Specialisation:
Heckscher-Ohlin, or
not?
An exploration of trade pattern specialisation
Daan Freeman S1865765 11403989
Abstract
This thesis seeks to test the Heckscher-Ohlin model and how it interacts with different types of traded goods. The emphasis lies on checking the models predictions for goods traded as intermediates and those as final consumption goods. To this end, a dataset was constructed containing bilateral trade data including 95 countries and 78 industries for the year 2000. Using a model designed to explain trade patterns, several regressions were ran employing methods consisting of simple OLS, a Tobit and instrumental variable specifications. The results reveal significant differences between the effect of Heckscher-Ohlin type endowments on trade patterns of intermediate and consumption goods.
Robert Inklaar Laura Birg Supervisors
1 Table of Contents Keywords ... 2 List of Tables ... 2 List of abbreviations ... 2 Introduction ... 3 Theory ... 4 Methodology ... 9 Data ... 12 Variables ... 14 Descriptive statistics ... 18 Tests ... 19 Heteroskedasticity ... 20 Outliers ... 21 Multicollinearity ... 21 Endogeneity ... 22 Results ... 24 Base dataset ... 24 Extended dataset ... 27
Instrumental variable approach ... 29
Discussion & Conclusion ... 31
References ... 33
Appendix ... 35
Data sources ... 35
Industry matching procedure ... 36
Compustat data variables ... 37
Leverage-Residual Plot ... 38
Country list... 38
2 Keywords
Trade; Heckscher-Ohlin; Ricardo model; Intermediate goods;
List of Tables
Table 1: Goods classifications ... 14
Table 2: Summary statistic... 18
Table 3: Correlations... 19
Table 4: Endogeneity Test ... 23
Table 5: Primary specification – Part 1... 25
Table 6: Primary specification - Part 2 ... 26
Table 7: Extended dataset ... 28
Table 8: 2SLS Specification ... 30
List of abbreviations Abbreviation Explanation
2SLS Two-Stage Least Squares BEC Broad Economics Categories
CEPII Centre d'Etudes Prospectives et d'Informations Internationales FREIT Forum for Research in Empirical International Trade
GDP Gross Domestic Product H-O Heckscher-Ohlin
HS2 Harmonised System 2002
IID Independent and Identically Distributed ISO International Organisation for Standardisation
IV Instrumental Variables
MLE Maximum Likelihood Estimation OLS Ordinary Least Squares
R&D Research & Development RFI Revealed Factor Intensities S&P Standard & Poor’s
SIC Standard Industrial Classifications
SITC Standard International Trade Classification
UNCTAD United Nations Conference on Trade and Development VIF Variance Inflation Factor
3 Introduction
Trade is one of the most researched topics in economics, and also boasts a rich theoretical background. A wide consensus has emerged that trade is generally good for economic growth as it allows countries to focus on activities in which they are most productive, which raises overall productivity. The most frequently used theory to explain differences in trade patterns is the Heckscher-Ohlin theory. This theory (explained more fully in the next section) predicts that economy-wide factor endowments govern the structure of the economy and with it trade. Although this theory is relatively dated, it is still widely used in the literature, albeit in expanded and/or altered form.
The analysis in this thesis uses insights of more recent literature that have identified a profound difference in various types of traded goods. The difference is supposed to be especially profound between intermediate goods1 and final goods.2 Trade patterns are expected to differ between these two goods, as the factor endowments are not necessarily related to the overall production of the goods. The aim of this thesis is to identify the specialisation patterns with regard to intermediate and consumption goods. Examining a large dataset that contains bilateral trade flows, allows a very detailed look at the patterns of trade across the world. The results are indicative of a clear difference between the specialisation on intermediate and consumption goods with regards to the standard Heckscher-Ohlin model of trade.
This thesis is structured as follows; the next section, theory features a description of the theoretical background used in this thesis, including the hypotheses that are tested. This is followed by the methodology, which discusses the methods that are used. The next section describes the database and the variables used in the investigation. The subsequent tests and
results sections ensure data quality through a series of tests, and present the primary results.
The latter also includes two alternative specifications of the primary model. Finally, the
4
discussion and conclusion section summarizes the results and interpretations and draws some
final conclusions.
Theory
The ideas that this paper uses build on a rich history of theoretical and empirical literature regarding the predictions of specialisation and comparative advantage. Firstly, at the core of this literature are two models that together have to a large extent structured the debate on international trade in the literature. These two models are the Ricardian trade model and, later, the Heckscher-Ohlin trade model. While both these models are relatively simple, they have aided greatly in forming thought on international trade.
The Ricardian model of trade was developed in the early 19th century. This theory uses differences in labour productivity to predict comparative advantage. Countries are predicted to export the products in which they have a comparative advantage (Helpman, 1999). While this model is very useful in theory, empirical tests of this model remained difficult. The reason is that labour productivity as the exclusive explanation of the structure of trade was revealed to be too limited and incomplete (Helpman, 1999). Furthermore, the model’s two-country assumption makes it difficult to employ in empirical research (Dornbusch, Fischer, & Samuelson, 1977). The problem was addressed by Dornbusch et al. (1977), who developed a Ricardian model allowing the inclusion of multiple countries. This model was later extended by Eaton and Kortum (1997, 2002).
5
greater flexibility than the Ricaridian model (Helpman, 1999; Romalis, 2004). For this reason, it will serve as basis for the analysis of the remainder of this thesis.
The main aim of this paper is to test whether this Heckscher-Ohlin prediction is reflected in the data.One way to test this is to compare the factor intensities of imports and with those of exports. If, for instance, a country’s exports are more capital intensive than its imports, domestic production is relatively capital intensive. When this corresponds to a country’s endowments, i.e. a high capital endowment, this is in line with Heckscher-Ohlin predictions (Helpman, 1999).
The study by Wassily Leontief (1954) is an early example of an empirical test seeking to prove Heckscher-Ohlin theory. He compared factor intensities of imports and exports to examine whether specialisation is in line with Heckscher-Ohlin theory. He famously found that specialisation for the U.S. ran against the Heckscher-Ohlin predictions in 1949. He uncovered that the U.S., although being considered the most capital-intensive country in the world, imported more capital intensively than it exported (Leontief, 1953). This result was aptly named the “Leontief paradox”.
Roughly 25 years later, the Vanek equation (Vanek, 1968) brought a possible solution to the Leontief paradox (Leamer, 1980). The Vanek equation is essentially a measure of factor abundance, which is calculated by subtracting the factor inputs of consumption from the factor inputs of production. Surpluses will be abundant factors, while shortages scarce factors. In this sense, a product that intensively uses the abundant factors in a country, will have surplus production, and this product will therefore be exported. The Vanek equation sometimes finds Heckscher-Ohlin relations in the data; however, it does not consistently agree with the theory. (Helpman, 1999)
6
endowments. The advantage of using this methodology is that it measures the specialisation independent of sector structure (Gerlagh & Mathys, 2011).
Recent literature has increasingly recognised that the landscape in which trade occurs is changing. For instance, Baldwin (2006) puts forward a theory called the “second unbundling”; the separation across geographical spaces of different stages of firms’ production processes. This theory argues that increasingly final products are produced using intermediaries from different countries, which themselves might also have been produced using inputs from different countries. In this sense, trade has been becoming more about the exchange of different production stages, rather than that of final products. One way to think about this shift in focus is by considering firms’ production as product value chains. These chains represent the entire production process from conception to final sale, of a single product. Traditional theory, assuming that the value chain is contained within one or perhaps two countries, suggests that specialisation will also occur at this level. ‘Old’ Heckscher-Ohlin theory predicts that countries will specialise on products, whose production, including the entire value chains, fit their endowments. However, the different stages within the value chains can be heterogeneous concerning the types of factors used intensively. For example, consider the value chain for a car, which requires to a great extent large-scale capital-intensive automated production, but also high-skilled labour for design and R&D, as well as for certain stages low-skilled labour production (Humphrey & Memedovic, 2003). In the end, although the production of a car uses both high-skilled as low-skilled labour, all stages taken together, the production of a car will be classified as mostly capital intensive (Humphrey & Memedovic, 2003).
