MASTER THESIS
Double Degree Programme
M.Sc. International Development and Globalization
University of Groningen, Faculty of Economics and BusinessM.Sc. International Economics
University of Göttingen, Faculty of Business and Economics
The impact of the Brexit referendum on the economic
performance of British industries
Mareike Godemann
Student number: S3918130Mail address: m.d.godemann@student.rug.nl
First supervisor: Prof. Dr. B. Los (University of Groningen) Co-assessor: Prof. Dr. F. Unger (University of Göttingen)
Abstract
This thesis studies the consequences of the 2016 Brexit referendum on the economic performance of British industries. This is done for two classifications of economic activities. The costs of the vote are first quantified for 11 broad UK sectors over the period July 2016 to December 2019, followed by a more detailed sectoral analysis examining the impact on industries within manufacturing and financial services between 2016 and 2018. Applying the synthetic control method, it is demonstrated that the public vote caused output losses in the majority of industries, which not only increased over time but also varied across industries. Furthermore, the results support the hypothesis that, for industries within manufacturing and financial services, the dependency on exports to the EU increases the likelihood of experiencing output losses.
i Table of contents List of figures ... ii List of tables ... iv List of abbreviations ... v 1. Introduction ... 1 2. Literature review ... 2
2.1 The referendum shock and its impact on the UK economy ... 3
2.2 Responses to the Brexit vote across industries ... 5
2.3 The synthetic control method and previous applications ... 8
2.4 Determinants of economic growth ... 11
3. Empirical analysis ... 13
3.1 The synthetic control method ... 13
3.2 Inference about the effects of the Brexit vote ... 17
3.3 Challenges in the construction of the synthetic controls ... 18
4. Data ... 19
5. Results ... 21
5.1 The economic cost of the Brexit vote for eleven major British sectors ... 21
5.2 The output effect in industries within manufacturing and financial services ... 26
ii List of figures
Figure 1: Sterling exchange rate with US Dollar (June 2016 –March 2020). ... 4
Figure 2: Distribution of value added across UK´s main sectors, 2015. ... 6
Figure 3: Stylized composition of the doppelganger of the Spanish Basque Country. ... 10
Figure 4: Quarterly average pre-and post-vote value added growth across 11 industries. ... 19
Figure 5: Annual average pre-and post-vote value added growth rates in manufacturing. .... 20
Figure 6: Evolution of quarterly VA growth in the manufacturing and trade sector: UK industry (solid line) versus doppelganger industry (dashed line). Results of 1st set-up. ... 23
Figure 7: Evolution of quarterly VA growth in the financial sector and in public services: UK industry (solid line) versus doppelganger industry (dashed line). Results of 1st set-up. ... 23
Figure 8: Annual output loss as percentage of sectoral value added. Results of 1st set-up. .... 24
Figure 9: Evolution of quarterly VA growth in the manufacturing and trade sector: UK industry (solid line) versus doppelganger industry (dashed line). Results of 2nd set-up. ... 25
Figure 10: Annual output loss as percentage of sectoral value added. Results of 2nd set-up. . 26
Figure 11: Evolution of VA growth in the food and metal industry: UK industry (solid line) versus doppelganger industry (dashed line). Results of 1st set-up. ... 27
Figure 12: Evolution of VA growth in the car and financial services industry: UK industry (solid line) versus doppelganger industry (dashed line). Results of 1st set-up. ... 27
Figure 13: Annual output loss as percentage of sectoral value added. Results of 1st set-up. .. 28
Figure 14: Evolution of VA growth in the car industry: UK industry (solid line) versus doppelganger industry (dashed line). Results of 2nd set-up. ... 29
Figure 15: Annual output loss as percentage of sectoral value added. Results of 2nd set-up. . 29
Figure 16: Testing of hypothesis one: First and second approach (Set-up 2). ... 30
Figure 17: Testing of hypothesis two: First and second approach (Set-up 2). ... 31
Figure 18: Robustness check: Fictitious Brexit vote in donor pool industries (1.1). ... 32
Figure 19: Robustness check: Fictitious Brexit vote in donor pool industries (2.2). ... 32
Figure 20: Robustness checks: Fictitious Brexit vote in 2013(Q1), (2nd set-up). ... 33
Figure 21: Sensitivity analysis. Differences in output losses (percentage points). 1st approach. ... 34
Figure 22: Sensitivity analysis. Differences in output losses (percentage points). 2nd approach. ... 35
iii Figure B2: Evolution of VA growth in the food and metal industry: UK industry (solid line)
versus doppelganger industry (dashed line). 2nd approach results. ... 52
Figure B3: Evolution of VA growth in financial services industry: UK industry (solid line) versus doppelganger industry (dashed line). 2nd approach results. ... 52
Figure B4: Testing of hypothesis one: First and second approach (Set-up 1). ... 53
Figure B5: Testing of hypothesis two: First and second approach (Set-up 1). ... 53
Figure B6: Robustness check: Fictitious Brexit vote in donor pool industries (1.2). ... 54
iv List of tables
Table 1: The four analyses. ... 16
Appendix Table A1: Structure of NACE Rev. 2. ... 42
Table A2: The two largest weights assigned to donor pool country-industries– Approach 1. 43 Table A3: The two largest weights assigned to donor pool country-industries– Approach 2. 44 Table A4: Annual output losses as share in sectoral VA (%) – Approach 1. Set-up 1. ... 44
Table A5: Annual output losses as share in sectoral VA (%) – Approach 1. Set-up 2. ... 45
Table A6: Annual output losses as share in sectoral VA (%) – Approach 2. Set-up 1. ... 45
Table A7: Annual output losses as share in sectoral VA (%) – Approach 2. Set-up 2. ... 46
Table A8: Regression results of the hypothesis testing. ... 47
Table A9: Placebo studies: RMSPE ratios. 1st Approach. ... 47
Table A10: Placebo studies: RMSPE ratios. 2nd Approach. ... 48
Table A11: Robustness to the placebo studies (1st approach). ... 48
Table A12: Robustness to the placebo studies (2nd approach). ... 49
Table A13: Sensitivity analysis: Output losses as share in sectoral VA (%) – Approach 1. .. 50
Table A14: Sensitivity analysis: Output losses as share in sectoral VA (%) – Approach 2. .. 50
v List of abbreviations
DID Difference in differences
EEA European Economic Area
EU European Union
FDI Foreign direct investment
GDP Gross domestic product
GVC Global value chain
ISIC International Standard Industrial Classification
OECD Organisation for Economic Cooperation and Development ONS Office for National Statistics
RDD Regression discontinuity design R&D Research and development
TFP Total factor productivity
UK United Kingdom
US United States
1
1. Introduction
“It is a disgrace that, more than two years after the result of the 2016 referendum, businesses are still unable to plan properly for the future” (Merrick 2019). This statement was given by former chief executive of Airbus, Tom Enders, in January 2019, revealing the firm´s frustration about Brexit uncertainty. It came along with a warning that British plants could be shut down if the situation remained unchanged. Other companies had already taken action. Preparing for Brexit, the British Bank Barclays moved £166 billion of assets to Ireland, and Credit Suisse announced the relocation of around 250 workers from London to other European cities. While some firms were able to adjust their operations relatively swiftly after the vote, others were not as lucky. In the aviation sector, the British regional airline Flymbi had to declare insolvency in February 2019, blaming uncertainty surrounding Brexit as a major reason for the collapse (Spero 2019). Just a few months later, British Steel filed for bankruptcy, communicating that a decline in demand from European customers as a response to the leave vote had put it in the difficult situation (Davies 2019).
