Roam Like at Home and its effects on consumption and pricing

(1)

(2)

2

Roam Like at Home

and its effects on consumption

and pricing

Master’s thesis

Mathijs Moons

MSc Public Administration – Economics & Governance

s2563614

(3)

3

Chapter 1 | Introducing Roam Like at Home

1.1. When in Roam...

Ever since the Agreement on the European Economic Area was signed in 1992, the EU Single Market included all EU members and all but one EFTA state (EEA, 2016). Amongst many other policy fields, the countries set out to harmonize regulation on telecommunication services. For instance, the imposition of a continental standard transmission technology (GSM) propelled network development and penetration rates in Europe (Cave et al., 2019). The rollout of subsequent generations (3G, 4G and currently 5G) are centrally coordinated by the European Commission and implemented throughout the EEA (EEA, 2016).

Now part of the Digital Single Market initiative, the common telecommunication policies aim to remove barriers to inner-EEA digital goods and services (European Commission, 2020). This creates economic efficiency, business opportunities and digital security (Eurostat, 2018).

When it came to using telecom services outside domestic borders – known as international roaming, consumers were confronted by excessive tariffs throughout the EEA. Despite the Commission’s attempts to regulate these roaming prices, rates remained persistently high (Spruytte et al., 2017). To address it more effectively, the EU implemented Regulation (EU) 2017/920 on June 15th of 2017 and established the principle known as Roam Like at Home (RLAH). Roaming sub charges were essentially abolished (EUR-Lex, 2017). According to the first reviews, the Commission deemed RLAH a great success as it increased the amount of telecom consumption significantly (European Commission, 2019). While this is an adequate starting point, the Commission might have used a biased methodology to support their evaluation. By solely monitoring consumption developments in the EEA, no potential confounding variables are controlled for. When developments like technological innovations, software evolutions and (digital) cultural shifts are not considered, can we truly ascribe the observed increase to RLAH alone? Or at all? On top of that, the Commission has yet to review additional economic consequences. As the popular economic saying goes, ‘there is no

(5)

5 such thing as a free lunch’. What if a loss of roaming revenue led to an increase in domestic prices, at the expense of the very consumer RLAH aims to help?

Taken this into consideration, the need for a broader and more systematic evaluation of RLAH arises. This thesis aims to take a first step in that direction .

1.2. Research question and goal

What was the effect of implementing the Roam Like at Home principle in June 2017, on the usage of roaming telecom services and domestic telecom prices throughout the EEA?

The thesis sets out to explain how RLAH affects the usage of roaming telecom services and the level of domestic telecom prices throughout the EEA. By controlling for unobserved variables and taking into account domestic prices, it aims to gain deeper understanding of the true effect of the 2017 roaming regulation.

1.3. Relevance

Academic relevance

Implementing the RLAH principle in multiple countries at once has never occurred before, which make theoretical predictions by field experts like Spruytte et al. (2017) and Falch & Tadayoni (2013) hard to verify. For this, post-implementation empirical analysis is as crucial as it is lacking (Spruytte et al., 2017). While BEREC (2020) and the European Commission (2019) provided the first step by monitoring post-implementation data, it is now up to academic researchers to apply a scientific analysis. The Difference-in-Difference method used in this thesis could be a valuable contribution to this process, as it controls for hidden confounding variables.

Societal relevance

According to the Juncker Commission, RLAH is ranked among the top achievements of its administration (European Commission, 2019). If its success could be scientifically verified, the policy could serve as the prototype for multilateral roaming market integration. On top of that, by applying scientific methodology to the evaluation of

(6)

6 RLAH, this thesis emphasizes the important role of academic research in validating policy assumptions and evaluations. This facilitates better informed policymaking.

1.4. Approach and thesis outline

Chapter 2 A Roam through Literature sets out to explore current literature. The focus is mainly directed at the development of EEA roaming markets, the imposition and effectiveness of EC price caps and the implementation of RLAH. With regards to this, available theoretical explanations and empirical evidence are explored. Chapter 3

Theories on Telecom & Markets continues the theoretical section of the thesis by introducing a theoretical framework. Relevant concepts are defined and relevant market structures are identified. The framework is then used to derive estimations and form the hypotheses. Chapter 4 Overall Research Design outlines the operationalization and the case selection. Furthermore, it justifies the use of a difference in differences methodology and elaborates on its strengths, assumptions and requirements. Chapter 5 Data and Models completes the methodological part of the thesis. After clarifying the methods of selection and processing of available data, the designs of regression models are laid out. Chapter 6 Estimated Effects of RLAH provides an overview of the regression results and describes each variable individually. Based on the fitness with regards to the diff-in-diff assumptions and the predictive power of within-country variance, the results are used to reject the hypotheses formulated in chapter 3. Chapter 7 Conclusion

summarizes the thesis and formulates the overall answer to the research question. Challenges to research, limitations, directions for future research and policy recommendations are also discussed.

(7)

7

Chapter 2 | A Roam through Literature

2.1. Roaming in the EEA

Roaming occurs when a subscriber makes use of a network which is not owned by his provider. While this can take place within domestic borders (Toet, 2010), usually there are no sub-charges involved due to national regulation or agreements. However, once a user crosses a national border, a different protocol is engaged (Spruytte et al., 2017).

As soon as a consumer makes use of mobile telecom services abroad, he consumes outside the network of his Domestic Service Provider (DSP). Instead, the user relies on the network of a different operator another country, a Foreign Service Provider (FSP). Through a technological method called international roaming, the FSP will charge a roaming fee to the DSP. This is regarded as a wholesale rate of international roaming. In turn, the DSP will charge the consumer an additional rate – which is a retail roaming rate.

Originally developed in Scandinavia (GSMA, 2012), international roaming became an EEA-wide technological standard with the second generation of GSM (Haug, 2002). Remaining to be part the subsequent generations (3G and 4G), the technology is now a global standard for cross-border intra-provider telecommunications (Spruytte et al., 2017) amount of literature seems to emphasize restricted competition and excessive pricing (Georgios & Scott, 2016; Infante & Vallejo, 2011; Salsas & Koboldt, 2004; Shortall, 2010; Spruytte et al., 2017).

Sutherland (2010) lays part of the blame at the Standard Terms for International Roaming Agreements (STIRA) facilitated by the GSM Association (GSMA). This global association represents all mobile network operators (MNO’s). MNO’s are large operators which are licenced by a government to maintain their own network. Most smaller operators have no such licence or network, and instead rely on agreements with MNO’s to provide services via their networks. Furthermore, they cannot be members of the GSMA and are therefore excluded from the STIRA framework. Sutherland (2010) assessed that because STIRA predominantly sets the terms for international roaming agreements, smaller telecom operators are disadvantaged or even excluded from these

(8)

8 negotiations. Stumpf (2001) made a similar assessment, but argued that specific technological disadvantages in network frequencies added to the restricted competition in the telecom sector.

