Размещено на http://www.allbest.ru/

The Impact of Internet Development on International Trade in Services

Kalyanov Nikita Alexandrovich

Introduction

The Internet greatly impacts everyday life of a modern man. It transformed a great deal of routine. In addition, Internet is a key enabler of many services from new means of communication to the completely new areas of business. As of its more global impact, there are some theoretical considerations which suggest that the Internet might be affecting the international trade. One of the key determinants of international trade is "trade-costs" - different barriers to trade. Trade costs that are associated with cultural differences and logistic problems limit and reduce the volume of the international services flows. One of the main trade costs is distance, which makes the delivery of the product to consumer expensive and slow. The Internet seems to significantly reduce or even completely remove this burden (in services in particular, take for example a whole range of online consultations/education/management). In addition, the Internet has created the completely new industries - like cloud computing and Software-as-a-Service (SaaS). However, it is not obvious that such changes are visible worldwide on the international trade-flows. internet international service

This research project aims to determine the impact of the Internet on the international service trade. The overall impact of the Internet on trade flows of services is analyzed with the use of several indicators and the possible reasons for such impact in more detail that some previous research. The focus is on the fixed entry barriers to trade. These barriers are found to be accurately predicted by the indicators of origin country (and not the destination one). In addition, the infrastructural indicators are proved to perform worse than the direct measure of users of the Internet. The notable trade pattern of developing countries export to the developed ones is examined and proved to be a significant determinator of trade, in line with previous research. In addition, the backed by data policy recommendations are produced.

Literature review

This project is related to several streams of the international trade literature. First, I review the classical gravity literature that explains the structure of the international trade flows and was primary developed for the manufacturing sector. Second, I look at the application of the gravity models in the services sector and, in particular, I review the closest mine studies that investigate the impact of the Internet on services trade.

Gravity model specification

The basic gravity model is the multiplicative model that explains the trade flow between two countries with some economic, geographic and cultural factors. The model is usually derived from the structural model of bilateral trade flows. In the simplified form of the model equation could be written as follows:

From this equation the title of the model becomes reasonable. The model regresses the bilateral trade flows on the product of `economic masses' (supply and demand) and the square of distance just as in the famous physics formula for the force of gravity, which is:

To conduct a meaningful research the equation is augmented with the variables of interest while the Supply and Demand is proxied by the respected GDPs:

The у is the elasticity parameter of the CES function commonly used as the demand function to derive the equation. The `trade costs between A and B' may include geographical distance, contiguity (common border), similarity of language/religion/legal system, historical ties (such as being in a colony-metropolis relationship or being the colonies of the same metropolis) and the state of being members of the same trade (such as NAFTA NAFTA - North American Free Trade Agreement and other FTA FTA - Free Trade Agreement's) or currency (such as the European Currency Union) unions, policy barriers (e.g. tariffs, quotas, technical standards). The selection of these variables highly depends on the goal of the paper, most frequently the effect of FTA's or currency unions is researched with controls for other factors employed. In addition, the proper specification of the model may require to account for time value of money and/or different currencies.

Gravity model was used to explain trade flows for the first time by Nobel laureate Tinbergen (1962), in a purely empirical paper. His results suggest that Free Trade Agreement (Benelux in particular) has no significant effect on bilateral trade.

An example of an early theoretical derivation of gravity equation is Bergstrand (1985). The author constructs a general equilibrium world trade model and derives a classic gravity equation from there making certain assumptions. However, the author provides empirical evidence that traditional gravity equation is incorrectly specified because it relies on the assumption of perfect international product substitutability. The derivation was further generalized in a subsequent paper Bergstrand (1989) to include deffirentiated-product industries.

Matyas (1997) made one of the earliest calls to use panel data instead of cross-sectional. His another important contribution was a suggestion that exporter, importer and time effects be included in the model to control for other factors. Later it was empirically proved that the omission of any of the three effects may lead to biased estimates and misleading results (Baltagi et al (2003)) and these three effects became a de-facto standard.

Further improving the specification, Egger and Pfaffermayr (2003) argued that the proper specification of a panel gravity model should include not only exporter, importer and time effects, but also time invariant exporter-by-importer (bilateral) interaction effects. Such specification is more general than traditional ones because it accounts for any time invariant bilateral effect. The results were empirically confirmed on panel data for 11 APEC countries from 1982 to 1998 (bilateral effects were significant and accounted for the largest part of variation).

A clean and minimalistic derivation of the gravity equation is presented in Baldwin and Taglioni (2006) along with the three most common errors in model specification. The authors also analyze the possible size of bias due to these mistakes and provide the ways to counter them, which makes the paper resemble a thorough Quick-Start guide to gravity models.

One of the most challenging (and probably most frequently cited) empirical research is the paper by McCallum (1995). The author uses the traditional gravity model and comes to an unrealistic result - Canada's provinces traded enormous 22 times (2200%) more with each other than with U.S. states (all else equal) in 1988. This result have fueled the development of gravity model by provoking future researchers to find errors in the traditional specification of the model.