These cases demonstrate that before Brexit actually took place, the economic consequences of the public vote were visible. In the days after the Brexit referendum, which was held on June 24, 2016, the pound fell by 11 percent against the US dollar. Even today, its value is still around 10 percent lower than it was immediately before the vote (Partington 2020; Breinlich et al. 2017). The sterling exchange rate against the dollar, a measure of confidence in the British economy, revealed a decline in optimism about the United Kingdom´s (UK) future economic performance. As a consequence, households reduced spending on consumption and investment and, fearing to lose access to the European single market, a significant number of companies decided to reorient towards the remaining 27 member states. Economic growth in the UK slowed down. Born et al. (2019) estimate that the Brexit vote led to a reduction in British gross domestic product (GDP) of 1.7 to 2.5 percent by the end of 2018.
This thesis aims to break down the costs of the referendum at the industry level. Previous studies have analysed the impacts on a national level (Born et al. 2019; Springford 2018), or have investigated the effects on inward and outward foreign direct investment (FDI), which are particularly sensitive to uncertainty and risk (Serwicka and Tamberi 2018; Breinlich et al. 2017). This paper complements literature by focusing on the economic consequences for British industries in terms of gross value added. It is expected that the effects of the leave vote differ across industries for a number of reasons. First, Brexit is a growing source of uncertainty for many firms, but insecurity is particularly high in industries that are heavily dependent on trade with the European Union (EU). It is therefore likely that, in anticipation of higher future trade costs, industries more exposed to the EU through exports made adjustments after the vote, while businesses that are less reliant on EU trade will not have reacted as strongly. Second, even if Brexit exposure is high, certain industries do not have the means to make rapid adjustments. Sectors with large investments in immobile capital may not be able to instantly delay investments, downsize production or relocate to other countries. This paper examines the role of these two factors in the responses of industries to the Brexit vote.
2 unit with the evolution of the same variable in absence of the treatment. It involves the creation of a synthetic group which represents estimates on how the variable of interest would have evolved if the treatment had not taken place. This means that the development of value added in British industries since the vote is compared to estimates on how it would have evolved without the 2016 referendum. The synthetic group, also referred to as the doppelganger, is constructed by using weighted combinations of unaffected control groups, so that the final group closely resembles the treated unit before the Brexit referendum. In this thesis, industries in member countries of the European Economic Area (EEA) as well as Switzerland, hereafter also referred to as European industries, will serve as the donor pool to compose the synthetic British sectors. The difference between the actual UK industries and the synthetic industries will then show the causal impacts of the Brexit referendum on value added.
Two approaches are chosen to study the effects of the vote across industries. First, the consequences are analysed for 11 major sectors in the United Kingdom, where quarterly data ensures a large amount of pre-Brexit vote observations and allows the start of the post-intervention period to be set close to the actual date of the Brexit vote. Second, annual data is used to determine the impact on industries within manufacturing and financial services. Also, two strategies are taken to construct the doppelgangers. In a first step, pre-Brexit value added growth rates are considered to identify industries that resemble the British ones. The more similar they are, the higher the weight assigned to them in the synthetic group. In a second stage, economic determinants of growth are included. Besides a selected number of value added data, these consist of an industry investment ratio, a measure of trade openness, labour productivity growth and levels as well as functional shares for research and development, management, marketing and fabrication in each industry. Data is gathered from the Eurostat National Accounts database, the UK Office of National Statistics, Timmer et al. (2019) and the World Input Output Database (WIOD) (Timmer et al. 2015).
This paper is organized as follows. After giving an overview of the Brexit vote effects on the UK economy, it is discussed why the impacts will vary across industries. In the same section (section 2), the synthetic control method, its previous applications and possible economic determinants are introduced. The next two sections deal more closely with the methodology and data. Section 5 then contains the empirical results. Concluding remarks are made in section 6.
2. Literature review
3
2.1 The referendum shock and its impact on the UK economy
When Great Britain voted to leave the UK on June 24, 2016, it came as a major surprise to the great majority of market participants and the public. In the weeks before the referendum, voter polls and betting markets had largely predicted that UK citizens would choose to remain in the European Union (Born et al. 2019). Even on the day of the vote, bookmakers´ odds suggested a 90 percent chance of the United Kingdom deciding against Brexit (Doyle 2016). The surprising victory of the “leave” campaign did not lead to an immediate withdrawal of the UK from the EU. Britain continued to remain an official member for more than three years before officially exiting the Union only recently, on February 1, 2020.
Its formal partnership with the EU stayed untouched during this period and, from this perspective nothing much changed. The country remained part of the EU´s customs union and single market, and the free movement of people was guaranteed. However, the referendum led to a shift in expectations which affected business decisions and the UK economy in two ways. First, uncertainty increased sharply and remained over the next years. As time passed without an agreement, market participants became increasingly uncertain about the UK´s future relationship with the EU, the timing of the exit or whether it would actually happen (Breinlich et al. 2017). In order to prevent a no-deal scenario, Brexit was delayed several times and even to date, it is still unclear how the future relations with the EU will look like. The exit of the UK in February 2020 has started an 11 months long transition period during which London and Brussels will engage in negotiations to clarify this very question. Given the uncertainty surrounding Brexit, the referendum likely prompted a reaction from businesses and consumers. Firms may have decided to postpone investments and projects until the nature of the future partnership between the EU and UK is clarified. Moreover, UK citizens may have curbed consumption for fear of future loss of income, thereby putting domestic companies in a difficult position.
Second, a shift in expectations regarding Britain´s future openness occurred, possibly making the UK a less attractive destination for businesses and investors (Breinlich et al. 2017). The UK has historically been a major recipient of foreign direct investment - in 2016 it was even the second largest in the world after the United States (US) (UNCTAD 2018, 2017). The country is appealing to foreign investors because of its free trade policy and stable institutional environment and, most importantly, because of its membership in the European Union. Being part of the EU enabled it to benefit from trade deals, research funding and access to the European single market. With the withdrawal from the EU, these arrangements will likely be abolished and it is expected that trade costs will rise. It is therefore likely that, given an increase in future barriers to trade, investment and skilled labour, companies reduced their investment in the UK.
4 of the pound was a clear signal that market participants took a more negative view on the UK´s future economic prospects (Stojanovic and Tetlow 2018). The sterling exchange rate has since been “a barometer of Brexit confidence” (Phillips and Tsang 2019), giving an indication about the trust foreign investors place in the country´s economic stability and strength. The pound did recover temporarily, especially in 2017, after Article 50 was triggered.1 Nevertheless, the value of the sterling in January 2020 was still around 10 percent lower than it was immediately before the referendum (Partington 2020; Breinlich et al. 2017).