Invante and Vallejo (2012) emphasized that the reciprocal characteristics of the roaming market damages competition. Each large operator is both a receiver as well as a provider of roaming traffic. This means that on top of prices, coverage, and quality of service, they also take into consideration the amount of revenue particular roaming contracts would generate. In line with Shortall’s (2010) reasoning, this incentivises large operators to strike deals amongst themselves – again excluding the smaller operators. Invante and Vallejo (2012) state that this is reflected in the structure of the international roaming market. Through mergers and acquisitions, the EEA saw a rise of large and internationally active operators. With this, nearly 80% of all mobile subscriptions in the EU were associated with either Vodafone, Orange, T-Mobile or O2/Movistar (Invante & Vallejo, 2012). These operators aim to facilitate international roaming through their own networks to save costs, which causes at least a third of all roaming traffic to stay within these large conglomerations (BEREC, 2010). This minimizes competition on the roaming market even further (Spruytte, 2017).

Both Sutherland (2010) and Invante & Vallejo (2012) remark that most these theoretical explanations for excessive pricing are rarely empirically tested, since data on wholesale roaming agreements is highly confidential. Paltridge et al. (2009) estimate that wholesale tariffs made up about 75%of the retail rate – meaning the wholesale market would be the main driver of excessive retail prices.

(9)

9

2.2. Price caps

Since the early 2000’s, the EC reviewed complaints about excessive pricing and came to the conclusion in 2003 that the EEA roaming market was subject to market failure (Spruytte, 2017). To combat excessive pricing, three consecutive rounds of regulation introduced increasing price caps on both retail and wholesale roaming rates between 2007 and 2015. In theory, price caps yield the most efficient results if prices do not exceed the operators marginal cost of production (Joskow, 2007). Capping prices any further would cause a negative profitability and therefore completely stop provision of service. Therefore, in order to find the most efficient price cap, the regulator needs to know the provider’s marginal cost curve. Knowing what the regulator will use this information for, providers have no incentive in sharing this information. This makes setting the right price cap quite challenging in practice (Joskow, 2007).

In the case of the EEA roaming market, the effectiveness of the Commissions price caps between 2006 and 2016 have been called into question. Spruytte et al. (2017) remark that the caps far exceeded the wholesale rates, which in turn far exceeded the actual costs of service – stating that there was a lot of room to further reduction of wholesale roaming caps. Sutherland (2010) adds that price caps were subject to significant opposition and were based more on political compromise rather than economic science. During each transition in EU Commission, alternatives to the price caps were explored but deemed unsuitable – ultimately postponing it to 2015 (Spruytte et al., 2017).

2.3. Roam Like at Home

In the winter of 2013 however, Commissioner Kroes proposed the imposition of Roam Like at Home (European Commission, 2013). This comprehensive principle dictates that roaming prices are no longer allowed to be higher than the respective service rates at home. While Parliament seemed in favour of completely eliminating roaming sub-charges, the Council of the EU and most stakeholders favoured maintaining the cap model – emphasizing the substantive negative effects on revenues (Spruytte et al., 2017). It eventually took three and a half years to reach consensus and implement the

(10)

10 regulation. An important addition were the fair use clauses, which prevent misuse of RLAH. For instance, providers are allowed to charge roaming rates when roaming consumption exceeds domestic usage (Shortall, 2019). This prevents West-European consumers from buying cheaper bundles in Eastern Europe, only to ‘roam’ from home . While the European Commission has already dubbed RLAH a great success (European Commission, 2019), the first few academic evaluations of RLAH are more cautious. Spruytte et al. (2017) warn for a possible water bed effect, in which domestic prices are raised to compensate for the loss of roaming revenues. Cave, Genakos, and Valletti (2019) point out that despite the Commission’s claims, no valid empirical evaluation of RLAH exists as of yet. Furthermore, they deem the theoretical economic underpinnings for the intervention to still be unclear. Spruytte et al. (2017) and Cave, Genakos and Valletti (2019) all emphasize the importance of further and more systematic research to the true effects of RLAH – which this thesis would hopefully be a valid first step.

(11)

11

Chapter 3 | Theories on Telecom & Markets

3.1. Defining telecommunication

In general terms, telecommunication services are defined as “the transmission and reception of signals by any electromagnetic means” (WTO, 1995). To narrow it down to the most used forms mobile telecommunication, this thesis analyses voice call, SMS and mobile data services. Voice calls are direct and strictly verbal user-to-user communications (WTO, 1995), measured in minutes of transmission. SMS too is user-to-user, but instead relies on textual messages of 140 characters per unit. Mobile data is all-round internet access from a mobile device, often measured in MB’s (or GB’s) downloaded and uploaded within a set amount of time.

On a retail level, roaming telecom services are divided into four categories: inbound voice call, outbound voice call, SMS and mobile data (GSMA, 2012). When a user makes a phone call from his local network to a foreign network, he is making an outbound voice call. In contrast, when a user receives a call from a foreign network to his local network, he is making an inbound voice call. For example, when a user calls from his French home network to a device in Germany (which has a German home network), the user in France pays an outbound roaming fee per minute while the user in Germany pays an inbound fee per minute. While the inbound/outbound principles also apply to SMS services, only the outbound service is charged on a retail level. This means that while sending an SMS to a foreign network is met with an additional roaming fee, users are not billed for receiving an SMS from a foreign network (GSMA, 2012). Lastly, when a user accesses the internet with his mobile device through the network of a foreign operator, he will be charged an additional roaming fee per MB or GB.

3.2. Defining the EEA telecom markets

Over the years, EEA markets have shown a trend towards bundling the three mobile services into monthly subscription packages (Yang & Ng, 2010). Because of this, they have largely merged into one mobile telecom market. These markets have undergone

(12)

12 significant change in the past 25 years, as they moved from monopolistic market structures to oligopolies with three to four large operators (Feasey & Cave, 2017; Holmes, 2017). BEREC (2015a) classifies three branches of oligopolist structures that can be found in the EEA telecom markets: effective oligopolistic competition, ineffective

oligopolistic competition due to tacit collusion and ineffective oligopolistic competition without tacit collusion.

The first form is often hailed as the ideal form (Valaskova, Durica, Kovacova, Gregova, & Lazaroiu, 2019), as it allows for effective competitive pressures even in markets with few competitors. For instance, a constant threat of new operators entering the market should deter current providers from raising their prices too much. Alternatively, competitive pressure could come from adjacent markets where consumers could switch to a comparable service. While the EEA telecom markets are generally defined by large sunk costs (Joskow, 2007) and therefore experience little threat of new market entrants, traditional telecom services are increasingly competed with by services like Skype, What’sApp and Facebook (Farooq & Raju, 2019). However, BEREC (2015a) recognizes that EEA telecom markets are in risk of ineffective oligopolistic competition due to tacit collusion. Unlike illegal cartels, tacit collusion is behaviour that providers follow without explicit agreement (BEREC, 2015a). Without coordination, operators settle for a certain strategy that involves uncompetitive behaviour to reach a higher joint profit. BEREC outlines various conditions (marked below) that could increase the risk of tacit collusion. When compared to the average domestic telecom market, most conditions seem to apply.

