The effect of terrorism on consumer confidence : an analysis of five European countries

(1)

The effect of terrorism on

consumer confidence

An analysis of five European countries

Bachelor Thesis Economics and Business

Name: A.B. Hoornweg (10537430) Supervisor: R.E.F. van Maurik Date: 29-01-2017

This paper analyses the effect of terrorism on the consumer confidence index in the Netherlands, Belgium, Germany, France and the United Kingdom. After conducting panel data regressions, two conclusions can be drawn. Firstly, when looking at the total number of casualties, terrorism in both the home country and a foreign country negatively affect the consumer confidence index of the home country. Secondly, when looking at dummy variables, the effects of terrorism are ambiguous.

(2)

2

Statement of Originality

This document is written by Student Anne Hoornweg who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

3

1. Introduction

Terrorism is a concept that is known for many years already. On 11 September 2001, a series of four coordinated terrorist attacks occurred in the United States. These terrorist attacks killed almost 3000 people, injured more than 6000 people and caused about $3.3 trillion in total cost. After the 9/11 attacks, terrorism, especially Islamic terrorism, has become more and more important in international politics. Besides the field of international politics, there exists a lot of scientific research on the economic consequences of international terrorism. Unfortunately, terrorism did not stop after the 9/11 attacks. With the arise of the Islamic State and its influences in Europe as well, the concept of

terrorism is still a hot debate. One of the questions that arises as a consequence of terrorism is what are its economic consequences?

Prior literature studied the economic effects of terrorism in multiple ways. Abadie and Gardeazabal (2003), for example, focused on the effects on GDP via a case study. On the other hand, Blomberg, Hess, & Orphanidesc (2004) investigated the macroeconomic consequences of 177 different countries. Another study, by Arin, Civerri, & Spagnolo (2008), observed the effects of terrorism on the stock market. Although the three researches mentioned before examined different economic consequences of terrorism, there are similarities between the studies. The similarity lies in the fact that terrorism negatively influenced the specific economic indicator in all three researches. This study aims to assess the consequences of terrorism on consumer confidence indexes. Consumer confidence is defined as the degree of consumer optimism on the current and expected future state of the economy. This optimism is expressed through savings and spending activities. The relation between terrorism and consumer confidence will be analysed by looking at five European countries, in the period from 1969 up to and including 2015. After collecting data of the terrorist attacks, consumer confidence indexes and other control variables, panel data regressions are conducted to estimate the consequences.

A contribution is made to existing literature by focusing on the economic indicator consumer confidence index. Despite the existence of a lot of literature regarding the economic consequences of terrorism, the effects on the consumer confidence index are relatively unexplored. Since existing literature shows that the consumer confidence index is a key determinant of economic growth (Ludvigson, 2004), it is interesting to investigate the effects of terrorism on this economic indicator. The paper of Garner (2002) already investigated the effect of the 9/11 attacks on the United States consumer confidence index. This study will enlarge the research of Garner in two ways. First of all, this study focuses on five European countries, instead of just one country. The second enlargement of this study, is the fact that it analyses terrorist attacks occurring in the period from 1969 up to and including 2015. Garner’s research only investigated the consequences of the 9/11 attacks. By looking at multiple countries and a lot of terrorist attacks, more general conclusions about the consequences of terrorism on the consumer confidence indexes can be drawn.

(5)

5

This research studies the effects of terrorism on the consumer confidence index in two ways. Firstly, by looking at variables regarding the total number of casualties of an attack. Secondly, by looking at dummy variables regarding terrorism. In short, two conclusions can be drawn. On the one hand, when looking at the variables regarding the total number of casualties, terrorism in both the home country and a foreign country negatively influence the consumer confidence index in the home country. On the other hand, when looking at the dummy variables, the effects of terrorism are ambiguous.

The outline of this study is as follows. In section 2, a review of prior literature on the topics of terrorism and consumer confidence is given. Explanations of the concepts used in this study are provided and existing views are discussed. Section 3 discusses the data that are used in this study. The databases of which the data are gathered are pointed out and the main variables are visualized. In section 4, the methodology is discussed. The regression equations are described and the econometric model that is used is explained. Section 5 presents the results of this research. The results of the regression models and robustness checks are discussed. Lastly, in section 6, this study will end with a conclusion and suggestion for further research.

2. Literature review

This section discusses existing literature on the topics of terrorism and consumer confidence. Firstly, the definition of terrorism and the impact of terrorism on the economy are discussed. Secondly, the concept of consumer confidence is discussed, together with its determinants and economic

implications. Finally, this section concludes with the hypotheses of this research.

2.1 Terrorism and its economic consequences

In the literature, there are different definitions for the concept of terrorism. Although there are different definitions for the concept of terrorism, in the quarterly report of the Dutch Central Bank (2005, September), three elements are mentioned, which are the main elements of most definitions of terrorism. The main elements consist of the target, the goal and the motives of terrorism. The target is intentionally and selectively chosen by the terrorists. The goal is to cause fear, intimidation or

uncertainty, whereby more than only direct victims are affected. Motives are of a political or religious background and not for personal gain. The definition of terrorism given by the Global Terrorism Database (GTD) is somewhat more specific. According to the codebook of the GTD (2016, June), terrorist attacks must have met certain criteria before they are included in the GTD database. All of the three following criteria must be met: the incident is intentional, the incident involves some level of violence or immediate threat of violence and the terrorists are sub-national actors, so state terrorism is not included. Moreover, two of the following three criteria must be met: the goal of the attack must be political, economic, religious or social, there must be an intention to coerce, intimidate or convey a message to more people than the direct victims, and the attack is not included in the context of

(6)

6

legitimate warfare activities.

International statistics capture terrorism by different indicators. According to Frey, Luechiner, & Stutzer (2007) the indicators ‘number of terrorist incidents’ and ‘number of casualties’ are

frequently used. However, Frey et al. state that these indicators have drawbacks. Whenever terrorist incidents are captured by the indicator ‘number of terrorist incidents’, terrorist activities of varying magnitude are put together. This indicator is only a proper indicator when the structure of terrorist activities remains unchanged. Another drawback is the fact that terrorism statistics are vulnerable to biased reporting. Only terrorist activities captured by official statistics or mentioned in the media are counted. Especially in less developed and authoritarian countries, the problem of biased reporting is of frequent occurrence. To limit the mentioned drawbacks, Frey et al. state, that it is favorable to restrict the research to developed countries. Moreover, most researchers limit their research to a small number of countries, where attack characteristics are comparable across countries and over time.

