Cloud computing and extensive margins of exports: Evidence for manufacturing firms from 27 EU countries

The use of cloud computing by firms can be expected to go hand in hand with higher productivity, more innovations, and lower costs, and, therefore, should be positively related to export activities. Empirical evidence on the link between cloud computing and exports, however, is missing. This paper uses firm level data for manufacturing enterprises from the 27 member countries of the European Union taken from the Flash Eurobarometer 486 survey conducted in February – May 2020 to investigate this link. Applying standard parametric econometric models and a new machine-learning estimator, Kernel-Regularized Least Squares (KRLS), we find that firms which use cloud computing do more often export, do more often export to various destinations all over the world, and do export to more different destinations. The estimated cloud computing premium for extensive margins of exports is statistically highly significant after controlling for firm size, firm age, patents, and country. Furthermore, the size of this premium can be considered to be large. Extensive margins of exports and the use of cloud computing are positively related.


Motivation
Digital technologies like artificial intelligence, cloud computing, the use of robots to automate processes, or big data analytics, are more and more widely applied by innovative firms.However, comprehensive empirical evidence on the links between the use of digital technologies and various dimensions of firm performance seems to be lacking.A case in point is the role of these technologies for export activities of firms.In their comprehensive discussion of artificial intelligence (AI) and international trade Goldfarb and Trefler (2018, p. 1) state that "even to the extent that progress has been made in understanding the impact of AI, we remain largely uninformed about its international dimensions.This is to our great loss." 1  This note attempts to contribute to closing one of these gaps by looking at differences in extensive margins of exports between firms that use or do not use cloud computing.To the best of my knowledge it is the first empirical study to do so. 2  We expect these differences to be positive for firms that use cloud computing for three reasons: First, Chen and Volpe Martincus (2022, p. 16) report that cloud computing has significantly boosted firm productivity over the last decade.Evidence in line with this is found in DeStefano, Kneller and Timmis (2024) and Haucap, Fritz and Thorwarth (2022).According to a large empirical literature that uses firm level data from many different countries productivity and export activities in firms are positively related (Ferencz, López-González and Garcıá 2022, p. 12; see Wagner 2007 for a survey of the empirical literature ).
Second, Haucap, Fritz and Thorwarth (2022, p.6) report evidence that a large share of companies stated in a survey that they were able to develop new innovative products and services through the use of cloud computing.It can be considered as a stylized fact that innovations are positively related to export activities of firms from developed industrial countries.
Third, Chen and Volpe Martincus (2022, p. 16) point out that cloud computing lowers the need for businesses to make upfront investments in hardware and software and maintain IT infrastructure.By making IT services available on-demand on a pay-as-you-go basis, cloud computing transforms IT input costs from large, centralized sunk costs into variable costs (Chen and Volpe Martincus (2022, p.1).This can be expected to lead to a cost advantage for firms that use cloud computing and to an increase in their international competitiveness.Furthermore, we know that international firm activities, including exports, tend to go hand in hand with presence of the firms on the web (see Wagner (2022)).If cloud computing lowers the costs of setting up and maintaining web presence, it lowers the costs of exporting, too.All this can be expected to increase export activities.
This note is the first empirical study that looks at differences in exports between manufacturing enterprises from 27 member countries of the European Union that use or do not use cloud computing.In doing so it adds to our understanding of the role of cloud computing in exports by presenting evidence for firms from a large number of countries.Furthermore, we report results for various extensive margins of exports beyond participation in exporting by looking at exports to seven distinct areas of the world market for goods.It should be pointed out that a new machine-learning estimator, Kernel-Regularized Least Squares (KRLS), is applied as a robustness check besides standard parametric econometric models.
To anticipate the most important result we find that firms which use cloud computing do more often export, do more often export to various destinations all over the world, and do export to more different destinations.The estimated cloud computing premium for exports is statistically highly significant after controlling for firm size, firm 1 See Ferencz, López-González and Garcıá (2022), Goldfarb and Trefler (2018) and Meltzer (2018) for a discussion of various aspects of the relations between artificial intelligence and international trade and Wagner (2023) for a study on the use of big data analytics and exports of firms from 27 EU countries. 2See the comprehensive recent survey of the economic literature on digitalization and globalization by Chen and Volpe Martincus (2022) -no existing evidence on the effects of cloud computing on trade (or foreign direct investment) is reported (see table 2 on p. 19).Exports are not looked at in the important study on cloud computing and firm growth by DeStefano, Kneller and Timmis (2024).104 age, patents, and country.Furthermore, the size of this premium can be considered to be large.The take-home message, therefore, is that extensive margins of exports and the use of cloud computing are positively related.
The rest of the paper is organized as follows.Section 2 introduces the data used and discusses the export activities that are looked at.Section 3 reports results from the econometric investigation.Section 4 concludes.

Data and discussion of variables
The firm level data used in this study are taken from the Flash Eurobarometer 486 survey conducted in February -May 2020.Note that while the data were collected at the start of the COVID-19 pandemic, information on export activities relates to the year 2019, the year before the pandemic.We use data for firms from the 27 member states of the European Union in 2020 (i.e., firms from the UK are no longer included in the sample).The sample covers 2,355 firms from manufacturing industries (included in NACE section C); the numbers of firms by country are reported in the appendix table.
In the survey firms were asked in question Q23_2 whether they introduced cloud computing, i.e. the use of remote servers via the internet for storage of files or processing of data.Firms that answered in the affirmative are classified as users of cloud computing.Descriptive evidence is reported in Table 1, showing a share of 44.8 percent of firms using cloud computing.
In the empirical study we look at various measures of export activity of firms:3 First, firms were asked in question Q11_1 whether they exported any goods (or not) in 2019.Firms are classified as exporters or non-exporters based thereon.Descriptive evidence is reported in Table 1, showing a share of 64.5 percent of exporters.Second, firms were asked in questions Q11_2 to Q11_8 whether they exported goods in 2019 to the following destinations: Other EU countries; other European countries outside the EU (including Russia); North America; Latin America; China; other countries from Asia and the Pacific; countries from the Middle East and Africa.Descriptive evidence is reported in Table 1, showing that 61.8 percent of firms exported to countries from the EU, while 29.2 percent exported to other European countries.The other destinations follow with shares between some 10 percent and about 16 percent.Exporters to each destination are investigated separately.
Third, from the evidence reported for exports to the seven destinations mentioned for each exporting firm the number of different destinations exported to is calculated.The share of firms by number of export destinations is reported in Table 2.Not surprisingly, most exporters serve one or two destinations only, but there are quite some firms that export to more (or even all) destinations.In the empirical investigation of the link between the use of cloud computing and exports we control for three firm characteristics that are known to be positively linked with exports: firm age (measured in years, based on the answer given to question Q1), firm size (measured as the number of employees -excluding the owners -at the time of the survey; see question Q2A), and whether the firms has a patent or a patent application pending (see question Q9_6). 4Descriptive statistics are again reported in Table 1.
Furthermore, in the empirical investigations the country of origin of the firms is controlled for by including a full set of country dummy variables.

Testing for cloud computing premium in export activities
To test for the difference in the types of export activities listed in section 2 between firms that do and do not use cloud computing, and to document the size of these differences, an empirical approach is applied that modifies a standard approach used in hundreds of empirical investigations on the differences between exporters and nonexporters that has been introduced by Bernard andJensen (1995, 1999).Studies of this type use data for firms to compute the so-called exporter premium, defined as the ceteris paribus percentage difference of a firm characteristic -e.g.labour productivity -between exporters and non-exporters.
Here we look at differences between firms that do and that do not use cloud computing (instead of differences between exporters and non-exporters) and are interested in the existence and size of a cloud computing premium in export activities (instead of an exporter premium in various forms of firm performance like productivity). 5 For export activities that are measured by dummy variables (the decision to export or not, and the decision to export to one of the seven export destinations listed in section 2) the empirical model in (1) is estimated by Probit.

106
(1) Indicatori = a + ß Cloud computingi + c Controli + ei where i is the index of the firm, Indicator is a dummy variable for the use or not of a type of export activity, Cloud computing is a dummy variable for the use of cloud computing by the firm (1 if the firm uses it, 0 else), Control is a vector of control variables (that consists of measures of firm age, firm size, and patents, and dummy variables for countries), and e is an error term.The cloud computing premium is computed as the estimated average marginal effects of the cloud computing dummy variable.
For the number of export destinations, (1) becomes (2) (2) Numberi = a + ß Cloud computingi + c Controli + ei where i is the index of the firm, number is the number of export destinations, Cloud computing is a dummy variable for the use of cloud computing by the firm (1 if the firm uses it, 0 else), Control is a vector of control variables (that consists of measures of firm age, firm size, and patents, and dummy variables for countries), and e is an error term.The model ( 2) is estimated by OLS.The cloud computing premium is the estimated coefficient ß; it shows the average difference between firms that use and do not use cloud computing, controlling for firm age, firm size, patents, and country of origin of the firm.

Results from standard parametric models
In a first step, the empirical models outlined above are estimated using standard parametric econometric models with Probit or OLS.Results are reported in the first columns of tables 3 -5.
The big picture that is shown is crystal clear: Firms that use cloud computing are more often exporters.Furthermore, firms with cloud computing do more often export to any of the seven different destinations looked at here, and do export to a larger number of destinations.Each estimated cloud computing premium is statistically highly significant ceteris paribus after controlling for firm age, firm size, patents, and country of origin of the firms. 6urthermore, the size of this premium can be considered to be large -the estimated marginal effects reported in the first columns of Table 3 and Table 4 are in the order of magnitude of five to eleven percent, and from Table 5 we see that the average difference in the number of destinations exported to is 0.455 in favour of firms that use cloud computing (with an average value of 1.544 destinations for all firms).