In markets with very few firms, it is easier to monitor and anticipate behaviour of competitors. With only three to four large operators dominating the EEA telecom markets (Holmes, 2017), this condition is certainly met. Furthermore, when operators are ensured of repeated interaction and are certain that colluders have long term interests, they can more safely assume that tacit collusion is sustainable. Barriers to entry or exit need to be high as well, because collusion can not be sustainable if new or small fringe firms can rapidly increase their output to compete with the colluding providers. Given the large sunk costs and considerable investments needed for providing telecom services (Alleman & Rappoport, 2006), repeated interaction and barriers to entry/exit seem persistent throughout EEA telecom markets. Very importantly, players need to be

(13)

13

capable to reach mutually acceptable equilibrium. This means that there is a mutual understanding of what a common policy or focal point should be, and each player finds this outcome acceptable. Given the obscure nature of such an equilibrium, it is difficult to assess whether or not current EEA operators maintain such an equilibrium price. Furthermore, each firm always has an incentive to deviate from the collusive outcome, for instance by lowering the market price to undercut all other colluders. It should therefore be easy to detect deviance, giving all colluders the opportunity to instantly

react. With retail prices being relatively transparent amongst providers in EEA telecom markets (Sutherland, 2010), potential deviance should be easily detectable. Lastly, to counter any deviant behaviour, compliance should be enforceable. In most cases, this is a price war which reduces revenues for all colluders – including the deviator. Since the price of telecom bundles can be changed in an instant (Xian-xiu & Xiu-qing, 2004), a price war could ignite instantly.

While the model of tacit collusion seems applicable to domestic telecom services, the excessive prices of roaming services are better explained by BEREC’s (2015a) last branch of oligopolies: ineffective competition without tacit collusion. Without any coordination or anticipation on the behaviour of competitors, these market structures still facilitate a stable form of ineffective competition. In such a case, undertakings unilaterally adopt a strategy, which in combination with the strategies of other undertakings in the market forms a self-sustaining reduction in competition.

This seems to have been the case with roaming charges. As roaming services were often overlooked by consumers when considering the purchase of their bundle, providers were able to set charges in relative obscurity (Sutherland, 2008). On top of that, roaming initially had a bilateral nature and created significant variation in fees – depending on location and foreign provider (Sutherland, 2010). These forms of asymmetric information made it quite hard for the consumer to be familiar with his options in roaming services. That allowed providers to set high roaming rates, regardless of the rates charged by other operators. Since roaming rates remained quite profitable and obscure, setting high roaming charges became a sustainable unilateral strategy (Spruytte et al, 2017). Salsas and Koboldt (2004) made a similar assessment. They claim that the excessive EEA roaming charges are not a product of collusion, but rise from underlying technological and institutional constraints on competition mechanisms.

(14)

14

3.3. Predicting the effects of RLAH

By implementing RLAH to combat restricted competition in the EEA roaming market, the Commission actively intervened to counter market failure (Cave, Genakos, & Valletti, 2019). Rather than setting a consistent and universal roaming price cap, RLAH forces providers to risk their position in the domestic market when raising prices for roaming services. In essence, each roaming price is now capped by the dynamics of the domestic market. Within the parameters of the Fair Use agreement and from the consumer’s point of view, RLAH essentially merges the domestic and roaming markets into one EEA telecom market. With roaming rates now as transparent as domestic rates, RLAH deals directly with the issue of asymmetric information.

Since most users own subscription bundles, roaming services are suddenly free of charge – given they stay within their bundles. This essentially means that users are able to maintain their domestic consumption pattern when abroad, which are on average considerably higher than their old roaming consumption patterns (BEREC, 2020). Because roaming fees per unit are replaced by (in most cases already paid for) subscription packages (Spruytte, 2017), it is to be expected that RLAH would pave the way for previously non-roamers to enter the market, as well as increase usage per roamer. In other words, both overall roaming as well as the consumption per individual roamer is expected to increase.

(15)

15 In contrast to predicting the response of users, it seems more complicated to foresee how operators will respond. Given the potential tacit collusion in the domestic telecom markets discussed above, the kinked demand model (graph 3.1 and 3.2) illustrate how strategic reactions to RLAH would not necessarily lead to an increase in price or reduction in service output.

When a provider sets his prices in a oligopolistic market, two important assumptions are made. Firstly, if he increases prices above the mutually acceptable equilibrium, his competitors would most likely not match him in order to attract more consumers (..., ...). As illustrated in the graph, demand will therefore not increase significantly for any price above equilibrium (E). Secondly, should the provider choose to lower his prices below equilibrium, his competitors would hastily match him in order not to lose market share (Pettinger, 2016). As BEREC (2015a) assessed, this would likely start a price war in domestic EEA telecom markets. This in turns significantly lowers marginal revenue for any price below equilibrium (E) – resulting in a kinked demand curve. Following his

(16)

16 marginal cost and marginal revenue lines, the operator would set the price at P* – equal to the mutually acceptable equilibrium.

The looming price war combined with the inelastic pricing above equilibrium, make price levels in oligarchic markets to be relativily stable. Even if costs increase, it would not necessarily have an effect on pricing. For example, were RLAH to shift the marginal cost curve from MC1 to MC2, maintaining the original price of P* would still be the most profitable course of action. Only a substantial increase in marginal costs, such as MC3, would lead to an increase in price to P(r) and the respective loss of purchased service bundles.

Given that before RLAH the average roaming revenue was only 4 % of overall mobile service revenue (BEREC, 2016; Infante & Vallejo, 2011), the loss of roaming fee revenues would not necessarily be costly enough to raise prices. Only if wholesale roaming tariffs increase substantially or overall roaming usage grows drastically, the increase in marginal costs of providing roaming services could be large enough to justify

(17)

17 an increase in domestic subscription fees.

3.4. Hypotheses

Since large per unit fees are replaced by (often already paid for) subscription fees, economic theory predicts a substantial increase in overall roaming consumption. Therefore, the following 𝐻₀ hypotheses are expected to be rejected:

1𝑎 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA overall inbound roaming voice traffic.

1𝑏 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA overall outbound roaming voice traffic.

1𝑐 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA overall roaming SMS traffic.

1𝑑 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA overall roaming mobile data traffic.

In favour of the following 𝐻1 hypotheses:

1𝑎 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA overall inbound roaming voice traffic.

1𝑏 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA overall outbound roaming voice traffic.

1𝑐 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA overall roaming SMS traffic.

1𝑑 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA overall roaming mobile data traffic.

This expectation extents to the amount of consumption per roaming individual. With that, the following 𝐻0 hypotheses too are expected to be rejected:

2𝑎 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA inbound roaming voice usage per roamer.