In the paper of Abadie and Gardeazabal (2008), the effects of terrorism are analysed in an integrated world economy. In their research, Abadie and Gardeazabal state the following: ‘from an economic standpoint, terrorism has been described to have four main effects’. Enumerated, the four main effects are: a reduction of human and physical capital stock, higher levels of uncertainty, increases in counter-terrorism expenditures and negative effects on specific industries, such as tourism. In an earlier paper, Abadie and Gardeazabal (2003) investigated the economic effects of political and terrorist conflict via a case study of the Basque Country. Before the start of the conflict, the Basque Country belonged to the richest regions of Spain, having the third highest per capita GDP of all 17 regions in Spain. After 30 years of conflict, the per capita GDP of the Basque Country declined to the sixth highest of the 17 regions in Spain. Abadie and Gardeazabal used a case study to estimate what percentage decline in per capita GDP is caused by the political and terrorist conflict. Using a combination of other Spanish regions, a control region has been constructed. This control region displays relevant economic characteristics of the Basque Country, before the terrorism

activities started. In the paper, the actual economic performance of the Basque Country is compared to the economic performance of this control region. The paper shows that political and terrorist conflict caused per capita GDP in the Basque Country to decline about 10 percentage points relative to the per capita GDP of the control region.

In the research of Frey et al. (2007), the consequences of terrorism are surveyed on the basis of other scientific papers, for eight different variables (tourism, foreign direct investment, savings and consumption, investment, stock markets, foreign trade, the urban economy and overall economic development). Frey et al. show that terrorism has a negative effect on six of those eight variables, namely tourism, foreign direct investment, the level and composition of investment, stock markets in some cases, foreign trade and national income and growth. However, according to Frey et al., these economic consequences do not capture the total effect of terrorism. They state that, in addition to economic losses, terrorism causes other utility losses, which are considerably larger. Non-market

(7)

7

values, like the fear of individuals are not included in economic consequences. Frey et al. show that terrorism has large utility losses as a consequence, due to a reduction in life satisfaction. This leads to the assumption that the total cost of terrorism may be underestimated.

Blomberg et al. (2004) observe the macroeconomic consequences of terrorism of 177 countries from 1968 to 2000. In their study, next to the impact of international terrorism, they investigate consequences of alternative forms of collective violence on economic growth. Their research is performed by means of cross-country growth regressions, panel data regressions and structural VARs. First of all, their research results show that terrorism may have an economic significant negative effect on growth. This negative effect is less persistent than the effect associated with external war or internal conflict. Furthermore, their paper suggests that there is a difference in the effect for developing countries and developed countries. Although there are more terrorist attacks in developed countries than in less developed countries, the negative impact on economic growth is more severe in developing countries. Besides the negative effect on economic growth, Blomberg et al. state that terrorism causes a shift of economic activity as well. Due to terrorist attacks, economic activity is shifted from investment spending to government spending.

The effect of terrorism on stock markets is, amongst others, investigated by Arin et al. (2008). In their paper, Arin et al. investigate the effect on stock markets and stock market volatility by looking at six different financial markets (Indonesia, Israel, Spain, Thailand, Turkey and the UK) in the period ranging from 2002 to 2006. Stock market returns and the terrorism index are modeled in a bivariate VAR-GARCH(1,1)-in-mean model. In this model, the domestic 90-days Treasury Bill rate and the US stock market returns are included to control for exogenous shocks. The results of their paper, show that terrorist attacks have a significant impact on stock markets and stock market volatility. Moreover, just like the earlier mentioned result of Blomberg et al. (2004), Arin et al. show that the effects of terrorism are bigger in emerging markets than developed markets.

After the September 11 attack in the United States, Garner (2002) examined the effect of this attack on the consumer confidence value of the United States. Garner investigated this effect by estimating the relation between the consumer confidence index and the lagged values of the inflation rate, unemployment rate, changes in stock prices and consumer confidence index, together with a dummy variable for the September 11 attack and dummy variables for other unique events. Using both the consumer confidence index of The Conference Board and the consumer sentiment index of the University of Michigan, the conclusion of his paper is that consumer confidence did not decline due to the September 11 attacks. The consumer confidence indexes declined in the fourth quarter of 2001, but this decline was due to weaker economic conditions in the previous quarters, instead of the attack itself. The consumer confidence indexes followed a normal relation relative to other economic

variables and started to recover at the end of 2001. Although the consumer confidence indexes did not include a lot of new economic information, the quick recovery of the indexes after the attack offered reassurance for the economic implications of the September 11 attack.

(8)

8

2.2 Consumer confidence: its determinants and its economic implications

As written in the paper of Ludvigson (2004), consumer confidence indexes measure the public confidence in the economy. These indexes are measured by taking surveys that include questions reporting two components: a present situation component and an expectations component. Ludvigson states that, nowadays, surveys of consumer confidence are treated as important economic information. Federal Reserve Chairman Alan Greenspan often calls consumer confidence ‘a key determinant of near-term economic growth’. Much of the variation in consumer confidence can, directly or indirectly, be explained by economic conditions. Fuhrer (in Ludvigson, 2004) argues that 70 percent of the variation of the consumer confidence index is explained by variation of the four economic variables national income, unemployment rates, inflation rates and real interest rates. However, De Boef and Kellstedt (2004) show that these economic variables do not account for the whole explanation. They argue that the remaining part comes from political conditions.

The research of Lemmon and Portniaguina (2006) focuses on the relation between investor sentiment and premiums for small-stocks. As a measure of investor optimism, the consumer confidence index is used. Before estimating the effect of investor sentiment on stock returns, they regress the consumer confidence index on a set of economic variables. The economic variables used in the regression model are: default spread, three-month treasury yield, dividend yield, GDP growth, consumption growth, labor income growth, unemployment rate, growth in unemployment rate, inflation rate and the consumption-to-wealth ratio. The regression results show that the R2_{has a value} of around 0.8, which means that about 80 percent of the variation in the consumer confidence index is explained by the macroeconomic variables. Thus, this means that about 20 percent of the variation in the consumer confidence index remains unexplained, even after controlling for the economic

variables.

Although consumer confidence indexes are treated as important economic information, Ludvigson (2004) states that the way in which consumer confidence influences the real economy is not well understood. Researchers have investigated whether consumer confidence indexes contain information about the economy, in addition to already existing economic information. Carroll, Fuhrer, & Wilcox (1994) tell us that ‘in the three months following the Iraqi invasion of Kuwait, the

University of Michigan's Index of Consumer Sentiment (ICS) fell an unprecedented 24.3 index points, to its lowest level since the 1981-1982 recession’. In their paper, Carroll et al. investigate whether an index of consumer sentiment has any predictive power and if so, whether this index includes extra information on top of the information included in other available indicators. The results of their paper, show that values of the ICS explain approximately 14 percent of the variation in growth of personal consumption expenditures in the period after 1954. This indicates that the ICS had predictive power in this period. Although the evidence for the second hypothesis is less certain, Carroll et al. state that the ICS provides at least some extra information, of about 3 percent, which is not provided by other indicators. This last result only holds for the longer sample period and not for their shorter sample

(9)

9

period.