Figure 1: Sterling exchange rate with US Dollar (June 2016 –March 2020).
Source: Bank of England database.
The shift in the value of the pound had several effects. On the one hand, the weaker currency made British products more attractive for foreign costumers as the prices of these goods fell. This gave exporters the opportunity to sell goods cheaper abroad, stimulating business. On the other hand, it resulted in an increase in the price of imports, making it more expensive for UK companies to source goods from around the world (Stojanovic and Tetlow 2018). Since exports nowadays require many imports, there is much to suggest that currency depreciations do not significantly improve exports (Fang and Miller 2007). UK businesses were therefore more likely to be negatively affected by the devaluation.
Breinlich et al. (2017) also show that, as a result of a surge in inflation after the leave vote, the living standards of British households declined. By October 2017, consumer price levels had risen by 2.6 percent and although price levels also increased in other parts of the world during the same period, the increase in the UK was significantly larger than in other currency areas. As inflation growth was reflected in a decline in real wage growth, it had a direct impact on living standards. British households subsequently cut back on consumption with the result that consumer spending and business investment only grew half as fast after the vote as it would have without the Brexit referendum (Born et al. 2017). In addition, expectations of future income losses and a decline in the country's economic performance led to further reductions in consumption by households and other market participants. Born et al. (2019) cite these channels as the reason for the slowdown in British economic growth since the vote. Fears of future job
1 By invoking Article 50 (a clause in the Lisbon Treaty of the EU that sets out the steps that a country that wants to leave the EU voluntarily must take) the UK declared its intention to leave the EU and set in motion the formal exit process. 1.2 1.25 1.3 1.35 1.4 1.45 1.5
01 Jun 16 16 Feb 17 06 Nov 17 27 Jul 18 15 Apr 19 06 Jan 20
Fo rwar d exc ha ng e r at e (12 m ), US$ int o S te rl
ing Brexit vote
Article 50 triggered
Brexit transition plans released
Cabinet approves
5 losses may have only accelerated this effect. Los et al. (2017) estimate that 2.5 million jobs are at risk from Brexit.
Another sign of a weaker UK economic performance as a consequence of the referendum is the decline in inward foreign direct investment flows. A report by the UK Trade Policy Observer estimates that while the UK remains a major destination for FDI, the referendum has led to a decline in inflows of around 16 to 20 percent by July 2018 (Serwicka and Tamberi 2018). The American electric vehicle company Tesla, for example, announced at the end of 2019 that the upcoming Brexit played a role in their considerations to set up a plant in Germany instead of Britain (Davey 2019). According to local newspaper articles, the new factory near Berlin could create up to 10,000 jobs (Metzner 2019). Due to its long maturity, FDI is particularly sensitive to uncertainty and risk (Singh and Jun 1995; Resmini 2000), which is why the Brexit vote most likely already had an impact on companies´ decision to invest in the UK.
At the same time, outward investments from the UK to mainland Europe increased. Breinlich et al. (2020) estimate that the Brexit vote led to a 17 percent rise by the first quarter of 2019. As significant transactions to other OECD (Organisation for Economic Cooperation and Development) countries did not take place during the same period, they conclude that maintaining access to the European market was the key driver of business decisions. There are many examples of firms that have chosen to offshore parts of their business to EU countries: Japanese electronics giant Sony moved its European headquarters to the Netherlands (Inagakki and Lewis 2019), as did Panasonic, which set up in Amsterdam (Sachgau 2018), or JPMorgan, which decided to transfer 200 billion euros of balance sheet assets to Frankfurt (Arons and Hadfield 2019). The banking and finance sector, in particular, appears to be taking action. One study concludes that by 2019 more than 250 firms had relocated jobs and parts of their business activities to the EU (Wright et al. 2019).
In sum, the public vote on Brexit had immediate consequences for the UK economy and ample evidence exists, that the economic performance of the UK was weaker than it would have been without the vote. The decline in confidence in Britain´s future economic outlook resulted in a devaluation of the pound and a subsequent increase in import costs. Uncertainty, the surge in inflation and simultaneous decline in wage growth affected households, who subsequently reduced their consumption expenditure. This was particularly felt by domestic businesses. Furthermore, some multinationals have decided to reorient themselves towards the remaining EU27 member states for fear of losing access to the single European market. Economic growth in the UK has therefore been lower than forecast before the vote (Stojanovic and Tetlow 2018). Born et al. (2019) estimate that the referendum caused a reduction in British GDP by 1.7 to 2.5 percent by the end of 2018, resulting in a loss of £55 billion in output.
2.2 Responses to the Brexit vote across industries
6 UK´s four main sectors to gross value added in 2015. The services sector is by far the largest, accounting for more than three quarters of gross value added and therefore representing the core of the economy. This is unlike other countries in the European Union, where dependence on services is considerably lower. In France and Germany, for example, the services sector amounted to 70.21 and 62.21 percent of GDP in 2015, respectively (Plecher 2020a, 2020b). With a share of approximately 79 percent, the services sector in the UK is one of the largest in the EU. Manufacturing, mining, public utilities and construction jointly contribute less than 21 percent to value added, while agriculture only adds 0.7 percent.
Figure 2: Distribution of value added across UK´s main sectors, 2015.
Source: Eurostat database.
As previously discussed, the Brexit vote created uncertainties and led to a loss of confidence among consumers and businesses. In a series of surveys conducted by the Bank of England in the period after the referendum until autumn 2018, around 40 percent of UK firms stated that Brexit was a major source of uncertainty for them. The proportions varied from sector to sector - in summer and autumn 2018, uncertainty was greatest in manufacturing, wholesale and retail, while it was lowest in human health and social work. The surveys also revealed that unease was particularly high in those industries that trade a lot with the EU and are more dependent on EU workers. A positive correlation was also found for firm size – larger firms were more likely to cite Brexit as a top source of uncertainty than smaller ones, possibly because they were more exposed to the EU through exports. Uncertainty persisted throughout the period and concerns about sales, export and labour costs remained strong (Bloom et al. 2019).
A high degree of uncertainty arose in sectors dependent on trade with the EU because of their relatively higher exposure to Brexit. Trading is expected to become more costly in the future, so that any prospective relationship with the EU is of great importance for those industries that trade a lot with continental Europe. British sectors that are most reliant on trade with the EU, both for both exports and imports, are manufacturing, wholesale and retail trade and financial services (Mor 2017). Due to its high degree of integration in global value chains (GVCs), where production is physically separated in different parts of the world (Gereffi et al. 2005), the manufacturing sector in particular procures many intermediate components and services from abroad, which are then processed and re-exported (Ijtsma et al. 2018). Overall, it is the sector with the highest level of participation in GVCs in terms of the use of imported inputs. Manufacturing sources 30.9 percent of all UK imported intermediates, 16.3 percent of which are from the EU (Ijtsma et al. 2018). Alongside mining and quarrying and financial services, it is also one of the three UK sectors most dependent on the EU for its exports. For other industries the exposure is more one-sided. The mining sector exports a lot to the EU, whereas public administration, accommodation and food services as well as health services are reliant on EU imports (Mor 2017). An industry that has received much attention in the news is the financial
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Value added
7 sector, which has profited immensely from passporting. This regulation allows British-based financial institutions to export their services to the rest of the EU without having to obtain regulatory approval (Ford 2017). This decreases the costs of cross-border transactions and investments and is a major reason why many financial institutions have chosen to locate in London. Financial activities are therefore mostly reliant on the EU as a source of income, while imports are not of great importance to them (Mor 2017).
