Political referenda and investment: evidence from Scotland

Working Paper Series

Andres Azqueta-Gavaldon Political referenda and investment: evidence from Scotland

No 2403 / May 2020

Disclaimer: This paper should not be reported as representing the views of the European Central Bank (ECB). The views expressed are those of the authors and do not necessarily reflect those of the ECB.

Abstract

We present evidence that referenda have a signicant, detrimental outcome on in- vestment. Employing an unsupervised machine learning algorithm over the period 2008- 2017, we construct three important uncertainty indices underlying reports in the Scot- tish news media: Scottish independence (IndyRef)-related uncertainty; Brexit-related uncertainty; and Scottish policy-related uncertainty. Examining the relationship of these indices with investment on a longitudinal panel of 3,589 Scottish rms, the evidence suggests that Brexit-related uncertainty associates more strongly than IndyRef -related uncertainty to investment. Our preferred specication suggests that a one standard- deviation increase in Brexit uncertainty foreshadows a reduction in investment by 8% on average in the following year. Besides we nd that the uncertainty associated with the Scottish referendum for independence while negligible at the aggregate level, relates more strongly with the investment of listed rms as well as those operating on the border with England. In addition, we present evidence of greater sensitivity to these indices among rms that are nancially constrained or whose investment is to a greater degree irreversible.

keywords| Political uncertainty, investment, machine learning, textual-data

JEL classications: C80, D80, E22, E66, G18, G31

ECB Working Paper Series No 2403 / May 2020

1

Non-technical summary

Scotland has recently experienced two signicant episodes when political uncertainty might have been especially pronounced: the Scottish referendum on independence in September 2014 (secession from the United Kingdom) and the Brexit referendum in June 2016 (on the UK leaving the European Union). Both of these events were preceded by extensive and intensive periods of national debate. These debates were often fractious and resulted in many claims that a 'Leave' vote (for Scotland to leave the UK or for the UK to leave the EU) would result in widespread economic uncertainty as they would usher in possibly protracted periods of political wrangling until trading regimes and the wider business environment were resolved.

The central aim of this paper is to quantify these two political uncertainty shocks, and to study their relationship with investment. To measure political uncertainty, we use an unsupervised machine learning algorithm to subdivide overall economic uncertainty reported in the news-media into dierent topics or themes. The unsupervised machine learning algorithm called Latent Dirichlet Allocation (LDA) studies the co-occurrences of words in news-media articles to frame two distributions: a distribution of words composing a topic and a distribution of topics for each document (news article). One can then track through time the evolution of the topics describing the uncertainty measures of interest. In other words, the LDA approach allows one to decompose economic policy uncertainty into endogenously determined sub-indices, without need to read the individual newspaper articles and apportion their content across pre-determinedsub-indices. Nonetheless, given that the topics uncovered by this approach are simply described by a set of words, it is left to the researcher to justify the labelling of each topic. However, it turns out that the LDA approach recovers indices that naturally comprise distinct political sources of uncertainty.

For example, in analyzing the Scottish press we label as 'IndyRef ' (Independence Ref- erendum) that index whose most representative words given by the LDA algorithm include: independence, SNP [Scottish National Party], referendum, party, vote, minister, Scotland and election. This index increased steadily from the moment when the UK Parliament approved the Scottish referendum for independence (January 2012), until its actual oc-

ECB Working Paper Series No 2403 / May 2020

2

currence in September 2014, rising again around mid-2016.Additionally, we label 'Brexit uncertainty' that index whose most representative words include: EU, Brexit, European, UK, negotiations, leave, country, membership, single and trade. That index peaked during the Brexit referendum in June 2016, and at the general election in June 2017. In addition, once we compare these two referendum-relateduncertainty indices, IndyRef and Brexit, with the proportion of individuals that Google searched Scottish Independence" and Brexit" in Scotland, we observe strong similarities: 0.78 and 0.81 correlation respectively. This reassured us that we are capturing uncertainty, understood as the second moment.

We then examine the relationship between the indices just described and rm investment by applying a standard investment regression to a longitudinal panel dataset of 3,589 Scottish rms during the period 2008-2017. Our baseline results suggest that a one standard- deviation increase in Brexit uncertainty foreshadows a reduction in investment by 8% on average in the following year. Besides we nd that the uncertainty associated with the Scottish referendum for independence while negligible for the overall rm network, relates more strongly with the investment of listed and border companies (those operating on the border with England).

We subject our baseline results to a battery of robustness tests. First, we incorporate a wide range of familiar variables into the empirical model that aims to explain the investment behaviour, such as cash-ows, sales growth rates, and GDP growth rates. Second, we add into the model alternative measures of uncertainty, such as the implied volatility index (VFTSE), election year dummies of various sorts, and an overall measure of UK Economic Policy Uncertainty index (EPU). Additionally, we ensure that the results are robust to several econometric approaches, including simple panel regressions both with and without xed e ects, a rst-di erence speci cation, as well as dynamic panel speci cations estimated using a System GMM estimator.

To study the most plausible mechanisms through which uncertainty impacts investment, we investigate whether investment across di erent types of companies respond equally to uncertainty. First, we distinguish between non-manufacturing and manufacturing rms. The Decision Maker Panel survey reported that rms in the manufacturing sector are the

ECB Working Paper Series No 2403 / May 2020

3

most likely to move part of their operations outside the UK due to the uncertainty produced by Brexit. Nonetheless, more recent evidence suggests that business con dence from the manufacturing sector has actually increased after Brexit. We nd evidence supporting this latter behaviour: Scottish manufacturing companies have been less negatively a ected by political uncertainty.

Second, we distinguish between listed and non-listed companies. Listed companies may be less likely to su er from nancing constraints than their non-listed counterparts to the extent that asymmetric information is less of a problem to them. That said, they may face more risk due to having a larger share of operations abroad, thus making them especially vulnerable to referendum uncertainties. We observe that investment from listed companies present greater sensitivity with political uncertainty, especially that uncertainty arising from the Scottish referendum for independence.

To further investigate to what extent the nancing constraints channel is behind these results, we construct two nancing constraints proxy variables commonly used in the litera- ture. Thus, we use company size and age to reect the possible impact of external nancial constraints whilst the 'coverage ratio' and 'cash-ows' to quantify the possible intensity of internal nancial constraints. We nd evidence that those rms that are more likely to be nancially constrained display higher drops in investment in the presence of uncertainty. This holds principally for rms with either internal or external nancing constraints and Brexit uncertainty. Finally, we study rms with potentially high degrees of irreversible investment. Consistent with priors, we nd a stronger negative relationship between rms whose investment is more likely irreversible and political uncertainty.

The resulting policy implications may be important, in particular to the current economic climate. Referenda are becoming a popular tool for politicians, yet their consequences as a source of uncertainty often escape the political debate. In this paper, we show not only that referenda are a signi cant source of political and policy uncertainty but also that they a ect private investment independently of their outcome.

ECB Working Paper Series No 2403 / May 2020

4

• Introduction

There is growing acknowledgement that economic policy uncertainty can have a signi- cant impact on economies, and in particular on rms' investment decisions. Scotland has recently experienced two signicant episodes where such uncertainty might have been especially pronounced: the Scottish referendum on independence in September 2014 (secession from the United Kingdom) and the Brexit referendum in June 2016 (on the UK leaving the European Union). Both of these events were preceded by extensive and intensive periods of national debate. These debates were often fractious and resulted in many claims that a 'Leave' vote1 (for Scotland to leave the UK or for the UK to leave the EU) would result in widespread economic uncertainty as they would usher in possibly protracted periods of political wrangling until trading regimes and the wider business environment were resolved.

As Figure 1 shows, the Brexit referendum campaign started o more nely balanced than the independence referendum campaign in Scotland. However, as the dates of both referenda drew near, the polls narrowed, in some measure as undecided voters decided which way to vote. The solid lines in the gure are a linear extrapolation of the Remain and Leave votes recorded in various polls through the campaigns (other extrapolative techniques tell the same story). That apparent convergence in the votes, may itself have been an additional source of uncertainty and we shall examine that possible eect later. Of course, in the end, Scotland voted to remain in the UK (55% to 45%) whilst the UK voted to leave the European Union (52% to 48%).

In the case of the Scottish referendum, it may be the case that much of the political (independence-related) uncertainty has resolved, or is at least somewhat diminished. On the other hand, signicant changes to 'devolved scal policy' (in particular to income tax raising powers) were introduced following the referendum and so policy uncertainty, a priori , need not have diminished. In other words, scal policy in Scotland may now diverge

• In the Scottish Independence Referendum (IndyRef for short) the question posed to voters was: 'Should Scotland be an independent country?' The political campaigns were organized around a Yes or No vote. For the EU Referendum the question was: 'Should the United Kingdom remain a member of the European Union or leave the European Union?' The political campaigns were organized around a vote to Remain or Leave. It is convenient simply to refer to Leave or Remain votes for either referendum.

ECB Working Paper Series No 2403 / May 2020

5

from rUK (the rest of the UK, excluding Scotland) in potentially signicant ways. And, of course, it is not clear that a second Scottish referendum on independence is o the political agenda. We will try to examine the extent to which this political (i.e., referendum-related) uncertainty has been resolved. As far as the EU referendum is concerned, it appears that much uncertainty, both political and policy related remains. The central aim of this paper is to attempt to identify the underlying sources of economic policy uncertainty (EPU) and to see which are more deleterious to investment: Are referenda an independent source of EPU and, if so, how costly are they? In doing this, we build on recent research which has established that economic policy itself can create an uncertain investment environment.

The principal challenge in extending the literature on policy uncertainty is isolating an appropriate measure of political/referenda-related uncertainty. In the literature, the overall economic uncertainty faced by a country has been measured using a variety of proxy variables, such as the dispersion in the forecast of GDP growth, implied volatility indices, or survey-based rm reports of investment uncertainty. A seminal development has been the news-based Economic Policy Uncertainty index developed by Baker, Bloom, and Davis (2016). Such indices describe primarily uncertainty concerning which or when economic policies the government will implement. However, measuring the portion of uncertainty attributable to the political system and in particular applicable to Scottish issues alone is rather challenging using their approach.

To ll this gap, we use an unsupervised machine learning algorithm to subdivide overall economic uncertainty reported in the news-media into dierent topics following the approach of Azqueta-Gavaldon (2017). The unsupervised machine learning algorithm called Latent Dirichlet Allocation (Blei, Ng, and Jordan (2003)) studies the co-occurrences of words in news-media articles to frame two distributions: a distribution of words composing a topic and a distribution of topics for each document (news article). One can then track through time the evolution of the topics describing the uncertainty measures of interest. In other words, the LDA approach allows one to decompose economic policy uncertainty into endogenously determined sub-indices, whilst the unsupervised machine algorithm makes the analysis feasible. Hence, there is no need to read the individual newspaper articles and apportion their content across pre-determinedsub-indices. Nonetheless, given that the

ECB Working Paper Series No 2403 / May 2020

6

topics uncovered by this approach are simply described by a set of words, it is left to the researcher to justify the labelling of each topic. However, as we describe briey now, and in more detail below, it turns out that the LDA approach recovers indices that naturally comprise distinct political and policy sources of uncertainty.

For example, in analyzing the Scottish press we label as 'IndyRef ' that index whose most representative words given by the LDA algorithm include: independence, SNP [Scot- tish National Party], referendum, party, vote, minister, Scotland and election. This index increased steadily from the moment when the UK Parliament approved the Scottish referendum for independence (January 2012), until its actual occurrence in September 2014, rising again around mid-2016.Additionally, we label 'Brexit uncertainty' that index whose most representative words displayed by the algorithm include EU, Brexit, European, UK, negotiations, leave, country, membership, single and trade. That index peaked during the Brexit referendum in June 2016, and at the general election in June 2017.

In addition, once we compare these two referendum-related uncertainty indices with the proportion of individuals that Google searched Scottish Independence" and Brexit" in Scotland, we observe strong similarities: 0.78 and 0.81 correlation respectively. The similarity between our referendum-related uncertainty indices and Google searches implies two things: i) IndyRef and Brexit indeed capture relevant events related to these two referenda;

1. given that internet users look for online information when they are uncertain (Casteln- uovo and Tran (2017)), it reassured us that we are capturing uncertainty, understood as the second moment, and not just the rst moment of beliefs. Furthermore, we label the index 'Scottish policy uncertainty' whose most representative words include: Scotland, Scottish,

government, budget, public, education, need, fund, report and tax. That index peaks when the Scottish Parliament approved the SNP's administration's budget at the second time of asking (Feb 2009); the Scottish public-sector strikes (November 2011) and Brexit (June 2016).

We then examine the relationship between the indices just described and rm investment by applying a standard investment regression to a longitudinal panel dataset of 3,589 Scottish rms. Our baseline results suggests that a one standard deviation increase in

ECB Working Paper Series No 2403 / May 2020

7

Brexit uncertainty foreshadows a reduction in investment by 8% on average in the following year. Besides we nd that the uncertainty associated with the Scottish referendum for independence, while negligible for the overall rm network, had a negative and signi cant outcome on the investment of listed and border companies (those operating on the border with England).

Our results appear signi cant and consistent with some important recent ndings in the literature. For example, Gulen and Ion (2015), in examining US rms over the period 1987:Q1-2013:Q4, found that a one standard deviation increase in policy uncertainty is associated with an average decrease in quarterly investment rates of 6%. In addition, Azzimonti (2018), studying the period 1987:Q1-2017:Q4, found that a one standard deviation increase in her Partisan Conict index over the period led to a drop in quarterly US investment of 13% of the sample mean. Regarding political uncertainty, Jens (2017) found that gubernatorial elections in the United States depresses investment by 5% on average while Dibiasi et al. (2018) found that the economic policy uncertainty induced by the 2014 referendum vote on Mass Immigration in Switzerland reduced irreversible investment by as much as 25-30% in exposed rms. In line with the Brexit referendum, Born et al. (2019) found that the Brexit vote caused a reduction in GDP by approximately 2% by the second quarter of 2018 and policy uncertainty accounts for 30% of this e ect.

We subject our baseline results to a battery of robustness tests. First, we incorporate a wide range of familiar variables into the empirical model that aims to explain investment behaviour such as cash-ows, sales growth rates, and GDP growth rates. Second, we add into the model alternative measures of uncertainty, such as the implied volatility index (VFTSE), election year dummies of various sorts, and an overall measure of UK Economic Policy Uncertainty index (EPU). Additionally, we ensure that the results are robust to several econometric approaches, including simple panel regressions both with and without xed e ects, a rst-di erence speci cation, as well as dynamic panel speci cations estimated using a System GMM estimator.