7
able to perform them more efficiently. This specialisation on production stages (or tasks) might make a measure exclusively focused on trade in final consumption goods to be of limited use in examining specialisation of countries.
To account for this trade in intermediates, some have argued that trade measured in gross exports no longer suffices (Timmer, 2013; Grossman, 2008). An alternative that is suggested, is to evaluate trade in terms of value added. A measure based on value added is indeed suitable for measuring trade under this ‘new paradigm’ (Baldwin, 2006). However, trade data based on value added is not available for broad set of countries or for many years. To this end, this thesis uses ‘standard’ bilateral trade data, but within this data it makes a general distinction between different kinds of traded goods. The two primary categories include those goods, which are traded as intermediate goods and those as final consumption goods. A further category is capital goods, and the remainder of traded goods are classified as a residual group ‘other’. This classification is made using the Broad Economic Category (BEC) industry classification system, which will be further elaborated on in the Data section.
Due to the changing structure of trade towards increasing importance of intermediate goods, and the characteristics of intermediate goods, specialisation patterns are expected to vary between intermediate goods and all other types of goods. Firstly, the theory predicts that due to international fragmentation of value chains, countries specialise on stages of production or tasks, rather than on whole value chains. This leads to hypothesis 1:
H1: Specialisation in intermediate goods is more in line with Heckscher-Ohlin theory than all other types of goods.
8
Due to this shifted focus in trade specialisation, and assuming that the different stages within value chains are in fact reflected by heterogeneous factor intensities, trade in final goods is expected to be no longer relevant to the specialisation patterns. Hypothesis 2 can be therefore stated as follows:
H2: Compared to other goods, specialisation in trade of final consumption goods is unrelated to Heckscher-Ohlin theory.
This hypothesis states that trade in final goods will adhere to the Heckscher-Ohlin prediction that specialisation occurs according to factor endowments, to a lesser extent than other types of goods. An example of early literature that concluded this is the previously mentioned Leontief paradox (Leontief, 1953). Illustrative is the case of exports from China, which contain very large amounts of electronics as consumer goods, whose total production generally uses large amounts of high-skilled labour and capital. As China is a cheap-labour abundant country, the exports of electronics are not in accordance with China’s factor endowments. This apparent contradiction is due to the assembly activities that occur in China, which are in line with Chinese factor endowments (Johnson, 2014). In all likelihood, this is related to the increasing prevalence of trade in intermediates, if final consumption goods are indeed a composite of imported intermediates from different countries, their factor content will quite possibly be completely unrelated, or even run against exporter factor endowments. This relation will be especially strong for goods whose final stages of production require factors that are different from those that make up the earlier stages of production.
9
to affect trade flow patterns. These three types of institutions can be considered as sources of Ricardian comparative advantage; they are modelled to explain productivity differences between different countries, allowing better tests of the Ricardian theory (Chor, 2010; Eaton & Kortum, 2002). Combining the Heckscher-Ohlin factors (human and physical capital) with the Ricardian forces of comparative advantage (institutional forces that govern productivity) make for a more complete model than either could produce separately (Chor, 2010).
While the institutional forces have been found to affect trade patterns, no previous research has investigated explicitly the effect of adding these institutional variables to the Heckscher-Ohlin specialization regressions. However, some scholars found that controlling for institutional variables invalidated the effects of Heckser-Ohlin specialization (Levchenko, 2007; Nunn, 2007), others found no such invalidating effect of different types of institutional forces (Chor, 2010; Manova, 2013; Cuñat & Melitz, 2012).
Methodology
To examine the relationship of interest that was outlined in the previous section, a two-stage regression is in order. Firstly, to find the type of industries that countries specialise on, bilateral trade is regressed on the factor intensities, for each industry. This will allow determination which type of industries are exporting relatively more than others, suggesting these are the industries that countries specialise on. Equation (1) shows this first regression
𝒍𝒏(𝑿𝒊𝒋𝒛) = 𝜷𝟎+ 𝜷𝟏𝒍𝒏(𝒉𝒛) + 𝜷𝟐𝒍𝒏(𝒌𝒛) + 𝜺𝒊𝒛 (1)
Where:
- 𝑋𝑖𝑗𝑧 are the exports of industry z from country i to country j. - ℎ𝑧 & 𝑘𝑧 are skill and capital intensities in industry z, respectively.
10
𝜷𝟏 = 𝜸𝟏+ 𝜸𝟐𝒍𝒏(𝒔𝒌𝒊𝒍𝒍𝒊) (2)
𝜷𝟐 = 𝜸𝟑+ 𝜸𝟒𝒍𝒏(𝒄𝒂𝒑𝒊𝒕𝒂𝒍𝒊) (3)
Where:
- 𝑠𝑘𝑖𝑙𝑙𝑖 & 𝑐𝑎𝑝𝑖𝑡𝑎𝑙𝑖 are the endowments of human and physical capital factors of country i, respectively.
Proceeding, equations (1), (2), and (3) are combined. Furthermore, following Eaton & Kortum (2002) and Chor (2010), the model is strengthened with gravity equation variables as important variables in explaining the bilateral trade flows. Additionally, institutional variables are added to the model as argued in the last part of the theory section. The result is equation (4).
𝒍𝒏(𝑿𝒊𝒋𝒛) = 𝜷𝟎+ 𝜷𝟏𝒍𝒏(𝒔𝒌𝒊𝒍𝒍𝒊) ∗ 𝒍𝒏(𝒉𝒛) + 𝜷𝟐𝒍𝒏(𝒄𝒂𝒑𝒊𝒕𝒂𝒍𝒊) ∗ 𝒍𝒏(𝒌𝒛) + 𝜷𝟑𝑫𝒊𝒋+ 𝜷𝟒𝑰𝒊+ 𝜺𝒛
(4)
Where:
- 𝐷𝑖𝑗 is a collection of gravity-equation variables characterising the bilateral relation between country i and j3.
- 𝐼𝑖 is a collection of institutional variables in country i expected to affect the specialisation of trade.
To deal with country specific characteristics, like country size, GDP, and geographical characteristics, country specific fixed effects are added to the model. These fixed effects keep all effects on trade flows that arise from country features constant. Fixing the country effects is achieved by adding a set of dummy variables, one for each exporting country.
In equation (4), the institutional variables are specified to explain the value of trade flows, by being added to the model directly. While it is certainly a possibility that a direct relationship
3𝐷
𝑖𝑗consists of ln(distance), a regional trade agreement dummy, a dummy for common borders, and another
11
exists, I am interested in the effects of institutions on trade patterns. This, in combination with the fact that the country fixed effects already account for the variation due to the country-specific institutional variables, calls for the institutional variables to be matched to relevant industry characteristics, and together enter the regression as interaction variables. Equation (5) shows the resulting regression.