In their paper, Los et al. (2017) assess the exposure to Brexit of UK industries, by evaluating the extent to which they rely on exports to and imports from the EU. Considering the role of global value chains, exposure is measured in terms of value added “that currently crosses a UK-EU border at least once, embodied in a product” (Los et al. 2017, p. 2).They find that up to 33 percent of value added is at risk in some manufacturing industries. Financial and insurance activities, however, only record an exposure level of up to eight percent of value added. Consequently, when considering the dependence on trade with the EU, exposure to the Brexit varies. While some industries face high risks, this is less important to others. Domestic businesses that only serve the local market and do not rely on imports from abroad, hairdressing services for example, are not affected by a future increase in trade costs and will therefore not have to make any adjustments in response to greater future trade costs. However, it is expected that companies where uncertainty is high and which trade heavily with the EU will have made adjustments and that these changes affected real economic outcomes. This is because, when confronted with uncertainty and risk, future expected transaction costs increase so that companies tend to shy away from costly investments (Bloom et al. 2019). Furthermore, customers increasingly anxious about the Brexit vote consequences become more hesitant about placing orders. In anticipation of greater costs and lower profits, investment decisions are postponed or businesses are downsized. These observations lead to the first hypothesis:
H1: The effects of the British referendum on the economic performance of British industries are more visible in industries that are highly dependent on exports to the EU.
8 the extractive industry, where machinery has to be acquired to retrieve natural resources, the energy sector, which requires large investment in infrastructure, the chemical industry (Economy Watch 2010), the transport sector, healthcare or construction. The extractive sector is also limited by geological issues – oil cannot be drilled and minerals are not found everywhere.
In contrast, the majority of service sectors can react more swiftly to economic shocks. In the finance and insurance industry as well as in other service activities, large investments in immobile capital are not necessary to maintain operations. They mainly require “offices and computers” (Serwicka and Tamberi 2018, p. 9) – investments that are easily reversible and that do not require large fixed costs. Activities are therefore more easily downsized or possibly even relocated to remaining EU countries as capital can simply be transferred or acquired in other locations. The possibility of making swift adjustments to a shock is much larger. Firms may have therefore decided to delay investment decisions until the situation is clarified or in the case of multinationals, they may have chosen to outsource to continental Europe. In addition, the decline in consumer demand might have translated faster into a reduction in real economic outcomes, as sectors without large investments in immobile capital have the means to react more quickly. It is also a possibility that some companies decided not to make any adjustments at all while waiting for the outcome of negotiations between Brussels and London. Yet, as evidence on FDI outflows suggests that firms did take action in response to the vote (Wright et al. 2019; Breinlich et al. 2020; Serwicka and Tamberi 2018), the former effect is expected to dominate and therefore that the short-term impact of the Brexit referendum on value added is larger in the sectors that are not dependent on large amounts of immobile capital. This leads to the following second hypothesis:
H2: The effects of the British referendum on the economic performance of British industries are more visible in industries that do not require large immobile capital investments.
The analysis of this paper will concentrate on testing these two hypotheses, but other factors exist that can also play a role in as to why the responses to the Brexit vote differ across industries. For example, differences may arise due to varying expectations regarding future trade barriers. As the British government has focused on reducing custom duties in the future rather than giving importance to remaining in the European single market, this may have induced the services sector to make more adjustments than manufacturing. Other factors may relate to differences in the reliance on skilled labour, as immigration is expected to be more restricted in the future (Breinlich et al. 2020), or to differing responses to the devaluation of the pound. In this case, especially firms that are dependent on inputs from several origins to produce their end products are sensitive to changes in the value of the sterling. Domestic businesses that do not export but are heavily dependent on imports are relatively more affected by the devaluation compared to those companies that also produce locally.
2.3 The synthetic control method and previous applications
9 expected that the effects differ across sectors. To measure the economic costs for UK industries, a statistical method increasingly used in comparative case studies is applied – the synthetic control method. This technique was first introduced by Abadie and Gardeazabal (2003), who examine the economic effects of the terrorist conflict in the Basque Country in Spain. This section gives an insight into the synthetic control method and explores previous applications. A more technical discussion then follows in section 3.
In studying the implications of terrorism on the economic performance of the Basque Country, Abadie and Gardeazabal (2003) were presented with the problem of finding a reasonable method for assessing the causal effects of the conflict. A pure time-series analysis would have led to biased results since Spain had suffered from an economic recession in the period of interest, making it difficult to separate the impacts of terrorism from those of the economic downturn. A simple comparison of the evolution of the Basque Country with a control group also posed a challenge as this would have required finding a Spanish region that resembled the Basque Country very closely before the outbreak of the conflict, particularly with respect to characteristics relevant for economic growth. However, Spanish regions are not driven by the same structural processes. While some of them are characterised by large metropolitan areas with a high proportion of companies in the services sector (e.g. Madrid), others have a relatively more developed agricultural sector (e.g. Galicia). The differences in sectoral composition have strong implications for economic outcomes, so that a comparison of the Basque country to a single other region would have led to results reflecting both the effects of terrorism as well as the heterogeneity in economic growth determinants. The synthetic control method firstly introduced in the paper by Abadie and Gardeazabal (2003) offered a solution. As opposed to simply comparing the effect of the intervention on the Basque Country to a single control unit, it involves the creation of a synthetic control which represents a weighted combination of several non-affected Spanish regions. The method is driven by the idea that a combination of several regions will be more similar to the Basque Country than if only a single one was chosen (Abadie et al. 2010). The causal effects of terrorism on the Basque Country in terms of real per capita GDP are then estimated by comparing the actual evolution of GDP with the development of the synthetic control.
10 one of 0.1492, whereas the weights of the remaining regions take the value of zero and do not form part of the synthetic control. The doppelganger, which demonstrates how the Basque Country would have evolved had the conflict not taken place, thus consists to approximately 85 percent of Catalonia and 15 percent of Madrid. These two regions are selected since their pre-terrorism characteristics are most similar to those of the Basque Country.
Figure 3: Stylized composition of the doppelganger of the Spanish Basque Country.
Source: Own illustration.