To study the most plausible mechanisms through which uncertainty relates to invest- ment, we investigate whether investment across di erent types of companies respond equally

ECB Working Paper Series No 2403 / May 2020

8

to uncertainty. First, we distinguish between non-manufacturing and manufacturing rms. The Decision Maker Panel survey reported that rms in the manufacturing sector are the most likely to move part of their operations outside the UK due to the uncertainty produced by Brexit (Bloom, Bunn, et al. (2017)). Nonetheless, more recent evidence suggests that business con dence from the manufacturing sector has actually increased after Brexit (see Born et al. (2019)). We nd evidence supporting this latter behaviour: investment from Scottish manufacturing companies appear less sensitive to political uncertainty.

Second, we distinguish between listed and non-listed companies. Listed companies may be less likely to su er from nancing constraints than their non-listed counterparts to the extent that asymmetric information is less of a problem (Carpenter and B. C. Petersen (2002)). That said, they may face more risk due to having a larger share of operations abroad, thus making them especially vulnerable to referendum uncertainties. We observe that investment from listed companies appears more sensitive to political uncertainty (in a negative way), especially to that arising from the Scottish referendum for independence.

To further investigate to what extent the nancing constraints channel is behind these results, we construct two nancing constraints proxy variables commonly used in the litera- ture. Thus, we use company size and age to reect the possible impact of external nancial constraints whilst the 'coverage ratio' and 'cash-ows' to reect the possible intensity of internal nancial constraints (see Guariglia (2008)). We nd evidence that those rms that are more likely to be nancially constrained also display higher drops in investment in the presence of uncertainty. This holds principally for rms with either internal or external nancing constraints and Brexit uncertainty.

Finally, we study rms with potentially high degrees of irreversible investment. Drawing on Chirinko and Schaller (2009), we use depreciation rates to proxy for investment irreversibility. This proxy is motivated by the fact that, in addition to selling capital, rms can reduce their capital stock through depreciation. Therefore, rms with low depreciation rates face higher risks when making capital purchases under uncertainty. Consistent with priors, we nd a stronger negative relationship between rms whose investment is more likely irreversible and political uncertainty.

ECB Working Paper Series No 2403 / May 2020

9

This paper relates to at least three strands of literature. The rst is research on the impact of uncertainty on investment. Theoretical work on this topic dates back to Bernanke (1983) who reveal that high uncertainty gives rms an incentive to delay investment when investment projects are costly to undo.2 Recent empirical literature (and which we closely follow) is Gulen and Ion (2015) who examine the impact of economic policy uncertainty on US rms investment over the period 1987:Q1-2013:Q4. They found a signicantly stronger eect of uncertainty on investment for rms with a higher degree of investment irreversibility and for rms that are more nancially constrained. Other empirical studies connecting political risk/uncertainty and economic activity are Azzimonti (2018) and Jens (2017).

Second, there are interesting studies examining explicitly the impact of referenda on the economy. Using a time-dummy approach (1 for when the referendum took place and 0 otherwise), Dibiasi et al. (2018) found that the economic policy uncertainty induced by the 2014 referendum vote on Mass Immigration in Switzerland reduced irreversible investment by as much as 25-30% in exposed rms. Also using a timeline approach, Darby and Roy (2019) examined the impact of the Scottish referendum on stock market volatility. They observed increases in the relative volatility of Scottish companies' stock returns compared to the rest of the UK when polls suggested the referendum result was too close to call. Finally, using a synthetic control method, Born et al. (2019) found that the Brexit vote caused a reduction in UK's GDP by approximately 2% by the second quarter of 2018 and that policy uncertainty accounts for 30% of this eect.

Finally, there is a rapidly growing literature on textual methods to measure a variety of outcomes. In their seminal contribution, Baker, Bloom, and Davis (2016) used newspaper coverage frequency and simple dictionary techniques to measure Economic Policy Uncertainty (EPU).3 Hansen, McMahon, and Prat (2017) used Latent Dirichlet Allocation on the Federal Open Market Committee talks to study communication patterns. Using simple text-mining techniques, Hassan et al. (2019) built a political risk measure as the share of

• R. K. Dixit and Pindyck (1994) oer a detailed review of the early theoretical literature.

3EPU indices have been replicated with more advanced methods (see Azqueta-Gavaldon (2017) or Saltz-

man and Yung (2018)).

ECB Working Paper Series No 2403 / May 2020

10

rm-quarterly conference calls that are devoted to the political risk for the USA.4 They found that increases in their rm-level measure of political risk are associated with significant increases in rm-specic stock return volatility as well as with signicant decreases in rms' investment, planned capital expenditures, and hiring. More recently, combining word-embedding and LDA algorithms, Azqueta-Gavaldon et al. (2020) built several EPU indicators for Spain, Italy, France and Germany.

The rest of the paper proceeds as follows: Section 2 describes the algorithm and news- media data used to produce the specic uncertainty indices for Scotland. Section 3 presents the data and econometric framework to study the eects of uncertainty on private invest- ment. Section 4 shows the empirical ndings of the average eect of uncertainty on invest- ment. Section 5 displays the analysis of cross-sectional rm heterogeneity under political uncertainty. Section 6 contains robustness tests and Section 7 concludes.

• Political and policy uncertainty in Scotland

2.1 LDA model

To obtain the distinctive narratives of political uncertainty embedded in the news media, we use the approach described in Azqueta-Gavaldon (2017). This approach applies an unsupervised machine learning algorithm to all news articles describing economic uncertainty (all news articles containing any form of the words economy and uncertainty) in order to unveil the wide range of themes or topics described on it. The unsupervised machine learning algorithm, called Latent Dirichlet Allocation (LDA) and developed by Blei, Ng, and Jordan (2003), reveals the themes across articles without the need for prior knowledge about their content. Intuitively, the algorithm studies the co-occurrences of words per articles to frame each topic as a composition of the most likely words (more likely to appear together) while each article is framed as a distribution of topics.

In other words, LDA is a generative probabilistic model that infers the distribution

4To come up with political topics, they rst lter political topics by correlating them to sources with a priori political vocabulary e.g. political sciences textbooks. They then count the number of instances in which these political-related words appear together with synonyms of risk or uncertainty.

ECB Working Paper Series No 2403 / May 2020

11

of words that denes a topic, while simultaneously characterizing each article with a distribution of topics. The model recovers these two distributions by obtaining the model parameters that maximize the probability of each word appearing in each article given the total number of topics K. The probability of word wi occurring in an article is:

K
Xj
P (wi) = P (wijzi = j)P (zi = j)	(1)
=1

where zi is a latent variable indicating the topic from which the ith word was drawn and P (wijzi = j) is the probability of word wi being drawn from topic j. Moreover, P (zi = j) is the probability of drawing a word from topic j in the current article, which will vary across dierent articles. Intuitively, P (wjz) indicates which words are important to a topic, whereas P (z) is the prevalence of those topics within an article. The goal is therefore to maximize P (wijzi = j) and P (zi = j) from equation (1). However, direct maximization turns out to be susceptible of nding local maxima and showing slow convergence (Griths and Steyvers (2004)). To overcome this issue, we use online variational Bayes as proposed by Homan, Bach, and Blei (2010). This method approximates the posterior distribution of P (wijzi = j) and P (zi = j) using an alternative and simpler distribution: P (zjw), and associated parameters.5

2.2 New article Data

We apply the LDA algorithm to three of the most read Scottish newspapers: The Herald (UK coverage and based in Glasgow), The Scotsman (UK coverage and based in Edin- burgh), and The Aberdeen Press and Journal (largely Scottish coverage). Because we are interested in building an aggregate political uncertainty index, that is, to what extent the general public (and in particular rms' CEOs) got exposed to news portraying the various sources of political wrangling, we do not dierentiate between political position or sympathy across these news outlets. For example, one could imagine that more conservative news outlets would tend to describe political uncertainty around Brexit to a lower degree than more liberal ones would. Nevertheless, provided that these news outlets are among the most read ones in Scotland, we are condent that they serve our purpose.

• For more details about the implementation see Rehurek and Sojka (2010).

ECB Working Paper Series No 2403 / May 2020

12

We use Nexis, an online database of journalistic documents to gather all news articles containing any form of the words 'economy' and 'uncertainty' from these three newspapers.6 The total number of news articles associated with any form of these two words from Jan- uary 1998 to June 2017 (inclusive) was 18,125. In this corpus, the aggregate of all articles, there are over one million words. Following usual practice in the literature, we preprocess the data (words). Stopwords are removed: that is, words that do not contain informative details about an article: e.g., that or me. All words are converted to lower case, and each word is converted to its root (known as 'stemming'). Finally, to nd the most likely number of topics K, we use a likelihood maximization method. This method consists of estimating empirically the likelihood of the probability of words for a dierent number of topics P (wjK). This probability cannot be directly estimated since it requires summing over all possible assignments of words to topics but can be approximated using the harmonic mean of a set of values of P (wjz; K), when z is sampled from the posterior distribution (Griths and Steyvers (2004)). This method indicates that the most likely number of topics in this corpus is K = 20 (see Table 1).

Table 2 displays all the 20 topics identied by the LDA algorithm in our corpus. Column 3 shows the most representative words for each topic given by the algorithm (in lower cases and root format). A useful method to further scrutinize how well LDA captures the essence of the corpus is to apply a visual representation of the sizes and distances between topics in the two-dimensional space. We use the LDAvis method developed by Sievert and Shirley (2014) to accomplish this task. Figure 2 represents each topic as a disc whose area denotes that topic's prevalence in the corpus; essentially, the bigger the disk, the more important the topic in the corpus. Furthermore, the inter-topic-distances between topics describe the similarities between them. These distances are given by the Jensen-Shannon divergence and are scaled by Principal Components in the two-dimensional space (see Sievert and Shirley (2014)); the closer the disks, the more the topics (words with a high probability of belonging to that topic) overlap. Furthermore, one observes that most of the information in this corpus lies within the top right-hand quadrant (top-right corner of Figure 2), indicating a

• Recall that news articles containing any form of the words economy and uncertainty describe overall economic uncertainty (see Baker, Bloom, and Davis (2016)).

ECB Working Paper Series No 2403 / May 2020

13

degree of similarity between most of the topics, as one would expect given that our corpus was constructed to focus on economic uncertainty. Recall, our interest is not so much in overall economic policy uncertainty, but in the constituent components of that uncertainty (policy uncertainty, Brexit, and so on). As we will discuss in more detail below, that quadrant is indeed mostly populated by policy uncertainty related topics.

One observes in Figure 2 that the two referendum topics (Topics 1 and 12) appear very close together and even overlap. Nonetheless, even though they are related by some of the most characteristics words associated with each topic, they are still distinct from each other according to the LDA (two di erent discs). Whether that distinctiveness is statistically or econometrically signi cant in explaining investment is, of course, of central importance. Also closely aligned are the topics related to Scottish policy uncertainty (Topic 6), monetary policy uncertainty (Topic 4) and agricultural policies (Topic 13). More distant to the core topics, but still of some signi cance in the overall corpus and still connected with Scottish policy uncertainty, we nd topics reecting labour policies (Topic 9), nancial regulation (Topic 10), and North Sea oil (Topic 8). From all these topics, we choose the three topics centrally related to political and Scottish policy uncertainty:7

• IndyRef: independ, snp, mr, referendum, parti, vote, labour, minist, scotland, elect, campaign, would, sturgeon
• Brexit Uncertainty: eu, brexit, european, britain, europ, union, uk, negoti, leav, countri, membership, singl, trade, brussel
• Scottish Policy Uncertainty: scotland, scottish, govern, budget, busi, univers,

public, educ, need, fund, council, report, tax

• Although there are other topics related to Scottish policy uncertainty we choose Topic 6 for our study for two reasons. First, it is the largest of the topics describing Scottish policy uncertainty (9% of the total news describing economic uncertainty) and, second, it is the closest to the two referendum Topics. Also note that while the topic Preferences (Topic 3) seems related to the two referendums, we do not take it into account for two reasons. In the rst instance, its meaning is highly ambiguous and hence dicult to map to observable economic variables. In addition, once transformed into a time series, see next paragraph, Topic 3 is only weakly correlated with the two referenda uncertainty indices: -0.01 with IndyRef and 0.17 with Brexit uncertainty.

ECB Working Paper Series No 2403 / May 2020

14

Building each time series requires a few extra steps. First, we label each article according to its most representative topic (the topic with the highest percentage in the article). Next, we produce a raw count of the number of news articles for every topic each month (20 raw time-series). Finally, since the number of news articles is not constant over time, we divide each raw time-series by the total number of news articles containing the word today each month (the proxy for the total number of news articles, see Azzimonti (2018)).

2.3 Uncertainty indices

Figure 3 shows the evolution of IndyRef, Brexit uncertainty and Scottish policy uncertainty indices from Jan 2008 through June 2017. IndyRef covers around 10 per cent of all news articles describing economic uncertainty. It shows spikes when the UK Government legally approved the Scottish referendum for independence (Jan 2012); when the chancellor of the Exchequer George Osborne argued that a 'Yes' vote meant Scotland giving up the pound (Feb 2014)8; the Scottish referendum for independence (Sept 2014); and Brexit (June 2016). 'Brexit uncertainty' (4 percent of all economic uncertainty news) shows its peak at the time of the Brexit referendum (June 2016); it also rises in the run-up to the general election of June 2017. Lastly, Scottish policy uncertainty (9 percent of all economic uncertainty news) peaks when the SNP budget was approved following initial rejection (Feb 2009); Scottish public sector strikes (Nov 2011)9, and, most notably in the run up to the Brexit vote (June 2016).

To validate that these indices are capturing periods of high uncertainty, we compare each uncertainty index with the implied volatility index of the FTSE (VFTSE) and Google searches. On the one hand, the VFTSE index uses implied option volatilities information which represents the market consensus of future UK stock market volatility. This index is based on market data, is forward-looking and is often referred to as the investor fear gauge; the higher the index, the greater the fear (Whaley (2000)). Signicantly, implied volatility indices are often used as a proxy for overall uncertainty (see for example Baker, Bloom, and Davis (2016) and Gulen and Ion (2015)).