𝒍𝒏(𝑿𝒊𝒋𝒛) = 𝜷𝟎+ 𝜷𝟏𝒍𝒏(𝒔𝒌𝒊𝒍𝒍𝒊) ∗ 𝒍𝒏(𝒉𝒛) + 𝜷𝟐𝒍𝒏(𝒄𝒂𝒑𝒊𝒕𝒂𝒍𝒊) ∗ 𝒍𝒏(𝒌𝒛) + 𝜷𝟑𝑫𝒊𝒋+
𝜷𝟒∑(𝒊,𝒛)(𝑰𝒊∗ 𝒁𝒛)+ 𝜷𝟓𝝌𝒊+ 𝜺𝒛 (5)
Where:
- 𝑍𝑧 are industry specific characteristic matched to the institutional variables - 𝜒𝑖 is a collection of exporter specific dummies
The regression design displayed in equation (5) is closely related to the regressions performed by Romalis (2004) and especially the ones by Chor (2010). However, to test hypotheses 1 & 2, one further addition needs to be made to the current regression equation. To examine the effects of intermediate and consumption goods on the Heckscher-Ohlin specialisation, a dummy for the appropriate BEC categories is interacted with both the human and physical capital specialisation terms. Equations (6) and (7)4 are the final regressions with which this thesis aims to test hypotheses 1 & 2.
𝒍𝒏(𝑿𝒊𝒋𝒛) = 𝜷𝟎+ 𝜷𝟏𝒍𝒏(𝒔𝒌𝒊𝒍𝒍𝒊) ∗ 𝒍𝒏(𝒉𝒛) + 𝜷𝟐𝒍𝒏(𝒄𝒂𝒑𝒊𝒕𝒂𝒍𝒊) ∗ 𝒍𝒏(𝒌𝒛) + 𝜷𝟑𝑫𝒊𝒋+ 𝜷𝟒∑(𝒊,𝒛)(𝑰𝒊∗ 𝒁𝒛) + 𝜷𝟓𝑩𝒎∗ 𝒍𝒏(𝒔𝒌𝒊𝒍𝒍𝒊) ∗ 𝒍𝒏(𝒉𝒛) + 𝜷𝟔𝑩𝒎∗ 𝒍𝒏(𝒄𝒂𝒑𝒊𝒕𝒂𝒍𝒊) ∗ 𝒍𝒏(𝒌𝒛) + 𝜷𝟕𝝌𝒊+ 𝜺𝒛 (6) 𝒍𝒏(𝑿𝒊𝒋𝒛) = 𝜷𝟎+ 𝜷𝟏𝒍𝒏(𝒔𝒌𝒊𝒍𝒍𝒊) ∗ 𝒍𝒏(𝒉𝒛) + 𝜷𝟐𝒍𝒏(𝒄𝒂𝒑𝒊𝒕𝒂𝒍𝒊) ∗ 𝒍𝒏(𝒌𝒛) + 𝜷𝟑𝑫𝒊𝒋+ 𝜷𝟒∑(𝒊,𝒛)(𝑰𝒊∗ 𝒁𝒛)+ 𝜷𝟓𝑩𝒄 ∗ 𝒍𝒏(𝒔𝒌𝒊𝒍𝒍𝒊) ∗ 𝒍𝒏(𝒉𝒛) + 𝜷𝟔𝑩𝒄∗ 𝒍𝒏(𝒄𝒂𝒑𝒊𝒕𝒂𝒍𝒊) ∗ 𝒍𝒏(𝒌𝒛) + 𝜷𝟕𝝌𝒊+ 𝜺𝒛 (7) Where:
- 𝐵𝑚 is a dummy identifying industries as producing intermediate goods.
4 Equations (1)-(7) use log-transformed variables following Chor (2010). Additionally, the institutional variables
12
- 𝐵𝑐 is a dummy identifying industries as producing consumption goods,
The addition of the new terms in the equations (6) and (7) allow the identification of the effect of the Heckscher-Ohlin forces, specifically with regards to intermediate and consumption goods. These two terms can then be interpreted as the difference in effect compared to all other goods.
Data
The two data sources that make up the core of the analysis in this paper provide data on bilateral trade and industry specific factor intensities. These two datasets use different industry classification systems. The Standard International Trade Classifications (SITC) Revision 2 classifications system is used for the bilateral trade data, which are obtained from the Feenstra international trade flow database (Feenstra, Lipsey, Deng, & Ma, 2005). This trade flow database contains bilateral trade data starting 1962 until 2000 on a great number of countries for large amount of industries on the SITC Rev. 2 4-digit level5.
On the other hand, the industry data uses the Standard Industry Classifications (SIC) system. Factor intensity data is obtained from the National Bureau of Economic Research (NBER) and U.S. Census Bureau's Center for Economic Studies (CES) (or NBER-CES) database (Becker, Gray, & Marakov, 2013). This dataset includes data on factor-rewards to capital and labour; production and non-production labour, for different countries at different years. The dataset goes as far back as the late 1950’s, which easily covers the available data on trade flows. The advantage of using this dataset is that the SIC classification shows factor intensities of manufacturing industries at a very dis-aggregated level. Additional industry specific data are obtained from the Standard & Poor’s (S&P) Compustat database, also use the SIC classification system.
13
To align the datasets, the industry specific data is re-classified into SITC Revision 2 categories, making it compatible with the bilateral trade data from the Feenstra (2005) dataset. This compatibility will allow the industry specific characteristics to be linked to the bilateral trade data at the industry level, which was done on the level of 3-digit SITC Rev. 2 classification, a detailed description of the matching procedure can be found in the appendix. Other datasets that could be used for factor intensity industry data are the UNCTAD Revealed Factor Intensity (RFI) dataset (Shirotori, Tumurchudur, & Cadot, 2010), or data from the World KLEMS project (O'Mahony & Timmer, 2009). The former is classified using the SITC Rev. 1 system; this makes matching to the Feenstra (2005) dataset straightforward. However, this dataset contains ‘revealed factor intensities’ that have been constructed from trade data coupled with data on factor endowments. Constructing factor intensities in this way requires the assumption that countries specialise on industries that use their abundant factors intensively. This is exactly the relation that is in question here, which makes the RFI dataset essentially useless for the purposes of this thesis. The other dataset, the KLEMS dataset is not used here as the data included is aggregated on a much higher level compared to the data in the NBER-CES, restricting intra-country industry factor intensity comparisons. However, an advantage of the KLEMS over the NBER-CES database would be through the presence of data on multiple countries. Data like that enables a better differentiation between countries when comparing industry factor intensities, i.e. for a more detailed inter-country view. The lack of country variation in the NBER-CES dataset requires the assumption that the ranking of measures is the same across countries, this a common assumption, often made in similar literature (Romalis, 2004; Cuñat & Melitz, 2012). Having made this assumption clearly makes the NBER-CES with detailed industry data the most suitable.
14 Table 1: Goods classifications
Class of goods BEC-codes Intermediate goods 111, 121, 21, 22, 31, 322, 42, 53 Consumption goods 112, 122, 522, 61, 62, 63 Capital goods 41, 521 ‘Other’ 7, 51, 321 Source: http://unstats.un.org/unsd/tradekb/Knowledgebase/Intermediate-Goods-in-Trade-Statistics Variables
To examine the effects of specialisation in a Heckscher-Ohlin setting, besides the factor intensities and trade flows datasets, data on factor endowments is required. The factor endowment data on labour (human capital) and physical capital are obtained from the Penn world tables. With this data, the core model can be specified, where the bilateral trade flows variable is regressed on the interaction effects between intensities and endowments of both labour and capital. In this core model the physical capital intensity is defined as the industry capital stock divided by its employment. In turn, the human capital intensity is the share of non-production workers in total employment. On the endowment side, physical capital is the total capital stock divided by the total employment. Finally, the human capital measure from the Penn World tables, already being on a per capita basis is used as such.
The data described above enables examination of whether export-specialisation is in line with factor endowments. Both the effects of labour and capital are required to be in line with Heckscher-Ohlin theory. This prediction is supported by evidence from Romalis (2004), who finds positive effects of both the physical and human capital specialisation. Cunat & Melitz (2012) find that, even when controlling for labour market institutional variables, the effects of both physical and human capital remain significant. Similarly, Manova (2008) find that when controlling for a collection of variables focused on financial development, the significance remains intact. Chor (2010), even when controlling for a collection of different institutional and gravity variables, finds consistent positive effects throughout a series of different regressions.
