Non-Tariff Barriers to Agricultural Trade between Turkey and the EU

Contents

ABSTRACT

Acknowledgements

List of Tables

List of Abbreviations

1 Introduction
1.1 Why research Non-Tariff Barriers to Trade
1.2 Structure of the Thesis
1.3 Ways of capturing NTBs
1.3.1 The Price Gap Method
1.3.2 Price-based Econometric Methods
1.3.3 Quantity-based Econometric Methods
1.3.4 Simulations Methods
1.4 The Gravity Equation in Physics and Economics

2 The Border Effect Approach
2.1 The Original Approach by McCallum
2.2 Multilateral Resistance Terms
2.2.1 Theoretical Derivation of the Gravity Equation
2.2.2 Use of Price Index Data
2.2.3 Use of estimated Border Effects
2.2.4 Use of Fixed Effects
2.3 The Trouble with Heteroskedasticity – Why OLS-Estimation of Log-linearised Models is problematic
2.3.1 Jensen’s Inequality and its Implications for Log-linearised Models
2.3.2 CES Models and Heteroskedasticity
2.3.3 Finding an Appropriate Estimator for the Non-Linear Model
2.3.4 Maximum Likelihood Estimation and Poisson Regressions
2.3.5 Application of the PPML Estimator in the Gravity Model
2.4 The Problem of Endogeneity

3 Non-Tariff Barriers to Agricultural Trade between Turkey and the EU: An Empirical Analysis
3.1 Aim and Approach of Analysis
3.2 Applied Gravity Model
3.3 Data
3.4 Estimation and Results
3.4.1 Single sectors estimation using Importer-Exporter-Dummies
3.4.2 Pooled and aggregated Regression using PPML
3.4.3 Comparison of aggregated / pooled regression results with literature
3.5 Calculation of AVEs

4 Concluding Remarks

5 References

Appendix: Calculation of Weighted AVEs

ABSTRACT

Non-Tariff Barriers to Agricultural Trade between Turkey and the EU

by Claus Mayer

This thesis reviews the border effect approach as an application of gravity models of trade and different methods of including multilateral resistance terms (MRTs) in it. Some focus is laid on the endogeneity problem of the approach. In an empirical application of the approach, agricultural trade between Turkey and the EU is analysed; the effect of data pooling and aggregation is studied; the conversion of estimated border effects into ad-valorem tariff equivalents (AVEs) reveals the crucial importance of a reliable measure of the elasticity of substitution when trying to separate the effects of NTBs of the total effect of a border

Keywords: Non-tariff barriers to trade; gravity model; border effect approach; agricultural trade; EU; Turkey; ad-valorem tariff equivalents;

Acknowledgements

First and foremost, I would like to thank my supervisor Prof. Dr. Martina Brockmeier for the unique opportunity to work on this topic and her continued support throughout the writing of this thesis. Her agreement to let me read and cite her unpublished work (Befus et al. 2012) enabled me to put my results into context

I wish to express particular gratitude to M.Sc. Tanja Engelbert for her significant contributions to this thesis in the form of numerous discussions and constructive comments, which have greatly improved this work

Further, I thank Prof. Dr. Harald Grethe for his time and resources involved in being the second supervisor of this thesis

Moritz Stollovsky has put a great deal of time and motivation into proof-reading this thesis and has provided helpful suggestions for corrections and improvements, which I highly appreciate

Finally, my special thanks go to my wife Ivana and our family for their ongoing support throughout the writing of this thesis and my entire Bachelor and Master Studies. I could not have made these achievements without you

List of Tables

Table 1: Results of Single Sector PPML and OLS Estimations

Table 2: Results of Pooled an Aggregated PPML Regressions

Table 3: Comparison of Pooled Regression Results

Table 4: Implied Ad-valorem Equivalents of Border Barriers (%)

Table 5: Unexplained Part of Implied AVEs

List of Abbreviations

illustration not visible in this excerpt

1 Introduction

1.1 Why research Non-Tariff Barriers to Trade

Ever since David Ricardo came up with his “Law of Comparative Advantage” (Ricardo 1817) it has been a central argument for the establishment of free trade agreements around the world. It states that even if one of two countries is less efficient in the production of each of its goods than the other one, it will still gain from trade with the other country as long as the two differ in their relative efficiency of producing different goods.