The example of a solution to McCallum puzzle is the paper by Anderson & van Wincoop (2003). They developed a rigid theoretical derivation of gravity equation (sometimes it is called structural gravity model) and applied this estimate to the puzzle. The authors argue that the so-called "multilateral resistance" (which is theoretically justified measure of average trade barrier of the region) should be included in gravity model. The method yields a much more realistic figure of 20-50% decrease in the trade as the effect of the national border.

There is another paradox (reviewed in Brun et al (2005)) concerning the traditional gravity model - estimated coefficient of distance was found to increase over time on the dataset of 130 countries for 1962-1996, which is counterintuitive (a decrease is anticipated as a result of globalization). This paradox remains even with the inclusion of "multilateral trade resistance" and is only solved by including "augmented" transport cost function (it includes indexes of infrastructure, oil price and composition of trade as arguments). Nevertheless, the authors argue that this decline is mainly due to increased bilateral trade between rich countries, while poor countries become more isolated - in accordance with more reliable customs data on transport cost for several countries.

Another paper in the line of improving gravity model specification is Baier & Bergstrand (2007). The authors criticize some previous papers which tried to analyze the impact of Free Trade Agreements (FTA) on the grounds that trade policies itself is an endogenous variable. They use panel approach and account for trade policy's endogeneity. The result of research suggests that the impact of FTAs is underestimated in most empirical papers and it (on average) doubles two members' bilateral trade after 10 years. A panel of five-year intervals from 1960 to 2000 for 96 countries was used.

The gravity model indeed became a workhorse for studies of international trade and there is an admirable wealth of empirical papers. To effectively grasp the variety of empirical papers the "meta-research" papers and reviews of recent developments was developed. One of the most complete ones is Anderson (2011) which illustrates "the bigger picture" of the gravity model. Another paper of that type is Kepaptsoglou et al (2010). The authors review 10 years of empirical research on the effect of Free Trade Agreements on international trade. In addition, the authors pinpoint some best practices of specifying a gravity model.

It is also important to be aware of another way of analyzing international trade - analysis from firm's perspective. Country-wide analysis has some limitations and several new firm-level theories were developed to explain some observable facts like the results of empirical research on micro datasets that has shown that firms-exporters are larger, more productive, more skill- and capital-intensive and pay higher salaries even before the exporting begins. A detailed review of new theories' insights and new stylized facts are presented in Bernard et al (2007). The authors also show that the effect of distance in a gravity equation is closely interconnected with the number of destinations for a firm and number of exported products for that firm.

All the above-mentioned papers apply the `direct approach' to modelling the trade flows (i.e. use the available indicators to explain the observed trade flows). However, an alternative approach was also developed. The `indirect approach' compares the observed trade flows to a theoretical benchmark trade flows (i.e. the size of trade flows if there were no trade costs) to indirectly estimate the trade costs. This approach could be used to estimate the border-effects as a whole by benchmarking the `phi-ness' of trade (the ratio of bilateral trade flows to inner trade flows between regions).

Gravity model in services

The trade in services is relatively less covered in the literature than trade in goods, but this situation is changing as more data becomes available to the researchers.

Kimura & Lee (2006) confirm the potential of gravity model for explaining services trade. The authors ran gravity models separately for goods and services on the data for 10 OECD member countries for 1999 and 2000. As a result, services trade was found to be better predicted by gravity model than goods trade.

An important paper which justifies the use of distance in modelling of consumption of (at least some) digital goods is Blum (2006). The dataset of every website visited by 2654 American households from Dec. 1999 to Mar. 2000 was used. The authors show that Americans are more likely to visit websites from closer countries, even controlling for a bunch of other variables. However, this effect seems to hold only for taste-dependent digital goods (music, games, pornography) and is statistically insignificant in non-taste-dependent goods (such as software). The authors suggest that distance proxies tastes and therefore it is not appropriate to drop out distance (at least without introducing some other proxy for similar tastes) even for the purely digital products without transportation costs.

Services trade and the Internet

One of the papers most closely related to the project is Freund & Weinhold (2002). The authors try to determine the significance of the Internet's impact on trade of services. Only U.S. data were used and Internet penetration was measured by the number of host sites in a country. The authors claim that a 10% increase in Internet penetration in a foreign country is associated with 1.7% increase in services export growth and 1.1% increase in services import growth (after controlling for GDP and exchange rate fluctuations). I use more data and some more direct indicators of the Internet use comparing to this research. In addition, only the export flows are directly used in model estimation as there are more incentive for agents to report them (i.e. the data should be more accurate) than the import flows.

This research was continued in Choi (2010). The author claims that a 10% increase in the number of the Internet users per hundred people leads to a 0.23-0.42% increase in the aggregated trade of services. Panel data for 151 countries from 1990 to 2006 was used, generalizing the results of Freud and Weinhold (2002). I use the disaggregated trade of services, but for a far less sample size (both in terms of countries and years). In addition, cross-sectional data is used instead of a panel.