(18)

18

2𝑏 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA outbound roaming voice usage per roamer.

2𝑐 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA roaming SMS usage per roamer.

2𝑑 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on intra-EEA roaming mobile data usage per roamer.

In favour of the following 𝐻1 hypotheses.

2𝑎 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA inbound roaming voice usage per roamer.

2𝑏 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA outbound roaming voice usage per roamer.

2𝑐 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA roaming SMS usage per roamer.

2𝑑 ȁ 𝐻₁ ȁ The implementation of RLAH had a positive effect on intra-EEA roaming mobile usage per roamer.

Lastly, given the often entrenched prices and dynamics of tacit collusion in the EEA domestic telecom markets, the kinked-demand model predicts a price change only in case of a considerable increase in marginal costs. Given that roaming revenue on average only accounts for 4% of overall telecom revenues, the following hypotheses are tested:

3 ȁ 𝐻₀ ȁ The implementation of RLAH had no effect on EEA domestic prices. 3 ȁ 𝐻₁ ȁ The implementation of RLAH had an effect on EEA domestic prices.

(19)

19

Chapter 4 | Overall Research Design

4.1. Operationalization

RLAH

The Commission implemented RLAH through Regulation (EU) 2017/920; which facilitated the immediate abolishment of roaming sub-charges throughout the entire EEA (including Iceland, Norway and Liechtenstein) as of the 15th_{of June 2017 (EUR-Lex,}

2017). Given that RLAH was implemented simultaneously and the telecommunication markets were able to react instantly (Spruytte et al., 2017), it is assumed that the effects of RLAH should be observable from the second quarter of 2017 onwards (BEREC, 2017).

Roaming service usage

The concept of roaming service usage is easily translated into the amount of minutes, textual units of 140 bytes or GB’s consumed over a set amount of time. However, a potential complication lies in interpreting its direction and country of interest. When a Greek sim-card is used in Croatia to make a call to Greece, to which country should this specific service be attributed? To prevent any confusion, telecom markets have adopted a common framework dividing roaming traffic into inbound voice call, outbound voice

call, SMS and mobile data (GSMA, 2012). Firstly, the point of reference is set at the

sim-card. When a Danish sim-card receives a call from a foreign network while being in Denmark, it is considered a Danish inbound voice call. When the Danish sim-card is used to make a call to a French sim-card in France, it is considered a Danish outbound voice call and a French inbound voice call. While voice traffic is always measured (and billed) for two directions, only the outbound side is tracked for SMS roaming services. This means that a Hungarian message send to Austria will only be considered a Hungarian unit of SMS roaming. Following this framework, roaming service usage can be operationalized into an overall roaming traffic indicator, which is measurable at the level of a country over a relevant period of time.

(20)

20 In addition to overall traffic, individual usage per roamer is also analysed using the same framework (BEREC, 2019). For example, the overall Italian SMS roaming units are divided by the amount of Italian roamers at that point in time. By tracking both overall usage as well as usage per roamer, a more comprehensive effect of RLAH can be estimated. For instance, if a small increase in roaming traffic is accompanied with a large increase in usage per roamer, it seems likely that RLAH mostly affected already existing roamers in the market. Were it to be the other way around, a steep rise in overall traffic but a small effect on usage per roamer, then RLAH seems to have mostly opened the market to new roamers.

Domestic prices

In establishing whether RLAH had any effects on domestic prices of telecommunication services, it would be ideal to focus on price developments of voice call per minute, SMS per message and data per GB. However, these services are quite often bundled together (Georgios & Scott, 2016), making it infamously difficult to allocate individual prices to specific services (BEREC, 2020). To approximate developments in domestic prices, telecommunications analysts make use of the Average Retail Revenue Per User (Desphande & Narahari, 2017). The ARRPU indicator makes no distinction between voice calls, SMS and data services, but assumes that the average consumer pays one price for all three. This assumption is congruent with the fact that most users pay one subscription fee for all three services (Yang & Ng, 2010). If a country sees an increase in their average ARRPU in the domestic telecom market, the average user is paying more for their mobile services – which likely indicates an increase in price per bundle. ARRPU’s are calculated by dividing the amount of revenue earned from mobile telecom services with the amount of active sim-cards within a specific country over a relevant period of time (BEREC, 2020).

The ARRPU is not without its flaws. BEREC (2020) emphasizes that the indicator does not account for factors like overall consumption volumes, handset subsidies and the sensitivity to the number of active SIM cards. However, it remains a widely used tool to at least approximate domestic price developments (McCloughan & Lyons, 2006).

(21)

21

4.2 Case selection

Since RLAH was implemented in all EEA and three EFTA countries, all 31 states are relevant for analysis. Due to data unavailability on Iceland, the following cases are selected: Austria, Belgium, Bulgaria, Croatia, Cyprus, the Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Liechtenstein, Lithuania, Luxembourg, Malta, the Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden and the United Kingdom.

In addition, Switzerland will serve as their counterfactual case – since RLAH was never implemented there. In many ways, Switzerland resembles an EEA country. Having a GDP per capita comparable of that the most developed EEA members, Switzerland’s economy and consumer base are similar (European Commission, 2020b). On top of that, the telecom market structure is oligopolistic and compares to those of EEA countries (OFCOM, 2020). From a cultural perspective, the Swiss share a European history and speak mostly French, German and Italian. Governmental organization is, like EEA countries, highly democratic and follow the rule of law/ Geographically, the country is completely surrounded by EEA countries and is located at the very heart of the area. On top of that, it receives 43 million visits from EEA tourists per year and sends about 13 million visits in return (EuroStat, 2020).

Most importantly, Switzerland is the most integrated non-member of the EEA. Since June 2002, Switzerland has entered a process of policy harmonization with the EEA (Federal Department of Foreign Affairs, 2016). From travel to agriculture to trade: in most policy fields it directly implements EEA regulation through a considerable collection of bilateral agreements. On top of that, by joining the Schengen Area in 2008, Switzerland is virtually indistinguishable for EEA visitors. The same is true for EEA workers seeking (seasonal) employment in Switzerland and vice versa. In most cases, no working permits are required (Ruedin, 2012), making Swiss-EEA labour migration flows comparable to those amongst EEA countries (Afonso, 2009). Given that labour migration and tourism are two strong drivers of international roaming usage (Spyratos et al., 2018; Tennekes, Offermans, & Heerschap, 2017), these policy integrations strengthen the Swiss case as an EEA counterfactual.