15
enforcement, invalidates the significant effects of Heckscher-Ohlin specialisation forces. This result is supported by Nunn (2007) who finds an insignificant effect of physical capital when controlling for different institutional variables focused on contract enforcement.
The two interaction terms covering the specific effects for consumption and intermediate goods in equations (6) and (7) respectively will be used to test the hypotheses. A positive sign would mean that Heckscher-Ohlin forces are stronger for its respective type of goods, compared to all other types of goods (combined). According to hypothesis 1, the signs for the coefficients representing intermediate goods should be positive, while those for consumption goods should be insignificant or even negative under hypothesis 2.
The additional data that is added to the model fulfils the role of controlling for additional effects. This data consists of the gravity equation variables physical distance, colonial ties, shared borders and languages, and secondly, of institutional variables aimed at exploring the effects of the institutional environment on trade specialisation.
The gravity variables are largely6 as in Chor (2010) and Eaton & Kortum (2002). These variables are well known in the literature to have important effects on the patterns of trade (Anderson & Marcouiller, 2002). The coefficients of these variables are expected to be positive for all but one variable; the bilateral distance, which is expected to negatively influence trade volumes (Eaton & Kortum, 2002; Chor, 2010).
A more recent stream of literature has found that the effects of ‘institutional comparative advantage’ are an important determinant of trade patterns. This literature describes different types of institutional variables that are found to affect trade patterns, an overview of some of the different types can be found in Chor (2010). To account for the institutional effects, a series of exporter specific institutional variables are interacted with relevant industry specific characteristics. This way, the effects of institutions on the pattern of trade are identified, as
16
the direct effect of country-specific institutional variables on trade flows is already captured by country-specific dummies that are added to the model (Nunn, 2007).
Firstly, financial development is proxied for by the credit-to-deposit ratio and by the ratio of
bank deposits to GDP (data from Worldbank), these ratios interacted with the industry
specific dependency on capital. This capital dependency is calculated using Compustat data in the same way as Chor (2010) and Manova (2008) did, as described in Rajan & Zingales (1998)7. These interaction variables show to what extend a country’s financial development is beneficial for exports in highly capital depend sectors. The interaction term of credit-to-deposits and capital dependence could be positive or negative. A positive coefficient would indicate that firms with a high dependency on capital thrive in environment in which credit is relatively well supplied. On the other hand, a negative coefficient could indicate a relative lack of bank services, if both credit and deposits are on a low level. The other term, the interaction between deposits to GDP (Worldbank data) and capital dependence, explains the relation of bank services prevalence (proxied by deposits to GDP) and its effect on more capital dependent firms. This effect is expected to be positive; capital dependent firms will benefit from the prevalence of widely available finance through high rates of deposits-to-GDP. Chor (2010), using a similar measure of financial development finds a positive relation in a similar regression.
Secondly, the labour market is considered through the addition of a labour-market flexibility
index, which is a compound of different variables available from the Worldbank’s
doing-business database. The index is interacted with industry sales volatility, which is calculated following Cunat & Melitz (2012) and consist of the standard deviation of the growth of labour weighted sales over the entire period of data availability; 1960-2000 (data are from Compustat). The variable consists of the interaction between labour market flexibility and industry business climate volatility. This effect is expected to be positive; firms facing a volatile business climate will benefit from flexible labour markets to absorb shocks. This
17
relation was confirmed by Cunat & Melitz (2012), who find the relation to be positively related to exports in all of their specifications.
Finally, a third institutional variable covers the legal environment through an indicator that measures the strength of the legal system and the property rights. This data is obtained from the Frasier institute (Gwartney & Lawson, 2004). An interaction is made with a variable representing the holdup problem. It is defined as the size of capital investments, relative to total output, calculated using Compustat data. Using the insight that in industries can suffer from holdup problems due to high investments, this interaction variable predicts that countries with a better legal environment will be more competitive in industries that are required to make large investments relative to output. Levchenko (2007) indeed found a positive effect of this variable, and Chor (2010) found positive effects using interaction terms with similar variables.8
The total final dataset contains exports of 95 countries9 for 78 different 3-digit SITC Rev. 2 industries10. However, the trade flow database only contains bilateral relations where actual trade took place; this means that the base dataset only contains positive trade values, and lacks data on no-trade relations. An additional dataset was constructed where the same trade data was used, but expanded by all potential trade relations for all exporter-importer-industry ‘observations’. This extension requires the assumption that for every missing exporter-importer-industry observation, no trade has in fact taken place. A problem with this larger dataset is that estimation by OLS becomes difficult since the logistic transformation of trade value is not defined for cases where trade value is zero. This problem can be resolved by adding one to every zero-value before the logistic transformation takes place (Cuñat &
8 Chor (2010) interacted the legal environment variable with indicators of relationship specificity and job
complexity.
18
Melitz, 2012), then transforming the data would return these values to zero, since log(1) = 0. An alternative would be a Tobit regression, which is elaborated on in results section.11
An important difference between the base and extended datasets deals with the intensive margin of trade (Cuñat & Melitz, 2012). The base dataset, containing only positive trade amounts, deals exclusively with the intensive margin of trade, i.e. the extent of trade when it in fact takes place, or the decision how much to export. On the other hand, the extended dataset also includes the extensive margin, i.e. it includes information on whether countries export at all in certain industries. This means that information on both the decision whether to export and on the one how much to export is included.12
Descriptive statistics
Table 2: Summary statistics (1) (2) (3) (4)
Mean Median Minimum Maximum
Ln(value) 6.847 6.522 0 17.35 Ln(cap. Int.) 4.461 4.402 2.773 6.891 Ln(skill int.) -1.231 -1.241 -2.126 -0.491 Ln(capital) 11.72 12.16 7.440 12.77 Ln(skill) 1.015 1.033 0.137 1.259 Ln(distance) 8.206 8.435 4.742 9.892 Credit/Deposits 119.6 123.1 9.320 663.8 Deposits/GDP 69.91 61.62 3.382 222.7 Legal 7.313 7.4 3.670 8.860 Sales Volatility 2.767 2.763 2.027 4.414 Capital Dependency 0.523 1.02 -20.16 4.555 Holdup Problem 0.0544 0.0483 0.0114 0.193 Flexibility 0.644 0.674 0.174 0.958
Before moving further, a broad overview of the data and an examination of the general relations between the different variables are in order.
11 Chor (2010) concludes that the addition of 1 to 0 trade values or using a Tobit estimation are both
insufficient and uses instead a SMM (simulated method of moments) estimation procedure, which is not used here.
12 A dataset consisting of only the extensive margin could be constructed by examining only whether export
19
Table 2 shows summary statistics for the main variables that are used in this thesis. The variables are presented in their most basic form, with logistic transformations but no interactions, to give an idea of the principle characteristics of the data.