• See http://www.bbc.co.uk/news/uk-scotland-scotland-politics-26166794.

• See http://www.bbc.co.uk/news/uk-scotland-scotland-politics-15938970.

ECB Working Paper Series No 2403 / May 2020

15

Setting the nancial and European debt crises aside (where the VFTSE shows its two most prominent peaks), the implied volatility index and the uncertainty indices display some similarities in the run up to the Scottish referendum for independence and Brexit. However, it is interesting to note that after Brexit the implied volatility index, in contrast to the three uncertainty indices, did not rise but remained somewhat subdued. This indicates that uncertainty perceived by nancial markets after Brexit was not as high as the one apparently being picked up by our three political/policy uncertainty indices. An interesting question, therefore, is whether these three measures of uncertainty are able to contribute in explaining investment whilst controlling for the VFTSE and other more standard measures of uncertainty.

To further validate these uncertainty indices, we compare them with Google searches available via Google Trends. The data provided by Google Trends is freely available in real time and it has been used before to construct uncertainty indicators. For example, Castel- nuovo and Tran (2017) use words associated to uncertainties about future economic conditions such as bankruptcy", stock markets", economic reforms" or debt stabilization" to construct an uncertainty index for the United States and Australia. The assumption is that economic agents, represented by internet users, look for online information when they are uncertain (Castelnuovo and Tran (2017)). This assumption implies that an increase in the frequency of terms associated to future, uncertain events results from high periods of uncertainty. With this in mind, we compare the proportion of individuals who searched Scottish Independence" and Brexit" in Scotland via Google with our political news-based uncertainty indices.

As can be seen by the discontinuous red line in Figure 4, developments in the proportion of individuals who searched Scottish Independence" via Google closely resembles the IndyRef uncertainty index (0.78 correlation). The rst notable increase in this particular Google search occurred when the UK Government legally approved the Scottish referendum for independence (Jan 2012). In addition, just like in the IndyRef index, the second most prominent spike takes place when the chancellor of the Exchequer George Osborne argued that a 'Yes' vote meant Scotland giving up the pound (Feb 2014) while the most

ECB Working Paper Series No 2403 / May 2020

16

prominent spike occurs during the Scottish referendum for independence (Sept 2014). Even though the No won the Scottish referendum, there are two important spikes in the Google search and in the IndyRef in the aftermath of the referendum. The rst one occurs in the month of Brexit: shortly after the Brexit referendum results, the SNP advocated for another Scottish independence vote on the justication that Scotland voted in favour of the UK staying in the EU by 62% to 38%. The second one takes place in March 2017; when the Scottish parliament voted to demand a second independence referendum (69 to 59 votes).10 Nonetheless, this proposition was rejected by the U.K. Prime Minister Theresa May and therefore a second Scottish independence referendum scheduled for Autumn 2018 was cancelled.

Besides, the dynamics in the proportion of individuals who searched Brexit" in Scot- land via Google and the Brexit uncertainty index are also very similar (0.81 correlation); both spiking in the month of the referendum and displaying high levels in the aftermath. Note, however, that the uncertainty indices created via the conventional press are preferred over those built using Google Trends for four main reasons. Firstly, we do not need to impose any query and therefore risking ad hocness. Secondly, the conventional press-media is likely to lead Google searches since agents react to what they read in the news by searching for additional information online. Thirdly, one can only retrieve Google Trends data as far back as 2004, limiting the time span. Finally, Google Trends does not provide an exact measure of the number of times a given query was formulated but oers a re-scaled time series from 0 to 100. In this regard, we do not know whether Scottish Independence" was searched by 2 million people at its peak (September 2014) or only a few thousands. In both cases, it would display a maximum peak of 100.

• Firm level data and methodology

3.1 Data

To perform the analysis, we extract the data from the prot and loss and balance sheet section assembled by the Bureau Van Dijk Electronic Publishing, and available in the Fi-

10See ttps://www.ft.com/content/195d9986-13d1-11e7-80f4-13e067d5072c.

ECB Working Paper Series No 2403 / May 2020

17

nancial Analysis Made Easy (FAME) dataset. This dataset provides yearly information on British and Irish companies for the period 2008-2017. To be consistent with the uncertainty measures, we include in the analysis only companies with registered oce address or primary trading address in Scotland. The companies selected perform in a wide range of economic sectors: agriculture, forestry and mining; manufacturing; construction; retail and wholesales; hotels and restaurants; and business and other services.11

We measure the investment rate as the purchase of xed tangible assets by the rm over its capital stock at t 1. Investment is the di erence between the book value of tangible xed assets at the end of year t and the end of year t 1, plus depreciation at t, whilst the capital stock is xed tangible assets at t 1.12 The other two variables of interest are cash-ows (CF) which is computed as the sum of rm's after-tax pro ts and depreciation, and sales growth rates (SG).

Finally, we exclude rms that do not have complete records on investment, cash-ows, or sales growth rates, as well as those companies with less than three years of observations. Also, to control for the potential inuence of outliers, we exclude observations in the 1% tails for each of the regression variables. These rules are common in the literature and also aid comparability with previous work (Guariglia (2008); Gulen and Ion (2015)). The nal data used in the estimation comprises 3,589 companies or 22,769 rm-year observations. Of these rms, 800 operate in the manufacturing sector and 43 are listed companies (see Table 3). Comparing column 1 and column 2 in Table 3, we can see that even after imposing these lters on the data, the nal sample is similar to the entire FAME universe for Scottish rms. On average over the period 2009 to 2017 our sample of companies account annually for around 40% of the total workforce of interest (total employment less those employed in banking and nancial services and the public sector).13

1. For standard reasons, we exclude companies operating in the nancial and regulated sectors.
2. Sometimes, the normalizing variable is not the capital stock but the replacement value of the capital

stock calculated using the perpetual inventory formula (Blundell, Bond, et al. (1992)). In our short sample,

the replacement value of the capital stock produced a signicant downward trend in the overall investment

(see Chirinko and Schaller (2009) for discussion). It is for this reason that we prefer using the capital stock.

13Specically, our rms employed annually on average over the sample 524,680 individuals (after re- moving outliers). The aggregate employment level in the economy, less that in banking and nancial services and the public sector, during the same time period was on average (annually) 1,342,422, see

ECB Working Paper Series No 2403 / May 2020

18

3.2 Econometric framework

To study the relationship between investment and uncertainty, we employ the classical investment regression augmented to include political and policy uncertainty measures and a set of macroeconomic variables:

	Ii;t	= i + 1P Ut1 + 2	CFi;t	+ 3SGi;t + 4Mi;t1 + 5Dt1 + i;t	(2)
Ki;t1	Ki;t1
where i = 1; 2; :::; N indexes the cross-section dimension and t = 1; 2; :::; T	the time
series dimension. Ii;t=Ki;t1	is the ratio between investment in xed tangible assets and

the capital stock at the beginning of the period; i is a rm xed e ect which captures rm-speci ctime-invariant omitted variables; P Ut1 indicates the yearly average news uncertainty indices; CFi;t=Ki;t1 corresponds to cash-ows scaled by the capital stock at the beginning of the period and SGi;t stands for sales growth rates. Dt1 and Mt1 contain a set of yearly dummy and macroeconomic variables meant to control for possible seasonality and other time-dependent factors of investment. Finally, standard errors are clustered at the rm level to correct for potential cross-sectional and serial correlation in the error term it (M. A. Petersen (2009)).

Because the uncertainty indices are rm invariant, time-xedeects cannot be incorporated into this basic econometric framework since doing so would entirely absorb the coecients of our uncertainty indices. So,to address concerns that results might be driven by time-dependentfactors such as business cycles or year-speciceects, we include a battery of macroeconomic variables (Mt1 ) to account for such eects. An important concern in the literature when studying the impact of uncertainty on investment comes in the form of countercyclical behaviour of political/policy uncertainty: during bad economic outcomes,

policy-makers often feel increasing pressure to make policy changes" (Gulen and Ion (2015)). To this end, we use Scottish GDP growth rates14 to control for business cycles (in line with Azzimonti (2018); Gulen and Ion (2015); Baker, Bloom, and Davis (2016)). Unfortunately,

https://www.gov.scot/Topics/Statistics/Browse/Labour-Market/Local-Authority-Tables.

14Available at http://www.gov.scot/Topics/Statistics/Browse/Economy/PubGDP.

ECB Working Paper Series No 2403 / May 2020

19

GDP growth rates during the sample are positively correlated with the IndyRef index, see Table 4. For this reason, we need to be particularly cautious when interpreting the coef- cient of IndyRef and this is why both results, with and without GDP growth rates, are discussed.15

There are a number of other issues which we try to address/control for in the subsequent analysis. These issues are largely concerned with whether or not our political and policy uncertainty indices are really justi ed in being so labelled. For example, our political uncertainty indices might be recording risk derived to a greater or lesser extent from election years, when investment tends to drop (see for instance Julio and Yook (2012)). In this case, we add a dummy variable which takes the value 1 if during that year a Scottish parliamentary election occurred and 0 otherwise (in line with Gulen and Ion (2015)).

Finally, note that we include the natural logarithm of the implied volatility index (VFTSE obtained from Bloomberg) which serves as a proxy for overall uncertainty. Recall that a graphical comparison of the three measures of uncertainty and the VFTSE suggested that after the Brexit referendum these measures diverged somewhat (see Figure 3). The uncertainty indices we construct indicate heightened uncertainty, in apparent contrast to the VFTSE.

It is worth mentioning that controlling for cash-ows and sales growth rates aim at capturing expected pro tability/investment opportunities, that is, the rst moment e ects (Gulen and Ion (2015)). In the case that these rst moment e ects are not properly accounted for by these variables, the rm xed e ects as well as other macroeconomic vari- ables, we might have biased coecients. Nonetheless, since we always use lagged values of the uncertainty variable with respect to the dependent variable, omitted variables bias is unlikely. This is because our uncertainty measures are predetermined, which means that their e ects are estimated consistently in our speci cations (see Hayashi (2000), p. 109). In addition, this lagging technique also helps to alleviate any reverse causality concerns.16

15We also tried di erent measures to control for business cycles such as dummy variables for when GDP

growth rates are positive/negative, and for the UK's GDP growth rates. Worth is mentioning that using

these alternative speci cations, the results remain unchanged.

16Note that the cash-ows and sales growth rates variables are not lagged while the uncertainty measures

ECB Working Paper Series No 2403 / May 2020

20

• The Average eect of Political Uncertainty on Investment

Table 5 shows our baseline empirical results from estimating equation (2). To facilitate interpretation, each uncertainty coecient has been normalized by its sample standard de- viation. Therefore, each coecient may be interpreted as the change in the investment rate associated with a one standard deviation increase in uncertainty. Panel A shows the results without controlling for business cycles while Panel B adds Scottish GDP growth rates to control for them. Overall, our results show that each of the three uncertainty indices is estimated to impact investment negatively and highly signicantly when entered separately. However, when we include the uncertainty indices jointly, the explanatory power becomes centred on Brexit uncertainty.

Columns (1) through (3) include only one of the three uncertainty indices. Column

1. reports the results including only IndyRef uncertainty. There we observe that a one standard deviation increase in uncertainty implies a decrease in investment in the following year of -0.077 when controlling for GDP growth rates (Panel B). That is equivalent to a decline of 23% in the average rm investment rate for the whole sample (I=K = 0:34, see Table 3). As mentioned, GDP growth rates and IndyRef uncertainty are positively corre- lated in the run-up to the referendum. Hence, when we exclude GDP growth rates (Panel A) we estimate the coecient of the IndyRef index to be -0.028, equivalent to a drop of 8% in the average rm investment rate for the whole sample. This change in magnitude when excluding GDP growth rates really aects only the coecient of the IndyRef index whereas other estimated coecients remain largely unchanged following the exclusion of GDP growth. Nevertheless, this suggests that multicollinearity is an issue between those two variables.17

are. This is done in order not to lose a year of observations. That said, results remain unchanged when these two variables are lagged and we con rmed that the uncertainty measures at t 1 have no predictive

power for cash-ows nor sales growth at t.

17The Variance Ination Factor (a tests to study multicollinearity), reveals values much greater than 10 for IndyRef when GDP growth rates are included in the regression equation.

ECB Working Paper Series No 2403 / May 2020

21

Column (2) reports the results with only Brexit uncertainty included. Here we see that the coecient of Brexit uncertainty remains pretty much unchanged when excluding/includ- ing GDP growth rates: -0.045 and -0.046 (Panel A and B respectively). These magnitudes are equivalent to a drop in the average investment rate of 13.2% and 13.5% respectively. Besides, when Scottish policy uncertainty is included alone (column (3)), it reports a coe- cient equivalent to a fall of 9% in the average investment rate when excluding the business cycles control (Panel A) and 10% when including it (Panel B).

Next, we challenge the explanatory power of each referendum uncertainty index by simultaneously controlling for Scottish policy uncertainty (columns (4) and (5)).18 It turns out that both coecients on the referenda uncertainty indices drop in value. That is especially so for IndyRef when excluding GDP growth rates, which is no longer signicant. This indicates a strong link between IndyRef and Scottish policy uncertainty: the explanatory power observed when IndyRef was set alone is absorbed completely by Scottish policy uncertainty. As we will see in the robustness tests below, IndyRef displays a negative and signicant coecient once we replace Scottish policy uncertainty with the UK policy un- certainty. This is not the case for Brexit uncertainty, which remains statistically signicant after controlling for Scottish policy uncertainty (column (5)). Nonetheless, the coecient on Brexit uncertainty drops from 13% to 8% but remains highly signicant. This indicates also a relationship between the uncertainty caused by Brexit and Scottish policy uncertainty (being the coecient of this latter uncertainty no longer signicant).

Overall these results expose the gravitational eect that Brexit uncertainty had on the other two indices. This comes as no surprise since Brexit, on the one hand, has induced policy changes at the Scottish level while, on the other hand, has fuelled the debate for a second Scottish referendum for independence. Indeed, shortly after the Brexit referendum results, the SNP advocated for another Scottish independence vote on the justication that Scotland voted in favour of the UK staying in the EU by 62% to 38%. In March 2017, the

18Note that due to multicollinearity problems that arise when placing the two uncertainty indices together, we exclude the implied volatility index (VFTSE). Using the Variance Ination Factors we detected values much higher than 10 for the VFTSE when all controls were placed, something which indicates pronounced multicollinearity.