Ludvigson (2004) measures the impact of consumer confidence on consumption, by distinguishing between five categories of household personal consumption expenditure. The results show that consumer confidence taken alone has both an economic and statistical significant predictive power for quarterly consumption growth, of 19-20 percent for four out of five categories. However, when a standard set of baseline economic indicators for consumption growth is added, the forecasting power of consumer confidence on consumer expenditure growth reduces to 5-10 percent. This indicates that consumer confidence, taken alone, has a strong predictive power for consumer

expenditure growth. However, in accordance with the research of Caroll et al. (1994), the incremental predictive power of consumer confidence, which is not contained by other economic indicators, is smaller.

In the paper of Acemoglu and Scott (1994), Acemoglu and Scott try to answer some empirical and conceptual issues, regarding the use of confidence indicators. The two issues that are attempted to solve, are whether consumer confidence contains information about current or future consumption and whether consumer confidence has predictive power in addition to other macroeconomic variables. To solve the above mentioned issues, a regression is conducted with the change in the logarithm of labour income as the dependent variable and the lagged value of the dependent variable, together with changes in the confidence index, changes in unemployment rates, inflation rates, real interest rates and changes in financial and housing wealth as independent variables. The regression results show that the variable ‘changes in the confidence index’ is one of the variables that predicts the change in labour income. Next to this regression, the same regressions are conducted with a wide range of other macroeconomic variables as dependent variable. The only macroeconomic variable for which the consumer confidence index did not have predictive power is the variable ‘changes in financial wealth’. Since the consumer confidence index contains predictive power, even when other macroeconomic variables are included in the regression, the suggestion is made that the consumer confidence index reflects households’ private information. Moreover, the research shows that the confidence index has predictive power for consumption growth, over and above other macroeconomic variables. The results of Acemoglu and Scott differ from the earlier mentioned results of Carroll et al. (1994) in two ways. First, the predictive power of the consumer confidence index in the study of Acemoglu and Scott is greater than the predictive power of the consumer confidence index in the study of Caroll et al.. The second difference is the magnitude of the predictive role of the consumer confidence index, once conditioned on income. In the paper of Acemoglu and Scott, this magnitude shows to be bigger than the magnitude found in the paper of Caroll et al.. The different results can perhaps be explained by the fact that the two papers use a different definition of consumption.

(10)

10

2.3 Conclusion and hypotheses

Prior studies show that terrorism influences the economy in different ways. Enumerated, the ways in which terrorism affects the economy are a reduction of human and physical capital stock, higher levels of uncertainty, increases in counter-terrorism expenditures and negative effects on specific industries. One of the aspects of the second effect, higher level of uncertainty, can be found in consumer

uncertainty or consumer confidence. Researches regarding consumer confidence show that a great part of the variation in consumer confidence can be explained by macroeconomic variables. Nevertheless, even after controlling for those macroeconomic variables, a part of the variation in consumer

confidence remains unexplained.

This study will contribute to this debate, by looking whether some of, and if so, what part of, the variation in consumer confidence can be explained by terrorism. In this research, two hypotheses will be studied. The first hypothesis that will be studied, is whether terrorism in the home country affects the home country’s consumer confidence index. The second hypothesis that will be studied, is whether terrorism in a foreign country affects the home country’s consumer confidence index. Since prior studies show different results about the economic predictive power of the consumer confidence index, the results of the hypotheses tests should be carefully analyzed. Although most studies show some predictive power of the consumer confidence index itself, the magnitude of this predictive power is not consistent between different studies.

3. Data

This section discusses the data that are used in this research. In order to investigate into the effect of terrorism on consumer confidence, using Excel, a dataset has been constructed. This dataset consists of quarterly data, ranging from the fourth quarter of 1969 up to and including the second quarter of 2015, for five West-European countries (the Netherlands, Belgium, Germany, France and the United Kingdom). The databases that are used, are The Global Terrorism Database (GTD) (2015) for all data regarding terrorism and the database of the Organisation for Economic Co-operation and

Development (OECD) (2016)regarding all economic data. Since there are missing observations for some variables, the data panel used is unbalanced.

Firstly, the dependent variable, the consumer confidence index, is discussed. Secondly, the independent variables of interest, the variables regarding terrorism, are discussed. Thirdly, the macroeconomic control variables are discussed.

3.1 Consumer confidence index

The dependent variable in this research is the consumer confidence index. According to the definition of the OECD database (2016), the consumer confidence index is a qualitative index on economic conditions, which displays households’ current and expected future economic situation. The index is composed on the basis of households’ opinions. Households’ opinions about their economic situation

(11)

11

are compared to a ‘normal’ state, after which the differences between positive and negative answers are taken. The OECD consumer confidence index is amplitude adjusted, with the long-term average equal to 100. The consumer confidence index is issued monthly by the OECD. To obtain the

consumer confidence data needed for this research, the monthly data are converted into quarterly data. Quarterly data are obtained by taking, for each quarter, averages of the data of the three corresponding months.

To gain more insight into the values of the consumer confidence indexes, different line plots and box plots are created. The figures of the line plots can be found in appendix A, figures A.1 till A.6. Firstly, separate line plots for the five countries are created. Secondly, a line plot for all countries combined is created. The line plots show that the values of the consumer confidence indexes range between 96 and 104. Moreover, in the figures, it can be noted that the consumer confidence indexes are volatile. For all five countries, the indexes include peaks and troughs. Although the consumer confidence indexes of the five countries are volatile, when the line plots are combined, all five indexes move more or less in the same way. Peaks and troughs in the consumer confidence indexes of the five countries occur in approximately the same periods. This result indicates that the five countries are comparable when looking at their consumer confidence indexes.

When looking at the box plots of the consumer confidence indexes in figure A.7 of the appendices, the same conclusion can be drawn. Although the lowest values of the indexes differ considerably among the five countries, when looking at the first, second and third quartiles the countries are more or less alike. The medians of the indexes of the five countries are around 100, which equals the long-term average. The differences between the five countries in the values of the first and third quartiles are about 0.5. From the box plots, it can be seen that two out of five countries contain outliers in the consumer confidence index. The outliers of Belgium are above the highest value, excluding outliers. The outliers of the Netherlands are below the lowest value, excluding outliers. Since the outliers are not very extreme and do not represent errors in the data, they will not be removed from the dataset.