Table 3: Correlations L n (v alu e) L n (s k ill in t.) L n (ca p . I n t.) L n (s k ill) L n (ca p ital) Dep o sits /GD P C red it/Dep o sits C ap ital Dep en d en cy Flex ib ilit y Sales Vo latili ty L eg al Ln(value) 1.0000 Ln(cap. Int.) 0.0358 1.0000 Ln(skill int.) 0.0275 0.2186 1.0000 Ln(skill) 0.1043 0.0712 0.0212 1.0000 Ln(capital) 0.1076 0.0717 0.0124 0.7153 1.0000 Deposits/GDP 0.0868 0.0652 -0.0018 0.2462 0.3911 1.0000 Credit/Deposits 0.0511 0.0104 -0.0182 -0.0790 -0.2077 -0.1430 1.0000 Capital Dep. 0.0003 0.0639 -0.0058 0.0135 0.0059 0.0219 -0.0011 1.0000 Flexibility 0.0786 0.0491 0.0019 0.3822 0.1867 0.3776 -0.0205 0.0103 1.0000 Sales Volatility 0.0191 0.1020 0.1647 0.0298 0.0135 0.0021 -0.0044 -0.0444 0.0118 1.0000 Legal 0.1131 0.0891 0.0207 0.5647 0.7127 0.5418 -0.1322 0.0057 0.4343 0.0221 1.0000 Holdup Problem 0.0060 -0.1072 0.4503 0.0085 0.0059 -0.0192 -0.0141 -0.1006 -0.0193 0.3636 -0.0097
Additionally, to examine initial relations in the data, table 3 shows the correlations between the variables. The highest correlations are between the logs of capital and skill, and the legal system and deposits. These high correlations could lead to problems in estimation and therefore a test is performed in the multicollinearity section to assess the influence of these high correlations.
Tests
20 Heteroskedasticity
Firstly, I test for heteroskedasticity; heteroskedasticity in the estimation can lead to problems in the results, if left unattended. The problem is, essentially, that the variance of the error is not constant. This violates a primary assumption of OLS estimation, and leads to inefficient estimation; while the coefficient estimates remain unbiased, the variance of the coefficient is not optimal (Wooldridge, 2012).
To test for heteroskedasticity, a test is run to determine the presence of heteroskedasticity in the data. Using the ‘standard’ chi-2 Breusch-Pagan/Cook-Weisberg test, as well as the Wooldridge’s F-test version, result in a strong rejection of the null-hypothesis, indicating that heteroskedasticity is indeed present in the data13. The difference between these two tests is that the former assumes that the errors are normally distributed, while the latter drops this assumption in favour of the assumption that the errors are IID. While the errors appear to approximate normality in a histogram (figure 1), the second test was run in the interest of completeness.
To counteract the effects of heteroskedasticity, robust standard errors are employed for all subsequent analysis (Wooldridge, 2012). An alternative would be to use specific cluster robust errors, like for example Chor (2010), who uses standard errors clustered by importer-exporter pair. A specification using this type of error-correction was also ran, but did not impact the results in any significant way compared to robust errors.
13 Chi2- and F-values are over 3500 and 5000 respectively, resulting in extremely small p-values, indicating
rather strongly that heteroskedasticity is present.
0 .0 5 .1 .1 5 .2 .2 5 D e n sit y -10 -5 0 5 10 Residuals
21 Outliers
Outliers can cause significant problems to estimation procedures employing error minimisation, increasing standard errors and altering coefficients. Outliers are observations that are outside the bounds defined by ‘normal’ observations within a dataset; these observations can differ on one or more dimensions, when compared to ‘normal’ observations. These disparities can have multiple origins, the worst of which is measurement error, this does not just affect the results adversely; it will do so unnecessarily. If however the outlier is a genuine observation on the basis of extraordinary circumstances, it might still carry valuable information. In this case the outlier’s adverse influence on the estimation might be less severe. Regardless of the type of outlier, the fact remains that they affect the estimation results considerably more than a regular observation due to the error-minimising characteristics of the OLS method (Wooldridge, 2012).
To test for outliers, the Cook’s distance measure was employed after a regression of equation (5), identifying outliers as those observations whose Cook’s distance is larger than four divided by the number of observations (Cook, 1977). This results in 7183 observations being selected as outliers, comprising about 2.6% of the total observations. Re-running equation (5) without these observations shows no real difference with the initial regression, sign and significance are totally unaffected; some coefficients are altered however, but only in the slightest way.
Additionally, a leverage-residual plot (see appendix) shows in some detail the specific countries that deviate from the common leverage-residual relation. The leverage measures how much a certain individual observation influences the outcome of the regression. The countries with large distance from the origin are a similar set of countries that have been identified by the Cook’s distance; the cook’s distance is essentially a combination of the leverage and the residuals. (Belsley, Kuh, & Welsch, 2005)
Multicollinearity
22
estimation assumptions, which limits it severity14. However, multicollinearity has the potential to affect standard errors upwardly, making estimation less reliable. (Wooldridge, 2012)
To explore this, the variance inflation factor (VIF) was calculated for the variables used in regression (5). Here, multicollinearity was only found to be present among the exporter dummy variables, leading to a high average VIF for the entire model exceeding 50. However, among the variables of interest, only one control variable exceeds the ‘cut-off’ value of 10. When the VIF is calculated for a specification without the exporter dummy variables, the VIF drops to 1.63 on average, not exceeding 2.5 for any variables.
Endogeneity
Endogeneity is a problem that can appear in several different flavours whose effects on the estimation are quite similar. The first problem of endogeneity is that of reverse causality, or bi-directional causality. This means that certain explanatory variables that might seem very successful in explaining variation in the dependent variable are actually not causally connected in the way expected. This happens when the direction of causality is uncertain, i.e. the explanatory variable explains variation in the dependent variable or the other way around. The second problem of endogeneity is that of omitted variables. In this case, an apparent successful explanatory variable’s observed effect on the dependent variable actually only works because a third (unknown or unavailable) variable determines them both.
Both these problems15, which are related to the direction and the presence of causal links between explanatory and dependent variables, cause results of regressions to become unreliable, reflecting mere correlations between variables rather than causal links. (Wooldridge, 2012)
23
To test for the presence endogeneity, the Durbin-Wu-Hausman test is used. This test uses an auxiliary regression where the suspected endogenous variable(s) is/are regressed on the other explanatory variables, plus some instruments; additional variables that are not included in the primary specification. These instruments should, ideally, be correlated with the dependent, but less so with the independent variables. The errors of the auxiliary regression are then used as an additional explanatory variable in the primary regression, if the errors are significant, endogeneity could be a problem.