Probably the most notable international free trade agreements were reached within the World Trade Organisation (WTO) and its predecessor, the General Agreement on Tariffs and Trade (GATT), which was founded in 1947. By 1994, the GATT had a total of 124 member countries (WTO 2012) that agreed to lower international tariffs on trade. For instance, between 1947 and 1994 the average tariff on industrial products traded between the members was lowered from 40% to 3.8% (Past et al. 2007, p. 6).

No longer able to apply tariffs, many nations resorted to an alternative way of impeding international trade in order to protect their less competitive industries, namely Non-Tariff Barriers to Trade (hereafter NTBs) (Past et al. 2007, p.3).

NTBs can be defined as “any measure (public or private) that causes internationally traded goods and services [...] to be allocated in such a way as to reduce potential real world income” (Baldwin 1970, p.5). While various ways of classifying NTBs can be found in literature, Baldwin (1970) divides them into twelve groups:

- Quotas and restrictive state-trading policies
- Export subsidies and taxes
- Discriminatory government and private procurement policies
- Selective indirect taxes
- Selective domestic subsidies
- Restrictive customs procedures
- Antidumping regulations
- Restrictive administrative and technical regulations
- Restrictive business practices
- Controls over foreign investment
- Restrictive immigration policies
- Selective monetary controls and discriminatory exchange-rate policies

The WTO prohibits the introduction or retention of NTBs (Past et al. 2007, p.15). However, proving the existence of an NTB is not easy (Past et al. 2007, p. 7). This is one of the reasons why there is a rising amount if international literature on how to detect and scale NTBs.

1.2 Structure of the Thesis

After this overview in part one on what NTBs are and how they can be measured in general, part two will deal with the border effect approach as a means of estimating NTBs in particular. In part three, I will apply the border effect approach to agricultural trade between Turkey and the EU, before drawing some conclusions in part 4.

1.3 Ways of capturing NTBs

1.3.1 The Price Gap Method

The method shows by how much NTBs within a country have lifted the domestic price of a good above the international price. This allows to take the “workings of policies” into account as well as to convert results into tariff equivalents. However, not only may price data availability be problematic, it is difficult to determine whether one of the prices is fully affected by an NTB whilst the other is completely unaffected. This method is sometimes called a handicraft, because it can literally be done manually with pen and paper. (Ferrantino 2006, p. 9.)

1.3.2 Price-based Econometric Methods

Price-based econometric methods extend the price-gap method to many countries and products at a time while taking systematic price differences into account, showing how much of a price difference may be attributable to NTBs. They enable us to compare how NTBs affect different countries or product groups. Price data unavailability as well as the method’s incapability to make use of “product- and policy specific detail”, which may cause the results to differ from those of a case specific analysis, as well as the influence of econometric specifications on results, are problematic. (Ferrantino 2006, p. 9.)

1.3.3 Quantity-based Econometric Methods

Data on traded quantities are not only better available internationally than price data, but also more standardised, which may make it possible to see how NTBs affect trade flows. Three types of models can be distinguished:

- Gravity models: explain trade by size of countries, distance between them and other factors
- Factor-content models: explain trade by different resource availability between countries
- Mixtures of the previous two

A significant amount of research is carried out to improve these methods. The non-includability of country- and product specific information and the influence of econometric specifications on results may be more problematic than with price-based methods. To convert results into tariff equivalents or price gaps, additional information and / or assumptions are required. (Ferrantino 2006, p. 10.)

NTBs are often introduced primarily to reduce trade flows, which lowers domestic supply and only secondarily leads to an increase in prices, along with a range of other factors that may influence prices. These other factors will not bias quantity-based methods (Ferrantino 2006, p. 21, point 64).

Therefore, I will apply a quantity-based approach and in particular a gravity model in this thesis when examining agricultural trade between Turkey and the EU.