3.2 Terrorism data

The independent variables of interest for this research are variables regarding terrorism. Statistics regarding terrorism are obtained via The Global Terrorism Database (2015), which is a database with information on terrorist events around the world from 1970 till 2015. The Global Terrorism Database contains data on domestic, transnational and international terrorist incidents, including specific information regarding the incidents. To make the terrorism data fit the variables of the regression, manually a new terrorism dataset has been created in Excel. In this dataset, only terrorist attacks that were of the type ‘Bombing/Explosion’ are included. As mentioned before, Frey et al. (2007) state that the terrorism indicators captured by international statistics have drawbacks. In reaction to the

(12)

12

terrorist attacks that took place in the Netherlands, Belgium, Germany, France or the United Kingdom are considered. The characteristics in the GTD ‘number of people killed’ and ‘number of people wounded’ are summed, to create the new variable ‘number of casualties’. In this research the effect of terrorist attacks in the home country and the effect of terrorist attacks in a foreign country are studied. These effects are studied in two ways. Firstly, by looking at the total number of casualties of a terrorist attack. Secondly, by looking at dummy variables that indicate whether or not a terrorist attack took place.

To investigate into the effects of terrorism on the consumer confidence index by looking at the total number of casualties, two independent variables are created. The first variable is the number of casualties of a terrorist attack in a given period in the home country. The second variable is the sum of the number of casualties of a terrorist attack in a given period in all of the four foreign countries. To gain more insight into the terrorist attacks that took place in the five countries, descriptive statistics, line plots and box plots are created. The descriptive statistics can be found in appendix B, table B.1. The line plots and box plots can be found in appendix B, figures B.1 to B.6. When looking at the line plots, it can be concluded that the number of attacks that took place in the five countries differs a lot between the five countries. Moreover, the number of casualties of a specific attack varies widely. Table B.1 shows that the attack with most casualties, 841 casualties, took place in the United

Kingdom. This attack occurred in the year 2005, which can be seen from the line plot in figure B.4. In 2005, a series of coordinated terrorist attacks occurred in London. With 841 casualties, this attack is known as the worst terrorist incident in the United Kingdom since 1988. On the other hand, in the Netherlands, the attack with most casualties only caused five casualties. These large differences in the number of casualties of a specific attack are confirmed by the box plots in figure B.6 of the

appendices. When looking at the box plots, it can be concluded that there are a lot of outliers. The largest outlier is the attack in the United Kingdom in 2005. The Netherlands has least outliers. Moreover, these differences in magnitude cause great differences in the mean values and standard deviations of the five countries. Namely, the United Kingdom has the highest mean value and highest standard deviation, 33.3388 and 77.97996 respectively. On the contrary, the Netherlands has the lowest mean value and lowest standard deviation, 0.0874317 and 0.537603 respectively. When looking at all countries combined, the terrorist attacks caused a mean value of 9.518033 casualties, with a standard deviation of 40.23447.

To investigate into the effects of terrorism on the consumer confidence index by looking at dummy variables, two dummy variables are created. The first dummy variable equals one if the number of casualties in the home country in a given period is higher than zero. The second dummy variable equals one if the sum of the number of casualties in all of the four foreign countries in a given period is higher than zero. To gain an understanding of the number of attacks that took place in the different countries, a frequency table of the first dummy variable has been created. This frequency table can be found in appendix B, table B.2. When looking at the frequency table of this dummy

(13)

13

variable, for every country, the number of periods with a terrorist attack with more than one casualty can be observed. From the table, it can be observed that the country with most periods in which there was a terrorist attack with more than one casualty, is the United Kingdom. In the United Kingdom, in 130 out of 183 periods, the dummy variable equaled one. This means that, in the United Kingdom, in 71.04% of all periods a terrorist attack with more than one casualty occurred. This percentage is far above the mean of the five countries combined. When looking at the five countries combined, only in 28.96% of all periods the dummy variable equaled one. The Netherlands has the lowest amount of periods with a terrorist attack that caused more than one casualty, namely six.

Thus, when combining the results of the descriptive statistics of the number of casualties and the frequency table of the dummy variable, the following two conclusions can be drawn. Firstly, most terrorist attacks occurred in the United Kingdom. Moreover, amongst the terrorist attacks occurring in the United Kingdom, are the terrorist attacks that caused most casualties. Secondly, least terrorist attacks occurred in the Netherlands. Moreover, the terrorist attacks that occurred in the Netherlands, caused little casualties.

3.3 Macroeconomic control variables

Next to the independent variables regarding terrorism, control variables will be added to provide internal validity. Control variables prevent the estimated effect of the variables of interest to suffer from an omitted variable bias (Stock & Watson, 2015). As written in the quarterly report of the Dutch Central Bank (2007, June), Dutch consumer confidence has economic determinants and psychological determinants. The five decisive economic determinants are changes in unemployment rates (-), fluctuations in stock market rates (+), housing prices (+), inflation rates (-) and the yield spread as an indication for monetary policy (+). The ways in which the control variables influence the consumer confidence index, can be explained as follows. Firstly, higher unemployment rates negatively influence the consumer confidence index. Higher unemployment rates cause a reduction in

households’ income. Reductions in income cause a decline in households’ spending activities. Since the degree of consumer optimism is expressed through savings and spending activities, higher unemployment rates will cause a reduction in consumer optimism. Secondly, fluctuations in stock market rates and housing prices influence the consumer confidence index positively. This result can be explained by the part that shares and homeownership take in the assets of consumers. In case that either stock market rates or housing prices increase, consumers will see an increase in their assets. This increase in assets positively influences households’ opinions about their economic situation. Thirdly, inflation rates have a negative influence on the consumer confidence index. The definition of inflation is an increase in the price level of goods and services. This increase in price level, results in a loss of purchasing power. Because of the loss of purchasing power, spending activities decline, which explains the decline in consumer confidence indexes. Finally, monetary policy influences the

(14)

14

authority of a country that influence the money supply and interest rates. A reduction in interest rates makes borrowing more interesting for households. This increase in borrowing results in an increase of households’ willingness to buy goods and services, which in turn increases the amount of spending activities. Besides the economic determinants, the research of the Dutch Central Bank reveals four psychological determinants of consumer confidence. The four psychological determinants are temperature, rainfall, terrorism and football performances of the Dutch national football team on European and World championships. Since this research uses a panel regression, the psychological variables temperature, rainfall and football performances are omitted. These variables can be captured by country and time fixed effects in the panel regression. In the research of Acemoglu and Scott (1994), the idea of the Dutch Central Bank about economic variables influencing consumer

confidence is confirmed. Moreover, in addition to the five economic variables indicated by the Dutch Central Bank, Acemoglu and Scott state that consumer confidence is influenced by changes in labour income as well. The six macroeconomic variables indicated by Acemoglu and Scott are real interest rates, inflation, changes in housing wealth, changes in financial wealth, changes in labour income and changes in unemployment. Next to these six macroeconomic variables, the lagged value of the consumer confidence index is added as an explanation for the current consumer confidence index. All macroeconomic variables are gathered from the OECD database (OECD, 2016). After collecting the data, if needed, the data are transformed into quarterly data, to fit the variables of the regression. Monthly data are transformed into quarterly data, by taking, for each quarter, averages of the data of the three corresponding months.