Since the first application of the synthetic control method by Abadie and Gardeazabal (2003), many studies have implemented this methodology to analyse the causal effects of several events on GDP. Examples include Cavallo et al. (2013), who examine the impact of large natural disasters in a cross-country comparative case study; Abadie et al. (2015), who measure the economic effects of the 1990 German reunification on West Germany; Campos et al. (2014), who consider the benefits from EU membership for countries that joined the EU in selected years; and Billmeier and Nannicini (2013), who analyse the effects of trade liberalizations for about 180 countries over the period 1963-2000. With respect to the Brexit vote, output costs are measured by applying the synthetic control method in Springford (2018) and Born et al. (2019). They exploit the referendum as a natural experiment, which had not been anticipated nor related to any macroeconomic events, and measure the economic impact on a national level. The methodology is also used to study the referendum effects on FDI inflows (Serwicka and Tamberi 2018) and outflows (Breinlich et al. 2020).
The advantages of the synthetic control method are manifold. It allows to study the dynamics of an intervention, such as the Basque terrorist conflict, and thus how its effects evolve and vary over time. Moreover, it permits to analyse the heterogeneous effects of economic shocks across several treated units – something, the analysis in this paper benefits from. Abadie et al. (2010) also stress that the method is transparent with respect to the weights matched to the donor units and that it provides a safeguard against extrapolation. The latter is possible as the weights are restricted to positive values between zero and one. This ensures that the estimates of the counterfactual do not rely on extrapolation “outside the support of the data” (Abadie et al. 2015, p. 498). If weights were not restricted, some might take negative values or values greater than one. This could result in counterfactuals that extrapolate in order to resemble the treated units
Donor pool The final
doppelganger Selection based on
pre-terrorism characteristics Murcia v v Madrid Catalonia Anda-lucia Aragon Catalonia Madrid 85% 15%
11 more closely. The synthetic control method therefore provides a way of finding a synthetic control that closely fits the outcome variables and characteristics of the treated units and does not rely on data other than those provided by the controls (Abadie et al. 2015).
The method also hinges on a number of assumptions. First, it is assumed that the intervention does not impact the treated unit prior to the treatment. When investing the economic costs for the Basque Country as in Abadie and Gardeazabal (2003), this would require that the economic performance of the region was not affected by the terrorist conflict before its start in 1968. Second, it is assumed that the control groups were not affected by the intervention, ruling out any interference between units (Abadie et al. 2010). Spanish regions other than the Basque Country are thus considered to be unaffected by the conflict. Not meeting these assumptions may produce biased results. For example, if the Spanish regions included in the doppelganger, Catalonia and Madrid, were also negatively affected by the conflict in the Basque Country and therefore performed worse than they would have without terrorism, the synthetic Basque region will show lower outcome levels. This will lead to an underestimation of the cost of terrorism.
2.4 Determinants of economic growth
In studying the impact of the Brexit referendum on British industries in this thesis, it is aimed to construct doppelgangers that reasonably demonstrate how the UK industries would have evolved had the vote not taken place. It is therefore essential to include industries in the doppelganger that are comparable to the British ones. The donor pools for each UK industry are consequently restricted to European industries with the same industrial classification code, hereafter also referred to as country-industries. When studying the impact of the vote on the British steel industry, for example, this implies that the synthetic UK steel industry is created only from a combination of other European steel sectors. Industries sharing different classification codes, such as the French pharmaceutical industry, are not included in the donor pool and thus also not in the respective doppelganger. The industrial classification scheme used in this paper is based on the EU classification system NACE revision 2 which categorizes industries by their business activities in two to four digit codes.2
The weights that donor pool country-industries are then assigned to in the UK doppelgangers, depend on how closely they match the British industries in terms of selected indicators, also referred to as predictors. Papers investigating the economic costs of a political intervention on national GDP with the synthetic control method have made use of a varying set of economic determinants in this respect. As previously discussed, Abadie and Gardeazabal (2003) use pre-conflict real per capita GDP, an investment ratio, population density, sectoral shares of certain industries and human capital indicators to identify donor pool regions that are comparable to the Basque Country (see Figure 3). Born et al. (2019), who study the effects of the Brexit vote at the national level, compose the UK doppelganger by using pre-Brexit vote GDP growth rates to assign country weights. These are complemented by a set of economic variables, including consumption, investment, an export and import ratio as well as labour productivity growth and the UK employment share (the ratio between total employment and the working-age
12 population). In this thesis, a similar but slightly different approach is adopted. To estimate the consequences of the vote for value added at the industry level, in a first step only pre-referendum growth rates of the outcome variable gross value added are used to determine the industry weights. In a second step, a number of economic characteristics are applied as predictors to match the synthetic control to the treated industries. Following both Abadie and Gardeazabal (2003) and Born et al. (2019), these are reported in averages over the pre-Brexit vote period.
Economic indicators are selected that are typically associated with economic growth and thus have a large predictive power for value added. This ensures that those country-industries are chosen to form part of the doppelganger, which follow the same growth paths as the British industries before the vote and can reasonably demonstrate how they would have performed without it. Some of the indicators applied in this paper are also chosen by Born et al. (2019). This is because the analysis conducted here is comparable to theirs, but only broken down to the industry-level - the sum over value added generated in British industries is equivalent to the GDP, the variable of interest of Born et al. (2019). Labour productivity, for example, plays a significant role for output growth and is an important variable for identifying comparable donor pool units in this thesis. Country-industries with similar levels and labour productivity growth are more likely to show a similar development of value added as the British. High levels of labour productivity can stem either from a large amount of production inputs such as human and physical capital or from high total factor productivity (TFP). In the first case, the opportunities for further growth are limited if an increase in capital stock has increasingly less impact on output growth (due to diminishing marginal returns to capital). In contrast, labour productivity growth driven by technological change is unbounded since higher growth can be achieved with the same amount of inputs. Important drivers of TFP are the innovation of new technologies or the acquisition and assimilation of knowledge about their effective use. In sum, it is not only vital to include industries in the donor pool that share the same industrial classification code and can follow very different growth trajectories, but it is important to base the doppelganger selection on economic characteristics such as labour productivity. Other important indicators, also proposed by Born et al. (2019), are investment and trade openness ratios. Investment is a key component for achieving growth - higher investment rates translate into an increasing capital stock, leading to a larger level of output (Weil 2013). Moreover, openness to foreign markets leads to higher growth rates for various reasons. Among others, it raises the incentive for firms to invest in technological innovation and it provides companies with a larger market for their goods and services. Also, it enhances firm productivity since openness facilitates technology transfer among companies and countries and increases competition for local firms, thereby raising efficiency (Weil 2013).
13 fabrication activities. Differences also exist across country-industries. The British chemicals industry, for example, tends to produce management services, the French and German industries research and development services and the Spanish chemical industry marketing activities. In the transportation equipment sector, the English, Spanish and German industry all specialise in fabrication activities, whereas the French sector concentrates more on R&D (Timmer et al. 2019). The type of activity carried out is important as it gives an indication of a country’s and industry´s potential for productivity growth, knowledge generation and of its tendency of being relocated. Furthermore, the amount of labour and returns to each activity differ due to their different usage in worker type and capital. R&D, services, marketing and design have the highest returns in terms of value added, while distributional, purchasing and production activities receive the lowest returns (Gereffi et al. 2005; Gereffi et al. 2001; Sturgeon and Gereffi 2009; Timmer et al. 2019). Therefore, one has to take into account that, although the industries in a donor pool share the same industrial classification code, they differ with respect to the type of activities they carry out. Whereas Born et al. (2019) aim to include countries with the same industry mix in their doppelganger, the analysis in this thesis consequently requires including functional shares as economic determinants. These reveal the extent to which an industry is dedicated to R&D, management, marketing and fabrication.