1.3.4 Simulations Methods

Simulating the effect of changes in tariffs and other border barriers on trade, price and other economic variables can be done either in a partial (to focus on specific products or sectors) or general equilibrium setting. To simulate the effects of the introduction or removal of a NTB, a tariff equivalent (TEV) of the NTB must have been estimated by one of the above methods in order to implement it into the simulation. It is possible to perform sensitivity analyses to see which assumptions influence the simulation outcome by how much; simulations are also capable of explaining causal relationships. (Ferrantino 2006, p. 10)

1.4 The Gravity Equation in Physics and Economics

This thesis will focus on gravity models as a quantity-based econometric method to measure NTBs. In physics, the Newtonian gravity model predicts the force with which two objects will pull at each other by using the formula:

illustration not visible in this excerpt

(Hammer et al. 2001)

Tinbergen (1962) was the first to apply the formula to predict trade flows between countries. In his work, F becomes the monetary measure of a trade flow between two countries, and represent the GDPs of the countries, and r is the distance between them.

While there is a variety of applications of the gravity model in economics, I will follow the development of the border effect approach as a particularly promising and researched way to capture the effect of NTBs.

2 The Border Effect Approach

A first attempt to apply the gravity equation to estimate the trade impeding effect of a border was carried out by McCallum (1995) when he compared Canadian interprovincial trade to international trade between Canada and the U.S. With this new methodology and his finding that without the border trade flows between Canada and the U.S. would be 2200% larger, he caused widespread interest in the topic and a large amount of related research to be carried out.

Improvements to the approach focused somewhat on the used data (McCallum did not take U.S. interstate trade into account), but mainly on the specification of the gravity equation and found still significant, if lower, border effects looking at the same data as McCallum.

The Border Effect Approach is particularly interesting when trying to quantify the effect of non-tariff barriers. If it was possible to get a consistent estimate of the total border effect, then deducting the known effect of tariffs from it would leave us with an estimate of the effects of non-tariff barriers.

Thus, chapter 2 will look into the original approach by McCallum and its most significant advancements by later contributions in order to obtain a consistent approach for the analysis of agricultural trade between Turkey and the EU.

2.1 The Original Approach by McCallum

In his 1995 paper “National Borders Matter: Canada-U.S. Regional Trade Patterns”, McCallum conducts a widely noted application of the following gravity equation to examine the determinants of international trade patterns:

illustration not visible in this excerpt

As apparent from equation (2), McCallum has logarithmised it in order to enable a standard ordinary least squares (hereafter OLS) estimation. He uses the equation to compare 1988 data on trade between Canadian provinces to trade between Canadian provinces and U.S. states in order to show the effect of the U.S.-Canadian border. Notably, 1988 is the year that the FTA, a free trade agreement between the U.S. and Canada, was signed. McCallum assumes that FTA did not yet impact the effect of the border during the examined year. Data on interprovincial trade flows between all ten Canadian provinces as well as trade between those provinces and the 30 U.S. states that account for 90% of U.S.-Canadian trade are taken into account. This means that the author did not include intra-national trade flows between U.S. states, the potential effect of which will be discussed in section 2.2. This leads to a total of 690 observations, out of which 7 are zero-values, the logarithm of which cannot be taken, which is why McCallum does not include them in the estimation. The potential bias arising from this neglect of zero-values will be dealt with in section 2.3.

The equation is then estimated by OLS. Assuming that the border effect for Canada can be expressed as the exponential of the coefficient of DUMMY, it results that trade is expected to be an average of 22 times larger between two jurisdictions that are not separated by a border. McCallum also identifies a number of potential econometric problems to this approach. For one, he acknowledges the biasing effect heteroskedasticity may have on results. To test if it does, he estimates the gravity equation three more times, once by only using data from large jurisdictions with large GDP, once again by using all available data but weighing the regression towards jurisdictions with large GDP, and finally by using all data and substituting minimal trade flows for the zero-values. Also, he realises the problem that the dependent variable (exports) is a component of one of the regressors (GDP), the latter of which will thus be correlated to the error term. To find out about the disturbing effect of this, he estimates the gravity equation two more times, once while using the logarithm of population as instruments of GDP, and once by replacing the logarithms of GDPS with logarithms of population altogether. In none of these additional regressions do the results defer significantly from the original ones, from which McCallum concludes that neither heteroskedasticity nor the correlation of exports and the error term have lead to a biased estimate of the border effect. Furthermore, results also remain stable when the following aspects are incorporated into the model:

- Distance represented as logarithm, natural number or square of the natural number
- Province-specific constant terms (indicator of a province’s overall exports) and province-specific coefficients (indicators of the degree to which each province’s trade is biased towards exports to other provinces)
- Provinces’ production structures (share of agriculture and manufacturing in GDP)
- Interaction term between DUMMY and the other explanatory variables, or the squares of yi, yj and dist

All in all, McCallum finds a significant hindering effect of the U.S.-Canadian border on trade flows.

2.2 Multilateral Resistance Terms

Anderson and van Wincoop (2003) argue that the paper of McCallum (1995) and related literature lack a theoretical foundation of the gravity equation, which has two negative implications: Firstly, an omitted variable bias can be expected and secondly, comparative static exercises, e.g. to show the effects of removing a trade barrier, cannot be conducted conclusively, because a countries resistance to trade with a bilateral partner also always depends on its resistance to trade with all other partners. The latter, however, is mostly the reason to estimate a gravity model in the first place.

The original gravity equation by Anderson 1979 is based on constant elasticity of substitution (CES) preferences and the differentiation of goods by region of origin. Anderson and van Wincoop (2003) were the first to manipulate the CES expenditure system in order to derive a gravity model from it. From this, three components of trade resistance can be derived:

- The bilateral trade barrier between i and j
- i’s resistance to trade with all regions
- j’s resistance to trade with all regions

I will now show how these components can be captured by the so-called Multilateral Resistance Terms (hereafter MRTs), and how these MRTs can be created from a theoretical derivation of the gravity equation.

2.2.1 Theoretical Derivation of the Gravity Equation

Feenstra (2002) shows how the gravity equation can be derived from a CES expenditure system: He points out that in the face of border effects, McCallum ‘s (1995) assumption that prices equalize across countries is no longer true and patterns of trade are more complex than reflected in his gravity equation. To deal with this, Feenstra (2002) develops his gravity equation from the assumed CES utility function:

illustration not visible in this excerpt

To simplify this, the author adopts the “iceberg transportation cost” approach by Samuelson (1952):

illustration not visible in this excerpt

Transport cost then could be written as . When transport cost is equal for all varieties of a good, consumption will be equal for all products k=1,…, in country j, so that

illustration not visible in this excerpt

Consumers from j will maximize this utility function subject to their budget constraint:

illustration not visible in this excerpt

From this maximization, the demand of each product is derived as:

illustration not visible in this excerpt

where is j’s overall price index:

illustration not visible in this excerpt

If we define the total value of exports from i to j as:

illustration not visible in this excerpt

then it follows from (7) and (8):

illustration not visible in this excerpt

The author summarizes three options of estimating this equation so that multilateral resistance is taken account of: The use of price index data, the use of estimated border effects as in Anderson and van Wincoop 2003 and the use of fixed effects, all of which will be presented in the following.

2.2.2 Use of Price Index Data

I follow Feenstra (2002) and, since the variable in the gravity equation (10) is unobservable, we apply the zero-profit assumption to infer that the output of each firm in i, , is constant. Therefore, the GDP of country i can be calculated as:

illustration not visible in this excerpt

or:

illustration not visible in this excerpt

Substituting this into the gravity equation (10) we get:

illustration not visible in this excerpt

Since , we can rewrite this as:

illustration not visible in this excerpt

This can be transformed into the estimation equation by taking logarithms and first differences of variables:

illustration not visible in this excerpt

As a data source for the price indexes and , GDP deflators can be used.