ECB Working Paper Series No 2403 / May 2020

22

Scottish parliament voted (69 to 59 votes) to demand a second independence referendum.19 Nonetheless, following the decline in SNP votes during the UK general election (June 2017), Nicola Sturgeon announced that the Scottish government would postpone legislation concerning a second referendum for independence.20

The overarching signicance of Brexit uncertainty is apparent when the three uncertainty indices enter jointly (column (6)). In this setting, only Brexit uncertainty remains negative and signicant.21 In this formulation, a one standard deviation increase in Brexit uncertainty foreshadows a drop in the average investment rate of 12% in the following year. That is barely unchanged in the case when Brexit uncertainty was postulated as the sole source of uncertainty. To further study how political uncertainty has evolved during and after the referenda took place, in what follows we incorporate a set of dummy variables aiming to isolate the two referenda events and to check also whether or not simple dummy variables have more explanatory power than our uncertainty indices.

We rstly undertake this latter exercise by incorporating simple year-dummy variables describing when the referenda took place. We label these year-dummy variables as SCOTreferendum and BREXITreferendum (1 in the year the referendum took place and 0 otherwise). To be consistent with our measurements of uncertainty, all dummy variables are lagged by one year. First, these dummy variables are considered on their own (columns (1) and (4) of Table 6). We observe that although both are negative (except for IndyRef when GDP growth rates are excluded, column (1) in Panel A), only the coecient associated with the Brexit referendum is statistically signicant. This seems to conrm the insight from Table 5 on the relevance of Brexit.22

More importantly, however, once we add our referenda uncertainty measures IndyRef and Brexit (columns (2) and (5) respectively), they prevail over the dummy variables; in

1. See https://www.ft.com/content/195d9986-13d1-11e7-80f4-13e067d5072c.
2. See https://www.bbc.co.uk/news/uk-scotland-40415457.

21Once again, we had to drop the implied volatility index and GDP growth rates from the regression

equation due to strong multicollinearity indicated by the Variance Ination Factors test. For this reason,

the results in both panels are the same.

22Note that even though these dummies are included individually, the results are unaltered even when the two dummy variables are included.

ECB Working Paper Series No 2403 / May 2020

23

all cases only the uncertainty indices are statistically signicant. This holds independently of whether or not we include/exclude GDP growth rates (Panels A and B). Therefore, we conclude that our uncertainty measures have important explanatory power over and above simple referendum-year dummies. These results also hold when incorporating a dummy variable for the period when the Scottish referendum was being legislated: 2012-2014.

Next, we investigate whether or not IndyRef displays any eect on investment once the uncertainty after the Scottish referendum is removed. In other words, we sought to isolate the uncertainty that may have been present in the run-up to the Scottish referendum from any post-referendum uncertainty. When IndyRef uncertainty is included on its own (column (1) of Table 5) the size of its estimated coecient was substantially larger than when put together with Brexit uncertainty. The implication, therefore, may be that IndyRef was picking up some of the eects of Brexit uncertainty. For this reason, we now interact IndyRef with a dummy variable that removes any post-Scottish referendum uncertainty (SCOT2014 = 1 from the beginning of the sample period up until the year of the referendum and 0 afterwards). To be consistent with our lagged uncertainty measure, this time dummy variable is also lagged by one year. Column (3) displays the results also controlling for Brexit uncertainty with the dummy variable BREXITreferendum. The interaction term IndyRef *SCOT2014 turns out to be negative although not signicant. In this scenario, a one standard deviation increase in IndyRef, once removing the uncertainty post-referendum, suggests a drop in investment of 4% in the following year.

All in all, the results presented in these two tables allow us to be condent of a strong relationship between Brexit uncertainty and rm investment. The most conservative results -including Scottish EPU and excluding GDP growth rates- foreshadows a drop in average investment rate in the following year by 8% (column (5) Panel A Table 5) while when Brexit uncertainty enters alone, this magnitude represents a drop of 14% of the average investment rate (column (2) Panel A Table 5). Taking into account that Brexit uncertainty rose by 2.65 standard deviations, the lower-bound Brexit uncertainty eect on investment adds to 21.5%.

Regarding the link between IndyRef and investment, results seem to indicate a weak and non statistically signicant relationship between investment and uncertainty related

ECB Working Paper Series No 2403 / May 2020

24

exclusively with the Scottish referendum for independence. Results when excluding the uncertainty period after the Scottish referendum for independence and business cycles indicate that only the uncertainty regarding the Scottish referendum for independence foreshadows a drop in average investment rate in the following year by 4% (although not statistically signicant).

It is worth to mention, however, that the results displayed in this section should be taken with certain cautiousness. Firstly, we have a relatively low amount of years in our sample period. As a result of this, we could only place a limited number of aggregate variables aiming at capturing time-dependent factors that could confound the eect of political uncer- tainty. Secondly, the time frequency comes at a yearly average which, as we have noted with our uncertainty measures, tend to move at a higher frequency. Nonetheless, we have seen that our measures of uncertainty prevail over a simple time-dummy approach (they remain statistically signicant once we incorporate a time dummy variable for when the referenda took place). To somewhat dissipate these concerns, in the following section we will be able to incorporate time-xed eects in our regression by interacting our uncertainty measures with rm characteristics known to be more sensitive to uncertainty. This should reassure us that any negative eect observed in investment comes indeed from political uncertainty.

• Heterogeneous eect of uncertainty

So far, we have assumed that the relationship between uncertainty and investment is equal across the dierent types of companies. However, there are reasons to believe that this may not be the case. For example, there might be cross-sectional heterogeneity among sectors, corporate structure, or balance sheets. In addition, investment decision is not equally costly for all rms in the economy since there might be variations in the degree of investment irreversibility or nancial constraints.

To study the plausible cross-sectional heterogeneity link between uncertainty and invest- ment, we include an interactive term for the uncertainty measure and a dummy variable describing dierent heterogeneous rm characteristics. Note that here we are no longer

ECB Working Paper Series No 2403 / May 2020

25

interested in estimating the average eect of political uncertainty on investment. This allows us to replace the uncertainty indices, macro and time dummy controls in equation (2) with a time xed eect. This has the added benet of controlling for any macroeconomic, cross-sectionally invariant forces which may confound the eect of political uncertainty:

Ii;t	= i + t + 1P Ut1 Hi + 2	CFi;t	+ 3SGi;t + i;t	(3)
Ki;t1	Ki;t1

where Hi stands for heterogeneity characteristics that are time invariant. In the case of having cross-section and time-variant heterogeneity Hi;t the econometric equation will look like the following:

Ii;t	= i + t + 1P Ut1 Hi + 2Hi;t + 3	CFi;t	+ 4SGi;t + i;t	(4)
Ki;t1	Ki;t1

In both of the above equations, the interactive coecient 1 is the coecient of interest. It allows us to evaluate whether or not the eect of uncertainty on investment is likely to have been equal across rms with specic characteristics and their counterparts.

5.1 Manufacturing and listed companies

Recent surveys indicate stronger adverse eects of the uncertainty derived from Brexit for the manufacturing sectors than the rest of industries. For example, the Decision Maker Panel survey reported that rms in the manufacturing sector are more likely to move part of their operations outside the UK on account of uncertainty due to Brexit (Bloom, Bunn, et al. (2017)). Conversely, as results presented in Panel A from Table 7 show, investment from the 800 Scottish manufacturing companies display a lower sensitivity with political uncertainty than their counterparts. While all the interacted coecients are positive, only those from IndyRef and Scottish policy uncertainty are statistically signicant. This could be explained by the fact that manufacturing-business' condence increased rapidly after an initial drop following the Brexit referendum (see Born et al. (2019)).

Another classication of rms that might be expected to be more sensitive to Brexit uncertainty is those that are listed (those whose stocks are publicly traded). Therefore we could expect them to be more negatively aected by referendum uncertainty. That might be

ECB Working Paper Series No 2403 / May 2020

26

because they are larger and more involved in international trade. On the other hand, they are also less likely to suer from nancial constraints compared to their unlisted counterparts since they may have fewer problems derived from asymmetric information (Carpenter and B. C. Petersen (2002)). Panel B of Table 7 shows that although all dummy-listed- variables interacted with each uncertainty index are negative, they are not signicantly dierent from zero.

5.2 Financing constraints

To further investigate to what extent the nancing constraints channel is responsible for any heterogeneous outcome of uncertainty on investment, we construct several proxy variables to account for nancing constraints. Recall that the nancing constraints channel states that an increase in uncertainty exacerbates any underlying asymmetric information problem. This, in turn, reduces credit access as it becomes more dicult for lenders to assess the probability of repayment (Gilchrist, Sim, and Zakrajsek (2013); Arellano, Bai, and Kehoe (2010); and Byrne, Spaliara, and Tsoukas (2016)). One would, therefore, expect that companies facing greater diculties in accessing credit might cut investment more sharply as uncertainty rises, compared to those with easier access to credit. As Doshi, Ku- mar, and Yerramilli (2017) suggest, the adverse eect of uncertainty on investment will be more powerful for nancially constrained rms as they reduce capacity in a bid to minimize possible ex-post costs of nancial distress.

Following the recent literature, we distinguish between internal and external nancial constraints. On the one hand, internal nancial constraints operate through restrictions to internal funds generated by the rm that could otherwise, in principle, be targeted towards investment. Thus, rms with lower levels of available internally generated funds (e.g., funds directed to debt service) will be more constrained. On the other hand, external nancial constraints operate through various forms of information asymmetries.

Following the approach of Guariglia (2008), we dene an external nancing constraints dummy variable based on size and age. The intuition is that younger and smaller rms are more likely to face problems of asymmetric information given their short track records

ECB Working Paper Series No 2403 / May 2020

27

and lower collateral levels (Schiantarelli (1995)).23 To this end, we rst de ne company i as Y oungi;t = 1, if its age falls within the lowest quartile of the distribution of the ages of all rms operating in her sector and zero otherwise. Similarly, we de ne company i as Smalli;t = 1, if its total assets fall within the lowest quartile of the distribution of total assets of all rms operating in her sector and zero otherwise. The external nancing constraints dummy variable is then represented by those young and small companies Y Si;t.24

We de ne an internal nancial constraints dummy variable based on the level of cash- ows and the coverage ratio. This latter variable is the ratio between rm's total pro ts before tax and before interest and their total interest payments. It is a measure of the number of times a company could make its interest payments relying on her earnings before interest payments and taxes (Guariglia (2008)). Cash-ows, on the other hand, is the total amount of money being transferred into and out of a business, primarily a ecting short- term liquidity. The intuition for using cash-ows to capture internal nancing constraints hinges on empirical evidence. Provided that cash-ows are the main source of variation in internal funds, rms with low cash-ows levels likely have low levels of internal funds (Cleary, Povel, and Raith (2007)). Therefore, those rms with low levels of cash-ows will nd it harder to raise internal funds to nance investment. Nonetheless, a company might have high levels of cash-ows by selling-o its long-term assets or assuming high debt levels (bringing interest payments up). Thus, we de ne an internally nancially constrained rm as one with low levels of cash-ows and a low coverage ratio: lowCF &CRi;t. Just as before, we create a dummy variable for companies with low levels of cash-ows and coverage ratio.25

Results regarding internal nancing constraints (Table 8) show that only Brexit uncertainty foreshadows a higher negative drop in investment among those companies with higher levels of nancing constraints. The link is particularly strong for Young and Small

23A recent empirical study by Hadlock and Pierce (2010) nds that size and age are the best predictors

of nancing constraints.

24The reason we combine these two variables is that size and age may cancel each other. For example,

large but young companies might not face nancing constraints due to a larger pool of assets available as

collateral while small but old companies may have a long track record of activity to inform credit institutions. 25Company i is lowCFi;t = 1, if its cash-ows level falls within the lowest quartile of the distribution

operating in their sector, while company i is lowCRi;t = 1, if its coverage ratio falls within the lowest

quartile of the distribution of the coverage ratio of all rms operating in her sector.

ECB Working Paper Series No 2403 / May 2020

28

rms (external nancially constrained) exposed to Brexit uncertainty (Panel A).

5.3 Irreversibility of investment

The real-option theory predicts that a rise in uncertainty will have a stronger negative impact on investment for those rms facing a higher degree of irreversibility of investment (Bernanke (1983); McDonald and Siegel (1986), A. Dixit (1989); and Bloom (2000)). When investment is irreversible (capital can only be resold at a lower price than its original purchase price), rms will only invest when demand for their products rise above some upper threshold level. Under uncertainty, this threshold level rises, causing a delay in investment. To proxy irreversibility of investment, we follow Chirinko and Schaller (2009) and use the depreciation to capital ratio. The use of this ratio to proxy irreversibility of investment is motivated by the fact that, in addition to selling capital, rms can reduce their capital stock through depreciation. As noted by Chirinko and Schaller (2009), in companies with low depreciation rates, this recourse is sharply limited.

To be consistent with the approach used to characterise nancing constraints, we dene now an irreversibility dummy variable IRRi;t = 1 for every company whose depreciation to capital ratio falls within the lowest quartile of the distribution of all rms operating in her sector. As predicted by the theory, those rms with a higher degree of investment irreversibility experience higher investment drops when facing political uncertainty compared to those rms with lower degrees of investment irreversibility (Panel A of Table 9). The interactive term between the dummy variable for investment irreversibility and political uncertainty is particularly high for Brexit uncertainty compared to IndyRef (-0.042 and 0-0.028 respectively), being both of them statistically signicant.

5.4 Isolating the Scottish Referendum for Independence eect

In this section, we study the possible outcome that the Scottish referendum for independence (Sept. 2014) might have had on investment by eliminating the last two years in our sample. In other words, we want to isolate the period of the running up to the Scottish referendum of independence until its instance. It should be recalled that Brexit, on the one

ECB Working Paper Series No 2403 / May 2020

29

hand, has induced policy changes at the Scottish level while, on the other hand, has fuelled the debate for a second Scottish referendum for independence. Just as in the previous sub- sections, we interact several heterogeneous variables with the IndyRef index.