To gain more insight into the control variables, descriptive statistics are obtained to evaluate the variables. Descriptive statistics can be found in appendix C, table C.1. The maximum number of observations for each variable is 915, namely five countries, each observed in 183 time periods. Since the panel is an unbalanced panel, not every variable includes all 915 observations. The variables with less than 915 observations are ‘lagged consumer confidence index’ (841), ‘yield spread’ (882), ‘share prices growth’ (848), ‘real house prices growth’ (903) and ‘unemployment rate’ (703). The mean value of the lagged consumer confidence index is equal to 99.97371, which is slightly lower than the long-term average of 100. The inflation rate has a mean value of 4.064068. However, the lowest value of the inflation rate and the highest value of the inflation rate are quite far from one another. The lowest inflation rate equals -1.218893 and the highest inflation rate equals 26.56581. Due to these large differences, the inflation rate has the highest standard deviation of the seven control variables, namely 3.7741. Moreover, for all control variables, the standard deviation within is higher than the standard deviation between. This result means that, for all variables, the variation over time within one country is higher than the variation between countries.

(15)

15

4. Methodology

This section discusses the methodology that is used to test the two hypotheses of this research. Firstly, an overview of the regression model and its components is given. Secondly, the model that should be used to test the hypotheses is discussed. Thirdly, the assumptions underlying the model of interest are tested. Finally, this section ends with a conclusion.

4.1 Regression models

In this research, the hypotheses are tested by performing panel data regressions for two different models. The first model includes the two variables regarding the number of casualties of an attack. The second model includes the two dummy variables regarding terrorism. Next to these terrorism variables, the discussed macroeconomic control variables are added to the model. In econometric literature, the three main approaches for panel data regressions are fixed effects regression, random effects regression and pooled regression. Moreover, the fixed effects regression can include two types of fixed effects. First of all, a fixed effects regression includes country fixed effects. Country fixed effects control for omitted variables that vary across countries, but do not change over time (Stock & Watson, 2015). Next to the country fixed effects, a fixed effects regression can include time fixed effects. Time fixed effects control for omitted variables that vary over time, but do not change across countries.

In this research, the regression equations for the three approaches, fixed effects regression, random effects regression and pooled regression, are as follows:

CCIit = (αi + ϒt or α) + β1*CCIit-1 + β2*CHit + β3*CFit + β4*GDPit + β5*INFit

+ β6*YIELDit + β7*SHAREit + β8*HOUSEit + β9* UNEit + µit + (εit) (1) CCIit = (αi + ϒt or α) + β1*CCIit-1 + β2*DHit + β3*DFit + β4*GDPit + β5*INFit

+ β6*YIELDit + β7*SHAREit + β8*HOUSEit + β9* UNE it + µit + (εit) (2) Where ‘i-subscript’ indicates the country variable, with i = 1, 2, …, 5 and ‘t-subscript’ indicates the time period, with t = 1, 2, …, 183. ‘αi’ represents a country specific intercept, for the five different countries. ‘ϒt’ represents a time specific intercept, for the 183 different periods. ‘α’ represents an intercept that does not depend on a specific country or time period. ‘µit’ is the error term between the different countries and ‘εit’ is the error term within the different countries. The parentheses are used to indicate the differences for the three different regression models. The fixed effects regression includes the country specific intercept ‘αi’ and can include the time specific intercept ‘ϒt’. The random effects regression model uses the intercept ‘α’ and includes the error term ‘εit’. The pooled regression model uses the intercept ‘α’. The definitions of the variables are as follows:

(16)

16

• CCI is the consumer confidence index. CCIit is the consumer confidence index in period t and is used as the dependent variable. CCIit-1 is the lagged value of the consumer confidence index and is used as an independent variable.

• CH is the number of casualties of an attack in the home country

• CF is the sum of the number of casualties of an attack in the four foreign countries • DH is the dummy variable for a terrorist attack in the home country

• DF is the dummy variable for a terrorist attack in a foreign country • GDP is the growth in GDP

• INF is the inflation rate, according to the consumer price index

• YIELD is the yield spread in percentages. The yield spread is measured by the difference in long term (10 year) interest rates and short term (three-month) interest rates on government bonds.

• SHARE is the growth in share prices • HOUSE is the growth in real house prices • UNE is the unemployment rate

4.2 The model: fixed, random or pooled?

As stated by Baltagi (2001), the choice between using the fixed effects or random effects model has generated a hot debate in panel data econometrics literature. As argued by Mundlak (in Baltagi, 2001), the fixed effects model allows for endogeneity of all regressors with the random individual effects. On the other hand, the random effects model assumes exogeneity of all regressors with the random individual effects. In 1978, Hausman proposed a specification test, based on the difference between the fixed effects and random effects models. In the Hausman test, it is tested whether the coefficients estimated by the efficient random effects model are the same as the coefficients estimated by the less efficient, but consistent fixed effects model. Rejection of the null hypothesis of the Hausman test is interpreted as an adoption of the fixed effects model and non-rejection is interpreted as an adoption of the random effects model. However, the Hausman test allows some of the regressors to be correlated with the random individual effects. The allowance for some regressors to be correlated with the individual effects, is in contradiction with the previous mentioned argument of Mundlak, which says that the random effects model assumes exogeneity of all regressors. This result means that one should be well aware of the limitations of the Hausman test, when choosing between the fixed effects or random effects model.

In order to determine whether to use the fixed effects model or the random effects model, the Hausman test is conducted. To run a Hausman test in Stata, both the fixed effects model and random effects model have to be estimated. After estimating both models, the Hausman test compares the models. The Hausman test is conducted two times, the first time for the model including the variables regarding the number of casualties and the second time for the model including the dummy variables.

(17)

17

The results of the two Hausman tests can be found in appendix D, table D.1 and D.2 respectively. For the model including the variables regarding the number of casualties, the test generates a Chi-squared value of 4.86 with a corresponding p-value of 0.6771. For the model including the dummy variables, the test generates a Chi-squared value of 8.97 with a corresponding p-value of 0.4398. These p-values imply that the null hypotheses are not rejected and, following the Hausman test, the random effects model should be used for both models.

As a result of the Hausman test, both hypotheses will be tested using the random effects model. However, as mentioned before, one should be well aware of the limitations of the Hausman test. Since the Hausman test allows some of the regressors to be correlated with the random individual effects, which is not allowed in the random effects model, next to the random effects model, the fixed effects model will be used. In the fixed effects model, endogeneity of all regressors with the random individual effects is allowed. To decide whether it is necessary to include time fixed effects in the fixed effects regression model, a Wald test in Stata is conducted. The Wald test tests whether the dummies for all quarters are equal to zero or not. Rejection of this null hypothesis implies that time fixed effects have to be included. For the model including the variables regarding the number of casualties, the test generates an F-value of 1.74 with a corresponding p-value of 0.0000. For the model including the dummy variables, the test generates an F-value of 1.81 with a corresponding p-value of 0.0000. Thus, since for both models the null hypothesis is rejected, time fixed effects should be included in the fixed effects model.