The link between Internet penetration and trade flows is also examined in Clarke & Wallsten (2006). The authors research the observed fact that developing countries export more when Internet penetration is higher but only to developed countries, and not to other developing nations. The direction of causation was clarified using instrumental variables and the result suggests that improved access to the Internet increases export in developing countries (but not in developed). The results of this project is close to the result of this paper: the Internet penetration were found to increase export of services and this effect is larger in developing countries but the export of services was considered.

Vemuri & Siddiqi (2009) use panel data for 64 countries from 1985 to 2005 and claim that IT infrastructure and the availability of Internet for commercial transactions positively impact the volume of international trade. The authors use the length of telephone lines per 100, number of PCs per 100 and number of Internet users per 100 as a measures for IT infrastructure and Internet availability. In this project, a more recent data on a subset of trade (trade in services) was used and the infrastructural indicators of the Internet was found to be insignificant to the international trade in services but the number of Internet users per 100 does positively impacts the international trade in services.

In even more recent and extensive research, Meijers (2014) confirms the previous results restating that the use of Internet has a positive impact on international trade and that this impact is more significant in developing countries. Data for 162 countries from 1990 to 2008 was used. The use of Internet was measured as per-capita number of Internet users. The result of this project is qualitatively the same: the use of Internet does have a positive impact on the international trade in services and this impact is higher for developing countries but the more recent data was used and only trade in services was considered.

One of the most recent papers combining rigid theoretical derivation of a gravity equation (i.e. structural gravity model) and empirical estimation of factors affecting service trade is the paper Dark Costs, Missing Data: Shedding Some Light on Services Trade by Anderson et al. (2018). As mentioned in the title authors address the widespread problem of missing data, which is common in services trade in particular. Authors use gravity model on mainly OECD countries to predict the missing trade flows and the predictions turned out to be quite accurate (model was estimated controlling for sectors, countries and time). Apart from these results, the authors also show trade costs tend to vary across different sectors of services. Another relevant result is that the Internet was estimated to be one of the key factors of importance for services trade (among other factors). The number of secure Internet servers per 1 million of population was used as an indicator of the development of the Internet. In this project the mainly OECD countries were also used and some variation was found in sectors of services. However, solely Internet-related factors was considered in this research and the number of secure Internet servers per 1 million of population turned out to be insignificant to international trade (instead, a more direct indicator of the percentage of Internet users was significant).

The Internet is rapidly changing the service industry of the world by increasing the diffusion of some services. An example of analysis of this process from economic viewpoint is Wallsten (2015). The author uses the dataset of over a billion New York City taxi rides, Google Trends data and taxi complaints data from New York and Chicago in 2009-2014 and analyzes the impact of Uber. The author found that Uber's growth is associated with a decline in consumer complaints per taxi trip in New York. This decline is also present in Chicago, but due to more elaborate data it is possible to even determine which type of complaints declined. This result suggest that Uber created (probably indirectly, by stimulating traditional taxis to provide better service) an alternative for consumers who would otherwise be dissatisfied with the service quality. Such massive impact of the Internet on national markets of services lets us make a hypothesis that some larger international impact also exists. This hypothesis of extrapolating the local markets to the international level were confirmed during the project, i.e. the diffusion of the Internet was found to positively affect the international trade in services.

Data

There are three main dimensions of data for gravity models: Countries (for what set of countries trade flows are modelled), Time (for what period(s) trade flows are modelled) and Sector (what particular type of trade flows are modelled. Trade flows are oriented data, later in the text these definitions are used interchangeably: origin - exporter (the country which is delivering services), destination - importer (the country which is receiving services). In addition, `gravity variables' refer to a number of commonly accepted variables for gravity model while `internet variables' refer to a number of variables describing the current state and some pre-requisites for the diffusion of the Internet (a more comprehensive information on these two groups of variables is provided in section `Data processing'). In this research `statistically significant' by default means `statistically significant at level 0.05' unless otherwise stated.

Overview

Gravity model is a kind of model which requires detailed data on trade flows for each country pair in the dataset. That is why availability of data was a serious restriction for the research. In addition, the data on services is historically less accessible due to the lack of common international classifications of services and the overall difficulties in registration of services (i.e. the absence of a physical form). After considering different data sources a limited subset of countries was selected (26 countries) to allow for comparison in time. These countries include mainly OECD members (European countries and several others) The list of countries: Denmark, Germany, Czech Republic, Belgium, Estonia, Latvia, Greece, Hungary, Italy, Luxembourg, Netherlands, Poland, Portugal, Spain, Slovenia, Sweden, United Kingdom, Australia, Canada, New Zealand, Korea (Republic of), Mexico, United States, Chile, Israel, Lithuania..