In addition to the heterogeneous variables displayed above, we consider whether or not those Scottish companies operating on the border counties with England are aected dier- ently by this particular referendum uncertainty compared to those established in the rest of Scotland. We believe that those Scottish companies nearer to the border with England have closer relationships with the English economy compared with those further away, and hence may be especially exposed to the political uncertainty derived from the Scottish Referendum for independence. To this end, we classify company i as being on the border if it is registered or its primary trading address falls in either of the three bordering counties with England: Berwickshire, Roxburgh, or Dumfries and Galloway. As expected results show a much stronger and signicant negative coecient across companies operating on the border than for the rest (columns (4) and (7) of Table 10).

Next, we consider whether or not listed companies have reacted more adversely to the Scottish referendum for independence alone. Recall that in subsection 5.1 we found weak evidence (statistically non-signicant) of higher negative links between political/policy uncertainty and investment in the case of listed companies. Nonetheless, previous studies have already documented a signicant impact of the Scottish independence referendum on Scot- tish listed companies. This is the case of Darby and Roy (2019), which observed increases in the relative volatility of Scottish companies' stock returns compared to the rest of the UK when polls suggested the referendum result was too close to call. As can be seen in the second and fth columns of Table 10 (excluding/including time-xed eects respectively), once we consider the uncertainty of the Scottish referendum of independence alone we nd signicant evidence that listed companies have cut to a greater extent on xed tangible investment as a result of IndyRef than their counterparts.26

26Given that these specications contain rm and time xed eects, little can be said regarding the average eect of uncertainty on investment. Nonetheless, once both are removed, we observe that the investment rate of listed companies is not signicantly dierent from that of the rest of companies. Taking into account that the average investment rate of listed companies during 2009-15 was 0.34, we could approximate the average eect to be around 29%.

ECB Working Paper Series No 2403 / May 2020

30

Besides, and in line with previous results, we nd that investment from manufacturing companies relate less negative adversely to IndyRef than their counterparts once the after Scottish referendum uncertainty is not taken into account (columns (3) and (6) of Table 10). Also in line with the ndings of the previous section, our results display a more detrimental link between the Scottish referendum of independence and investment in the case of companies with higher levels of nancing constraints (internal and external) and irrereversibility of investment than their counterparts (although only this latter is statistically signicant, see Panel B of Table 10).

• Robustness

6.1 Uncertainty Indices

We consider the implications, if any, of solving the Latent Dirichlet Allocation algorithm (LDA) with a dierent number of topics. Recall that the log-likelihood approach suggested 20 as the optimal number of topics. However, this measure might lead to over-tting since we are computing the within sample likelihood. In addition, empirical ndings suggest that in some cases, models which perform better on likelihood may infer less semantically meaningful topics (Chang et al. (2009)). Therefore, we want to examine whether it is possible to identify the two referenda topics plus the policy uncertainty in Scotland when using alternative number of topics closer to 20: i.e. K = 15, 25, and 30.

Figure 5 shows the word-clouds of political related topics for dierent values of K. Their sizes represent the probability of the word occurring in the topic, that is, the larger a word is, the most representative it is for a given topic. The rst thing we notice when moving further away from the optimal number of topics indicated by the log-likelihood approach (K = 15 and K = 30) is that there is no longer a separation between Brexit-related uncertainty and that related to the Scottish referendum for independence. For example, when K = 15 we nd a single topic containing words such as independend, scotland, referendum, eu, and brexit.27 Similarly, when K = 30 there is no detachment between the two referendum top-

27Even though this topic could be labelled as overall referendum uncertainty, it would be detrimental for our purpose since we want to isolate the uncertainty produced by each referendum.

ECB Working Paper Series No 2403 / May 2020

31

ics: words such as referendum, scotland, independence, eu, brexit or membership assemble a unique topic. For this reason, selecting K = 15 or K = 30 renders no validity in our analysis.

However, when we set K = 25 the two referendum-related uncertainty topics emerge again as two separate topics: one topic clearly characterizes Brexit uncertainty: brexit, european, uk, negotiation, membership, leav and vote while a dierent topic characterizes the IndyRef uncertainty: scotland, independ, referendum, snp. Worth is noting that when we compare the three uncertainty indices (IndyRef, Brexit and Scottish policy uncertainty) produced when K = 20 and K = 25, we observe a degree of high correlation among them:

0.97 between the two IndyRef indices; 0.95 between the two Brexit indices; and 0.69 between the two Scot:EP U indices (see Figure 6). For this reason we believe that even though having 25 topics is also reasonable, results connecting uncertainty and investment will remain almost unaltered.

6.2 Econometric Framework

In addition, we test the robustness of the baseline results to several alternative methodological specications and additional control variables. First, we incorporate additional controls at the rm and macro level to further address concerns of endogeneity. Second, we apply dierent econometric models to ensure that results are not driven by the particular modelling choices.

The robustness tests are introduced solely to the results that appear in Table 5 because they are likely the most vulnerable ones to endogeneity issues given that we do not include time xed eects. Recall that including time xed eects has the added benet of controlling for any macroeconomic, cross-sectionally invariant forces which may confound the eect of political uncertainty not previously accounted for.

Concerning additional variables, we rst consider the incorporation of Economic Policy Uncertainty for the whole UK. After all, any policy implications that shape the Scottish landscape carry consequences to the whole UK. We borrow from Baker, Bloom, and Davis

ECB Working Paper Series No 2403 / May 2020

32

(2016) their economic policy uncertainty index28 and include it into the baseline regression (equation (2)). Results in Panel A of Table 11 show that our three uncertainty indices remain negative and signicant. Moreover, when the UK's EPU is considered alone (column (6)), its coecient indicates that a one standard deviation rise in UK's EPU foreshadows a drop in average investment by 6 to 7 per cent in the following year (including or excluding GDP growth rates respectively).

Next, we consider controlling for rm's size trough the number of employees. Duchin, Ozbas, and Sensoy (2010) have stressed the importance of rm's size for accessing external nance during recession times. Also, we have presented evidence of nancing constraints aecting investment in the previous section. Therefore, we wish to test whether or not the average eect of political uncertainty presented in Section 4 substantially changes once we control for the rm's size. To this end, we add the natural logarithm of the total number of employees. Results in Panel B of Table 11 show that the uncertainty coecients are barely unchanged and remain statistically signicant. When the coecients enter alone (columns

1. to (3) in Table 11) a one standard deviation increase in IndyRef and Brexit uncertainty foreshadow a drop in average investment of 6% and 11% respectively whereas the baseline

ndings displayed an 8% and 13.5% reduction in investment respectively.29

Regarding the econometric model, we rst consider removing rm xed eects from the baseline regression (equation (2)). The within-group transformation (xed eects transformation) could induce correlation between the lagged political uncertainty variables and the current error term, something that would render strict exogeneity invalid (i.e., E(P Ut1 i;t) 6= 0, see Gulen and Ion (2015)). Therefore, estimating our baseline specication without the within-group transformation would not suer from this problem and would, therefore, yield a consistent estimate for our political/policy uncertainty variables. We do this in Panel A of Table 12 and we nd that the estimate for our political uncertainty indices remains almost unchanged. Therefore, we can conclude that controlling for rm xed eects

1. http://www.policyuncertainty.com/.
2. In addition, we have also tried controlling for government expenditure (Scottish government expenditure

per capita) and rm-specic uncertainty (the cross-industry standard deviation of the growth rate in prots). Because results remain unchanged when we incorporate these two additional controls and for clarity purposes, the results are not presented. Nonetheless, results are available upon request.

ECB Working Paper Series No 2403 / May 2020

33

Ii;t3

Ii;t2

does not signicantly alter the coecient estimate of our political uncertainty indices and that a possible violation of strict exogeneity will have a negligible impact on our main results.

We then consider the model in rst dierences rather than using rm xed eects to remove any rm time-invariant omitted variable bias. This approach addresses the concerns that results may be driven by a spurious correlation due to a common trend in the uncertainty and investment variables. The rst dierences approach deals with these two concerns by taking the rst dierences of every variable (including the error term):

	Ii;t	= 1	P Ut + 2		CFi;t	+ 4Mi; t + i;t	(5)
Ki;t1
				Ki;t1

Nonetheless, one of the downsides of the rst dierences approach is that it removes a whole year of observations in the sample and two observations per rm when there is a gap in the series of observations. Be it as it may, once again our main result remains virtually unchanged (Table 12, Panel B).

Next, we examine the investment regression in a dynamic panel format by incorporating the lag dependent variable as a control (see for example Bloom, Bond, and Van Reenen (2007)):

Ii;t	= i +	Ii;t1	+ 1logP Ut + 2	CFi;t	+ 3SGi;t + 4Mt + i;t	(6)
Ki;t1	Ki;t2
		Ki;t1

Because the within-group and rst-dierence transformation needed to eliminate the rm xed eects mechanically correlates the lagged investment variable with the error term, we estimate this specication using the system GMM methodology (Blundell and

Bond (1998)). Following the approach of Gulen and Ion (2015), we use Ki;t3 and Ki;t4

as instruments for	Ii;t1	in the dierence equation, and	Ii;t2	as an instrument for
Ki;t2		Ki;t3

Ii;t1	in the level equation. This set up rejects AR(1) errors while not AR(2) errors. As
Ki;t2

can be seen, the coecients for the uncertainty indices remain negative but only Brexit uncertainty retains statistical signicance (Table 12, Panel C). Nonetheless, the inclusion of these instruments reduces the sample data considerably by removing four years of ob- servations, and this could induce sample bias. This fact, together with the impossibility to cluster standard errors by rms or year makes the static specication our preferred approach.

ECB Working Paper Series No 2403 / May 2020

34

• Conclusion

In this study, we analyse the eect of three distinctive uncertainty narratives embedded in the Scottish press, namely IndyRef, Brexit uncertainty, and Scottish policy uncertainty on private investment of Scottish rms. To frame these distinctive sources of uncertainty, we use an unsupervised machine learning algorithm able to classify news articles with a range of themes without prior knowledge regarding their content. On analysing these narratives trough time, we observe that they co-move strongly with the Google search queries Scot- tish Independence" and Brexit". For example, IndyRef and the Google query Scottish Independence" display prominent spikes when the chancellor of the Exchequer George Os- borne argued that a 'Yes' vote meant Scotland giving up the pound (Feb 2014) and during the Scottish referendum for independence (September 2014). Besides, the Google query Brexit" and Brexit uncertainty both ramp-up during the month of the Brexit referendum and maintain high levels in the aftermath.

We then examine the relationship between the indices just described and rm investment by applying a standard investment regression to a longitudinal panel dataset formed of 3,589 Scottish rms. Our baseline results suggest that a one standard-deviation increase in Brexit uncertainty foreshadows a reduction in investment by 8% on average in the following year. Besides we nd that the uncertainty associated with the Scottish referendum for independence while negligible for the overall rm network, had a negative and significant outcome on the investment of listed and border companies (those operating on the border with England). These results are robust to controlling for alternative measures of investment opportunities and macroeconomic uncertainty as well as to several identifying econometric frameworks.

Nonetheless, given the relative low amount of years in our sample, a certain caution is warranted regarding these results. Given the relatively low number of years in our sam- ple, we could only place a limited amount of aggregate variables aiming to capture time- dependent factors that could confound the eect of political uncertainty. To somewhat reduce these concerns, in further analysis we incorporate time-xed eects in our regression by having our uncertainty measures interact with rm characteristics known to be more

ECB Working Paper Series No 2403 / May 2020

35

sensitive to uncertainty. This should reassure us that any negative eect observed in investment indeed comes from political uncertainty.

To this end, we examine the hypothesis of whether manufacturing, unlisted, more nan- cially constrained and those with higher irreversible investment rates companies cut down on investment more severely than the rest of companies as a result of an increase in uncer- tainty. In line with the literature, we nd evidence that those rms that are more likely to be nancially constrained display higher drops in investment in the presence of uncertainty. This holds principally for rms with either internal or external nancing constraints and Brexit uncertainty. Also consistent with priors, we nd a stronger negative relationship between rms whose investment is more likely irreversible and political uncertainty.

In addition, we distinguish between non-manufacturing and manufacturing rms. The Decision Maker Panel survey reported that rms in the manufacturing sector are the most likely to move part of their operations outside the UK due to the uncertainty produced by Brexit (Bloom, Bunn, et al. (2017)). Nonetheless, more recent evidence suggests that business condence from the manufacturing sector has actually increased after Brexit (see Born et al. (2019)). We nd evidence supporting this latter behaviour: investment from Scottish manufacturing companies appear less sensitive to political uncertainty.

The resulting policy implications may be important, in particular to the current economic climate. Referenda are becoming a popular tool for politicians, yet their consequences as a source of uncertainty often escape the political debate. In this paper, we show not only that referenda are a signicant source of political and policy uncertainty but also that they aect private investment in a negative way independently of their outcome.

ECB Working Paper Series No 2403 / May 2020

36

References

Arellano, Cristina, Yan Bai, and Patrick Kehoe (2010). Financial markets and uctuations in uncertainty". Federal Reserve Bank of Minneapolis Working Paper.

Azqueta-Gavaldon, Andres (2017). Developing news-based economic policy uncertainty index with unsupervised machine learning". Economics Letters 158, pp. 47{50.

Azqueta-Gavaldon,Andres, Dominik Hirschbuhl, Luca Onorante, and Lorena Saiz (2020). Economic policy uncertainty in the euro area: an unsupervised machine learning ap- proach". Working Paper Series 2359, European Central Bank.

Azzimonti, Marina (2018). Partisan conict and private investment". Journal of Monetary Economics 93, pp. 114{131.

Baker, Scott R, Nicholas Bloom, and Steven J Davis (2016). Measuring economic policy uncertainty". The quarterly journal of economics 131.4, pp. 1593{1636.

Bernanke, Ben (1983). Irreversibility, uncertainty, and cyclical investment". The Quarterly Journal of Economics 98.1, pp. 85{106.

Blei, David M, Andrew Y Ng, and Michael I Jordan (2003). Latent dirichlet allocation". Journal of machine Learning research 3.Jan, pp. 993{1022.

Bloom, Nicholas (2000). The real options eect of uncertainty on investment and labor demand". IFS working paper 00/15.

Bloom, Nicholas, Stephen Bond, and John Van Reenen (2007). Uncertainty and investment dynamics". The review of economic studies 74.2, pp. 391{415.

Bloom, Nicholas, Philip Bunn, Paul Mizen, Pawel Smietanka, Gregory Thwaites, and Garry Young (2017). Tracking the views of British businesses: evidence form the Decision Maker Panel". Bank of England Quarterly Bulletin 2017 Q2.