The pooled regression model is the simplest regression of the three regression models, since it does not include individual-specific fixed or random effects. The pooled regression model assumes that all regression coefficients are the same across countries. Since prior literature showed evidence for psychological determinants of the consumer confidence index, which vary across countries, it is not likely that all regression coefficients are the same across countries. This causes the standard errors of the pooled regression model to be highly inaccurate. However, to see what the influences of the fixed and random effects are, next to these two regressions, a pooled regression will be conducted.

4.3 The assumptions

In order to use the three regression models, the data need to meet certain assumptions underlying these models. The assumptions that are tested are:

1. strict exogeneity 2. stationarity 3. homoskedasticity

4. no autocorrelation within panels 5. no perfect multicollinearity

(18)

18

The first assumption implies that the error term must have a conditional mean of zero, given all T values of X for that country. This assumption is met if the current error term is not correlated with past, present or future values of X (Stock & Watson, 2015). The assumption is tested via a Kernel Density Estimation. With the Kernel Density Estimation, the probability density function of the error term of a model is estimated. The results of the Kernel Density Estimations can be found in appendix E, figures E.1 till E.6. The figures show that the error terms of both models have approximately conditional means of zero for all three regressions. This implies that the assumption of strict exogeneity is met.

In this research, a regression analysis including panel data of 183 time periods is performed. In time series regressions, data from the past are used to quantify historical relationships and forecast the future. However, predicting the future based on data from the past is only reliable if the future is like the past. If the future fundamentally differs from the past, data from the past are not reliable for predicting the future. This idea if formalized by the concept of stationarity. Time series data are stationary if its probability distribution does not change over time (Stock & Watson, 2015). One way to test for stationarity in panel data is via a Fisher-type test. This test is a unit-root test, which allows for unbalanced panels and a finite number of panels. The Fisher-type test resulted in a p-value of 0.0000. This p-value implies that the null-hypothesis is rejected and therefore, it can be concluded that the data are stationary.

Homoskedasticity of the error term, µit, means that the variance of the conditional distribution of µit, given Xit, is constant over time and does not depend on Xit. If this is not the case, the error term is called heteroskedastic (Stock & Watson, 2015). To test for heteroskedasticity, the LR-test can be used. To conduct an LR-test, both the model with level heteroskedasticity and without panel-level heteroskedasticity are regressed using a feasible generalized least squares regression. After fitting both models, the LR-test can be conducted. For the model including the number of casualties, the LR-test resulted in a Chi-squared value of 41.97 with a corresponding p-value of 0.0000. For the model including the dummy variables, the LR-test resulted in a Chi-squared value of 42.61 with a corresponding p-value of 0.0000. For both models, the null hypothesis of the LR-test is rejected, which indicates that the error terms are heteroskedastic. Since the error terms are heteroskedastic, the usual standard errors are not valid. To cover this problem, heteroskedasticity-consistent standard errors have to be used.

The fourth assumption says that autocorrelation within panels is not allowed. Autocorrelation occurs when Xit is correlated over time for a given entity. In other words, autocorrelation occurs when Xit is correlated with Xis for different values of s and t. Since, in time series data, events in one year are often correlated with events in the next year, Stock and Watson (2015) state that autocorrelation is of frequent occurrence in time series data. Autocorrelation in panel data can be tested by conducting a Woolridge test. The null hypothesis of the Woolridge test states that the dataset contains no

(19)

F-19

value of 69.461 with a corresponding p-value of 0.0011. For the model including the dummy

variables, the Woolridge test resulted in an F-value of 67.587 with a corresponding p-value of 0.0012. Since for both models the test statistics are significant, it can be concluded that the dataset contains autocorrelation. To control for the autocorrelation in both models, autocorrelation-consistent standard errors will be used.

The last assumption that is tested, states that perfect multicollinearity is not allowed. Stock and Watson (2015) describe a regressor to be perfectly multicollinear, if it is a perfect linear function of one of the other regressors. To test for perfect multicollinearity, the variance inflation factor (VIF) is measured. By calculating the VIF, an index is created that measures how much the variance of a regressor is increased due to collinearity. In statistics, a VIF score larger than 10 is considered as a case of multicollinearity. The variance inflation factors are measured for the model including the number of casualties and the model including the dummy variables. The results can be found in appendix F, table F.1 and F.2 respectively. Since, in table F.1 the highest VIF is 1.52 and in table F.2 the highest VIF is 1.56, there is no perfect multicollinearity. This result indicates that the last

assumption is met and no variables have to be removed.

Since the dataset is suffering from both heteroskedasticity and autocorrelation,

heteroskedasticity- and autocorrelation-consistent (HAC) standard errors have to be used. One type of HAC standard errors are clustered standard errors (Stock & Watson, 2015). Clustered standard errors are valid in cases of heteroskedasticity, autocorrelation or both. Because of the validity of clustered standard errors in case of both heteroskedasticity and autocorrelation, clustered standard errors will be used in this research.

4.4 Conclusion

The two hypotheses of this study will be tested via panel data regressions. As a result of the Hausman test, one of the regressions that will be conducted is the random effects regression. However, in reaction to the limitations of the Hausman test, next to the random effects regression, a fixed effects regression will be conducted. Due to the results of the Wald test, the fixed effects regression will include time fixed effects. After conducting the random effects and fixed effects regressions, a pooled regression will be conducted, to evaluate the influences of the random and fixed effects. Since the panel contains heteroskedasticic error terms and autocorrelation, the three regressions will be performed using clustered standard errors.

5. Research results

In this section the research results are discussed. The outline of this section is as follows. Firstly, the regression results for model 1 are discussed. Secondly, the regression results for model 2 are

(20)

20

discussed. Thirdly, the results of different robustness checks are evaluated. Finally, this section ends with a conclusion of the regression results.

5.1 Results for model 1

This section presents the regression results for model 1. In model 1, the two hypotheses of this research are answered by looking at the variables regarding the number of casualties of an attack. Firstly, this section will provide insights into the different components of the regression output table. Secondly, the results for the first hypothesis are discussed. Thirdly, the results for the second

hypothesis are discussed. Lastly, the regression results of the control variables are discussed. To answer the two hypotheses, three different regressions are conducted. Table 1 shows the regression output for model 1. In the regression output, the values displayed next to the variable, represent the estimated coefficient of the variable. The values displayed in parentheses represent either the z-statistic or the t-statistic. The random effects regression model uses z-statistics. The fixed effects regression model and the pooled regression model use t-statistics. The stars indicate whether or not the variable is significant at the 10 percent, 5 percent or 1 percent significance level. In the regression output, the fixed effects coefficients are not included. However, two notes, Country FE and Time FE, are added, to make clear that the model includes fixed effects.