3. Empirical analysis
In this section, the motivation for choosing the synthetic control method over other techniques in comparative case studies is laid out and the construction of the doppelganger as well as the measurement of the Brexit vote impacts are explained in more detail. Finally, this section presents inferential techniques and discusses how problems that arise in the creation of the synthetic controls are dealt with.
3.1 The synthetic control method
15 value added in the pre-treatment period, whereas similar shocks were not experienced in the post-Brexit vote period. To deal with this issue, countries hit especially hard by the crisis, such as Greece or Spain, are excluded from the donor pool in a sensitivity analysis.
The second approach also has a number of other drawbacks. First, since only annual data can be taken into account, the timing of the Brexit vote is artificially shifted to the beginning of January 2016. Ideally, the synthetic controls should not differ from the real UK industries before the intervention and a gap should only be perceived afterwards, also because the outcome of the referendum was largely unexpected by market participants. However, by shifting the end of the pre-treatment period to the beginning of 2016, it is possible that the counterfactuals already deviate from the real outcome variable before the vote, indicating changes in value added that did not actually happen. Second, data for 2019 and in many cases also for 2018 is not available on such a disaggregated level. Therefore, the post-treatment period is reduced to the years 2016 to 2018 or even 2017, which unfortunately only gives a relatively short observation period. Given these constraints, it is difficult to argue that the analysis results in the second part will provide strong and convincing evidence of the causal effects of the Brexit vote. However, they will give an important indication of the varying responses of British industries to the referendum.
In the following, the approach which is largely based on Abadie and Gardeazabal (2003) and Abadie et al. (2010, 2015), is explained for the case that only one British industry is subject to the intervention. Let 𝐽 + 1 be the observed number of country-industries indexed by 𝑗, where 𝑗 = 1 is the UK industry exposed by the Brexit referendum and the remaining 𝐽 units (𝑗 = 2, … , 𝐽 + 1) refer to the unexposed country-industries that constitute potential doppelganger units. When, for example, the impacts for the UK automotive industry are analysed, 𝐽 is the number of control units from the donor pool of European car industries which are unaffected by the Brexit vote. A distinction is made between two periods: 𝑇0 refers to the number of pre-intervention periods, and 𝑇1 to the post-intervention period, with 𝑇 = 𝑇0+ 𝑇1. Industry 1 is affected by the intervention, the Brexit referendum, during periods 𝑇0+ 1, … 𝑇. A balanced data
set is used, where all units are observed in the same periods 𝑡 = 1, … , 𝑇. The aim is to find a control that approximates the British industry in absence of the referendum. To ensure this, the synthetic control is constructed from a weighted combination of the 𝐽 comparison industries such that it resembles the British industry in the pre-intervention period accurately. Let the synthetic control be a (𝐽 × 1) vector of non-negative weights 𝑊 = (𝑤2, … , 𝑤𝐽+1)´ that sum to
16 economic determinants as predictors to match the synthetic control to the treated industries. Overfitting may arise if the counterfactual is based on too many control industries. Then the synthetic control might consist of a combination of unaffected units that experienced idiosyncratic variations in the outcome variable in the pre-Brexit period. The counterfactual can resemble the treated industry before the intervention very well, but it runs the risk that it does not do so in absence of the treatment (Abadie Forthcoming). Therefore, in a second set-up, economic determinants of the outcome variables are included. Besides a selected number of value added data (growth rates of 2010Q4, 2012Q4 and 2015Q4 as percentage deviations from 2010Q1 in the first approach and growth rates of 2000, 2007, 2009, 2012 and 2015 as percentage deviations from 2000 in the second approach)3, these consist of an industry investment ratio, a measure of trade openness, labour productivity growth and levels as well as functional shares for research and development, management, marketing and fabrication in each industry. In total, four analyses are carried out in this thesis. Table 1 provides an overview. Table 1: The four analyses.
Unit of analysis Predictor variables
Approach I: 11 sectors Quarterly data
Observation period: 2010Q1 – 2019Q4
Set-up I: Value added growth rates Quarterly percentage deviations
from 2010Q1
Set-up II: Economic determinants Approach II: Disaggregated industries
Annual data
Observation period: 2000 – 2017/18
Set-up I: Value added growth rates Annual percentage deviations
from 2000
Set-up II: Economic determinants
In order to identify a combination of country-industries that track the trajectory of the British industry very closely before the vote, it is aimed to find a vector 𝑊∗ that minimizes the
difference between the pre-treatment characteristics of the treated industry and those of its counterfactual. Therefore, let 𝑋1 be a (𝐾𝑥1) vector which contains the pre-Brexit vote characteristics for the exposed industries, where 𝐾 = (𝑘1, … , 𝑘𝑇0)′ indicates a linear combination of pre-treatment outcomes. Also, let the (𝐾𝑥𝐽) matrix 𝑋0 contain the values of the
same variables for the non-British industries. Following Abadie and Gardeazabal (2003), 𝑊∗ is then selected by finding 𝑊 that minimizes (𝑋1− 𝑋0𝑊)′𝑉(𝑋1− 𝑋0𝑊). The relative importance of each predictor variable is reflected in the elements of 𝑉, a diagonal matrix with positive values. Accordingly, variables with large predictive power are given higher weights. Thus, if the functional mix is more important in predicting value added growth than trade openness, the method ensures that the doppelganger contains industries that share the same functional shares. Conversely, if the trade openness index has a larger predictive power, country-industries that are as open as the British are selected. In case of the car manufacturing
17 sector, the vector 𝑊∗ then comprises the weighted combination of unexposed EU car industries which match the British car industry as close as possible before the referendum.
The impact of the Brexit referendum is then determined by the difference between the development in value added of the doppelganger and the respective industry in the period after the vote. Therefore, the effect on industry 𝑗 at time 𝑡 > 𝑇0, denoted by 𝛼𝑗𝑡, is measured by
𝛼𝑗𝑡 = 𝑌𝑗𝑡𝐼 − 𝑌𝑗𝑡𝑁, (1)
where 𝑌𝑗𝑡𝐼 is the observed outcome for the industry exposed to the Brexit vote and 𝑌𝑗𝑡𝑁 is the outcome of the doppelganger. Since 𝑌𝑗𝑡𝐼 can be observed, it remains to estimate 𝑌𝑗𝑡𝑁. This is done by constructing synthetic controls from the donor pool, as explained in the previous paragraphs.