OECD Stat was used as a source of data related to trade flows. In particular, EBOPS 2010 - Trade in services by partner economy database was used.

Time period of 2010-2015 and 7 sectors of services (Education, Construction, Telecommunications, Insurance, Financial, Health and Advertising services) were selected taking data availability restrictions into account.

Data processing

The following steps were taken for each year in 2010-2015 during the data processing:

1) Construction of the bilateral trade flows

Every trade flow possible between the selected countries was constructed. Existing trade flows were converted to relative basis (export from a particular origin / overall import of destination (including self-flow)) and log-trasformed. Non-existing trade flows were assumed to be zero.

2) Construction of self-to-self trade flows

The OECD STAN database was used for construction of self-to-self trade flows (self-flow = output - export). The services sectors was chosen by the condition of matching the EBOPS classification (it turned out that only 7 of them exactly match). Output was converted to US dollars using average yearly exchange rate for the given year to match the bilateral trade flows originally measured in US dollars.

3) Adding of `gravity variables'.

A number of commonly used `gravity variables' was added to the dataset from CEPII gravity database:

a. Common official language. A dummy-variable that represents whether the two countries have common official (legally proclaimed) language. Is 1 if they do and 0 otherwise. Is a proxy for cultural distance.

b. Great Circle distance between the two countries. A measured in kilometers variable that represents the distance between the capitals of two countries (taking the shape of the Earth into account). The variable was logged before adding to the model.

c. Contiguity. A dummy-variable that represents whether the two countries are direct neighbors (have common border). Is 1 if they do and 0 otherwise.

d. `Mass variable' - a specially constructed variable representing the multiplication of economic masses of partners. Following the common practice, the GDPs of the trade partners are used as a proxies for respective economic variables (supply and demand). Basically, equals GPD(origin) * GDP(destination). The variable was logged before adding to the model.

4) Adding of `internet variables'.

Several variables presumably explaining the diffusion of the Internet and the potential of such a diffusion was added:

a. Number of secure Internet servers per 1 million people in the given year. The variable has three forms: indicator measured at origin, indicator measured at destination and the multiplication of the former two, using the same logic as with `Mass variable'. The variable was logged before adding to the model.

b. Percentage of Internet users in the population, the most direct measure of Internet diffusion in the dataset. The variable has three forms: indicator measured at origin, indicator measured at destination and the multiplication of the former two, using the same logic as with `Mass variable'. The variable was logged before adding to the model.

c. Number of mobile cellular subscriptions per 100 people, infrastructural indicator which indirectly measures the potential of the diffusion of the Internet. The variable has two forms which are measured at origin and destination respectively.

d. Number of fixed telephone subscriptions per 100 people, infrastructural indicator. The variable has two forms which are measured at origin and destination respectively.

e. Broadband_subscriptions - the indicator representing the number of fixed broadband subscriptions per 100 people, infrastructural indicator. The variable has two forms which are measured at origin and destination respectively.

5) Some variables describing the use of different ICT technologies by businesses were added:

a. Percentage of businesses with a broadband connection (both fixed and mobile) for a given year measured at the destination country. The variable indirectly measures the use of Internet by businesses and the potential of such use.

b. Percentage of businesses with a website or homepage at destination country. The variable measures the use of a technology which is highly related to the Internet and is used as a proxy for actual Internet use by businesses.

c. Percentage of businesses with a website or homepage at origin country. The same indicator but measured at origin.

d. Percentage of businesses using an ERP (Enterprise Resource Planning) software at destination for a given year. The selling of a subscription-based EPR and the related services (consulting, maintaining & support, customization) both could be conducted internationally and/or remotely. Thus, the indicator can serve as the proxy for the attractiveness of the destination's market for services conducted via the Internet.

e. Percentage of businesses using a CRM (Customer Relationship Management) software at destination. The selling of a subscription-based CRM and the related services both could be conducted internationally and/or remotely. Thus, the indicator can serve as the proxy for the attractiveness of the destination's market for services conducted via the Internet.

All the variables related to ICT use by businesses were logged before adding to the model.

6) The construction of control variables

The exporter-, importer- and industry-(services sector-) dummies were constructed to control for other factors. In addition, the dummies for origin being a developed country and destination being a developed country were constructed to differentiate the influence of the Internet on developed countries and developing ones separately.

Zeroes handling

The gravity model is specified on every possible trade flow while actually very few of them are non-zero - the percentage of non-zero trade flows was 27.8% in 2010 and 32.5% in 2015. The selection problem (prevalence of zeroes) could significantly bias the result, as most estimation methods do not address this. Thus, a special treatment of zeroes is needed which is achieved by applying a two-step model (described in detail in the next section). This way the fixed entry effects of trading between countries are analyzed. The first step of modelling regresses the dummy `trade flows are positive (not 0)' on some factors thus describing the way countries select to trade or not to trade with other countries. However, an alternative specification could have been applied by going down to the firm level and modelling the self-selection of firms into exporting activities. This alternative approach is applied for instance in Helpman, Melitz, Rubinstein (2008) and Baldwin, Harrigan (2011).