Here, the two interaction variables between factor endowments and intensities, being central in this thesis, are tested for endogeneity. Testing is done using equation (5), the results are
Table 4: Endogeneity Test
Ln(value) (1) (2) (3) (4) (5) (6)
Base dataset Base dataset Base dataset Ext. dataset Ext. dataset Ext. dataset
Ln(skill) * Ln(skill int.) 0.181*** 2.072*** 0.181*** 0.218*** 0.504 0.218***
(0.00923) (0.201) (0.01000) (0.00589) (0.392) (0.00675)
Ln(capital) * Ln(cap. int.) -0.00285*** -0.00285*** 0.00883*** -0.0117*** -0.0117*** -0.0111***
(0.000557) (0.000570) (0.00112) (0.000308) (0.000322) (0.000720)
Ln(distance) -0.111*** -0.146*** -0.110*** -0.397*** -0.398*** -0.397***
(0.00513) (0.00627) (0.00505) (0.00321) (0.00363) (0.00348)
Regional Trade Agreement 0.596*** 0.616*** 0.595*** 2.213*** 2.214*** 2.213***
(0.0102) (0.00980) (0.00956) (0.00816) (0.0106) (0.0106)
Shared Border 0.772*** 0.761*** 0.768*** 1.209*** 1.211*** 1.208***
(0.0132) (0.0141) (0.0140) (0.0124) (0.0172) (0.0168)
Language 0.232*** 0.234*** 0.233*** 0.0556*** 0.0553*** 0.0556***
(0.0104) (0.0103) (0.0103) (0.00589) (0.00604) (0.00603)
Deposits * Cap. Dep. 0.000167*** 5.34e-05* 0.000154*** -0.000277*** -0.000303*** -0.000277***
(2.72e-05) (2.95e-05) (2.74e-05) (2.74e-05) (5.23e-05) (3.88e-05)
Credit * Cap. Dep. -3.20e-05* -4.59e-05*** -3.72e-05** -0.000417*** -0.000452*** -0.000418***
(1.63e-05) (1.61e-05) (1.61e-05) (1.50e-05) (4.99e-05) (1.74e-05)
Flexibility * Sales Volatility 0.437*** 0.234*** 0.432*** 0.234*** 0.170* 0.230***
(0.0212) (0.0361) (0.0217) (0.0133) (0.0914) (0.0162) Legal * Holdup 0.216*** 0.00806 (0.0219) (0.0103) Skill Residuals -1.891*** -0.286 (0.201) (0.392) Capital Residuals -0.0117*** -0.000582 (0.00124) (0.000797) Constant 6.156*** 9.828*** 5.699*** 4.085*** 4.623*** 4.072*** (0.252) (0.459) (0.204) (0.0489) (0.746) (0.0492) Observations 275,584 275,584 275,584 1,331,371 1,331,371 1,331,371 R-squared 0.139 0.139 0.139 0.432 0.432 0.432
Exporter Dummies yes yes yes yes yes yes
Countries 95 95 95 95 95 95
Industries 78 78 78 78 78 78
24
presented in table 4. For the auxiliary regression, the extra instrument that is used for both variables is GDP per capita, in following of Romalis (2004), who uses GDP per capita as an instrument for both physical capital and human capital endowments. Unfortunately, this is not a perfect instrument, but for the purposes of this thesis it will suffice. The result shows that both the auxiliary regressions’ error terms show a significant effect when added to the primary regression using the base dataset (columns 1-3). This is an indication that endogeneity might be a problem. If however, the extended dataset is used, the same test does not yield significant auxiliary-regression error terms (columns 4-6). This suggests that endogeneity is in fact not a large issue in this dataset. Regardless of this second result, an instrumental variable regression is added to the results section using both the base and extended datasets.
Results Base dataset
Equation (6) and (7) have been used for the primary regression analyses. The results are reported in table 5 and 6. The first part shows the results for equation (6) which deals with testing hypothesis 1, on the effect of intermediate goods; columns 1-6 (table 5). The second part shows the results using equation (7), testing hypothesis 2, showing the effects of consumption goods; columns 7-12 (table 6). Tables 5 and 6 show analyses using the base dataset, which means that all interpretations are related to the intensive margin of trade. Firstly, the effects of the gravity model variables, distance, and the dummies for trade agreement, shared borders, and shared language are all significant at the 1% level, with the expected signs. This is in concurrence with Eaton & Kortum (2002) and Chor (2010). This result holds across both tables, with coefficients being all but identical. Additionally, gravity variables boost the explanatory power of the model, almost doubling it, from just over 7% without, to almost 14% when included.
25
legal environment and the debt/GDP measures all show expected signs and significance at the 1% level. However, the credit/deposit measure is again anomalous, in this case showing up equally significant, but with a negative coefficient. This would suggest that more capital dependent industries thrive in countries with low levels of credit compared to deposits. An explanation could be that relatively high deposits suppress interest rates, and could therefore allow capital dependent firms access to cheap finance.
Table 5: Primary specification – Part 1
(1) (2) (3) (4) (5) (6)
Base model Institutional Forces Full Model
Ln(value) Interm. Goods Interm. Goods Interm. Goods Interm. Goods Interm. Goods Interm. Goods
Ln(skill) * Ln(skill int.) -0.0420*** -0.0119 -0.00459 -0.00942 0.00389 -0.00886
(0.0148) (0.0145) (0.0145) (0.0144) (0.0143) (0.0144)
Ln(capital) * Ln(cap. int.) 0.00640*** 0.00406*** 0.00395*** 0.00460*** 0.000315 0.00166**
(0.000654) (0.000637) (0.000637) (0.000637) (0.000663) (0.000671)
M * Ln(skill) * Ln(skill int.) 0.161*** 0.199*** 0.200*** 0.137*** 0.180*** 0.136***
(0.0169) (0.0164) (0.0164) (0.0167) (0.0164) (0.0167)
M * Ln(capital) * Ln(cap. int.) 0.00105** 0.00155*** 0.00156*** -0.000193 0.000679* -0.000481
(0.000418) (0.000406) (0.000406) (0.000417) (0.000410) (0.000418)
Ln(distance) -0.111*** -0.111*** -0.112*** -0.111*** -0.112***
(0.00505) (0.00505) (0.00505) (0.00505) (0.00505)
Regional Trade Agreement 0.598*** 0.598*** 0.598*** 0.599*** 0.598***
(0.00957) (0.00957) (0.00957) (0.00956) (0.00956)
Shared Border 0.770*** 0.770*** 0.770*** 0.770*** 0.770***
(0.0141) (0.0141) (0.0141) (0.0140) (0.0140)
Language 0.231*** 0.231*** 0.233*** 0.232*** 0.234***
(0.0103) (0.0103) (0.0103) (0.0103) (0.0103)
Deposits * Cap. Dep. 0.000135*** 0.000317***
(2.71e-05) (4.16e-05)
Credit * Cap. Dep. 3.79e-06 -0.000109***
(1.59e-05) (2.44e-05)
Flexibility * Sales Volatility 0.334*** 0.252***
(0.0203) (0.0214) Legal * Holdup 0.342*** 0.272*** (0.0215) (0.0228) Constant 5.868*** 6.853*** 6.870*** 6.036*** 6.925*** 6.271*** (0.193) (0.195) (0.195) (0.202) (0.196) (0.203) Observations 275,584 275,584 275,584 275,584 275,584 275,584 R-squared 0.072 0.139 0.139 0.139 0.139 0.140
Exporter Dummies yes yes yes yes yes yes
Countries 95 95 95 95 95 95
Industries 78 78 78 78 78 78
Robust standard errors in parentheses; *** p<0.01, ** p<0.05, * p<0.1
26
Hypothesis 1 predicts that intermediate goods are more in line with Heckscher-Ohlin specialisation than are other goods. To test this, equation (6) included a term interacting the two Heckscher-Ohlin specialisation terms with a dummy variable representing intermediate goods. If this term is positive (and significant), this means that Heckscher-Ohlin specialisation is stronger concerning intermediate goods than for the other types of goods. The results show that for human capital, the interaction term is positive, and strongly significant across all columns, including in the full model. When considering physical capital, the term is positive and significant in most columns; however, turns insignificant (and negative) when the labour market flexibility is controlled for, individually and in the full
Table 6: Primary specification - Part 2
(7) (8) (9) (10) (11) (12)
Ln(value) Base Model Institutional Forces Full Model
Cons. Goods Cons. Goods Cons. Goods Cons. Goods Cons. Goods Cons. Goods
Ln(skill) * Ln(skill int.) 0.178*** 0.239*** 0.246*** 0.224*** 0.260*** 0.229***
(0.0108) (0.0107) (0.0107) (0.0108) (0.0105) (0.0109)
Ln(capital) * Ln(cap. int.) 0.00621*** 0.00409*** 0.00397*** 0.00304*** 0.000183 0.000429
(0.000518) (0.000506) (0.000506) (0.000509) (0.000571) (0.000571)
C * Ln(skill) * Ln(skill int.) -0.457*** -0.492*** -0.487*** -0.438*** -0.477*** -0.450***
(0.0245) (0.0238) (0.0238) (0.0240) (0.0238) (0.0240)
C * Ln(capital) * Ln(cap. int.) -0.0110*** -0.0123*** -0.0122*** -0.0107*** -0.0113*** -0.0106***
(0.000657) (0.000631) (0.000631) (0.000640) (0.000634) (0.000641)
Ln(distance) -0.111*** -0.111*** -0.111*** -0.111*** -0.112***
(0.00506) (0.00506) (0.00506) (0.00505) (0.00505)
Regional Trade Agreement 0.594*** 0.594*** 0.594*** 0.595*** 0.595***
(0.00957) (0.00957) (0.00957) (0.00957) (0.00957)
Shared Border 0.769*** 0.770*** 0.769*** 0.769*** 0.769***
(0.0141) (0.0141) (0.0141) (0.0141) (0.0141)
Language 0.232*** 0.232*** 0.233*** 0.232*** 0.234***
(0.0103) (0.0103) (0.0103) (0.0103) (0.0103)
Deposits * Cap. Dep. 0.000183*** 0.000325***
(2.72e-05) (4.16e-05)
Credit * Cap. Dep. 3.54e-05** -8.74e-05***
(1.60e-05) (2.44e-05)
Flexibility * Sales Volatility 0.264*** 0.195***
(0.0201) (0.0213) Legal * Holdup 0.271*** 0.219*** (0.0213) (0.0227) Constant 6.034*** 7.022*** 7.037*** 6.429*** 7.125*** 6.633*** (0.192) (0.195) (0.195) (0.201) (0.195) (0.202) Observations 275,584 275,584 275,584 275,584 275,584 275,584 R-squared 0.072 0.138 0.138 0.139 0.138 0.139
Exporter Dummies yes yes yes yes yes yes
Countries 95 95 95 95 95 95
Industries 78 78 78 78 78 78
27
model. These results provide moderate support for hypothesis 1. Especially the results regarding human capital are strongly in favour, those for physical capital slightly less so. Hypothesis 2, on the other hand deals with consumption goods, which are dealt with in equation (7). The corresponding results are shown in table 6, in the columns 7-12. Here the Heckscher-Ohlin forces are interacted with a dummy for final consumption goods, to examine whether there exists a difference between these goods and all others when it comes to Heckscher-Ohlin type specialisation. The results of the interaction terms are uniformly negative and significant, suggestive of a counter-specialisation effect when compared to other types of goods. In other words, on average, compared to the other types of goods, the Heckscher-Ohlin specialisation is less important for consumption goods.