3.2 Inference about the effects of the Brexit vote
In order to quantify whether the estimated gap between the outcome of the exposed industry 𝑌𝑗𝑡𝐼 and the synthetic control 𝑌𝑗𝑡𝑁 is truly attributable to the intervention and not a result of factors other than the Brexit vote, a series of placebo tests are carried out. In a first step, synthetic controls are constructed for industries in the donor pool that are very similar to the British ones. Following Abadie and Gardeazabal (2003), this is done for those country-industries which are assigned the largest weights in the synthetic industries. This approach is adopted to assess whether the estimated gap for the affected industries is large relative to comparable country-industries that are not exposed to the actual Brexit vote. If the placebo tests show that the magnitude of the effects is similar to the one estimated, this would indicate that the constructed doppelgangers do not produce reliable forecasts of the outcomes without the vote. Conversely, small gaps suggest that the estimated ones truly capture the causal effects of the Brexit vote. The results of the placebo analysis are quantified by computing a statistic of pre- and post-intervention fits of the synthetic controls, the root mean squared prediction error (RMSPE) (Born et al. 2019). The RMSPE is the rooted average of the squared differences between actual value added and its synthetic counterpart in the pre- or post-intervention period. For industry 𝑗 = 1 exposed to the vote, the pre- and post RMSPE are calculated as follows:
𝑅𝑀𝑆𝑃𝐸𝑝𝑟𝑒 = √ ∑ (𝑌1𝑡𝐼 −𝑌1𝑡𝑁) 2 𝑇0 𝑡=1 𝑇0 ; 𝑅𝑀𝑆𝑃𝐸𝑝𝑜𝑠𝑡= √ ∑𝑇𝑡=𝑇0+1(𝑌1𝑡𝐼 −𝑌1𝑡𝑁)2 𝑇1 . (2), (3)
For comparison purposes, the ratio of the post-period and the pre-period is then taken (𝑅𝑀𝑆𝑃𝐸𝑝𝑜𝑠𝑡/𝑅𝑀𝑆𝑃𝐸𝑝𝑟𝑒), with large values indicating big post-intervention doppelganger gaps. A statistical test on the significance of the RMPSE ratio does not exist, thus no definitive statements about the statistical significance of the output effects of the Brexit vote can be provided. However, the RMSPE ratio gives a good indication of whether the results can be relied upon.
18 difference in growth performance between the doppelganger and the actual UK industries prior to the actual Brexit vote, thus if the two lines already start diverging from 2013 onwards, this would indicate that the estimated effects do not truly respond to the intervention. Conversely, if no large differences are observed, this would make confident that the causal effects of the referendum are estimated.
3.3 Challenges in the construction of the synthetic controls
To ensure the applicability of the synthetic control method in examining the causal effects of the Brexit vote on the economic performance of British industries, a number of issues in the construction of the synthetic control need to be addressed.
First, it is important to have only industries in the donor pool that share similar characteristics and are thought to be driven by the same structural processes as the treated units in order to reduce issues such as overfitting. This motivated the choice of European industries for the donor pools, since these would be most comparable to the UK industries. Second, the method assumes that there is no interference between the industries during the post-treatment period. In Abadie and Gardeazabal (2003) this meant that Spanish regions other than the Basque Country were considered not to be affected by the terrorist conflict. In this paper the non-interference assumption is violated as this would require industries in the European Union to not be exposed to the Brexit vote. However, the referendum also led to the creation of uncertainties in other parts of the world and certain European industries, like their British counterparts, will have made adjustments in response to the vote. First and foremost, those industries dependent on trade with the UK. When measured as the ratio of a sector´s exports to the UK in the country´s total exports to the world, exposure is, for example, especially high in Spain and Germany in the transport vehicles sector, while in the mechanics industry high levels of risk are found for Ireland and the Netherlands (European Committee of the Regions 2018). Moreover, the Brexit vote also led to an outflow of foreign direct investment, primarily in the financial sector. Here the main recipient countries were Ireland, Luxembourg, France, Germany and the Netherlands (Wright et al. 2019). Certain European industries may therefore have benefitted from the Brexit vote, while others in turn may have been negatively impacted. If this is truly the case, it will contaminate the donor pool and lead to biased results. If some European industries performed better than they would have without the vote, the synthetic group will show higher outcome levels resulting in an overestimation of the Brexit cost. Conversely, if the Brexit vote caused a reduction in the performance of European industries, the estimated effect will be smaller in magnitude. The violation of the non-interference assumption therefore requires that industries in the European Union strongly affected by the Brexit vote are excluded from the donor pool. This will be done in a sensitivity analysis.
19
4. Data
The bulk of the data is retrieved from Eurostat, which provides national accounts aggregates by industry for the remaining 27 member states of the European Union, the United Kingdom, the additional members of the European Economic Area (Iceland, Liechtenstein and Norway) as well as for Switzerland, summing to a total of 32 countries (Eurostat 2020). Thus, the donor pool for each treated industry consists of 31 country-industries with the same two-digit industrial classification code which are expected to have similar characteristics to the UK units. Industries are classified at the NACE Revision 2 level, an EU system dividing economic activities into 21 main groups.
The outcome variable, gross value added, is gathered in chain linked volumes (with 2010 as the base year) in million euros. This is done to ensure comparability across countries, as not all members of the EU are part of the Eurozone. Moreover, it puts the value added data in real terms, thereby removing any effects of price changes (i.e. deflation or inflation) so that variations in value added only reflect changes in production volumes. Gross value added data can be retrieved at the annual and quarterly level. However, quarterly data is only provided by Eurostat for 11 sector aggregates, whereas a more detailed industry breakdown is only available for values at the annual level. This motivates the adoption of two approaches to analyse the effects of the Brexit vote across British industries. As discussed in section 3, in a first step, quarterly data is used to estimate the impacts for 11 broad sectors of the UK economy. This estimation approach covers the period from 2010 to 2019, where the pre-treatment period is set from the first quarter (Q1) of 2010 to 2016Q2. The post-Brexit vote period then starts in 2016Q3 and ends in 2019Q4. Figure 4 gives an illustration of the quarterly average growth rates across 11 different UK sectors before and after the vote. It is clearly visible that the growth rates of the British sectors did not change homogenously after the referendum. Many recorded a decline in their growth, whereas others, such as construction or the public sector (public administration and defence, education, human health and social work activities) did not experience any large changes in value added after 2016.
Figure 4: Quarterly average pre-and post-vote value added growth across 11 industries.
Source: Eurostat National Accounts database.
Quarterly value added growth declined by approximately 0.34 percent in the manufacturing sector after the vote, but a more comprehensive picture is given when examining the changes at a more disaggregated level. Figure 5 depicts the variations in the outcome variable for 12
20 industries within the manufacturing sector. As only annual data is taken into account in this figure, the growth rates of the post-vote period also contain the pre-vote months of 2016. The illustration clearly indicates a heterogeneity in growth rate changes after the vote. Whereas some industries do not seem to be negatively affected and even experienced large increases (e.g. pharmaceutical products or computer, electronical and optical products), others recorded large declines in gross value added growth. Growth in the car manufacturing sector (motor vehicles), for example, declined by 6.6 percent or in the electrical equipment industry by nearly three percent. These heterogeneous growth rates experienced by industries at a more disaggregated level, motivate the adoption of the second approach.