Model specification

From the theoretical perspective the models used in the research are mainly based on `structural gravity model' concept, developed by Anderson (Anderson et al, 2018). The theory is based on the assumption of product differentiation by place of origin (Armington (1969) and commonly used CES preferences to derive a system of gravity equations.

*i - origin, j - destination, k -services sector. E - expenditure (demand), Y - sales (supply), t>=1 - variable trade costs, у - trade elasticity of substitution. П and P - theoretical constructs (capture general equilibrium trade cost effects).

The estimated coefficients are actually coefficients itself multiplied by elasticity of demand (sigma parameter from CES). Consequently, the estimated coefficients bear the influence of elasticity, however elasticity could be indirectly estimated using the "usual" real coefficients found in literature.

There are two main model types referenced in the literature as `gravity model', these are `panel gravity model' and `cross-sectional gravity model'. For this research the cross-sectional approach was chosen on the grounds of the Internet being one of the most fast-paced technologies of our time. The fast and sweeping trends of the Internet diffusion could get lost and stay unnoticed with panel approach. Instead, the trends in time are analyzed separately using the estimation result of separate years. In addition, as Baldwin and Taglioni (2006) prove, the panel approach is more likely to suffer from bias due to omission of `gravity unconstant' because this bias could not be fixed only with nation dummies as panel data also has time dimension. Although this can be fixed with pair dummies, such fix makes imposible to estimate the time invariant effects (such as distance).

The models are specified using the two-step approached which proved to yield meaningful results on the dataset. In the preliminary stage the tradeflows are `dummified', i.e. 1 if a trade flow exists, 0 otherwise. In the first step of estimation these dummified trade flows are regressed on a number of traditional gravity variables using probit estimation. The particular equation of the model is

Such specification could capture two effects: fixed entry barriers so that countries choose not to trade (these are proxied by distance) and the reporting problems when the trade flows do exist in reality but are not reported for some reason (this is captured by dummies). Then the predictions of the probit model are obtained and logged. On the second step of estimation the relative trade flows in logs are regressed on the (logged) predicted probability of trade (i.e. prediction of a probit model), a number of gravity variables and some internet-related variables of interest. The particular form of the model may vary in each case. Also control for exporter-, importer- and industry- fixed effects are used in order to account for other country-specific and industry-specific fixed effects (such as the variation of the Internet use by industry).

GATS breakdown

GATS is the international classification of international trade; it divides all international trade into 4 `modes' by the type of interaction between buyer and seller. Mode 1 represents the cross-border supply without any juridical presence at destination (e.g. selling via the Internet), Mode 2 covers the cases where customers travel to a country other than their home country and buy services there, Mode 3 covers the cases when there is a commercial presence of seller at the country where the buyer is located (e.g. a subsidiary) and Mode 4 represents the case when natural persons from seller are physically present at the buyers country (e.g. travel to provide services). The breakdown of services traded via the Internet by GATS modes is another point of interest in this topic.

According to Rueda-Cantuche, Kerner et al (2016) the Health services and Education services are primary traded at GATS Mode 2 (at least from EU-28 to other countries). These type of services does not solely represent the online medicine and online education but rather the classical forms of such services (studying abroad and going for health trips) account for the majority of trade flows. The other type of services research are primary traded at GATS mode 1.

Key results

The key results are discussed consequently, from the simpler model specifications to more advanced ones.

Baseline model, (1). One of the simpliest models of international trade flows, is aimed at determining if internet's diffusion influences international trade flows. The model is estimated using 2-step approach described earlier. On the first step the dummy "Trade Flow exists" is regressed on distance (logarithmic) and exporter-, importer- and industry- control dummies. On the second step the relative logged trade flows are regressed on the predicted probability from the first step, logarithmic distance and two of `internet variables' most directly representing the Internet's diffusion. These two variables are log("number of secure internet servers per 1 million population at origin" * "number of secure internet servers per 1 million population at destination") and log("% of internet users in population at origin" * "% of internet users at destination"). The method of combining variables for origin and destination with multiplication allows for carefully considering both sides of interaction at the same time and is analogous to the approach commonly used to construct so called "Mass variable" (i.e. log(GDP at origin * GDP at destination)). Model is estimated on the data for 2010 year using OLS estimation method.

The estimated equation:

Baseline model

Table 1. Model 1 estimation results.