Table 1: Regression Results - Model 1

(1) (2) (3)

Variables Random Effects Fixed Effects Pooled

Lagged CC index 0.882*** 0.865*** 0.882*** (52.25) (65.33) (52.25) Casualties home -0.000415** -0.0471*** -0.000415 (-2.073) (-12.93) (-2.073) Casualties foreign -0.000264*** -0.0467*** -0.000264** (-2.863) (-12.30) (-2.863) GDP growth 0.153*** 0.0971* 0.153** (4.308) (2.648) (4.308) Inflation rate -0.0319** -0.0199 -0.0319* (-2.540) (-0.844) (-2.540) Yield spread 0.0644*** 0.0450 0.0644** (3.531) (1.107) (3.531)

Share prices growth 1.722*** 0.737 1.722***

(7.245) (1.834) (7.245)

Real house prices growth 3.181** 1.883 3.181

(1.961) (1.218) (1.961) Unemployment rate -0.0122** -0.0122 -0.0122* (-2.442) (-1.561) (-2.442) Constant 11.81*** 13.74*** 11.81*** (6.870) (9.946) (6.870) Observations 676 676 676 R-squared 0.893 0.933 0.893 Number of country 5 RE YES NO NO

(21)

21

Country FE NO YES NO

Time FE NO YES NO

Robust z-statistics (1) and t-statistics (2 and 3) in parentheses *** p<0.01, ** p<0.05, * p<0.1

To answer the first hypothesis, the question whether or not an attack in the home country influences the consumer confidence index in the home country, the coefficient of the variable ‘casualties home’ is of interest. Table 1 shows a small negative value of this coefficient for all three regression models, namely -0.000415, -0.0471 and -0.000415. When looking at the significance levels of the variable ‘casualties home’, it can be concluded that the variable is significant, at least at the five percent significance level, in the random effects and fixed effects regression models. However, the variable ‘casualties home’ is not significant in the pooled regression model. This could be due to the inaccurate standard errors of the pooled regression model, mentioned before. The results imply that, for model 1, the first null hypothesis can be rejected in the random effects and fixed effects regression models. This means that, in the random effects and fixed effects regression models, a terrorist attack in the home country significantly negatively influences the consumer confidence index of the home country. In the pooled regression model, the first null hypothesis cannot be rejected for model 1. The results of the random effects and fixed effects regression models are in line with the negative effects of terrorism in the Basque Country (Abadie & Gardeazabal, 2008). Just like the first hypothesis of this study, Abadie and Gardeazabal studied the effects of terrorist conflict in the home country. However, Abadie and Gardeazabal looked at the economic indicator GDP per capita instead of consumer confidence index. Although the studied economic indicator differs, both studies conclude that terrorism in the home country has negative effects. However, the magnitude of the negative effects differs between the two studies. Since the magnitudes of the coefficients of the variable ‘casualties home’ are small for all three regressions in this study, an increase in the number of casualties in the home country does not affect the home country’s consumer confidence index a lot. On the other hand, terrorism in the Basque Country caused per capita GDP in the Basque Country to decline about 10 percentage points, which is a relatively large effect.

To answer the second hypothesis, the question whether or not an attack in a foreign country influences the consumer confidence index in the home country, the coefficient of the variable

‘casualties foreign’ is of interest. Table 1 shows a small negative value of this coefficient for all three regression models, namely -0.000264, -0.0467 and -0.000264. Moreover, the variable ‘casualties foreign’ is significant in all three regression models, at least at the five percent significance level. This result implies that, for model 1, the second null hypothesis can be rejected in all three regression models. This means that, a terrorist attack in a foreign country significantly negatively influences the consumer confidence index in the home country. The results of this hypothesis can be compared with the results in the papers of Blomberg et al. (2004) and Arin et al. (2008). In their papers, Blomberg et

(22)

22

paper. However, instead of the effects on the consumer confidence index, Blomberg et al. investigated into the effects on economic growth and Arin et al. investigated into the effects on stock markets. All three studies showed a significant negative impact of international terrorism on the specific economic indicator. Moreover, Blomberg et al. and Arin et al. found different results for emerging markets and developed markets. Since this study only investigated into the effects of international terrorism in developed markets, these results cannot be compared with this study.

When looking at the control variables in table 1, it can be concluded that, although the values and significance levels of the coefficients differ between the three regressions, the signs of the coefficients are the same. The regression results show that the variables ‘inflation rate’ and ‘unemployment rate’ have negative coefficients, which indicates that the variables negatively influence the consumer confidence index. The variables ‘lagged consumer confidence index’, ‘GDP growth’, ‘yield spread’, ‘share prices growth’ and ‘real house prices growth’ have positive

coefficients, which indicates that the variables positively influence the consumer confidence index. These signs are in line with the signs indicated by the Dutch Central Bank (2007, June) in section 3.3. Table 1 shows that, in the random effects regression model, all control variables are significant, at least at the five percent significance level. These results are in accordance with the results of the Dutch Central Bank (2007, June) and Acemoglu and Scott (1994). However, The Dutch Central Bank did not indicate the variable ‘GDP growth’ as a decisive economic variable, which in the random effects model is a decisive economic variable. When looking at the fixed effects regression model, table 1 shows that only the control variables ‘lagged consumer confidence index’ and ‘GDP growth’ are significant. This result could be due to the fact that the fixed effects regression model includes both country fixed effects and time fixed effects. In case that the control variables are correlated with those fixed effects, the reliability of the control variables will be reduced. When looking at the pooled regression model, the only control variable that is not significant is ‘real house prices growth’. However, all other control variables, except of ‘lagged consumer confidence index’ and ‘share prices growth’, have lower significance levels than in the random effects regression model. These results could be due to the inaccurateness of the standard errors in the pooled regression model, mentioned before.

5.2 Results for model 2

This section presents the regression results for model 2. In model 2, the two hypotheses of this research are answered by looking at the dummy variables regarding terrorism. To answer the two hypotheses, three different regressions are conducted. Table 2 shows the regression output for model 2. Firstly, the results for the first hypothesis are discussed. Secondly, the results for the second hypothesis are discussed. Lastly, the regression results of the control variables are discussed.