Figure 5: Annual average pre-and post-vote value added growth rates in manufacturing.
Source: UK Office for National Statistics.
In this step, annual data is used to estimate the impact of the referendum on a smaller sectoral breakdown. The pre-Brexit period is extended to the years 2000 to 2015 and due to use of annual data the timing of the Brexit vote is artificially shifted to the beginning of January 2016. Moreover, Eurostat does not provide data for 2019 and in many cases also not for 2018 on such a disaggregated level. Therefore, the post-treatment period is reduced to the years from 2016 to 2017 or 2018. Furthermore, data for the UK is not available from 2017 onwards on Eurostat. Thus, gross value added by industry is retrieved from the British Office for National Statistics (ONS) (Office for National Statistics 2020). This does not have any implications for the data quality since the ONS reported exactly the same data to Eurostat before. The outcome variable is gathered in chained volume measures in 2016 money value. To ensure comparability with the Eurostat data, the series are rebased to 2010 volumes and converted to euros using the average 2010 exchange rate of the British pound to the euro (1 pound equals 1.166 euros). Data for the predictor variables is also mainly retrieved from the Eurostat database. In the first approach pre-intervention growth rates of the outcome variable are relied on to determine industry weights. As explained in section 3, the risk of overfitting motivates the adoption of the second approach. Here, a series of economic determinants are included as predictor variables. In addition to a number of pre-Brexit vote value added growth rates, that are reported as percentage deviations from 2010Q1 in the first approach (i.e. the growth rates of 2010Q4, 2012Q4 and 2015Q4 as deviations from 2010Q1), and as percentage deviations from 2000 in the second approach (i.e. the growth rates of 2000, 2007, 2009, 2012 and 2015 as deviations from 2000), the industry investment ratio, labour productivity growth and levels, an indicator
21 for trade openness and a functional index for research and development, management, marketing and fabrication in each industry is added. All variables are reported in averages over the pre-treatment period. The investment rate is calculated as the ratio of gross fixed capital formation to gross value added in constant prices. Labour productivity is obtained by taking the difference between the logarithm of (quarterly) value added and (quarterly) total employment and labour productivity growth by calculating the (quarterly) growth rates of the ratio of gross value added to total employment in each industry. Also, the trade openness index is computed taking the ratio of the sum of exports and imports to value added. The data on imports and exports by economic activity is gathered from 2000 to 2014 from the WIOD database (Timmer et al. 2015). To obtain trade data for the last pre-intervention year 2015, the WIOD values are extrapolated using growth rates from the Eurostat database. Whenever no sufficient data is available, extrapolation occurs using the average growth rates of the pre-Brexit vote period. Lastly, the functional shares are obtained from Timmer et al. (2019), who have uploaded replication files on the website of the World Input-Output Database. For each industry at the International Standard Industrial Classification (ISIC) Revision 3 level, business function shares for four stages of the production process research and development, management, marketing and fabrication, together summing to one, are available. Although reported in an industrial classification system that is not perfectly comparable to NACE Revision 2, the indices are matched as best as possible to the industries using correspondence tables between ISIC Revision 4 (comparable to NACE Revision 2) and ISIC Revision 3. The shares are reported for the years 1999 to 2011, and since they are only given for industries at a more disaggregated level, this variable is only included as a predictor in the second part of the analysis. This does not pose a problem, as production fragmentation is in any case more interesting for the manufacturing sector, an industry which is analysed in more detail in the second part. Functional shares are reported as averages over the 1999 to 2011 period. Finally, it should be noted that, given limitations of data availability for some predictor variables, the size of the donor pool varies across industries and therefore does not always consist of 31 country-industries.
5. Results
This section presents the main findings of the analysis. In a first step, the results are discussed for the first approach, where, using quarterly data, the effects of the Brexit vote are analysed for 11 major sectors in the UK. In a second step, the findings are presented for the second set-up, in which the responses to the referendum are studied across industries within manufacturing and financial services. Section 5.3 contains the results of the hypothesis testing. The section concludes with a discussion of the outcomes of the robustness checks and the sensitivity analyses.
5.1 The economic cost of the Brexit vote for eleven major British sectors
22 period following the referendum shock. Since two set-ups are taken to assign weights to the country-industries (see Table 1), this section first presents the outcomes of the analysis using pre-intervention values of value added, before discussing the findings when economic determinants are included as predictor variables.
Figure 6 and 7 plot the results of the synthetic control method of the first approach for four British sectors: manufacturing, trade, transport and accommodation (also referred to as trade sector), finance and insurance as well as public services. With exception of the financial and insurance services, where the pre-treatment fit of the doppelganger is slightly looser than in the other sectors, the synthetic counterparts track the value added growth over the entire pre-Brexit vote period 2010Q1 to 2016Q2 very well. The two country-industries receiving the largest weights in the doppelganger of the UK manufacturing are the Norwegian (0.3565) and German (0.1082) manufacturing industries and in the trade sector these are the Hungarian (0.3786) and Portuguese (0.1759) trade industries. In the synthetic counterpart of the British finance and insurance sector, the Belgian (0.3619) and French (0.3072) industries are assigned the largest weights, whereas the Belgian (0.7508) and Czech (0.1037) public sectors compose the largest share in the synthetic UK public services.4 When visually inspecting the data and comparing the pre-Brexit vote economic growth determinants of the doppelgangers with those of the UK sectors, the composition of counterfactuals seems reasonable for the majority of industries. In manufacturing, for example, the pre-referendum characteristics of the Norwegian and German sectors are very similar to those of the British. Only the investment ratio in Germany (18.7 percent) is slightly higher than that of the UK (13.4 percent). Moreover, the number of country-industries receiving weights is very large – a sign for the existence of overfitting. In manufacturing as much as 11 units form part of the synthetic control and in the trade and finance sector nine country-industries are assigned positive weights.
In accordance with expectations, the impact of the vote varies across sectors. As displayed in Figure 6, it is estimated that both the manufacturing and trade sector would have grown stronger had the referendum not taken place, yet the magnitude and timing of the effect differs. In manufacturing, an impact is clearly visible shortly after the vote. Although real value added growth did not begin to slow until early 2018, the two series start to diverge substantially just a few months after the referendum and the gap continues to widen over the post-intervention period. It is found that, by the end of 2019, gross value added of the doppelganger is about 10 percentage points higher than of the real UK industry. The effect is even more pronounced in the trade sector, where by the end of the post-intervention period, a difference of approximately 13 percentage points is estimated. Here, the effects of the vote are not perceivable immediately after the treatment – the lines of the synthetic counterpart and the real trade industry first start diverging in 2018. The reduction in consumer spending by British households as well as the devaluation of the pound which resulted in an increase in the price of imports, might give an explanation for the strong impact of the vote on this sector. The depreciation of the sterling may have also supported the adverse effects on manufacturing, where, in addition, uncertainty and exposure to the Brexit is very high due to the strong dependence on trade with the EU.