====================================================================

Dependent variable:

---------------------------

Trade flow

--------------------------------------------------------------------

Predicted Probability of trade 0.161*** (0.021)

Distance -0.101*** (0.036)

Number of secure Internet servers per 1M -0.177 (0.121)

Percentage of Internet users -3.362*** (1.275)

Constant 29.579*** (10.966)

--------------------------------------------------------------------

Observations 2,655

R2 0.387

Adjusted R2 0.376

Residual Std. Error 2.943 (df = 2606)

F Statistic 34.335*** (df = 48; 2606)

====================================================================

Note: *p<0.1; **p<0.05; ***p<0.01

The model is overall significant and has an adjusted R-squared of 0.38. As a result of model's estimation, The percentage of Internet users at origin and destination combined is proved to be a negative and statistically significant factor of international trade in the 7 services sectors in a year 2010 among the selected countries. I.e. the higher the percentage of Internet users in the partner countries the lower trade in services between them. At the same time, the number of secure Internet servers per 1 million population at origin and destination combined is also negatively affecting the international trade but is statistically insignificant. The coefficients for distance and predicted probability of trade are both significant and have expected signs. The results are somewhat unusual and suggests a hidden factor in data.

In an effort to explain the unusual results of estimating model 1 the `internet variables' were decomposed to separate origin- and destination-specific values and the control for developed-developing country was introduced. Some earlier literature cite the developed-developing division as a major differentiator of international trade (e.g. Clarke & Wallsten, 2006). Also the other standard `gravity variables' are added to the model for proper control together with industry dummies controlling for sectoral effects.

Table 2. Model 2 estimation results.

==============================================================================

Dependent variable:

---------------------------

Trade flow

------------------------------------------------------------------------------

Predicted Probability of trade 0.166*** (0.020)

Distance -0.025 (0.061)

Common official language 1.300*** (0.347)

Contiguity -0.637** (0.280)

Mass variable 8.208*** (1.107)

Origin is developed 37.738*** (5.066)

Destination is developed 29.070*** (3.847)

Number of secure Internet servers per 1M at origin -20.936*** (2.779)

Percentage of Internet users at origin 45.155*** (4.866)

Constant -544.469*** (67.134)

------------------------------------------------------------------------------

Observations 2,655

R2 0.391

Adjusted R2 0.379

Residual Std. Error 2.935 (df = 2604)

F Statistic 33.452*** (df = 50; 2604)

==============================================================================

Note: *p<0.1; **p<0.05; ***p<0.01

After estimating the updated model a number of interesting insights could be discovered. The model is statistically significant as a whole and has an adjusted R-squared of 0.35, which is higher than the initial model's value. The dummy-variable `origin country is developed' along with the dummy-variable `destination country is developed' are both statistically significant and have positive coefficients which suggests increased trade between developed countries. However this could be the effect of data (only 4 countries out of 26 are developed) and further analysis is needed. The variable `number of secure internet servers per 1 million population at origin' is statistically significant but has a negative coefficient while `percentage of internet users at origin' is also significant but has a positive coefficient. This result is contradicting the common intuition; one possible explanation is that the percentage of the internet users is a more straightforward way to measure the actual diffusion of Internet at origin that the number of secure internet servers per 1 million population which is likely highly influenced by outsourcing/cloud servers when the number of servers in a country does not necessary represent the actual internet diffusion. As for the standard `gravity variables', they have the expected sign with the exception of contiguity. In addition distance turned out to be insignificant predictor of trade.

The multicollinearity tests were also performed using several criteria: VIF (Variance Inflation Factor), TOL (tolerance, is reverse of VIF), Klein's rule, Farrar and Glauber F-test (Wi), F and R-squared relation (Fi), Leamer's method and CVIF (Corrected VIF).

Table 3. Multicollinearity test statistics

All Individual Multicollinearity Diagnostics Result

VIF TOL Wi Fi Leamer CVIF Klein

Predicted Probability 1.1460 0.8726 48.3016 55.2227 0.9341 -10.5251 0

Distance 2.7652 0.3616 583.8389 667.4966 0.6014 -25.3953 1

Common language 2.1959 0.4554 395.5393 452.2157 0.6748 -20.1668 1

Contiguity 2.3264 0.4298 438.7086 501.5707 0.6556 -21.3655 1

Mass variable 1.4149 0.7067 137.2441 156.9097 0.8407 -12.9947 0

Origin developed 2.0592 0.4856 350.3468 400.5476 0.6969 -18.9119 0

Destination developed 1.0539 0.9488 17.8317 20.3868 0.9741 -9.6790 0

Servers (origin) 6.3145 0.1584 1757.7558 2009.6230 0.3980 -57.9913 1

Internet users (origin) 5.9043 0.1694 1622.1097 1854.5402 0.4115 -54.2248 1

1 --> COLLINEARITY is detected by the test

0 --> COLLINEARITY is not detected by the test

As the results suggest, the number of secure Internet servers per 1M population at origin is somewhat correlated with the percentage of Internet users at origin, but the VIF's are not higher that the commonly accepted threshold of 10. It means that the two variables indicate characteristics that are strongly interconnected with each other and probably only one of the variables should be used in the model (or an artificial composition of them).

Table 4. Multicollinearity tests results.

All Individual Multicollinearity Diagnostics in 0 or 1

VIF TOL Wi Fi Leamer CVIF Klein

Predicted Probability 0 0 1 1 0 0 0

Distance 0 0 1 1 0 0 1

Common language 0 0 1 1 0 0 1

Contiguity 0 0 1 1 0 0 1

Mass variable 0 0 1 1 0 0 0

Origin developed 0 0 1 1 0 0 0

Destination developed 0 0 1 1 0 0 0

Servers (origin) 0 0 1 1 0 0 1

Internet users (origin) 0 0 1 1 0 0 1

1 --> COLLINEARITY is detected by the test

0 --> COLLINEARITY is not detected by the test

Only Klein's rule detects multicollinearity in the two particular Internet variables.

These two facts could be a part of some larger geographical trends in services trade or the result of the developed-developing trade pattern as developing countries are usually not contingent and far away geographically from the developed countries. This hypothesis could be verified by introducing some interaction effects between `internet variables' and `developed dummies'. Other variables are significant and have expected signs.

A similar model was constructed for `internet variables' at destination.

Table 5. Model 3 estimation results.

===============================================================================

Dependent variable:

-----------------------

Trade flow

-------------------------------------------------------------------------------

Predicted Probability of trade 0.166*** (0.020)

Distance -0.025 (0.061)

Common official language 1.300*** (0.347)

Contiguity -0.637** (0.280)

Mass variable 0.092 (0.064)

Origin is developed -3.509*** (0.358)

Destination is developed 0.601 (0.777)

Number of secure Internet servers per 1M at destination -0.509* (0.260)

Percentage of Internet users at destination -3.087 (2.277)

Constant 13.555 (10.248)

-------------------------------------------------------------------------------

Observations 2,655

R2 0.391

Adjusted R2 0.379

Residual Std. Error 2.935 (df = 2604)

F Statistic 33.452*** (df = 50; 2604)

===============================================================================

Note: *p<0.1; **p<0.05; ***p<0.01

Overall, coefficients significance and signs are very much resembling the previous model, but the variables `number of secure internet servers per 1 million population at destination' and `percentage of internet users at destination' are insignificant, which suggests that destination's internet diffusion is not the key determinant of international trade in services but rather origin's internet diffusion is (or some more complex relationship between them).

Infrastructural factors of the diffusion of the Internet

A search for more indirect factors of internet's diffusion was conducted. The UN's ICT Development index (UN ITU, 2014) was used as a guideline for such factors. The expert-constructed index states three main categories affecting the ICT development of a country: Access (ICT readiness, includes infrastructure and access indicators), Use (ICT intensity, includes intensity and usage indicators) and Skills (measures ICT capability or skills needed to use the technologies). After taking data availability restrictions into account the three potential indicators were obtained: Number of mobile cellular subscriptions per 100 people, Number of fixed telephone subscriptions per 100 people and Number of fixed broadband subscriptions per 100 people.

Table 6. Model 4 estimation results.

===============================================================================

Dependent variable:

------------------------

Trade flow

-------------------------------------------------------------------------------

Predicted Probability of trade 0.166*** (0.020)

Distance -0.025 (0.061)

Common official language 1.300*** (0.347)

Contiguity -0.637** (0.280)

Mass variable 0.221** (0.093)

Origin is developed 0.796** (0.392)

Destination is developed 1.953*** (0.505)

Percentage of Internet users at origin 24.538*** (2.537)

Number of mobile cellular subscriptions per 100 at origin 8.132*** (1.079)

Constant -156.172*** (16.510)

-------------------------------------------------------------------------------

Observations 2,655

R2 0.391

Adjusted R2 0.379

Residual Std. Error 2.935 (df = 2604)

F Statistic 33.452*** (df = 50; 2604)

===============================================================================

Note: *p<0.1; **p<0.05; ***p<0.01

Number of mobile cellular subscriptions per 100 people is a significant and positive factor of international trade in services, the other model's coefficients qualitatively did not change. The number of mobile subscriptions is a more infrastructural and indirect indicator of the internet's diffusion because mobile subscriptions often (but not always) contain a data usage allowance. At the same time, the other two infrastructural variables seem to be the unlikely predictors of international trade in services (their coefficients are insignificant and/or take the explanatory power from other variables making them insignificant). This result highlights the importance of mobile users in the modern Internet (so-called `mobile revolution').

Sectoral-specific impact of the Internet

Next, the interaction effects for the sectors of services are constructed. The initial hypothesis is that all interaction effects would be negative comparing to the baseline sector of the information services, which is the easiest one to trade via the Internet.

Table 7. Model 5 estimation results.

===============================================================================

Dependent variable:

-------------------------

Trade flow

-------------------------------------------------------------------------------

Predicted Probability of trade 0.159*** (0.021)

Distance -0.029 (0.061)

Common official language 1.295*** (0.348)

Contiguity -0.634** (0.281)

Mass variable 6.766*** (0.614)

Origin is developed 31.035*** (3.000)

Destination is developed 33.265*** (3.175)

Percentage of Internet users at origin 40.732*** (3.804)

Percentage of Internet users at destination 30.947*** (3.642)

Number of secure Internet servers per 1M at origin -17.335*** (1.601)

Number of secure Internet servers per 1M at destination -17.407*** (1.600)

Number of mobile cellular subscriptions per 100 at origin 1.195 (0.913)

Number of mobile cellular subscriptions per 100 at destination 0.723 (0.813)

Education industry X Internet servers at origin 0.426*** (0.151)

Construction industry X Internet servers at origin 0.149 (0.150)

Telecom industry X Internet servers at origin 0.035 (0.150)

Insurance industry X Internet servers at origin -0.017 (0.150)

Financial industry X Internet servers at origin -0.051 (0.150)

Health industry X Internet servers at origin 0.454*** (0.151)

Constant -519.182*** (50.166)

-------------------------------------------------------------------------------

Observations 2,655

R2 0.387

Adjusted R2 0.375

Residual Std. Error 2.945 (df = 2604)

F Statistic 32.869*** (df = 50; 2604)

===============================================================================

Note: *p<0.1; **p<0.05; ***p<0.01

After estimating model 5 a few interesting results could be noticed. First, even after decomposing the combined Percentage of Internet users and Number of secure Internet servers per 1M variables into separate values for origin and destination qualitatively the results remain unchanged. Both Percentage of Internet users at origin and Percentage of Internet users at destination have significant and positive coefficients, while both Number of secure Internet server per 1M at origin and the same indicator at destination have significant and negative coefficients. Both Number of mobile cellular subscriptions at origin and the same number at destination are insignificant suggesting that this variable is not a significant predictor by itself, but possibly is significant in combination with some other variables. The interaction effects between industries and the relative number of internet servers at origin yielded interesting results. Every interaction effect is insignificant except the effects for health services and education services, i.e. only these two kinds of services differ from the base sector of Information services. This result could be connected to GATS breakdown. The Health and Education services are often traded via GATS mode 2 (consumption abroad) while the other types of services are mostly mode 3 (at least from EU-28 to other countries) according to Rueda-Cantuche, Kerner et al (2016). This suggests that the Internet influences the GATS mode 2 more than the other modes probably by making the cross-border search for services and deal negotiating easier. The other coefficients are qualitatively the same as in the previous model.

Table 8. Model 6 estimation results.

===============================================================================

Dependent variable:

---------------------------

Trade flow

-------------------------------------------------------------------------------

Predicted Probability of trade 0.167*** (0.021)

Distance -0.025 (0.061)

Common official language 1.300*** (0.347)

Contiguity -0.637** (0.280)

Mass variable 6.944*** (0.614)

Origin is developed 31.892*** (3.002)

Destination is developed 34.156*** (3.177)

Percentage of Internet users at origin 41.639*** (3.814)

Percentage of Internet users at destination 31.953*** (3.644)

Number of secure Internet servers per 1M at origin -17.621*** (1.593)

Number of secure Internet servers per 1M at destination -17.853*** (1.601)

Number of mobile cellular subscriptions per 100 at origin 1.298 (0.911)

Number of mobile cellular subscriptions per 100 at destination 0.812 (0.812)

Education industry X Internet users at origin 0.679*** (0.234)

Construction industry X Internet users at origin 0.245 (0.234)

Telecom industry X Internet users at origin 0.095 (0.234)

Insurance industry X Internet users at origin 0.018 (0.234)

Financial industry X Internet users at origin -0.040 (0.234)

Health industry X Internet users at origin 0.703*** (0.234)

Constant -534.925*** (50.228)

-------------------------------------------------------------------------------

Observations 2,655

R2 0.390

Adjusted R2 0.378

Residual Std. Error 2.938 (df = 2604)

F Statistic 33.298*** (df = 50; 2604)

===============================================================================

Note: *p<0.1; **p<0.05; ***p<0.01

In model 6 the interaction effects between industry type and percentage of Internet users at origin are estimated. The results do not qualitatively differ from the model 5 (interaction effects between industry type and number of internet servers per 1 million population).

Table 9. Model 7 estimation results.

===============================================================================

Dependent variable:

---------------------------

Trade flow

-------------------------------------------------------------------------------

Predicted Probability of trade 0.165*** (0.020)

Distance -0.026 (0.061)

Common official language 1.298*** (0.347)

Contiguity -0.637** (0.280)

Mass variable 6.900*** (0.612)

Origin is developed 31.674*** (2.991)

Destination is developed 33.934*** (3.166)

Percentage of Internet users at origin 41.623*** (3.792)

Percentage of Internet users at destination 31.700*** (3.632)

Number of secure Internet servers per 1M at origin -17.508*** (1.587)

Number of secure Internet servers per 1M at destination -17.743*** (1.595)

...