Extended dataset
The extended dataset includes observations for which the base dataset does not have any data available. These missing observations have been set to zero to reflect the instances where trade did not occur. As explained earlier, analysis based on this dataset combines effects based on the intensive margin of trade (like the base dataset), but also on the extensive margin. However, a major obstacle to analysing this dataset is the prevalence of zero-trade-observations, as a log transformation renders these observations useless. To solve this problem, the value of 1 has been added to all zero-trade. Equations (6) and (7) are re-estimated using OLS and Tobit specifications, the corresponding results are presented in table 7. The first two columns deal with the dataset in which zero values have been replaced by ones and estimation proceeded using OLS, the third and fourth columns with the Tobit estimation.
28
and the quantity of trade. The main variables of interest shows a similar pattern as those in table 6, hypothesis 2 seems to be strongly confirmed, with negative effects for both interactions with physical and human capital specialisation. For intermediate goods, there is once again one effect that is positive, and one that is insignificant; however, the effects are reversed from the table 4 results. Now the interaction term between intermediates and physical capital is significant, while that of human capital is insignificant.
Table 7: Extended dataset
(1) (2) (3) (4)
Ln(value) OLS - Extended
Dataset OLS - Extended Dataset Tobit - Extended Dataset Tobit - Extended Dataset Full Model Full Model Full Model Full Model Interm. Goods Cons. Goods Interm. Goods Cons. Goods
Ln(skill) * Ln(skill int.) 0.209*** 0.152*** 0.778*** 0.527***
(0.00855) (0.00723) (0.0336) (0.0266)
Ln(capital) * Ln(cap. int.) -0.0111*** -0.00568*** -0.0491*** -0.0240***
(0.000383) (0.000334) (0.00165) (0.00140)
C * Ln(skill) * Ln(skill int.) -0.0383*** -0.537***
(0.0127) (0.0600)
C * Ln(capital) * Ln(cap. int.) -0.0102*** -0.0559***
(0.000298) (0.00162)
M * Ln(skill) * Ln(skill int.) 0.0137 0.0539
(0.00951) (0.0416)
M * Ln(capital) * Ln(cap. int.) 0.00182*** 0.00895***
(0.000223) (0.00104)
Ln(distance) -0.397*** -0.397*** -1.914*** -1.911***
(0.00348) (0.00348) (0.0137) (0.0137)
Regional Trade Agreement 2.215*** 2.210*** 4.227*** 4.198***
(0.0106) (0.0106) (0.0271) (0.0270)
Shared Border 1.209*** 1.209*** 1.562*** 1.561***
(0.0168) (0.0168) (0.0367) (0.0366)
Language 0.0556*** 0.0556*** 1.228*** 1.227***
(0.00603) (0.00603) (0.0283) (0.0283)
Deposits * Cap. Dep. 2.87e-05 2.02e-05 0.000115* 7.27e-05
(2.44e-05) (2.44e-05) (5.97e-05) (5.93e-05)
Credit * Cap. Dep. -3.40e-05*** -4.38e-05*** -0.000148*** -0.000185***
(1.12e-05) (1.12e-05) (3.66e-05) (3.63e-05)
Flexibility * Sales Volatility 0.447*** 0.455*** 1.975*** 1.971***
(0.0141) (0.0140) (0.0593) (0.0589) Legal * Holdup -0.196*** -0.380*** -1.027*** -1.708*** (0.0138) (0.0141) (0.0579) (0.0597) Constant 3.568*** 3.370*** 0.163 -0.681* (0.0501) (0.0493) (0.374) (0.371) Observations 1,331,371 1,331,371 1,331,371 1,331,371 R-squared 0.431 0.433
Exporter Dummies yes yes yes yes
Countries 95 95 95 95
Industries 78 78 78 78
29
The Tobit model estimations presented in column 3 & 4 show very similar results to the OLS results in columns 1 & 2. This model uses a type of MLE (maximum likelihood estimation) combined with the principle of a censored dataset to estimate the parameters in an alternative way. Censored data means that the data is essentially censored at a certain max- or minimum, in this case a minimum at zero. The Tobit regression then uses MLE to estimate as if the censored values are in fact values beyond the point of censoring, given their distribution relative to the model (Wooldridge, 2012).
When comparing the results of the OLS with the Tobit estimations, no differences stand out. All the signs and significant variables are the same in both cases; however the effects themselves tend to be of larger magnitude (both positive and negative) in the Tobit regression16. These large magnitudes could be due to the larger variation in the value variable that is allowed in the Tobit specification.
Instrumental variable approach
This section presents the results for a final regression using an instrumental variable approach. A 2-stage least squares method is employed. Human, and physical capital endowments are instrumented for by the GDP per capital level. The required data is obtained from the Penn world tables (Feenstra, Inklaar, & Timmer, 2015).
Table 8 reports results for both the base (columns 1 and 2) and extended (columns 3 and 4) datasets. Examining the first two columns, it becomes obvious that there is little difference with respect to signs and significance between the instrumental variable approach and the initial regressions presented in tables 4 and 5. The most important difference is that both the interaction terms including intermediate goods are now actually positive, and significant. This gives additional support to hypothesis 1, while leaving the support with regards to hypothesis 2 unchanged. Comparing the results of column 3 and 4 to the initial results of the
16 The interpretation of the Tobit coefficients are different to those of the OLS, therefore the numbers don’t
30
estimates using the extended dataset (table 7) little difference can be observed. The signs and significance are almost identical; the only difference being the significant positive effect of the interaction between intermediates and human capital.