There are two problems with this approach. Firstly, aggregate price indexes cannot really capture the full extent of cross-border transaction cost and the risk (money, time, currency). Secondly, price indexes refer to a certain base period for which the index is set to equal 100. This base period frequently differs from country to country, making indexes incomparable to each other. Therefore, an alternative approach is introduced:

2.2.3 Use of estimated Border Effects

This approach to estimating the gravity equation is the one originally introduced by Anderson and van Wincoop (2003). As described by Feenstra (2002), they work around the problem of incomparable price indexes by modeling as a function of and trade costs, using the equation:

illustration not visible in this excerpt

If we substitute this back into the gravity equation (14), we will receive a set of equations where exports depend on and . However, those equations are not linear and thus difficult to estimate; therefore, the market clearing condition may be used to further simplify the gravity equation. The market clearing condition, assuming that output is greater than consumption due to “iceberg” transportation cost, is:

illustration not visible in this excerpt

Multiplying this by (from (4)), we can rewrite:

illustration not visible in this excerpt

Assuming symmetric bilateral trade cost (i.e. ), the market clearing condition (17) can implicitly be solved as:

illustration not visible in this excerpt

Multiplying this by , we get:

illustration not visible in this excerpt

Substituting (20) into the price index (8) yields:

illustration not visible in this excerpt

This gives us the gravity equation:

illustration not visible in this excerpt

Due to the relationship expressed in (21), , depend on the estimated border effect and can be called “Multilateral Resistance Terms”. While they cannot be observed directly, they can be estimated using (21) together with formula (16) to solve for the estimated border effect.

In order to estimate this equation, the GDP terms may be moved to the left, equation (21) may be substituted for transportation cost, before taking natural logarithms:

illustration not visible in this excerpt

This equation needs to be estimated iteratively; while minimizing the sum of squared residuals, (16) must be applied to include the border effect at each iteration and these are used to solve for the multilateral resistance terms from (21). Also, it is necessary to be more explicit about the border effects in (23): Replacing with where is the same dummy variable that McCallum 1995 used (i.e. unity for interprovincial trade and zero otherwise), and using and as coefficients on distance and , respectively , we can rewrite:

illustration not visible in this excerpt

When interpreting the results of this estimation, a problem arises: Simply taking the exponent of and interpreting it as the border effect is no longer viable because a removal of the border would also affect the MRTs. A solution is to compare trade in presence ( ) and absence ( ) of the border effect: If we say are the MRTs in absence of the border effect, calculated from formula (21), but obtaining in (16) as the logarithm of distance times its estimated coefficient only, take exponents of (24) and compare this equation with and without border effects, we get the ratio of trade in these two cases:

illustration not visible in this excerpt

Applying this to the original data of McCallum 1995, Anderson and van Wincoop (2003) find intra-Canadian trade to be 4.3 times and intra-US trade to be 1.05 times larger due to the presence of the border than what they would be without it, whereas international trade is 0.41 times smaller. From this, they calculate intra-Canadian trade to be 4.3 / 0.41 = 10.5 times higher than cross-border trade; the analogous number for intra-US trade is 1.05 / 0.41 = 2.6. The average border effect can then be calculated as the geometric mean of the two border effects, i.e. . Feenstra (2002) points out that 5.2 is the exact same value one would obtain by taking the exponent of , the coefficient of the cross-border dummy variable, and proves mathematically that doing so is a completely valid way of inferring the average (over countries) impact of the border on intra-national relative to international trade.

One may wish to note that in order to do these calculations, it is necessary to assume a value for the elasticity of substitution, . Feenstra 2002 assumes it to take the value 5, but hints that Anderson and van Wincoop 2003 found it to be irrelevant if the value is altered between 2 and 20 in their sensitivity analysis.

In their original paper on gravity models using estimated border effects, Anderson and van Wincoop (2003) also argue that McCallum ‘s (1995) estimate of the border parameter for Canada is too large for two reasons:

Firstly, in the relatively small Canadian economy the border raises inter-province trade more relative to international trade than it raises inter-state trade in the U.S relative to international trade. With the U.S. being a large economy, international trade makes up a smaller proportion of total U.S. trade than it does of total Canadian trade.