Blundell, Richard and Stephen Bond (1998). Initial conditions and moment restrictions in dynamic panel data models". Journal of econometrics 87.1, pp. 115{143.

Blundell, Richard, Stephen Bond, Michael Devereux, and Fabio Schiantarelli (1992). In- vestment and Tobin's Q: Evidence from company panel data". Journal of Econometrics 51.1-2, pp. 233{257.

Born, Benjamin, Gernot J Muller, Moritz Schularick, and Petr Sedlacek (2019). The costs of economic nationalism: evidence from the Brexit experiment". The Economic Journal 129.623, pp. 2722{2744.

ECB Working Paper Series No 2403 / May 2020

37

Byrne, Joseph P, Marina-Eliza Spaliara, and Serafeim Tsoukas (2016). Firm survival, un- certainty, and nancial frictions: is there a nancial uncertainty accelerator?" Economic Inquiry 54.1, pp. 375{390.

Carpenter, Robert E and Bruce C Petersen (2002). Is the growth of small rms constrained by internal nance?" Review of Economics and statistics 84.2, pp. 298{309.

Castelnuovo, Efrem and Trung Duc Tran (2017). Google it Up! A Google trends-based uncertainty index for the United States and Australia". Economics Letters 161, pp. 149{ 153.

Chang, Jonathan, Sean Gerrish, Chong Wang, Jordan L Boyd-Graber, and David M Blei (2009). Reading tea leaves: How humans interpret topic models". Advances in neural information processing systems, pp. 288{296.

Chirinko, Robert S and Huntley Schaller (2009). The irreversibility premium". Journal of Monetary Economics 56.3, pp. 390{408.

Cleary, Sean, Paul Povel, and Michael Raith (2007). The U-shaped investment curve: Theory and evidence". Journal of nancial and quantitative analysis 42.1, pp. 1{39.

Darby, Julia and Graeme Roy (2019). Political uncertainty and stock market volatility: new evidence from the 2014 Scottish Independence Referendum". Scottish Journal of Political Economy 66.2, pp. 314{330.

Dibiasi, Andreas, Klaus Abberger, Michael Siegenthaler, and Jan-Egbert Sturm (2018). The eects of policy uncertainty on investment: Evidence from the unexpected acceptance of a far-reaching referendum in Switzerland". European Economic Review 104, pp. 38{67.

Dixit, Avinash (1989). Entry and exit decisions under uncertainty". Journal of political Economy 97.3, pp. 620{638.

Dixit, Robert K and Robert S Pindyck (1994). Investment under uncertainty. Princeton university press.

Doshi, Hitesh, Praveen Kumar, and Vijay Yerramilli (2017). Uncertainty, capital invest- ment, and risk management". Management Science 64.12, pp. 5769{5786.

Duchin, Ran, Oguzhan Ozbas, and Berk A Sensoy (2010). Costly external nance, corporate investment, and the subprime mortgage credit crisis". Journal of nancial economics 97.3, pp. 418{435.

ECB Working Paper Series No 2403 / May 2020

38

Gilchrist, Simon, Jae Sim, and Egon Zakrajsek (2013). Uncertainty, nancial frictions, and irreversible investment". Boston University and Federal Reserve Board working paper.

Griths, Thomas L and Mark Steyvers (2004). Finding scientic topics". Proceedings of the National academy of Sciences 101.suppl 1, pp. 5228{5235.

Guariglia, Alessandra (2008). Internal nancial constraints, external nancial constraints, and investment choice: Evidence from a panel of UK rms". Journal of Banking & Finance 32.9, pp. 1795{1809.

Gulen, Huseyin and Mihai Ion (2015). Policy uncertainty and corporate investment". The Review of Financial Studies 29.3, pp. 523{564.

Hadlock, Charles J and Joshua R Pierce (2010). New evidence on measuring nancial con- straints: Moving beyond the KZ index". The Review of Financial Studies 23.5, pp. 1909{ 1940.

Hansen, Stephen, Michael McMahon, and Andrea Prat (2017). Transparency and deliberation within the FOMC: a computational linguistics approach". The Quarterly Journal of Economics 133.2, pp. 801{870.

Hassan, Tarek A, Stephan Hollander, Laurence van Lent, and Ahmed Tahoun (2019). Firm- level political risk: Measurement and eects". The Quarterly Journal of Economics 134.4, pp. 2135{2202.

Hayashi, Fumio (2000). Econometrics. 2000". Princeton University Press. Section 1, pp. 60{ 69.

Homan, Matthew, Francis R Bach, and David M Blei (2010). Online learning for latent dirichlet allocation". advances in neural information processing systems, pp. 856{864.

Jens, Candace E (2017). Political uncertainty and investment: Causal evidence from US gubernatorial elections". Journal of Financial Economics 124.3, pp. 563{579.

Julio, Brandon and Youngsuk Yook (2012). Political uncertainty and corporate investment cycles". The Journal of Finance 67.1, pp. 45{83.

McDonald, Robert and Daniel Siegel (1986). The value of waiting to invest". The quarterly journal of economics 101.4, pp. 707{727.

Petersen, Mitchell A (2009). Estimating standard errors in nance panel data sets: Comparing approaches". The Review of Financial Studies 22.1, pp. 435{480.

ECB Working Paper Series No 2403 / May 2020

39

Rehurek, Radim and Petr Sojka (2010). Software framework for topic modelling with large corpora". In Proceedings of the LREC 2010 Workshop on New Challenges for NLP Frameworks.

Saltzman, Bennett and Julieta Yung (2018). A machine learning approach to identifying dierent types of uncertainty". Economics Letters 171, pp. 58{62.

Schiantarelli, Fabio (1995). Financial constraints and investment: a critical review of methodological issues and international evidence". Conference Series-Federal Reserve Bank of Boston. Vol. 39. Federal Reserve Bank of Boston, pp. 177{214.

Sievert, Carson and Kenneth Shirley (2014). LDAvis: A method for visualizing and interpreting topics". Proceedings of the workshop on interactive language learning, visualiza- tion, and interfaces, pp. 63{70.

Whaley, Robert E (2000). The investor fear gauge". The Journal of Portfolio Management 26.3, pp. 12{17.

ECB Working Paper Series No 2403 / May 2020

40

Figure 1: Scottish and Brexit Referenda Polls

Notes: Scottish Referendum polls information is obtained from YouGov, Survation, Panelbase, Ipsos, BMG and TNS. Brexit Referendum polls information is obtained from the Financial Times (see https://ig.ft.com/sites/brexit-polling/)

ECB Working Paper Series No 2403 / May 2020

41

Figure 2: Global view of the LDA topics

2020 May / 2403 No Series Paper Working ECB

42

Notes: This Figure shows how large and semantically erentclose/di economic uncertainty topics produced by the LDA are. The gure was produced using the library LDAvis developed by Sievert and Shirley (2014). The three topics of interest are in bold (IndyRef, Brexit uncertainty and Scottish policy uncertainty). To see the 30 main words of each topic please see Table 2.

Figure 3: Evolution of Uncertainty indices in Scotland (continuous line, left legend) and

the implied volatility index, VFTSE (dotted line, right legend)

Scottish Political Uncertainty

Financial Crisis

Scottish referendum

for independence

60

0.05

●●

UK approval

"Yes vote

50

0.02

● ●●●

●●●●●●● ●

●●●

●

●●

means leaving

●

●●

●

VFTSE

0.04

●●

of the referendum

the pound"

Brexit

40

0.03

●●

●

●●

●●●

●

●

●

●●

30

0.01

●

● ●●

●●●

●●● ●●●

●● ●

●

●

●

● ●

●

● ● ●

● ●

●

●

20

●

●

●

●●

●

●

●

●●

●

●

0.00

●●

●●

●

●

●●●●●●●

10

●●●●●●

●●●●

●●●●●● ●

● ●

●●

2008

2010

2012

2014

2016

Time

Brexit Uncertainty

0.025

Financial Crsis

Brexit

60

●●

European

0.020

50

VFTSE

●●●●

debt crisis

0.015

General

40

●●●

●

0.010

●●●

●

●

●

●●

●

●

election

30

0.005

●●●●●

●●●●●●●●

●

●●●●●●

●

●●

●●● ●●●

●

●●

●

●

●

●●

20

0.000

●

●●● ●●

●

●●●●●

●●●

● ●●● ● ●●

●●● ●●● ●

10

●●●●●●

●●●● ●●●●●● ●

● ●

●●

●●●●●●●

2008

2010

2012

2014

2016

Time

Scottish Policy Uncertainty

0.020

Financial Crisis

Brexit

60

●●

SNP Budget

Scottish public

0.015

approval

50

VFTSE

strikes

●●●●

40

0.010

●

●●●

●●●

●

●

●

●●

●

●

30

0.005

●●●●●

●●●●●●●●

●

●●●●●●

●

●●

●●● ●●●

●

●●

●

●

●

20

●

●●● ●●

●

●●●●●

●●●

● ●●● ● ●●

●●● ●●● ● ●●

0.000

●●●●●●

●●●● ●●●●●● ●

● ●

●●

●●●●●●●

10

2008

2010

2012

2014

2016

Time

Notes: IndyRef, Brexit Uncertainty and Scottish Policy Uncertainty indices are built by computing the monthly ratio between news articles describing these uncertainty topics and the total number of news articles. The newspapers used are The Aberdeen Press & Journal, The Glasgow Herald and The Scotsman. Time period from Jan 2008 to June 2017. The implied volatility index, VFTSE, in levels is extracted from Bloomberg.

ECB Working Paper Series No 2403 / May 2020

43

Figure 4: Evolution of Uncertainty indices in Scotland (continuous line, left legend) and

the Google searches of Scottish Independence and Brexit (right legend)

Scottish Political Uncertainty (IndyRef)

		IndyRef			"Yes vote	Referendum	4
	0.04	'Scottish Independence' Google search	UK approval	means leaving
	the pound"
IndyRef			of the Referendum	I	*	Brexit
0.02			I		2 3
							1
	0.00
		2008	2010	2012	2014		2016

				Brexit Uncertainty		Brexit
	0.020	Brexit News					4
	'Brexit' Google search
Brexit						*	3
0.010					2
							1
	0.000						0
		2008	2010	2012	2014	2016

	0.020		Scottish Policy Uncertainty (Scot. EPU)	Brexit
	SNP budget
		approval
			Scottish public			*
ScotEPU	0.010		strikes		Scottish Referendum
			I
			*
	0.000
	2008	2010	2012	2014	2016
			Time

Notes: IndyRef, Brexit Uncertainty and Scottish Policy Uncertainty indices are built by computing the monthly ratio between news articles describing these topics and the total number of news articles. The newspapers used are The Aberdeen Press & Journal, The Glasgow Herald and The Scotsman. Google searches of the terms Scottish Independence" and Brexit" only looked in the region of Scotland and their series are presented in natural logs. * indicates when the referendums took place.

ECB Working Paper Series No 2403 / May 2020

44

Figure 5: Word clouds of political topics for di erent values of K. For each word cloud the size of a word reects the probability of this word occurring in the topic

(a) IndyRef K = 20 (b) IndyRef K = 25

(c) Brexit K = 20

(d) Brexit K = 25

(e) Political K = 15 (f) Political K = 30

Figure 6: Evolution of the uncertainty measures computed using 20 and 25 topics

Scottish Political Uncertainty 'IndyRef' (0.97 Correlation)

	105	IndyRef_20				105
IndyRef 20	101 103	IndyRef_25				101 103
	99					99
		2008	2010	2012	2014	2016

Brexit uncertainty (0.95 Correlation)

	106	Brexit_20				106
20	104	Brexit_25				104
Brexit	102					102
	100					100
		2008	2010	2012	2014	2016

Scottish Policy Uncertainty 'Scot.EPU' (0.69 Correlation)

	103	ScotEPU_20
20		ScotEPU_25				102
ScotEPU	101					100
	99					99
		2008	2010	2012	2014	2016

Notes: All series are standardize to mean 100 and 1 standard deviation.

ECB Working Paper Series No 2403 / May 2020

45

Table 1: Number of topics and log-likelihood scores

10	20	30	40	50	60
log P(w j K) -24502056	-24465226	-24477848	-24485771	-24581108	-24609611

De nitions of the variables used:

Investment: It is constructed as the dierence between the book value of tangible xed assets (which include land and building; xtures and ttings; plant and vehicles; and other xed assets) of end of year t and end of year t-1 while adding depreciation of year t.

Capital stock: tangible xed assets.

Cash-ows: It is dened as the sum of after tax prot and depreciation.

Coverage ratio: It is dened as the ratio between the rm's total prots before tax and before interest (also referred as Operating Prot or EBIT) and its total interest payments.

Total assets: It is dened as the sum of xed assets and current assets.

