The result of this two-step procedure provides some useful insights about the relationship of time invariant-variables and trade. Its interpretation and use, particularly on forecasting, is limited and requires special attention. 5.3 FTC Sector and Policy Makers The following conclusions and recommendations are based on the results of this study. They are aimed at being practical and applicable to FTC sector or used by policy makers to promote the sector. 73 The study showed that FTC is highly income elastic. SME should diversify its markets in terms of country income growth. The portfolio should include countries with high income growth rate, to take advantage of the effect of this rate over exports, but it should also include countries with income growth rates that are more moderate but sound over time. The approach should focus on reducing any negative impact that the high income growth countries may have, given that these economies are likely to be less stable. This study reinforces the relevance of building trade relationships over time. The volume of export/import from the previous year will generate positive effects on the present year trade. Once a FTC export firm has penetrated a market, it will be important that this firm consolidates its presence in order to take advantage of the effect of this new market. SME could explore some of the fixed effects trade relationships that the model showed as positive and significant where trade value is low or even non-existent. For instance Canadian SME could explore Australian markets. According to the model FTC trade between these two countries is expected to be 14 percent higher than the value assigned for those countries that did not trade. Eastern European countries are another good example; FTC trade among them is low, but their trade value could be increased based on time invariant variables. The U.S. bases its FTC trade on factors that change over time. This capacity could be seen as positive, since the U.S. SME are taking advantage of opportunities that appear over time, however, the U.S. market is also very susceptible to any production shifts, 74 such as the 2002 tart cherries shortage. In this year Poland took advantage to introduce itself to this new market. The U.S. SME needs to develop mechanisms to compensate for the lack of fixed factors such as establishing strategic alliances with companies overseas. U.S. policy makers should consider creating stronger trade links with the main FTC U.S. markets; this could be made by establishing special trade agreements for this product. In addition, policy makers should promote that FTC producers diversify their exports markers. 5.4 Future Research Future research on this field could address the following issues that were not covered in this study: 1. Analyze the behavior of time-variant dummy variables in the contexts of the model used in this study. The literature suggests that a model with random effects would produce unbiased and consistent parameters for these sorts of variables. 2. By using simulations it could be possible to establish the reliability of the two step procedure to estimate the effect of time-invariant variables (distance, border). 3. The hypothesis testing procedures for a Tobit model with fixed effects is a field that has not been explored. The adequacy of log-likelihood ratio test is of special interest for GM for specific agri-food products that are estimated by MLE. 75 m" "-7“. r“: .--:'t h ? tr AN NEXES 76 98859—va 5:358:— 98 ooEom Sea ”bouzom Hedi mom. 2 owe $2. 533 8nd— omm - Nam no.1. Sc _ ._ A. _ fin. a3..— awoé avoé Nag: Ev.— EN mp—mem ENS - - 3 Need 3: - - - om - w o _ we: 3. m N. .m.D 3?: mom - one; 53mm v3; - - m - pm 2a SN god Nom 2. m. m v5 m3 - - - - Z - - - _ - M: - - - ~18 - - MD... 3m - 2 - - S - - w on - 2: o - M: - E - 40m _mv.m_ 2%.— E omm woos - - - E. o; no 8:. 3: S: um nm em 2 RE 9.: e3 - - - - - - - - - - - - - NE ._ - o XmE we: 2:.— - - - n - - Z..— N - - Z - 2N - - 2: as 3nd a: - 3v oow 3m - - - 3 mm. MEN ”mm - m. - cmm - an: mmm - - N_ m. _ w v - - - - E - - mm of N - 2:: m3 - 2 new «mm v - - - - - mm_ c - - I - - 55 Shoe mow; 9mm 3?». com—N. mans - - omm Rod 9% - ohm 3:. S: R moo; v ”EC 3:: mom 3: man owmd «2 .N - - 3 mm 3 31m - n2 - mad me am <5— :fiv Tm an ac new; mum - - - - M: can 0 - - t: .N v _ ZmQ «New ~36 - mm wmm am - - - m - _N m - - mm. - o Z .88 I mom: £65550 386—om mo Eons—€3.85 .0550 Ed. 585 we 62? 3:8.» owfio>< ._ 552 77 owed 363 BED REED :oozzom 36.—SEQ 29:0 320 HooSom and MD Nab.» vaN MDH ms. _ .N. a3: 2%.— . 40m N2: own mmm.~ Noe; mil :3...” go.” new. 2 3.2: m _ md sz 25.2 ammd New.» Rm.” wwmd 3N: m5. ANNA. ”34.— N:.._ NNMA hon; cmNA: mmwd <5 mmms vac.— hwmé 3m 3;.— 52: Sad 3w 2:: 3N.» and own :5.