Table 8: 2SLS Specification (1) (2) (3) (4) 2SLS - Base Dataset 2SLS - Base Dataset
2SLS - Ext. Dataset 2SLS - Ext. Dataset Full Model Full Model Full Model Full Model
Ln(value) Interm. Goods Cons. Goods Interm. Goods Cons. Goods
Ln(skill) * Ln(skill int.) 0.343*** 0.0787*** 0.502*** 0.463***
(0.0119) (0.0157) (0.0111) (0.0135)
Ln(capital) * Ln(cap. int.) -0.000189 -0.000580 -0.00871*** -0.0175***
(0.000655) (0.000765) (0.000576) (0.000687)
M * Ln(skill) * Ln(skill int.) 0.00118** 0.00268***
(0.000497) (0.000455)
M * Ln(capital) * Ln(cap. int.) 0.202*** 0.0873***
(0.0194) (0.0177)
C * Ln(skill) * Ln(skill int.) -0.0134*** -0.0242***
(0.000745) (0.000553)
C * Ln(capital) * Ln(cap. int.) -0.503*** -0.391***
(0.0271) (0.0214)
Ln(distance) -0.114*** -0.114*** -0.397*** -0.398***
(0.00506) (0.00505) (0.00349) (0.00349)
Regional Trade Agreement 0.595*** 0.600*** 2.206*** 2.216***
(0.00957) (0.00956) (0.0106) (0.0106)
Shared Border 0.769*** 0.769*** 1.211*** 1.213***
(0.0141) (0.0141) (0.0168) (0.0169)
Language 0.234*** 0.234*** 0.0552*** 0.0553***
(0.0103) (0.0103) (0.00603) (0.00603)
Deposits * Cap. Dep. 0.000293*** 0.000282*** -1.45e-05 -6.37e-06
(4.17e-05) (4.16e-05) (2.43e-05) (2.43e-05)
Credit * Cap. Dep. -0.000105*** -0.000128*** -7.60e-05*** -7.05e-05***
(2.44e-05) (2.44e-05) (1.12e-05) (1.12e-05)
Flexibility * Sales Volatility 0.145*** 0.189*** 0.341*** 0.342***
(0.0214) (0.0216) (0.0141) (0.0143) Legal * Holdup 0.221*** 0.288*** -0.366*** -0.0477*** (0.0236) (0.0237) (0.0160) (0.0154) Constant 6.972*** 6.683*** 4.244*** 4.429*** (0.205) (0.206) (0.0554) (0.0577) Observations 275,584 275,584 1,331,371 1,331,371 R-squared 0.138 0.140 0.431 0.430
Exporter Dummies yes yes yes yes
Countries 95 95 95 95
Industries 78 78 78 78
Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1
31 Discussion & Conclusion
The previous sections have presented results from various regressions all aimed at answering the main research question of this thesis. This question is whether, if goods are traded as intermediates or as consumption goods, affects their interaction with Heckscher-Ohlin type specialisation forces. Using these regressions I have sought to show that while the Heckscher-Ohlin relations hold for intermediate type traded goods, this is not the case for final consumption goods. The two hypotheses that I used were presented in the theory section. The tests for hypothesis 1, I examined intermediate goods and interacted them with Heckscher-Ohlin forces, to examine their combined effects on bilateral trade flows. The results varied across the primary specification and robustness checks, with respect to both datasets that represent the intensive margin and the more extensive dataset including both the intensity and incidence of trade. However, the final regression, using an instrumental variables (IV) approach finds exclusively positive evidence for hypothesis 1. In the IV case, both effects interacted with the intermediate goods dummy are positive and significant. Considering the different results leads me to cautiously accept hypothesis 1, stating that trade patterns of intermediate goods are more strongly influenced by Heckscher-Ohlin specialisation forces than other types of trade goods.
Hypothesis 2 on the other hand, shows more clear and consistent results. In all specifications the effects of those terms featuring the consumption goods interaction dummy are negative and significant. Hypothesis 2 states that compared to other goods, trade patterns for consumption goods are affected to a lesser extent by the forces that govern specialisation according to Heckscher-Ohlin theory. The results of the primary specification show strongly significant negative effects, as does the IV. Therefore, on the aggregate, the proof for hypothesis 2 is strong, and can be accepted with relative confidence. This is to say that Heckscher-Ohlin specialisation forces have significantly less effect on the trade patterns of consumption goods than on those of other types of goods.
32
trade patterns of different types of trade goods. Most notably, the trade patterns of intermediate and final consumption goods seem to differ significantly, suggesting that trade patterns of different types of traded goods are affected by factors endowments in different ways. The evidence shows that the specialisation on intermediate goods is influenced by endowments, to a greater extent than are other goods. However, for consumption goods, the opposite is true.
However, despite my best efforts, certain improvements could be made to increase the certainty of the results. Most importantly, changes with respect to data could certainly benefit the quality of the research. As briefly touched upon in the initial sections of the paper, a dataset featuring value added trade data could help improve the accuracy of the trade pattern identification. Most salient would be the additional accuracy with which the distinction between trade in intermediate and consumption goods could be made.
Additionally, it is a common assumption in the literature (Helpman, 1999) that a single country’s industry intensities can be generalised to reflect those of all countries in general. However, data identifying industry intensities per country can provide the model with extra information to allow for more detail in estimation.
33 References
Anderson, J. E., & Marcouiller, D. (2002). Insecurity and the pattern of trade: an empirical investigation. Review of Economics and statistics, 84(2), 342-352.
Baldwin, R. (2006). Globalisation: the great unbundling (s). Economic Council of Finland, 20, 5-47. Becker, R., Gray, W., & Marakov, J. (2013). NBER-CES Manufacturing Industry Database:
Technical Notes . Technical report, National Bureau of Economic Research.
Belsley, D. A., Kuh, E., & Welsch, R. E. (2005). Regression diagnostics: Identifying influential data and sources of collinearity. John Wiley & Sons.
Chor, D. (2010). Unpacking sources of comparative advantage: A quantitative approach. Journal of International Economics, 82(2), 152-167.
Cook, R. D. (1977). Detection of Influential Observation in Linear Regression. Technometrics, 19(1), 15-18.
Costinot, A. (2009). On the origins of comparative advantage. Journal of International Economics, 77(2), 255-264.
Cuñat, A., & Melitz, M. J. (2012). Volatility, Labor Market Flexibility, and the Pattern of Comparative Advantage. Journal of the European Economic Association, 225–254. Dornbusch, R., Fischer, S., & Samuelson, P. A. (1977). Comparative Advantage, Trade, and
Payments in a Ricardian Model with a Continuum of Goods. The American Economic Review, 67(5), 823-839.
Eaton, J., & Kortum, S. (1997). Technology and bilateral trade. No. w6253. National Bureau of Economic Research.
Eaton, J., & Kortum, S. (2002). Technology, geography, and trade. Econometrica, 70(5), 1741-1779. Feenstra, R. C., Inklaar, R., & Timmer, M. P. (2015). The Next Generation of the Penn World Table.
forthcoming American Economic Review, available for download at www.ggdc.net/pwt. Feenstra, R. C., Lipsey, R. E., Deng, H., & Ma, A. C. (2005). World trade flows: 1962-2000. (No.
w11040). National Bureau of Economic Research.
Gerlagh, R., & Mathys, N. (2011). Energy Abundance, Trade and Industry Location. FEEM Working Paper No. 3.2011 , Available at SSRN: http://ssrn.com/abstract=1756688.
Grossman, G. M.-H. (2008). Trading Tasks: A Simple Theory of Offshoring. The American Economic Review, 1978-1997.
Gwartney, J., & Lawson, R. (2004). Economic Freedom of the World. Annual Report.
Heckscher, E. F. (1919). The Effect of Foreign Trade on the Distribution of Income. Ekonomisk tidskrift, 1-32.