ECB Working Paper Series No 2403 / May 2020

46

		Table 2: Topics unveiled by the LDA
Label	%	Top words
Scot. Political	9.9	independ, snp, mr, referendum, parti, vote, labour, minist, scotland, elect, campaign, would, sturgeon,
		tori, ye, salmond, polit, scottish, voter, poll, westminister, govern, conserv, leader, parliament, cameron
FTSE	9.8	cent, per, share, 5p, 1, ftse, stock, index, 2, 3, fell, 4, 2017, 5, 6, rose, close, analyst, 100, 7, 8, gbp, 9, 0
		market, gain, group, biggest, trade, us
Preferences	9.6	say, peopl, thing, one, get, work, think, time, go, feel, like, way, know, realli, someth, lot, make, seem,
		much, look, art, mani, want, want, always, idea, old, good, even, di er, women
Monetary Policy	9.3	rate, monetari, economi, bank, interest, mpc, inat, market, polici, cut, recess, econom, us, central,
		governor, euro, commite, risk, global, england, crisi, dollar, recoveri, would, king, fed, low, carney
Economy	9.2	cent, per, growth, month, survey, quarter, uk, rise, gur, year, manufactur, sector, show, 0, increas, retail,
		consum, 2, forecast, said, economi, 1, output, rate, economist, report, sale, latest, spend, fall
Scottish Policy	9	scotland, scottish, govern, budget, busi, univers, public, educ, need, fund, council, report, tax, local,
		commun, support, work, enterpris, plan, organis, sevic, challeng, sector, develop, research, student, econom
Business	7.2	compani, busi, pro t, year, rm, group, sale, oper, acquisit, 2016, brand, turnov, execut, million, said,
		market, pre, revenu, whiski, custom, scotch, half, chief, trade, manag, deal, continu, murgitroyd, base
Oil	4.8	oil, ga, invest, sea, north, asset, investor, barrel, price, equiti, fund, trust, bp, eld, compani, industri, shell,
		explor, aberdeen, portfolio, product, bond, manag, yield, drill, opec, crude, wood, return, petroleum
Jobs	4.7	job, said, moray, sta , sh, closur, raf, mr, worker, highland, trourism, employ, redund, plant, visitor, base,
		workforc, industri, 000, app, announc, futur, visitscotland, paterhead, sheri, island, defenc, factori, buchan
Banks	4.4	bank, rb, nanci, lloyd, mortgag, load, lend, lender, debt, credit, hbo, insur, clydesdal, tsb, custom, hsbc,
		barclay, taxpay, repay, billion, borrow, sharehold, royal, save, money, fund, gdp, deposit, branch, pay
America	3.6	obama, trump, centuri, world, american, human, bush, church, america, clinton, man, histori, donald, death,
		burn, republican, presid, barack, sdg, white, father, detent, polit, woman, supper, live, africa, nation, god
Brexit	3.5	eu, brexit, european, britain, europ, union, uk, negoti, leav, countri, membership, singl, trade, brussel,
		immigr, agreement, vote, greec, member, deal, want, referendum, free, hammond, exit, relationship
Farmers	3.3	pension, farm, farmer, agricultur, incom, scheme, ubi, payment, rural, pay, retir, nfu, crop, annuiti, milk,
		cap, beef, legisl, employe, dairi, sheep, food, fee, 2019, meat, bene t, tonn, wheat, employ, lamb
Transport	2.9	citi, airport, aberdeen glasgow, transport, passeng, rail, council, airlin, road, project, centr, rout, councillor,
		trac, bu, ferri, site, local, inver, plan, skinner, baa, heathrow, develop, travel, edinburgh, east, rstgroup
Geopolitical	2.3	war, militari, iraq, armi, presid, polic, russian, russia, hester, attack, hamon, ministri, un, prision, iran,
		weapon, islam, afghanistan, troop, protest, marshal, holland, socialist, ukrain, egypt, bomb, sanction, arab
Other Topics
Sports	2.1	club, footbal, ranger, game, leagu, cup, sport, celtic, player, hotel, season, murray, team, golf, spl, fan
Real Estate	2	properti, hous, home, buyer, estat, rent, market, tenant, oc, housbuilding, land, build, edinburgh
Energy	1.5	energi, wind, electr, carbon, edf, o shor, emiss, nuclear, turbin, coal, power, googl, onshor, rivaz, water
Unknown	0.8	scotsman, com, http, www, facebook, click, scotsmanbusi, read, mail, link, page, parcel, lossiemouth, kinloss
Cars	0.2	car, motor, ford, cc, q, bmw, walsh, diesel, gsk, poundland, glaxo, atlanti, mudoch, handbag, uber, barnard

Notes: This table displays the most representative words per topic unveiled by the Latent Dirichlet Allocation algorithm (3rd column), the proportion of the given topic with respect to all topics (2nd column), and the label given to each topic (1st column)

ECB Working Paper Series No 2403 / May 2020

47

2020 May / 2403 No Series Paper Working ECB

Table 3: Descriptive statistics rm level data

	FAME universe		Sample used		Manufacturing		Listed		YS		lowCF&CR		IRR		Border
Ii;t=Ki;t1	0.36	0.34	0.27	0.32	0.46	0.25	0.20	0.24
	(0.98)	(0.85)	(0.65)	(0.68)	(1.07)	(0.69)	(0.67)	(0.46)
CFi;t=Ki;t1	2.52	2.36	1.17	1.86	3.06	-0.6	0.37	0.87
	(10.41)	(9.26)	(5.39)	(8.93)	(10.98)	(2.36)	(2.17)	(3.23)
SGi;t	0.075	0.07	0.069	0.07	0.12	0.012	0.068	0.08
	(0.301)	(0.27)	(0.267)	(0.27)	(0.36)	(0.29)	(0.26)	(0.24)
n	4,238	3,589	800	43							65
N	24,006	22,769	5,480	337	1,652	2,280	5,525	405

Notes: This table reports sample means and standard deviations (in parenthesis) for the variables of interest and di erent subgroups. The

subscript i indexes rm, and the script t represents time, where t = 2009:::2017. Ii;t=Ki;t1 represents investment rate, where Ii;t is investment in xed assets and Ki;t1 the capital stock at t 1; CFi;t=Ki;t1 indexes cash-ows over the capital stock and SGi;t represents sales growth. FAME universe include Scottish companies operating in all sectors, whereas Sample used omits the regulated and nancial sectors and include only companies with at least three years of observations. Manufacturing and listed companies are those operating in the manufacturing sector and which are traded in a listed stock exchange respectively. YS stands for young and small companies (companies whose age and size falls within the lowest quartile of the distribution of the ages and sizes of all rms operating in their sector). Similarly, lowCF&CR stand for low cash-ows and Coverage ratio (companies whose cash-ows and Coverage Ratio fall within the lowest quartile of the distribution of all rms operating in their sector). IRR stands for high irreversibility of investment while Border stands for those companies operating on the three Scottish counties bordering England.

48

	Table 4: Descriptive statistics uncertainty indices
	IndyRef	Brexit	Scot. EPU	VFTSE	EPU UK	GDP Growth
IndyRef	1
Brexit	0.43	1
Scot. EPU	0.27	0.44	1
VFTSE	-0.34	-0.17	0.11	1
EPU UK	0.35	0.85	0.49	0.06	1
GDP Growth	0.21	-0.01	-0.12	-0.43	-012	1

Correlation matrix between the three measures of uncertainty: IndyRef, Brexit

uncertainty and Scottish policy uncertainty and other macro/uncertainty measures: the implied volatility index (VFTSE), UK's economic policy uncertainty index, Scottish GDP growth rates. All variables are in monthly frequency except GDP growth rates (quarterly frequency) from Jan 2008 until June of 2017. Variables are obtained from Scottish government statistics, Bloomberg, Economic Policy Uncertainty and own calculations.

ECB Working Paper Series No 2403 / May 2020

49

2020 May / 2403 No Series Paper Working ECB

50

Table 5: Baseline regression Results

Panel A

Panel B

(1)

(2)

(3)

(4)

(5)

(6)

(1)

(2)

(3)

(4)

(5)

(6)

IndyRef

0:028

0:001

0:014

0:077

0:045

0:014

t1

(0:011)

(0:007)

(0:009)

(0:013)

(0:014)

(0:009)

Brexitt1

0:046

0:027

0:040

0:045

0:031

0:040

(0:007)

(0:010)

(0:013)

(0:008)

(0:010)

(0:013)

Scot. EPU

0:031

0:029

0:015

0:009

0:034

0:015

0:011

0:009

t1

(0:007)

(0:007)

(0:009)

(0:010)

(0:007)

(0:008)

(0:010)

(0:010)

CF

=K

i;t1

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

it

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

SGit

0:202

0:205

0:205

0:205

0:204

0:206

0:202

0:205

0:207

0:205

0:206

0:206

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

V F T SEt1

0:029

0:018

0:002

0:024

0:017

0:019

(0:010)

(0:007)

(0:006)

(0:010)

(0:010)

(0:009)

Local Elections

0:044

0:020

0:028

0:028

0:025

0:012

0:104

0:021

0:037

0:078

0:003

0:012

(0:016)

(0:012)

(0:012)

(0:014)

(0:012)

(0:015)

(0:019)

(0:013)

(0:013)

(0:020)

(0:012)

(0:015)

GDPt1

3:675

0:075

1:422

3:022

0:770

(0:731)

(0:648)

(0:608)

(0:906)

(0:446)

R2

0.045

0.046

0.045

0.045

0.046

0.046

0.046

0.046

0.046

0.046

0.046

0.046

N

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

Fixed ects E

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

Clustered id

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

Notes: In this table, we regress investment rate Iit=Ki;t1

(Investment in xed assets scaled by the stock of xed assets at the beginning of period)

on the three types of uncertainty at time t 1 (IndyRef, Brexit uncertainty or Scottish policy uncertainty). Additional controls are cash-ows scaled

by the stock of xed assets at the beginning of period (CFi;t=Ki;t1 ), sales growth rate (SGi;t), the Scottish GDP growth rate (GDPt), the

implied volatility index (V F T SE), and local election dummy to control for elections uncertainty. All regressions include rm xed ects, e and

standard errors are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate statistical cance signi at the

10%, 5%, and 1% level, respectively.

2020 May / 2403 No Series Paper Working ECB

51

Table 6: Baseline regression Results and referendum dummies

Panel A

Panel B

Scottish Uncertainty

Brexit Uncertainty

Scottish Uncertainty

Brexit Uncertainty

(1)

(2)

(3)

(4)

(5)

(1)

(2)

(3)

(4)

(5)

SCOTreferendum

0:012

0:027

0:006

0:017

(0:019)

(0:019)

(0:022)

(0:022)

IndyRef

0:033

0:078

t1

(0:012)

(0:014)

IndyRef

*SCOT

2014

0:012

0:052

t1

(0:012)

(0:022)

BREXITreferendum

0:138

0:130

0:114

0:132

0:127

0:113

(0:023)

(0:022)

(0:108)

(0:023)

(0:023)

(0:108)

Brexitt1

0:081

0:081

(0:035)

(0:035)

CF

=K

i;t1

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

0:024

it

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

(0:003)

SGit

0:207

0:203

0:205

0:207

0:203

0:208

0:202

0:203

0:207

0:203

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

(0:031)

V F T SEt1

0:005

0:028

0:020

0:012

0:023

0:004

0:024

0:020

0:009

0:022

(0:007)

(0:010)

(0:011)

(0:006)

(0:008)

(0:009)

(0:010)

(0:011)

(0:009)

(0:011)

Local Elections

0:019

0:044

0:034

0:027

0:014

0:028

0:110

0:072

0:029

0:014

(0:013)

(0:016)

(0:015)

(0:012)

(0:014)

(0:014)

(0:021)

(0:022)

(0:013)

(0:014)

GDPt1

1:124

4:028

2:522

0:248

0:033

(0:723)

(0:866)

(1:158)

(0:641)

(0:650)

R2

0.045

0.046

0.046

0.046

0.046

0.045

0.046

0.046

0.046

0.046

N

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

22,769

Fixed Eects

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

Clustered id

yes

yes

yes

yes

yes

yes

yes

yes

yes

yes

Notes: In this table, we regress investment rate Iit=Ki;t1 (Investment in xed assets scaled by the stock of xed assets at the beginning of period) on the three types of uncertainty at time t 1 (IndyRef, Brexit uncertainty or Scottish policy uncertainty); Scottish referendum time dummy at

t 1, Scottish referendum legislation period at t 1 and Brexit dummy at t 1. In addition, we include a lagged year-dummy variable for the Scottish and Brexit referendums (SCOTreferendum and BREXITreferendum respectively), and a time dummy variable removing the post scottish referendum for independence period SCOT2014 (see Section 4). For information on additional controls see Table 5. All regressions include rm xed

ects,e and standard errors are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate statistical

cancesigni at the 10%, 5%, and 1% level, respectively.

Table 7: The Heterogeneous eect of policy uncertainty on investment

			Dependent variable: Investment rate (Iit=Ki;t1	)
		Panel A: Manufacturing versus non-manufacturing companies
			IndyReft1	Brexitt1	Scot. Policyt	1
			(1)	(2)	(3)
	Uncertainty*Manufacturing	0:028	0:014	0:026
			(0:012)	(0:013)	(0:014)
	R2	0.044	0.044	0.044
		Panel B: Listed versus non-listed companies
	Uncertainty*Listed	0 :068	0:004	0:019
			(0:043)	(0:025)	(0:026)
	R2	0.044	0.044	0.044
	N	22,769	22,769	22,769
	Firm Fixed E ects	yes	yes	yes
	Time Fixed E ects	yes	yes	yes
	Clustered id	yes	yes	yes
Notes: In this table, we regress investment rate Iit=Ki;t1	(Investment in xed assets scaled by

the capital stock at the beginning of period) on the three types of uncertainty (IndyRef, Brexit uncertainty or Scottish policy uncertainty) interacted with dummy variable for manufacturing and listed rms (panel A and B respectively). Additional controls are cash-ows scaled by the capital stock at the beginning of the period (CFi;t=Ki;t1 ) and sales growth rate (SGi;t). All regressions include rm and time xed e ects, and standard errors are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate statistical signi cance at the 10%, 5%, and 1% level, respectively.

ECB Working Paper Series No 2403 / May 2020

52

Table 8: Financial Constraints

Dependent variable: Investment rate (Iit=Ki;t1 )

Panel A: Young and Small rms (externally constrained)

		IndyReft1	Brexitt1		Scot. Policyt 1
		(1)	(2)		(3)
	YS	0:105	0:128		0:109
		(0:062)	(0:061)	(0:063)
	Uncertainty*YS	0:004	0:080		0:011
		(0:028)	(0:028)	(0:028)
	R2	0.043	0.043		0.043
	N	22,290	22,290		22,290
	R2	0.043	0.043		0.043
	Panel B: Low cash-ows and coverage ratio rms (internally constrained)
	lowCF&CR	0:084	0:077	0:080
		(0:022)	(0:021)	(0:021)
	Uncertainty*lowCF&CR	0:0002	0:032		0:021
		(0:018)	(0:019)	(0:018)
	R2	0.046	0.046		0.046
	N	14,774	14,774		14,774
	Firm Fixed E ects	yes	yes		yes
	Time Fixed E ects	yes	yes		yes
	Clustered id	yes	yes		yes
Notes: In this table, we regress investment rate Iit=Ki;t1	(Investment in xed assets scaled by

the capital stock at the beginning of period) on the three types of uncertainty (IndyRef, Brexit uncertainty or Scottish policy uncertainty) interacted with dummy variables for Young and small rms and those with low levels of cash-ows and coverage ratio (panel A and B respectively). Additional controls are cash-ows scaled by the capital stock at the beginning of the period (CFi;t=Ki;t1 ) and sales growth rate (SGi;t). All regressions include rm and time xed e ects, and standard errors are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate statistical signi cance at the 10%, 5%, and 1% level, respectively.