(22)

22

4.3. Method of analysis

The selected data was analysed through a difference-in-difference method. Compared to the methodology used in the Commission’s evaluation of RLAH (European Commission, 2019), this quasi-experimental design has a valuable advantage. By comparing the pre-implementation and post-implementation data of EEA countries alone, RLAH cannot be validly distinguished from other relevant unobserved variables shared by EEA countries. Phenomena like hardware innovations, software efficiency developments and general shifts in telecom demand could have affected all EEA countries in a similar way – alongside RLAH’s implementation. The diff-in-diff methodology includes a counterfactual to compare EEA countries to a contemporary and comparable case in which RLAH has not been implemented. While doing so, the model accounts for stable differences between Switzerland and the EEA countries. Systematic differences in network capacity, type of consumer base and market structures for example, are filtered out if they remained stable over the analysed period of time.

In order to validly estimate causal effects through a diff-in-diff model, four important assumptions are to be made (Ricardo & Iliana, 2012). The positivity assumption states that all possible treatments are included in each segment of the analysis. With only one intervention in the model, this essentially means that observations of both intervention and non-intervention cases are incorporated. Furthermore, the exogeneous treatment assumption demands that the outcome variable cannot determine the assignment of the intervention. Moreover, the stable unit treatment value assumption requires that the intervention and comparison group are stable for the repeated cross-sectional design, meaning that the composition of the groups cannot be altered throughout the model. On top of that, there cannot be any spill-over effects where the intervention also affects the outcome variable of the control group. Lastly and very importantly, the parallel trend assumption forms the bedrock of the counterfactual. It assumes that, in absence of the intervention, the intervention and comparison groups would have evolved parallelly.

(23)

23 Resting on these assumptions, a diff-in-diff model sets out to emulate an experimental design by comparing the intervention group to the counterfactual control group – before and after the intervention (Angrist & Pischke, 2015). Firstly, the values of the control and intervention group are measured in the pre-implementation period (first difference). This is repeated for the values in the post-implementation group (second difference), from which the first difference is subtracted. Under the assumption that the outcome variable of the treatment group would have evolved parallel to that of the comparison group, this difference in differences serves as an estimation the actual effect of the intervention. t1 t2 t3 t4 t5 V ari ab le o f I nt ere st Time

4.1. Difference in Differences

Intervention group Control group

(24)

24

Chapter 5 | Data and Models

5.1. Data selection

EEA

As of 2020, the Body of European Regulators for Electronic Communication (BEREC) has conducted a total of 22 annual rounds of monitoring EEA roaming traffic, prices and revenues. While the data has traditionally been published through bi-yearly reports (BEREC, 2020), BEREC set up an online database in 2019 (BEREC, 2019). However, most of its benchmarks cover a period from 2017 to 2019, which made manual extraction from the original reports necessary. In total, 2.320 additional datapoints were manually extracted from 15 different BEREC reports (BEREC, 2013a, 2013b, 2014a, 2014b, 2015a, 2015b, 2016a, 2016b, 2016c, 2017b, 2017a, 2018, 2019b, 2019a, 2020).

Average monthly roaming service usage per roamer is a benchmark BEREC tracks for voice call, SMS and mobile data since q2 2016. For each country and quarter, the benchmark shows the average monthly amount of minutes, messages and GB’s consumed per roaming subscriber.

Furthermore, BEREC tracks the overall traffic of intra-EEA roaming voice calls, SMS and data. The public database provided traffic indexes for each EEA country besides Iceland, starting in q4 2015 while taking 2012 as the baseline. Through manual extraction from the reports, the dataset could be expanded to range from q1 2013 to q4 2018. Instead of 2012, these indexes used 2008 as their baseline, rendering them incomparable. Luckily, two reports from 2012 used 2008 as a baseline, making it possible to reconfigure the entire dataset into 2012-indexes.

BEREC calculates the ARRPU for each EEA country by dividing the total amount of revenue from mobile telecom services by the total amount of active SIM-cards in the market. For each quarter since q2 2016, BEREC reports the average monthly ARRPU for each EEA country besides Iceland.

Switzerland

Data for the counterfactual case was extracted from the statistical observatory maintained by the Swiss Federal Office of Communications (OFCOM, 2020). Because

(25)

25 OFCOM uses slightly different formats for its benchmarks, its data needed to be adjusted in order to match that of BEREC.

Unlike BEREC, OFCOM does not publish the average monthly roaming service usage per roamer (in minutes, SMS and GB). Based on their available benchmarks, the following equation was used to calculate the Swiss average monthly roaming service usage:

𝑄_{𝑆𝑤𝑖𝑠𝑠} =1 3∗

𝑄_𝑞 𝑆_𝑞

Where 𝑄_𝑞 is the average roaming consumption to EEA countries (in minutes, messages or GB) per quarter and 𝑆𝑞 the average active roaming SIM-cards per quarter. While

OFCOM has data from q1 2013 to q4 2018, only data from q2 2016 and onwards was used in order to match BEREC’s somewhat less extensive dataset.

Furthermore, to calculate the Swiss overall roaming traffic,the following equation was used:

𝑇_{𝑆𝑤𝑖𝑠𝑠} = 𝑄𝑞 0,01𝑄_2013𝑞1

𝑇_{𝑆𝑤𝑖𝑠𝑠} is a traffic index comparable with those published by BEREC, calculated by dividing quarterly EEA roaming consumption (in minutes, messages or GB) by a hundredth of the quarterly EEA roaming consumption in q1 2013. To harmonize the dataset, the EEA traffic indexes were also rescaled to have q1 2013 as their baseline. Lastly, to calculate the Swiss average monthly ARRPU, the following equation was used:

𝐴𝑅𝑅𝑃𝑈_{𝑆𝑤𝑖𝑠𝑠}=1 3∗

𝑅_𝑞 𝑈𝑞

With Swiss quarterly revenue from mobile telecom services 𝑅_𝑞 divided by the quarterly total active Swiss SIM-cards 𝑈𝑞, an average monthly ARRPU was calculated. It is

important to note that OFCOM publishes the total active Swiss SIM-cards per year, not per quarter. Since it is safe to assume that the amount of Swiss SIM-cards will not show any significant seasonality, yearly data can be disaggregated into quarterly estimates (Grudkowska, 2015). By making use of JDemetra+, the European Central Bank’s

(26)

26 disaggregation software, quarterly datapoints were estimated based on OFCOM’s yearly dataset.