Secondly, the omitted variable bias can be attributed to two sources:

(a) Estimation bias (“ordinary econometric emitted variable bias”): While in Anderson and van Wincoop ’s (2003) paper the MRTs are calculated as

illustration not visible in this excerpt

: Border dummy that equals one for intra-national trade and zero

for international trade

it is only a dummy variable that equals one for international trade flows in McCallum ’s (1995) paper. Since the MRTs, however, are correlated with both distance and the dummy variable in his equation, McCallum ‘s (1995) estimation produces a biased estimate.

(b) Computation bias, resulting from comparative statics based on omitted terms:

Even if McCallum’s (1995) method would produce unbiased estimates of the coefficients of distance and the dummy variable, comparative statics based on these estimates would still be biased since MRTs will also change when a border effect is newly or no longer considered. This effect will inflate the border effect of Canada while reducing that of the U.S.

This demonstrates that when conducting the comparative statics experiment of eliminating a border, one must take into account its effect on multilateral resistance. It shows that the “the more resistant to trade with all others a region is, the more it is pushed into trading with a given bilateral partner” (Anderson 1979). Anderson and van Wincoop (2003) find that almost all the omitted variable bias can be attributed to the computation bias, whereas the estimation bias plays no significant role.

As seen, this method requires custom programming to perform the constrained minimization, and thus is tedious to conduct. A possible solution to this problem may be the use of fixed effects.

2.2.4 Use of Fixed Effects

As a solution to this, Feenstra (2002) suggests to, instead of calculating the MRTs in (17) by using (12) and applying (9), measure the MRTs as coefficients of source and destination region fixed effects, setting:

[illustration not visible in this excerpt]: Dummy variable, unity if i is the exporter, zero otherwise.

[illustration not visible in this excerpt]: Dummy variable, unity if j is the importer, zero otherwise.

This gives us the new gravity equation:

illustration not visible in this excerpt

in which the coefficients on the source and destination indicator variables estimate the multilateral indexes resistance.

Applying this to the McCallum (1995) data, the author interprets the exponential of as the average border effect and estimates a value of 4.7, which he views as reasonably close to the 5.2 calculated with the use of estimated border effects. It is also very close to the value of 4.8 which he receives when estimating the McCallum (1995) gravity equation but, unlike McCallum, includes intra-US trade in the data. This latter approach apparently corrects itself by overstating the border effect for Canada (15.7 instead of 10.5) and understating the effect for the US (1.49 instead of 2.56).

Feenstra (2002) concludes that the fixed effects approach, while being simple to implement, leads to consistent estimates of the border effect and may be preferred over the other two.

2.3 The Trouble with Heteroskedasticity – Why OLS-Estimation of Log-linearised Models is problematic

2.3.1 Jensen’s Inequality and its Implications for Log-linearised Models

Jensen’s Inequality states if is a convex function defined on and the expected value of the dependent variable, , it implies: . (Wooldridge 2010, p. 31)

Santos Silva and Tenreyro (2010) argue that this implies . According to them, with economic models like the gravity equation not holding with the accuracy of physical laws but on average only, they can only be used to calculate the (conditional) expected value of the dependent variable, i.e. E[yi|x]. Therefore, the widely spread habit of estimating the gravity equation and, in fact, a multitude of econometric models in their log-linear form can lead to biased estimates. The authors also explain why this bias arises in the face of heteroskedasticity. They begin the derivation of this result by a general consideration of Constant Elasticity of Substitution (hereafter CES) models:

2.3.2 CES Models and Heteroskedasticity

Assuming y and x are linked by a CES model of the form , which holds on average and has the error term , the stochastic model can be formulated as:

illustration not visible in this excerpt

As an example of this, the multiplicative gravity equation could be rewritten as:

illustration not visible in this excerpt

The standard practice of log-linearising this model (28) and then estimating by OLS does not lead to appropriate results for two reasons:

Firstly, observations of zero-values cannot be logarithmized and thus are generally either omitted or replaced by artificial minimal values, both of which will cause bias in results.

For the second reason, consider transforming the model (28) into:

illustration not visible in this excerpt

where:

illustration not visible in this excerpt

and . If , this model can be made linear:

illustration not visible in this excerpt

[...]