— NEN. unfl: wand ”me; we; mMO 25.0 mmo omoN 2m Rm vmhd m3.» cg; oao vow; mmO N3.» N3” bond com; 2:. mad 356 _N_._ 3cm.— god 2.» (9E NE.» ”no mamd «8 ms ”and one.” Sum.— :o._ mm _ .N Nmm 08.— ZmQ m3. emmw m2.» mend cmed 89m Ema: 53.6 :56 ”mug. MUN—.6 949m mood Zm toaxo 1358 65 .«o Beacon mm wommoaxo 3380a H.898 wouoomxm .m 552 79 REFERENCES Anderson, J. E. (1979). "A Theoretical Foundation for the Gravity Equation." American Economic Review 69(1): 106-16. Anderson, J. E. and E. van Wincoop (2001). Gravity with Gravitas: A Solution to the Border Puzzle. Working Payer. National Bureau of Economic Research. Cambrige, MA. Ash, M. (2004). Analysis of Canadian Agri-Food Trade, Agriculture and Agri-Food Canada. Babula, R. A. and R. L. Corey (2005). "U.S. Canned Tuna Supply and Demand." Journal of International Food & Agribusiness Marketingl6(2). Baker, R. G. V. (2000). "Towards a dynamic aggregate shopping model and its application to retail trading hour and market area analysis." Papers in Regional Science 79(4): 413-434. Bergstrand, J. H. (1985). "The Gravity Equation in lntemational Trade: Some Microeconomic Foundations and Empirical Evidence." The Review of Economics and Statistics 67(3). Brenton, P., F. Di Mauro, et a1. (1999). "Economic Integration and F DI: An Empirical Analysis of Foreign Investment in the EU and in Central and Eastern Europe." Empirica 26(2): 95. Carrillo, C. and C. A. Li (2002). "Trade Blocks and the Gravity Model: Evidence from Latin American Countries." Discussion Papers. University of Essex. Cheng, I.-H. and H. J. Wall (2004). Controlling for Heterogeneity in Gravity Models of Trade and Integration. Federal Reserve Bank of St. LouisWorking Paper 1999- 010E. ST. LOUIS, MO. Cherry Marketing Institute (2003). Tart Cherry Industry Lansing, MI: 2. Data Service and Information DSI-Eurostat. (2005). from http:// 1 95. 145.59. 164/_shop2/index2.php?lang=en. Deardorff, A. V. (1995). Determinants of Bilateral Trade: Does Gravity Work in a Neoclassical World? NBER Working Papers No. 5377. Cambridge, MA. Doherty, B. A. and J. C. McKissick (2000). Value Added of Processing Vegetables, Cotton and Peanuts in Georgia Report No. 12. Center for Agribusiness and Economic Development, The University of Georgia. 80 Egger, P. (2000). "A note on the proper econometric specification of the gravity equation." Economics Letters 66(1): 25-31. Egger, P. (2002). "An Econometric View on the Estimation of Gravity Models and the Calculation of Trade Potentials." The World Economy 25(2): 177-319. Eichengreen, B. and D. A. Irwin (1998). The Role of History in Bilateral Trade Flows. The Regionalization of the World Economy. J. A. Frankel. Chicago, University of Chicago Press: 33-57. Evenett, S. J. and W. Keller (2002). "On Theories Explaining the Success of the Gravity Equation." Journal of Political Economy 110: 281—3 16. Food and Agriculture Organization. (2005). "FAOSTAT ", from http://faostat.fao.org/. Frankel, J ., E. Stein, et a1. (1995). "Trading Blocs and the Americas: The Natural, the Unnatural and the Supernatural." Journal of Development Economics 47(1). Frankel, J. A. (1997). Regional Trading Blocs in the World Economic System. Washington, DC, Institute for lntemational Economics. Frankel, J. A. (1998). The Regionalization of the World Economy. Chicago, University of Chicago Press. Frankel, J. A. and E. A. Cavallo (2004). Does Openness to Trade Make Countries More Vulnerable to Sudden Stops or Less? Using Gravity to Establish Causality. NBER Working Paper. Fuller, W. A. and G. E. Battese (1974). "Estimation of linear models with crossed-error structure." Journal of Econometrics 2(1): 67-78. Greene, W. (2004a). "The behaviour of the maximum likelihood estimator of limited dependent variable models in the presence of fixed effects." The Econometrics Journal 7(1): 98-119. Greene, W. (2004b). "Fixed Effects and Bias Due to the Incidental Parameters Problem in the Tobit Model." Econometric Reviews 23(2): 125. Halicka, E. (2001). Links Between Participants of the Fruit Supply Chain- Case Study Poland. lntemational Conference on Quality in Chains. Handy, S. L. and K. J. Clifton (2001). "Local shopping as a strategy for reducing automobile travel, Transportation." Transportation 28(4): 317 - 346. Harrigan, J. (1994). "Scale Economies and the Volume of Trade." The Review of Economics and Statistics 76(2): 321-328. Heckman, J. J. and T. E. Macurdy (1980). "A Life Cycle Model of Female Labour 81 Supply " Review of Economic Studies - Econometrics Issue 47(1): 47-74. Hirschberg, J ., I. Sheldon, et a1. (1994). "An Analysis of Bilateral Intra-Industry Trade in the Food Processing Sector." @plied Economics 2(26): 159-167. Honore, B. E. (1992). "Trimmed LAD and Least Squares