5.2. Overview of datasets

Average monthly usage per roamer

Voice Call (inbound) Voice Call (outbound) SMS Data Amount of countries 31 31 31 31

Time span (in quarters) Q2 2016 – Q4 2018 Q2 2016 – Q4 2018 Q2 2016 – Q4 2018 Q2 2016 – Q4 2018 Amount of observations 338 339 339 332

Overall roaming traffic

Voice Call (inbound) Voice Call (outbound) SMS Data Amount of countries 29 29 29 29

Time span (in quarters) Q1 2013 – Q4

2018 Q1 2013 – Q4 2018 Q1 2013 – Q4 2018 Q1 2013 – Q4 2018 Amount of observations 605 605 605 603

Average monthly ARRPU

Amount of countries 31

Time span (in quarters) Q2 2016 – Q4 2018

(27)

27

5.3. Regression models

Since the models aim to explain the effects of only one intervention, two possible ‘treatments’ are included to satisfy the positivity assumption: a country has either an implementation of RLAH in the second quarter of 2017 (intervention group), or has no intervention at all. To satisfy the stable unit treatment value assumption, the composition of these groups do not change throughout the time series. Furthermore, spill-over effects are assumed to be negligible as the EEA’s roaming regulation did not come into effect in Swiss territory. While it is possible that certain consumers initially believed Switzerland to be included, graphs 5.1-5.10 imply that this type of spill-over had no notable effect on the comparison group. Furthermore, the ‘assignment’ of the intervention to a country, or in this case the implementation of RLAH, was not determined by any of the outcome variables (overall traffic, usage per roamer or ARRPU) – satisfying the exogenous treatment assumption. Rather, EEA membership was a necessary condition for treatment assignment, which in turn was in no way dependent on the variables of interest. While the beforementioned assumptions appear to be congruent with the case data, the important parallel trends assumption seems more problematic and demands differing approaches.

Seasonality 0 100 200 300 400 500 600 700 800 900 1000

5.1. Outbound Voice Traffic (index)

(28)

28 As displayed in graph 5.1, roaming traffic appears to be highly seasonal as it includes a high season, low season and two intermediate seasons. On top of that, considering that the pique season of Poland occurs in q1, Portugal in q2, and Slovenia in q3, the observed seasonality seems to be asynchronous between countries. If this is not accounted for, the assumption that the post-implementation outcome of Switzerland serves as the counterfactual to that of a country with a differing seasonality, will produce biased results. To prevent this, each value of the overall traffic index for country 𝑐 at time 𝑡 will be replaced with a moving average (MA). This new variable displays the average of the sum of the current value of a country in a particular quarter, the previous two quarters and one quarter ahead. Because the MA-value always covers four consecutive quarters, it guarantees that each value incorporates a high season, a low season, and two intermediate seasons. This dampens within-year seasonality and ensures that each value is comparable to values of other countries in the same period of time. Because the MA-values partly depend on values in the past, the first two (and last one) quarters of observations are not included in the regression. Incorporating three ‘lagging’ and no ‘leading’ quarters in the MA instead of two ‘lagging’ and one ‘leading’, would chronologically make more sense. However, the MA would then substantially lean towards the past and risks to delay observability of the treatment effect directly after implementation. Therefore, a more balanced approach to MA, which includes one leading value, is preferred. A formal expression of the MA-values of overall inbound voice roaming usage per roamer (𝑌_2𝑎) for country 𝑐 at time 𝑡 as follows:

MA_Y2a_ct =𝑌2𝑎𝑐(𝑡−2)+ 𝑌2𝑎𝑐(𝑡−1)+ 𝑌2𝑎𝑐𝑡+ 𝑌2𝑎𝑐(𝑡+1)

(29)

29 Which would transform graph 5.1 into the following:

As visible in graph 5.3-5.10 below, within-year seasonality seems to occur in both usage per roamer as well as overall roaming traffic. To adjust for this by regressing on the MA-values, the diff-in-diff model equates as follows for country 𝑐 at time 𝑡:

MA_Y1act = 𝛼 + ෍ 𝛽𝑘 United Kingdom 𝑘=Bulgaria 𝐶𝑂𝑈𝑁𝑇𝑅𝑌𝑘𝑐 + ෍ 𝛾𝑗𝑇𝐼𝑀𝐸𝑗𝑡+ 2014q2 𝑗=2018q4 𝛿(𝑇𝑅𝐸𝐴𝑇𝑐 ∗ 𝑃𝑂𝑆𝑇𝑡) + 𝜀𝑐𝑡

Where MA_Y1a is the moving average of overall outbound voice call traffic and 𝛼

represents the starting value of Austria, the reference country. To capture group effects relative to this reference country, a dummy will activate for all other countries (where

𝐶𝑂𝑈𝑁𝑇𝑅𝑌 = 1 if 𝑐 = 𝑘 and 0 otherwise) and calculate group effect 𝛽_𝑐. Similarly, common time effects relative to reference period Q1 2014, are captured by individual time dummies (where 𝑇𝐼𝑀𝐸 = 1 if 𝑡 = 𝑗 and 0 otherwise). Furthermore, an interaction dummy is activated to estimate the average treatment effect on the treated, where

𝑇𝑅𝐸𝐴𝑇_𝑐 = 1 and 𝑃𝑂𝑆𝑇_𝑡 = 1 for values in the treatment group and post-implementation 0 100 200 300 400 500 600 700 800

5.2. Outbound Voice Traffic (MA)

(30)

30 period respectively, and 0 otherwise. The ATT is then captured by 𝛿, followed by residuals in the error term 𝜀_𝑐𝑡.

Pre-implementation divergence

The following graphs display the seasonality adjusted datasets for usage per roamer and overall roaming traffic. The light-coloured dotted lines represent the original values, while the darker lines display the new MA-values.

0 50 100 150 200 250 300 350 400 450

5.3. Inbound Voice Call Traffic (index)

EEA total Switzerland

0 50 100 150 200 250 300 350 400 450

5.4. Outbound Voice Call Traffic (index)

(31)

31 0 20 40 60 80 100 120 140 160 180 200

5.5. SMS Traffic (index)

EEA total Switzerland

0 5000 10000 15000 20000 25000

5.6. Data Traffic (index)

(32)

32 0 2 4 6 8 10 12 14 16 18 2016q4 2017q1 2017q2 2017q3 2017q4 2018q1 2018q2 2018q3

5.7. Inbound Voice Call (minute per roamer)

EEA average Switzerland

0 2 4 6 8 10 12 14 16 18 20 2016q4 2017q1 2017q2 2017q3 2017q4 2018q1 2018q2 2018q3

5.8. Outbound Voice Call (minute per roamer)

(33)

33 0 0,05 0,1 0,15 0,2 0,25 0,3 0,35 0,4 0,45 0,5 2016q4 2017q1 2017q2 2017q3 2017q4 2018q1 2018q2 2018q3

5.10. Data (GB per roamer)

EEA average Switzerland

0 1 2 3 4 5 6 7 2016q4 2017q1 2017q2 2017q3 2017q4 2018q1 2018q2 2018q3

5.9. SMS (unit per roamer)

(34)

34 While the incorporation of moving averages has substantially decreased the severity of the violation, graphs 5.3-5.10 illustrate that there is still no strict parallelism. At the risk of (partly) decreasing the explanatory power of the diff-in-diff methodology (Angrist & Pischke, 2015), a common practice to deal with this problem is to include time trends for individual countries (Egami & Yamauchi, 2019; Freyaldenhoven, Hansen, & Shapiro, 2019; Mora & Reggio, 2012, 2017). With these, the model aims to separate the treatment effect from the estimated effects of time-variant confounding variables, which cause diverging trends in the pre-implementation period. Even if the trend of a country in the intervention group is not parallel to that of Switzerland, the model could still find evidence for a treatment effect by capturing sharp deviations from otherwise smooth trends (Angrist & Pischke, 2015). Applying both seasonality adjustment and individual time trends to the extended diff-in-diff model for the effect of RLAH on the overall inbound voice traffic 𝑌_{1𝑎𝑐𝑡},* forms the following equation for country 𝑐 at time 𝑡:

MA_Y2a𝑐𝑡 = 𝛼 + ෍ 𝛽_𝑘 United Kingdom 𝑘=Bulgaria 𝐶𝑂𝑈𝑁𝑇𝑅𝑌𝑘𝑐+ ෍ 𝛾𝑗𝑇𝐼𝑀𝐸𝑗𝑡 2018q4 𝑗=2016𝑞4 + 𝛿(𝑇𝑅𝐸𝐴𝑇𝑐∗ 𝑃𝑂𝑆𝑇𝑡) + ෍ 𝜃𝑘(𝐶𝑂𝑈𝑁𝑇𝑅𝑌𝑘𝑡∗ (𝑡 − 𝑡0) ) United Kingdom k=Bulgaria + 𝜀𝑐𝑡

The reference country is still Austria and the first quarter is taken as a reference time period, in this case Q3 2016. Compared to the previous diff-in-diff equation, most variables operate similarly. Contrarily, the original outcome variable Y_2𝑎 is replaced with its corresponding moving average variable MAY2a. On top of that, an additional

collection of interaction dummies capture the time trend effect 𝜃 for country 𝑐 at time

(35)

35 Compared to both the EEA average as well as individual EEA countries, the development of the Swiss ARRPU seems tremendously different. This appears to apply to the intervention as well as the post-intervention period. While the diverging pre-implementation trends of the previous outcome variables could still be controlled for, the divergences in the ARRPU dataset seems to great to overcome. Because of this, the parallel trend assumption is considered to be violated, making a true diff-in-diff model impossible. In order to still offer some insight in the developments of EEA telecom prices, a ‘lighter’ model with significantly less explanatory power is used as an alternative:

𝑌_3𝑐𝑡 = 𝛼 + ෍ 𝛽𝑘 United Kingdom 𝑘=Bulgaria 𝐶𝑂𝑈𝑁𝑇𝑅𝑌𝑘𝑐+ ෍ 𝛾𝑗𝑇𝐼𝑀𝐸𝑗𝑡+ 𝛿𝑌𝑧𝑡 q2 𝑗=q4 + 𝜂𝑃𝑂𝑆𝑇𝑖𝑡+ 𝜆(𝑃𝑂𝑆𝑇𝑖𝑡∗ 𝑌𝑧𝑡) + 𝜀𝑐𝑡

The model estimates values in ARRPU Y₃ for country 𝑐 at time 𝑡. The constant 𝛼 again captures the starting value of Austria, the reference country. Variable 𝑌_3𝑧𝑡 is the ARRPU

0 5 10 15 20 25 30 35 40

5.11. Average Retail Revenue per User (Euro per user)

(36)

36 of Switzerland at time 𝑡, which is used to estimate 𝛽 – the difference in ARRPU of country 𝑐 compared to Switzerland at time 𝑡. Through four time dummies representing

the four quarters per year, 𝛾𝑗captures seasonal time shocks. The first quarter is taken

as a reference. In terms of the hypotheses, coefficient 𝜂 is of interest as it estimates a sharp increase or decrease in EEA ARRPU’s in the post-implementation period. If 𝜂 is statistically significant, a potential effect of RLAH on ARRPU’s has been found. Similarly,

𝜆 would capture a post-implementation divergence from Switzerland over time. If significant, this could indicate a longer term increase or decrease of ARRPU’s within the EEA compared to a similar non-EEA country.

Fixed Effects and clustered standard errors

All three models make use of Fixed Effects instead of pooled OLS, as they account for systematic differences between countries over time through the use of various dummy variables. On top of that, the models make use of robust standard errors, clustered by country. Executed through STATA’s time-series software,1_{this controls for potential}

sources of variation that come from differences across countries (McKenzie, 2017).

(37)

37

5.4. Notes on validity

In determining the validity of these model designs, it is important to note that each of the adjustments bring both advantages and disadvantages. For instance, while seasonality adjustment reinforces the assumption of parallel trends, it replaces the outcome variable with an artificial substitute that could dampen (and therefore underestimate) a potential implementation shock. Similarly, introducing individual time trends could control for time-variant confounding variables which cause pre-implementation developments to diverge – such as different developments in telecom networks or changing purchasing power of the consumer base. However, individual time trends ‘compete’ with the treatment estimator and could potentially absorb (parts) of the actual treatment effect (Angrist & Pischke, 2015). Therefore, combining both adjustments would in theory provide the best model with regards to the diff-in-diff assumptions, at the risk of underestimating the actual treatment effect on the treated. Since the hypotheses are mainly focused on finding causal relationships rather than estimating causal effects, these adjusted models are preferred – under the condition that they sufficiently and significantly explain within-country variation in outcome variables.

Given that the ARRPU model could not be adjusted to fit the diff-in-diff assumptions at all, its estimations cannot directly support causal claims. Rather, this model can only serve a more exploratory purpose and form a starting point for further research into domestic price developments.

(38)

38

Chapter 6 | Estimated Effects of RLAH

6.1. Overall usage of roaming services

(1) (2) (3) (4) (5) (6) (7) (8)

6.1a. OLS with fixed time and group effects

Inbound Voice Traffic Outbound Voice Traffic

Unit Traffic

Index MA Traffic Index Traffic Index Traffic MA

Index

Traffic

Index MA Traffic Index Traffic Index Traffic MA

Index

Seasonality adjustment

No Yes No Yes No Yes No Yes

Individual trend adjustment

No No Yes Yes No No Yes Yes

RLAH 169,921*** (43,138) 162,030* (55,571) 56,914* (21,550) 56,183* (20,878) 91,733*** (24,393) 95,598** (31,589) 61,226*** (14,198) 69,089*** (12,387) Constant 100,254*** (25,441) 109,579*** (20,990) 99,676*** (9.488) 88,467*** (12,928) 100,241*** (13,401) 100,510*** (11,932) 99,775*** (7,485) 83,404*** (7,223) Quarterly time dummies

Yes Yes Yes Yes Yes Yes Yes Yes

Individual time trends

R2 _0,351 _0,378 _0,769 _0,910 _0,261 _0,569 _0,764 _0,917 Observations 688 605 688 605 688 605 688 605 Clustered groups (by country) 29 29 29 29 29 29 29 29

Coefficients are marked with *, **, and *** if they are statistically significant to the degree of 𝑃 < 0,05; 𝑃 < 0,01; and 𝑃 < 0,001 respectively. The corresponding robust standard errors are displayed below their coefficient and placed in brackets.