Results from Kernel-Regularized Lest Squares (KRLS) models
In the standard parametric models used in section 3.1 the firm characteristics that explain the export margins enter the empirical model in linear form.This functional form which is used in hundreds of empirical studies for margins of exports, however, is rather restrictive.If any non-linear relationships (like quadratic terms or higher order polynomials, or interaction terms) do matter and if they are ignored in the specification of the empirical model this leads to biased results.Researchers, however, can never be sure that all possible relevant non-linear relationships are taken care of in their chosen specifications.In a robustness check of the results from the standard parametric models, therefore, this note uses the Kernel-Regularized Least Squares (KRLS) estimator to deal with this issue.KRLS is a machine learning method that learns the functional form from the data.It has been introduced in Hainmueller andHazlett (2014) andFerwerda, Hainmueller andHazlett (2017), and used to estimate empirical models for margins of trade for the first time in Wagner (2024) 7 .
While a comprehensive discussion of the Kernel-Regularized Least Squares (KRLS) estimator is far beyond the scope of this applied note, a short outline of some of the important features and characteristics might help to 109 understand why this estimator can be considered as an extremely helpful addition to the box of tools of empirical trade economists (se Wagner (2024b)).For any details the reader is referred to the original papers by Hainmueller andHazlett (2014) andFernwerda, Hainmueller andHazlett (2017).
The main contribution of the KRLS estimator is that it allows the researcher to estimate regression-type models without making any assumption regarding the functional form (or doing a specification search to find the best fitting functional form).As detailed in Hainmueller and Hazlett (2014) the method constructs a flexible hypothesis space using kernels as radial basis functions and then finds the best-fitting surface in this space by minimizing a complexity-penalized least squares problem.Ferwerda, Hainmueller and Hazlett (2017) point out that the KRLS method can be thought of in the "similarity-based view" in two stages.In the first stage, it fits functions using kernels, based on the assumption that there is useful information embedded in how similar a given observation is to other observations in the dataset.In the second stage, it utilizes regularization, which gives preference to simpler functions (see Ferwerda, Hainmueller and Hazlett (2017), p.3).
KRLS works well both with continuous outcomes and with binary outcomes.It is easy to apply in Stata using the krls program provided in Ferwerda, Hainmueller and Hazlett (2017).Instead of doing a tedious specification search that does not guarantee a successful result, users simply pass the outcome variable and the matrix of covariates to the KRLS estimator which then learns the target function from the data.As shown in Hainmueller and Hazlett (2014), the KRLS estimator has desirable statistical properties, including unbiasedness, consistency, and asymptotic normality under mild regularity conditions.An additional advantage of KRLS is that it provides closedform estimates of the pointwise derivatives that characterize the marginal effect of each covariate at each data point in the covariate space (see Ferwerda, Hainmueller and Hazlett (2017), p. 11).These estimates can be used to examine the heterogeneity of the marginal effects.
Therefore, KRLS is suitable to estimate empirical models when the correct functional form is not known for sure -which is usually the case because we do not know which polynomials or interaction terms matter for correctly modelling the relation between the covariates and the outcome variable.
Results for an application of KRLS to the models for margins of exports are reported in the second to fifth columns of tables 3 -5.
The big picture that is shown is again crystal clear, and it is identical to the one shown by the standard parametric models: Firms that use cloud computing are more often exporters, do more often export to any of the different destinations, and do export to a larger number of destinations.Each estimated premium is statistically highly significant ceteris paribus after controlling for firm age, firm size, patents, and country of origin of the firms.8Furthermore, the size of this premium can again be considered to be large, although the estimated average marginal effects tend to be somewhat smaller here than in the standard parametric models.The difference in the size of the average marginal effects can be explained by the fact that the parametric model in column 1 imposes a restrictive functional form in the shape of the estimated relationships, while KRLS estimated this relationship without imposing a functional form.
An additional advantage of KRLS compared to the parametric models used in the original estimation is that it provides closed-form estimates of the pointwise derivatives that characterize the marginal effect of each covariate at each data point in the covariate space (see Ferwerda, Hainmueller and Hazlett (2017), p. 11).The last three columns of tables 3 -5 report the marginal effects estimated by KRLS at the 1 st quartile, at the median, and at the 3 rd quartile.We can clearly see the heterogeneity in the marginal effects.The estimated marginal effects differ widely over the quartiles and tend to increase for all variables considered here.This shows the nonlinearity and heterogeneity of the relationship between the covariates and the share of exports in total sales.
Source: Own calculation based on data from Flash Eurobarometer 486.

Table 2 .
Share of Firms by Number of Export Destinations.
Source: Own calculation based of data from Flash Eurobarometer 486.

Table 3 .
Empirical results, Part I: Export participation.

Table 4 .
Empirical results, Part II: Exporter by destination.Note: Probit reports average marginal effects from a model estimated by ML Probit.KRLS reports average marginal effects and marginal effects at the 25th, 50th and 75th percentile estimated by kernel-based regularized least squares.P-values are reported in parentheses.All models include a set of country dummies.For details, see text.

Table 5 .
Empirical results, Part III: Number of export destinations.
Note: OLS reports the estimated regression coefficients from a linear model.KRLS reports average marginal effects and marginal effects at the 25th, 50th and 75th percentile estimated by kernel-based regularized least squares.P-values are reported in parentheses.For details, see text.