Estimation of Truncated and Censored Regression Models with Fixed Effects." Econometricg 60(3): 533. Honore, B. E. and L. Hu (2004). "Estimation of cross sectional and panel data censored regression models with endogeneity." Joumgl of Econometrics 122(2): 293. Honore, B. E. and E. Kyriazidou (2000). "Panel data discrete choice models with lagged dependent variables." Econometrica 68(4): 839. Karemera, D., V. I. Oguledo, et a1. (2000). "A gravity model analysis of international migration to North America." Applied Economics 32(13): 1745 - 1755. Kennedy, P. (2003). A_guide to econometrics. Cambridge, Mass, MIT Press. Koo, W. W. and D. Karemera (1991). "Determinants of world wheat trade flows and policy analysis." Cfladian Journal of Agricultural Economics 39(3): 439-445. Lee, J .-W. and P. Swagel (1997). "Trade Barriers and Trade Flows Across Countries and Industries." The Review of Economics and Statistics 79(3): 372-382. Linnemann, H. (1966). An econometric study of intemgtional trade flows. Amsterdam, North-Holland Pub. Co. Maddala, G. S. (1983). Limited-dependent and qualitative variables in econometrics. Cambridge Cambridgeshire ; New York, Cambridge University Press. Martinez-Zarzoso, I. and F. Nowak-Lehmann (2003). "Augmented Gravity Model: An Empirical Application To Mercosur-European Union Trade Flows." Journal of Applied Economics VI(2): 291 -3 16. Matyas, L. (1997). "Proper Econometric Specification of the Gravity Model." The World Economy 20(3): 363-368. Matyas, L. (1998). "The Gravity Model: Some Econometric Considerations." The World Economy 21(3): 397-401. McCallum, J. (1995). "National Borders Matter: Canada-U.S. Regional Trade Patterns." American Economic Review 3(85): 615 - 625. McPherson, M. A., M. R. Redfearn, et a1. (2000). "A Re-Examination of the Linder Hypothesis: A Random-Effects Tobit Approach." Internationgfl Economic Journgl 14(3): 123-136. McPherson, M. A., M. R. Redfearn, et a1. (2001). "International trade and developing 82 countries: an empirical investigation of the Linder Hypothesis." Applied Economics 33: 649 - 657. Michigan Agricultural Statistics (2004). Michigan Agricultural Statistics 2003-2004. M. D. 0. Agriculture. National Agricultural Statistics Service - NASS - (2004). Michigan Agricultural Stadistics 2003 - 2004 Nerlove, M. (2002). Essays in panel data econometrics. New York, Cambridge University Press. OECD (2002). The OECD small and medium enterprise outlook. Paris, OECD. Paas, T. (2003). Regional Integration and lntemational Trade in the Context of EU Eastward Enlargement. Hamburgisches Welt-Wirtschafts-Archiv - Discussion Paper. Hamburg Institute of lntemational Economics. Poyhonen, P. (1963). "A tentative model for the flows of trade between countries." Weltwirtschlaftliches Archiv 90(1): 93-100. Pravorne, I., N. Skorohoda, et a1. (2003). The Gravity model and prediction of trade flows between Baltic States and their trade partners. EuroFaculty Working Papers in Economics. The University of Latvia and Eurofaculty. Prentice, B. E., Z. Wang, et a1. (1998). "Derived Demand for Refrigerated Truck Transport: A Gravity Model Analysis of Canadian Pork Exports to the United States." Canadian J oumal of Agricultural Economics 46(3): 317-328. Reynolds, P. D. (1997). "New and Small Firms in Expanding Markets." Small Business Economics 9(1): 79-84. Rose, A., K (2002). "Do We Really Know that the WTO Increases Trade?" NBER Working Paper 9273. Soloaga, I. and L. A. Winters (1999). How Has Regionalism in the 19908 Affected Trade? World Bank - Policy Research Working Paper. Washington, DC. Tinbergen, J. (1962). S_haping the world economy; suggestions for an internatiogl economic policy. New York,, Twentieth Century Fund. U.S. Small Business Administration (1999). America's Sm_all Businesses and lntemational Trade: A Report. Washington, DC, U.S. Small Business Administration. United State Census Bureau, Ed. (2004 a). Frozen Fruit, Juice, and Vegetable Manufacturing: 2002. Manufacturing Industry Series. Washington, DC. United State Census Bureau, Ed. (2004 b). Fruit and Vegetable Canning: 2002. 83 Manufacturing Industry Series. Washington, DC. Wall, H. J. (2000). Gravity Model Specification and the Effects of the Canada-U.S. Border. Working Paper Federal Reserve Bank of St. Louis. Wooldridge, J. (2005). The estimation of the expected value: Tobit model. C. Aguilar. East Lansing. Wooldridge, J. M. (2000). Introductory Econometrics: A Modern @pyrogch, South - Western College Publishing. Wooldridge, J. M. (2002). Econometric analysis of cross section and panel data. Cambridge, Mass, MIT Press. World Bank. (2005). "World Development Indicators." from http://web.worldbank.org/WBSITE/EXTERNAL/DATASTATISTICS. 84 llilllllljjlljlllljlllllllllllll