(23)

23

Table 2: Regression Results - Model 2

(1) (2) (3)

Variables Random Effects Fixed Effects Pooled

Lagged CC index 0.882*** 0.864*** 0.882*** (50.68) (68.79) (50.68) Dummy home -0.0505* -0.0237 -0.0505 (-1.648) (-0.859) (-1.648) Dummy foreign 0.0114 0.201* 0.0114 (0.898) (2.776) (0.898) GDP growth 0.152*** 0.0906* 0.152** (4.090) (2.409) (4.090) Inflation rate -0.0303** -0.0184 -0.0303* (-2.558) (-0.761) (-2.558) Yield spread 0.0649*** 0.0526 0.0649** (3.675) (1.226) (3.675)

Share prices growth 1.696*** 0.753 1.696***

(7.058) (1.764) (7.058)

Real house prices growth 3.229* 1.909 3.229

(1.915) (1.256) (1.915) Unemployment rate -0.0113*** -0.0131 -0.0113* (-2.653) (-1.574) (-2.653) Constant 11.76*** 12.38*** 11.76*** (6.623) (9.211) (6.623) Observations 676 676 676 R-squared 0.893 0.935 0.893 Number of country 5 RE YES NO NO Country FE NO YES NO Time FE NO YES NO

Robust z-statistics (1) and t-statistics (2 and 3) in parentheses *** p<0.01, ** p<0.05, * p<0.1

To answer the first hypothesis, the question whether or not an attack in the home country influences the consumer confidence index in the home country, the coefficient of the variable ‘dummy home’ is of interest. Table 2 shows a small negative value of this coefficient for all three regression models, namely -0.0505, -0.0237 and -0.0505. However, when looking at the significance levels of this variable, it can be concluded that the variable ‘dummy home’ is only significant in the random effects regression model, at the ten percent significance level. In the fixed effects and pooled regression models, the variable ‘dummy home’ is still not significant at the ten percent significance level. This result implies that, for model 2, the first null hypothesis can only be rejected in the random effects regression model. In the fixed effects and pooled regression models, the first null hypothesis cannot be rejected. Thus, when using model 2, the results for the first hypothesis are not clear. The random effects regression model shows that a terrorist attack in the home country significantly negatively influences the consumer confidence index of the home country. However, the fixed effects and pooled regression models do not show the same result. These results differ from the results of

(24)

24

model 1. In model 1, the first null hypothesis could be rejected in both the random effects and fixed effects regression models, at least at the five percent significance level. This difference could be explained by one of the drawbacks stated by Frey et al. (2007). In model 2, the hypothesis is answered by looking at a dummy variable instead of the number of casualties. However, the dummy variable does not take account for terrorist activities of different magnitudes. Frey et al. stated, that this is only allowed when the structure of terrorist activities remains unchanged. Since, in this research, not all terrorist attacks are of the same magnitude, the results of the model including the dummy variables could be biased. Although the results of model 2 differ from the results of model 1, the results of model 2 are more in line with the results found by Garner (2002). Garner concluded that the 9/11 attacks in the United States did not influence the consumer confidence index of the United States. However, since Garner only evaluated one terrorist attack in one country, his conclusions are not completely comparable with the results of this study.

To answer the second hypothesis, the question whether or not an attack in a foreign country influences the consumer confidence index in the home country, the coefficient of the variable ‘dummy foreign’ is of interest. Table 2 shows a small positive value of this coefficient for all three regression models, namely 0.0114, 0.201 and 0.0114. However, the variable ‘dummy foreign’ is only significant in the fixed effects regression model, at the ten percent significance level. In the random effects and pooled regression models, the variable ‘dummy foreign’ is not significant. This result implies that, for model 2, the second null hypothesis can only be rejected in the fixed effects regression model. In the random effects and pooled regression models, the second null hypothesis cannot be rejected. Thus, when using model 2, the fixed effects regression model shows that a terrorist attack in a foreign country significantly positively influences the consumer confidence index of the home country. This result is against the expectations and not in line with the results of model 1. Moreover, for the random effects and pooled regression models, model 2 shows no significant effect, which is again not in line with the results of model 1. The differences in results for the second hypothesis can be explained in the same way as the differences in results for the first hypothesis. Model 2 answers the second hypothesis by looking at a dummy variable instead of the number of casualties. This dummy variable is the same whether there are few or many casualties. Since the dummy variable does not take account for terrorist activities of different magnitudes, the results of model 2 could be biased.

When looking at the control variables in model 2, the same conclusions can be drawn as for the control variables in model 1. Although the values of the coefficients differ between model 1 and model 2, the signs and significances of the coefficients do not differ between the two models. So, the results of the random effects regression model are again more or less in accordance with the results of the Dutch Central Bank (2007, June) and Acemoglu and Scott (1994). However, in the fixed effects regression model, only the control variables ‘lagged consumer confidence index’ and ‘GDP growth’ are significant. The pooled regression model shows lower significance levels than in the random effects regression model for all control variables, except of ‘lagged consumer confidence index’ and

(25)

25

‘share prices growth’. Moreover, table 1 and table 2 show that the R-squared values do not differ a lot between the two models. This result indicates that the model including the number of casualties explains the same amount of the variance in the consumer confidence index as the model including the dummy variables.

5.3 Robustness checks

This section provides several robustness checks. Robustness checks are commonly exercised in empirical studies. The goal of a robustness check is to examine how certain regression coefficients behave when the regression model is modified. Typically, the behaviour of regression coefficients is examined when certain regressors are added or removed. In case that the regression coefficients do not change a lot, it is said that the coefficients are robust (Lu & White, 2014). This study contains two different robustness checks. The first robustness check removes a random variable, the variable ‘real house prices growth’, from the model. The second robustness check removes a random country, Belgium, from the sample. The robustness checks are conducted for model 1, for the random effects and fixed effects regression models. Since section 5.1 and 5.2 showed that the limitation of model 2, the fact that it includes dummy variables instead of number of casualties, caused significantly different results, model 2 will not be taken into consideration for the robustness checks. Furthermore, since the pooled regression model is only used to evaluate the influences of the random and fixed effects, no robustness checks are conducted for this model.

The results of the robustness checks for the random effects regression model can be found in appendix G, table G.1. When comparing the regression results of the model without the variable ‘real house prices growth’ with the regression results of the original model, it can be concluded that the regression coefficients do not differ a lot. Next to small differences in values of the coefficients, there are two differences in the significance levels of the variables. Firstly, the significance level of the variable ‘casualties home’ is reduced from the five percent significance level to the ten percent significance level. Secondly, the significance level of the variable ‘inflation rate’ is increased from the five percent significance level to the one percent significance level. When comparing the regression results of the sample excluding Belgium with the regression results of the original sample, again it can be concluded that the regression coefficients do not differ a lot. Besides small differences in values of the coefficients, the only difference is a reduction in the significance level of the variable ‘casualties foreign’ from the one percent significance level to the five percent significance level. Since the regression coefficients do not differ a lot between the three different regressions, it can be concluded that the coefficients are robust in the random effects regression model.

The results of the robustness checks for the fixed effects regression model can be found in appendix G, table G.2. When comparing the regression results of the model without the variable ‘real house prices growth’ with the regression results of the original model, it can be stated that the

The effect of terrorism on consumer confidence : an analysis of five European countries