ECB Working Paper Series No 2403 / May 2020

53

Table 9: Irreversibility of investment

Dependent variable: Investment rate (Iit=Ki;t1 )

	IndyReft1	Brexitt1	Scot. Policyt1
	(1)	(2)	(3)
IRR	0:490	0:484	0:490
	(0:048)	(0:047)	(0:048)
Uncertainty*IRR	0:028	0:042	0:008
	(0:014)	(0:015)	(0:016)
R2	0.078	0.078	0.077
N	21,843	21,843	21,843
Firm Fixed E ects	yes	yes	yes
Time Fixed E ects	yes	yes	yes
Clustered id	yes	yes	yes

Notes: In this table, we regress investment rate Iit=Ki;t1 (Investment in xed assets scaled by the capital stock at the beginning of period) on the three types of uncertainty (IndyRef, Brexit uncertainty or Scottish policy uncertainty) interacted with a dummy variable for irreversibility of investment. Additional controls are cash-ows scaled by the capital stock at the beginning of the period (CFi;t=Ki;t1 ) and sales growth rate (SGi;t). All regressions include rm and time xed e ects, and standard errors are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate statistical signi cance at the 10%, 5%, and 1% level, respectively.

ECB Working Paper Series No 2403 / May 2020

54

2020 May / 2403 No Series Paper Working ECB

55

Table 10: Scottish referendum for independence uncertainty and investment (excluding years 2015-16)

							Dependent variable:
						Investment Growth rate
				(1)	(2)	(3)	(4)	(5)	(6)	(7)
IndyReft1		0:003	0:002	0:011	0:001
				(0:014)	(0:014)	(0:015)	(0:014)
IndyRef	*Listed		0:097			0:098
		t1			(0:055)			(0:055)
IndyRef	*Manufact.			0:035			0:035
		t1				(0:014)			(0:014)
IndyRef	*Border				0:089			0:089
		t1					(0:051)			(0:051)
CF	=K	i;t1		0:025	0:025	0:025	0:025	0:025	0:025	0:025
it			(0:003)	(0:003)	(0:003)	(0:003)	(0:003)	(0:003)	(0:003)
SGit				0:201	0:201	0:204	0:201	0:201	0:204	0:201
				(0:035)	(0:035)	(0:035)	(0:035)	(0:035)	(0:035)	(0:035)
V F T SEt1		0:012	0:012	0:011	0:012
				(0:011)	(0:011)	(0:011)	(0:011)
Local Elections	0:005	0:005	0:005	0:005
				(0:022)	(0:022)	(0:022)	(0:022)
R2				0.041	0.042	0.042	0.042	0.041	0.041	0.041
N				18,906	18,906	18,906	18,906	18,906	18,906	18,906
Firm Fixed ectsE	yes	yes	yes	yes	yes	yes	yes
Time Fixed ectsE	no	no	no	no	yes	yes	yes
Clustered id	yes	yes	yes	yes	yes	yes	yes
Notes: In this table, we regress investment rate Iit=Ki;t1	(Investment in xed assets scaled by the capital stock at the beginning of period) on the

Scottish referendum uncertainty (IndyReft1 ). By considering the period from 2009 until 2015 we isolate the uncertainty developed by the Scottish

Referendum for independence alone. In addition, we interact IndyReft1 with a dummy variable for listed and Manufacturing companies as well as

those companies operating on the border with England. For information on additional controls see Table 5. All regressions include rm xed ects,e

and standard errors are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate statistical cancesigni at the

10%, 5%, and 1% level, respectively.

Continuation of Table 10

Dependent variable:

Investment Growth rate

			(1)	(2)	(3)
	Financing Constraints and Investment Irreversibility
	IndyReft1	*YS	0:013
			(0:026)
	IndyReft1	*lowCF&CR		0:0131
				(0:015)
	IndyReft1	*IRR			0:021
					(0:013)
	R2		0.036	0.038	0.057
	N		11,911	17,944	17,972
	Firm Fixed E ects	yes	yes	yes
	Time Fixed E ects	yes	yes	yes
	Clustered id	yes	yes	yes
Notes: In this table, we regress investment rate Iit=Ki;t1	(Investment in xed assets scaled by

the capital stock at the beginning of period) on IndyRef interacted with dummy variables for

external nancial constraints (Young and small rms, YS); internal nancial constraints (those with low levels of cash-owsand coverage ratio, lowCF&CR); and high investment irreversibility (IRR). Additional controls are cash-owsscaled by the capital stock at the beginning of the period (CFi;t=Ki;t1 ) and sales growth rate (SGi;t). All regressions include rm and time xed eects, and standard errors are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate statistical signicance at the 10%, 5%, and 1% level, respectively. Sample from 2009-2015.

ECB Working Paper Series No 2403 / May 2020

56

Table 11: Baseline regression robustness: additional control

Dependent variable: Investment rate (Iit=Ki;t1

)

(1)

(2)

(3)

(4)

(5)

(6)

Panel A: Adding UK's Economic Policy Uncertainty index

IndyReft1

0:067

0:055

(0:018)

(0:020)

Brexitt1

0:050

0:042

(0:012)

(0:016)

Scot:EP Ut1

0:025

0:016

0:009

(0:007)

(0:008)

(0:010)

UK:EP Ut1

0:007

0:013

0:008

0:008

0:011

0:021

(0:010)

(0:011)

(0:008)

(0:010)

(0:011)

(0:008)

R2

0.046

0.046

0.046

0.046

0.046

0.045

N

22,769

22,769

22,769

22,769

22,769

22,769

Panel B: Adding log(No. of employees)

IndyReft1

0:062

0:029

(0:012)

(0:012)

Brexitt1

0:036

0:018

(0:007)

(0:009)

Scot:EP Ut1

0:027

0:017

0:014

(0:006)

(0:007)

(0:008)

log(No Employeesi;t)

0:005

0:007

0:009

0:006

0:007

(0:017)

(0:017)

(0:017)

(0:017)

(0:017)

R2

0.040

0.040

0.039

0.039

0.039

N

19,747

19,747

19,747

19,747

19,747

Controls

yes

yes

yes

yes

yes

yes

Fixed Eects

yes

yes

yes

yes

yes

yes

Clustered id

yes

yes

yes

yes

yes

yes

Notes: In this table, we regress investment rate Iit=Ki;t1

(Investment in xed assets scaled by

the stock of xed assets at the beginning of period) on the three types of uncertainty at time t

1

(IndyRef, Brexit uncertainty or Scottish policy uncertainty). For information on additional

controls see Table 5. Panel A introduces Baker, Bloom, and Davis (2016) UK Economic Policy

uncertainty index to the baseline regression. Panel B includes rm size in the form of the natural

log of the total number of employees. All regressions include rm xed eects, and standard errors

are clustered at the rm level. Standard errors are reported in parentheses. *, **, and *** indicate

statistical signicance at the 10%, 5%, and 1% level, respectively.

ECB Working Paper Series No 2403 / May 2020

57

Continuation of Table 11 (without GDP growth rates)

	Dependent variable: Investment rate (Iit=Ki;t1	)
	(1)		(2)	(3)	(4)	(5)	(6)
Panel C: Adding UK's Economic Policy Uncertainty index
IndyReft1	0:027				0:002
	(0:011)			(0:008)
Brexitt1			0:051			0:037
			(0:012)			(0:015)
Scot:EP Ut1				0:025	0:027	0:013
				(0:009)	(0:007)	(0:010)
UK:EP Ut1	0:022		0:007	0:009	0:008	0:010	0:023
	(0:008)	(0:012)	(0:010)	(0:008)	(0:011)	(0:007)
R2	0.046		0.046	0.046	0.046	0.046	0.045
N	22,769		22,769	22,769	22,769	22,769	22,769

Panel D: Adding log(No. of employees)

IndyReft1	0:021			0:008
	(0:009)			(0:005)
Brexitt1		0:038			0:014
		(0:007)			(0:009)
Scot:EP Ut1			0:025	0:028	0:020
			(0:006)	(0:006)	(0:007)
log(No Employeesi;t)	0:009	0:007	0:008	0:009	0:001
	(0:017)	(0:017)	(0:017)	(0:017)	(0:017)
R2	0.038	0.040	0.039	0.039	0.039
N	19,747	19,747	19,747	19,747	19,747
Controls	yes	yes	yes	yes	yes	yes
Fixed Eects	yes	yes	yes	yes	yes	yes
Clustered id	yes	yes	yes	yes	yes	yes

Notes: Panels C and D presents the same specication as those in Panel A and B (respectively)

but excluding GDP growth rates from the regression controls.

ECB Working Paper Series No 2403 / May 2020

58

Table 12: Baseline regression econometric model robustness

		Dependent variable: Investment rate (Iit=Ki;t1	)
	(1)	(2)	(3)		(4)	(5)
Panel A: Without Fixed-Eects
IndyReft1	0:083				0:035
	(0:029)			(0:043)
Brexitt1		0:035				0:011
		(0:014)			(0:024)
Scot:EP Ut1			0:021		0:017	0:019
			(0:010)	(0:012)	(0:013)
R2	0.054	0.046	0.046		0.046	0.046
N	22,769	22,769	22,769		22,769	22,769
Fixed Eects	No	No	No		No	No
Clustered id	yes	yes	yes		yes	yes
Panel B: First-dierences estimation
IndyReft1	0:066				0:038
	(0:020)			(0:024)
Brexitt1		0:037				0:025
		(0:011)			(0:016)
Scot:EP Ut1			0:024		0:012	0:008
			(0:009)	(0:011)	(0:013)
R2	0.040	0.040	0.040		0.040	0.040
N	19,180	19,180	19,180		19,180	19,180
Fixed Eects	Yes	Yes	Yes		Yes	Yes
Clustered id	yes	yes	yes		yes	yes
Panel C: GMM estimation
IndyReft1	0:123				0:062
	(:041)			(0:153)
Brexitt1		0:064				0:029
		(0:021)			(0:142)
Scot:EP Ut1			0:052		0:026	0:028
			(0:018)	(0:066)	(0:118)
N	18,349	18,349	18,349		18,349	18,349
Fixed Eects	yes	yes	yes		yes	yes
Clustered id	No	No	No		No	No
Controls	yes	yes	yes		yes	yes

Notes: In this table, we regress investment rate Iit=Ki;t1 (Investment in xed assets scaled by the stock of xed assets at the beginning of period) on the three types of uncertainty at time t 1

(IndyRef, Brexit uncertainty or Scottish policy uncertainty). For information on additional

controls see Table 5. Panel A shows the baseline results without rm xed eects while Panel B

shows it using rst dierences xed-eects. In both specications, standard errors are clustered at

the rm level. Panel C shows a dynamic investment regression model estimated using a system

GMM model (see Section 6). Standard errors are reported in parentheses. *, **, and *** indicate

statistical signicance at the 10%, 5%, and 1% level, respectively.

ECB Working Paper Series No 2403 / May 2020

59

Continuation of Table 12 (Excluding GDP growth rates)

		Dependent variable: Investment rate (Iit=Ki;t1	)
	(1)	(2)	(3)		(4)		(5)
Panel D: Without Fixed-Eects
IndyReft1	0:020				0:022
	(0:012)			(0:018)
Brexitt1		0:040					0:053
		(0:009)				(0:018)
Scot:EP Ut1			0:027		0:038		0:013
			(0:008)	(0:012)	(0:016)
R2	0.054	0.054	0.054		0.054		0.054
N	22,769	22,769	22,769		22,769		22,769
Fixed Eects	No	No	No		No		No
Clustered id	yes	yes	yes		yes		yes
Panel E: First-dierences estimation
IndyReft1	0:015				0:038
	(0:016)			(0:021)
Brexitt1		0:029					0:052
		(0:012)				(0:018)
Scot:EP Ut1			0:004		0:034		0:012
			(0:011)	(0:013)	(0:015)
R2	0.040	0.040	0.040		0.040		0.040
N	19,180	19,180	19,180		19,180		19,180
Fixed Eects	Yes	Yes	Yes		Yes		Yes
Clustered id	yes	yes	yes		yes		yes
Panel F: GMM estimation
IndyReft1	0:038				0:133
	(0:035)			(0:064)
Brexitt1		0:071					0:086
		(0:019)				(0:009)
Scot:EP Ut1			0:048		0:105		0:017
			(0:018)	(0:033)	(0:030)
N	18,349	18,349	18,349		18,349		18,349
Fixed Eects	yes	yes	yes		yes		yes
Clustered id	No	No	No		No		No
Controls	yes	yes	yes		yes		yes

Notes: Panels D, E and F presents the same specication as those in Panel A, B and C

(respectively) but excluding GDP growth rates from the regression controls.

ECB Working Paper Series No 2403 / May 2020

60

Acknowledgements

I thank Charles Nolan, Campbell Leith, Michael McMahon, Diego Rodriguez Palenzuela, Diego Azqueta-Oyarzun, Spyridon Lazarakis, Daphne Aurouet, Andreas Dibiasi, Elisa Castagno, as well as participants at the Scottish Fiscal Commission Seminar (Feb 2018), the International Conference on Applied Theory, Macro and Empirical Finance (April 2018), the XXI Applied Economics Meeting (June 2018), the Asian Meeting of the Econometric Society (June 2018), and 3rd Essex Conference in Banking, Finance and Financial Econometrics (July 2018) for valuable comments. I alone am responsible for errors. This work was partly completed while at the European Central Bank. The views expressed in this paper are solely those of the author and do not necessarily represent the views of the ECB.

Andres Azqueta-Gavaldon

University of Glasgow, Glasgow, United Kingdom; European Central Bank; email: [e-mail]

© European Central Bank, 2020

Postal address 60640 Frankfurt am Main, Germany

Telephone [phone]

Website www.ecb.europa.eu

All rights reserved. Any reproduction, publication and reprint in the form of a different publication, whether printed or produced electronically, in whole or in part, is permitted only with the explicit written authorisation of the ECB or the authors.

This paper can be downloaded without charge from www.ecb.europa.eu, from the Social Science Research Network electronic library or from RePEc: Research Papers in Economics. Information on all of the papers published in the ECB Working Paper Series can be found on the ECB's website.

PDF

ISBN 978-92-899-4046-7

ISSN 1725-2806

doi:10.2866/961370

QB-AR-20-055-EN-N