MA stands for Moving Average, which controls for within-year seasonality. R2_{indicates the degree of variation within}

(39)

39

(9) (10) (11) (12) (13) (14) (15) (16)

6.1b. OLS with fixed time and group effects

SMS Traffic Data Traffic

Unit Traffic Index Traffic MA Index Traffic Index Traffic MA Index Traffic

Index MA Traffic Index Traffic Index MA Traffic Index

Seasonality

adjustment No Yes No Yes No Yes No Yes

RLAH 69,509*** (14,998) 66,173*** (18,995) 33,164** (12,609) 37,473** (10,961) 31410,95** (9716,455) 25512,85* (10356,85) 25168,24* (7929,073) 19472,05** (5907,601) Constant 100,159*** (10,273) 110,995*** (7,175) 100,124*** 5,355 112,201*** (4,788) 232,653 (3121,297) 231,390 (3911,971) -19,739 (3339,535) -2739.919 (2900,136) Quarterly time dummies

Individual

time trends No No Yes Yes No No Yes Yes

country units that is captured by the model.

Table 6.1 shows the results of the diff-in-diff regression estimating the effect of RLAH on overall roaming traffic. For each of the four roaming services, the result of all combinations of seasonality and/or individual trend adjustment are shown. For inbound voice traffic for example, (1) represents the unadjusted FE regression, (2) introduces seasonality adjustment, (3) introduces individual trend adjustment instead, and (4) incorporates both. In general, it seems that seasonal adjustment slightly decreases the estimated effects of RLAH, but increases the overall fitness of the model by providing higher within R2_{values. This comes at a cost of statistical precision of the treatment}

(40)

40 estimator, as the variable loses a degree of significance in three out of four cases. Independently of seasonal adjustment, the introduction of individual time-trends substantially decreases the estimated effect of RLAH, but greatly increases the overall fitness of the model. From a purely empirical point of view, combining seasonal adjustment with individual time trends provides the most predictive models – accounting for between 73 % and 92% of the variation within country clusters. Whether this also means it is the best model to explain the causal relationship between RLAH and overall roaming traffic, remains to be discussed the next section.

The estimated effect of RLAH on inbound voice traffic lies between 56,18 and 169,92 index points. In other words, the average percentual increase of inbound voice calls assumed to be caused by RLAH, was between 56 % and 170 % since Q1 2013. Translated into the underlying values provided by BEREC (2020), this would entail an overall increase of roughly 640 to 1935 million minutes throughout the EEA between 2013 and 2018. Considering the fact that inbound voice traffic seemed to violate the parallel trends assumption the most, it was to be expected that incorporating individual trends in the model had the strongest effect on its estimated treatment coefficient. Furthermore, while significant, the regression for inbound voice traffic estimates the least precise coefficients.

In contrast, outbound voice traffic regressions provide the highest significance and show the least deviations when the adjustments are introduced. The effect of RLAH is estimated to be between 61,2 and 95,6 index points, translating into an overall increase of 630 to 990 million minutes. Similarly, SMS increased by 33,2 – 69,5 index points or 710 to 1300 million messages. Lastly, the model shows an increase of 19472,0 to 31411,0 index points for mobile data traffic, representing an increase of 130 to 210 million GB due to RLAH.

(41)

41

6.2. Roaming service usage per roamer

(1) (2) (3) (4) (5) (6) (7) (8)

6.2a. OLS with fixed time and group effects

Inbound Voice Call per Roamer Outbound Voice Call per Roamer

Unit Minutes per Roamer MA Minutes per Roamer Minutes per Roamer MA Minutes per Roamer Minutes per Roamer MA Minutes per Roamer Minutes per Roamer MA Minutes per Roamer Seasonality adjustment

RLAH 6,608** (2,253) 5,432* (1,763) 1,336 (1,528) 0,061 (0,918) 10,152*** (2,672) 8,329*** (2,202) 3,871 (1,952) 0,672 (0,182) Constant 12.642*** (1,853) 13,222*** (1,415) 0,668 (1,941) -5,701** (1,706) 10,674*** (1,940) 11,395*** (1,724) -2,772 (3,152) -14,481*** (2,440) Quarterly time dummies

(42)

42

(9) (10) (11) (12) (13) (14) (15) (16)

6.2b. OLS with fixed time and group effects

SMS per Roamer Mobile Data per Roamer

Unit Units per

Roamer MA Units per Roamer Units per Roamer MA Units per Roamer GB per Roamer MA GB per Roamer GB per Roamer MA GB per Roamer Seasonality adjustment

RLAH -0,012 (0.604) 0,219 (0,458) 0,5923 (0,0545) -0,0116 (0,363) 0,211*** (0,0335) 0,167*** (0,0245) 0,026 (0,0399) -0,009 (0,0217) Constant 5,494 0,605 5,139*** (0,379) 5,980*** (0,0841) 2,721*** (0,518) 0,045 (0.022) 0,067** (0,021) -0,405*** (0,060) -0,667*** (0,040) Quarterly time dummies

(43)

43 Table 5.2 uses the same format as 5.1 and shows the results of four models for each of the four roaming services. With SMS as an exception, the initial FE regressions find significant positive effects of RLAH on the usage per roamer. Where seasonality adjustment seems to again somewhat reduce these, it does not cause a loss of significance of the treatment estimator. However, when individual trends are introduced, all significance is lost and no effect can be observed on any of the four roaming services. Again, combining seasonal adjustment with individual time trends produce the strongest predictive models. While the overall models account for 84 % to 93 % of the within country variance, the treatment estimator loses all statistical significance.

6.3. Domestic telecom prices

Coefficients are marked with *, **, and *** if they are statistically significant to the degree of 𝑃 < 0,05; 𝑃 < 0,01; and 𝑃 < 0,001 respectively. The corresponding robust standard errors are displayed below their coefficient and placed in

brackets.. R2_{indicates the degree of variation within country units that is captured by the model.}

(1) 6.3. OLS with fixed time and group effects

Domestic Prices

Unit ARRPU

Implementation shock 4,887

(3,192)

Divergence over time -0,132

(0,0937) Q1 -0,422* (0,180) Q2 0,078 (0,310) Q3 0,126 (0,310) Q4 Reference R2 _0,052 Observations 335

(44)

44 Even with the relaxed parallel trends assumption and the more interpretative approach, no significant effect was found on the monthly average ARRPU in domestic EEA markets. The post-implementation dummy would have captured an implementation shock and the divergence dummy would have estimated a divergence from the Swiss domestic ARRPU over time. Alas, neither coefficients are significant. Given that the overall model only accounts for 5 % of the variance within country clusters, it seems that it has negligible predictive power.

6.4. Hypotheses

While the estimated size of the effect varies, all models find a significant positive relationship between the implementation of RLAH and the outcome variables of overall intra-EEA inbound, outbound, SMS and mobile data traffic. With that, the following four

𝐻₀hypotheses are rejected: