The Influence of Consumer Confidence on Housing Prices A study of the housing market in the European Union, 2007-2017

(1)

Master Thesis International Economics & Business

2018-2019.2

The Influence of Consumer Confidence on Housing Prices

A study of the housing market in the European Union, 2007-2017

University of Groningen

Faculty of Economics and Business

July 14

th

_{, 2019}

Name:

Joriël David Johannes Koops

Date of birth:

May 8

th

_{, 1991}

Student number: 1878522

Email address:

j.d.j.koops@student.rug.nl

(2)

1

Abstract

This thesis is centered around the question whether a change in annual consumer confidence has an effect on the growth rate of housing prices. To answer this question a study regarding countries within the EU, in the time period of 2007-2017, is conducted. By means of an OLS linear regression model, the final conclusion in this paper is that the relationship indeed exists, and is both significant and positive. Besides the main effect, the results show that changes in household debt and inflation too have a significant effect on the growth rate of housing prices within the given data sample.

(3)

2

1. Introduction

“We do not always understand why prices fall or increase. Economists and market specialists who follow the housing market full-time, seem to get surprised time and again” (Bezemer, 2019). The characterized inexplicability of the development in the housing market seems to be an almost fundamental truth: more often than not, we simply do not understand price developments in housing markets from one period to another as they happen. As the residential real estate market is the main source of wealth for many households (Lusardi & Mitchell, 2007), it is worrisome that we are still very much unsure about the determinants of housing prices. As perceived housing wealth of households is an important determinant of consumption levels in an economy (Carroll et al., 2006), the economic importance of understanding what influences housing prices can hardly be overstated.

Surely, there are those rare occasions where we seem to have a full grip on what causes changes in housing prices, within a given short period of time. Think for example of the period preceding and following the most recent housing boom and crash, for which e.g. subprime lending forms a comprehensive and widely accepted explanation for the housing price developments (Shiller, 2015; Rogoff & Reinhart, 2009). These success stories of accurately, and confidently explaining events in the housing markets are however rather limited and seem to be confined largely to such extreme events as bubbles and crashes. As the housing market is such an important part of virtually every economy in the world, and it has real and direct effects on both individual wealth and well-being, and the overall economy (Rahal, 2016; Lowe, 2017), it is vital that we get a better understanding of what causes developments in it. This sentiment is echoed by research that shows how developments in the housing market, including events such as booms and crashes, can have long-lasting and pronounced effects on many economies (Rogoff & Reinhart, 2009). One important economic effect that rising house prices can cause, is attributable to fact that existing home owners see a rise in their asset value, while their liabilities remain unchanged. As a result of a shifting balance between the value of households’ assets and liabilities, an effect called the ‘wealth effect’ appears: people will feel as if they have earned money and become richer by simply owning their property (Holt, 2009). This additional ‘income’ households perceive to have gained, makes them more likely to increase their spending on a number of economic activities, including consumption, and can drive up housing prices (Lambertini et al., 2013; An de Meulen, et al., 2014; Muellbauer & Murphy, 2008; Cooper & Dynan, 2016). On the other hand and by the same logic, the wealth effect can have negative consequences too: if housing prices decrease, households will see a drop in their asset value, increasing the relative and perceived height of the value of their liabilities. This situation will make households feel less wealthy, and they will be more likely to decrease a number of economic activities, such as consumption (An de Meulen, et al, 2014).

(6)

5

been a reliably predictor of financial and banking crises (Rogoff & Reinhart, 2009; Leamer, 2008; Bolt et al., 2019; Zhang, 2019). This impact has the potential to have further reaching consequences than a single economy. The housing market has the potential to cause a crisis spreading globally, both in developed and developing countries, as happened after the 2008 crash (Reinhart & Reinhart, 2018). It is especially because of these potential far reaching and severe consequences, that research contributing to the overall understanding of the housing market is so important.

This research therefore aims to contribute to the body of knowledge regarding housing prices, by taking a non-fundamentals approach. Most studies looking at the housing market have done so from a rather mechanical, fundamental approach, based on the assumption that economic actors behave rationally given the economic circumstances (An de Meulen et al., 2014; Agnello et al., 2015). This general economic principle is underlined by Bovi (2009), who states that ‘[t]he standard economic literature merely assumes that the representative agent is an unemotional computer’. As these standard economic studies have not been able to arrive at a comprehensive explanation about what actually determines housing price developments, this research will take a different approach. The approach in this research will be more in line with Bovi’s (2009) conceptions that ‘economics is a behavioral science and people expectations play a pivotal role in it’, and Lambertini et al.’s (2013, 2017) conclusions that households expectations about future economic events are very important in determining future economic developments, including the development of housing prices. The potential consequences of these expectation include events such as booms and bubbles, illustrated by e.g. Holt (2009) and Shiller (2015) stating that the housing bubble causing the 2008-2009 financial crisis would not have occurred without the widespread belief that home prices would continue to rise.

In recent years, and especially after the most recent financial crisis, the theory that fundamentals are not sufficient to explain the price behavior in the housing market has gained in popularity. The realization that fundamentals cannot reasonably account in full for the large price swings in home prices has steadily gaining momentum in recent years (Shiller, 2015; Glaeser & Nathanson, 2015; Bolt et al., 2019). One of the most prominent non-fundamentals that is studied as an alternative to fundamentals, is related to consumers’ expectations and consumer confidence. Studies by Case et al. (2012) and Cheng et al. (2014) have for example found that unrealistic expectations about future developments, was one of the main causes of the housing bubble. Another prominent researcher, and winner of the Nobel Prize, Shiller has identified the potential (harmful) effects of consumer confidence and overconfidence.

(7)

6

high level of confidence, the more they are willing to increase their spending in the present. On the other hand, pessimistic consumers are theoretically likely to save more and delay their spending until a later moment in time (Gelper et al., 2007; Reinhart & Rogoff, 2009). The authors stress based on this, that consumer confidence can have significant economic consequences. The relative recent attention paid to psychological factors within general economic models, is preceded by studies as e.g. behavioral economics, marketing and psychology, where effects of subjective indicators – like consumer confidence is – have previously been proven to have relevance in many every day and experimental settings.

The main goal of this research will be to make a contribution to this growing body of literature regarding the role that non-fundamentals and its impact on actual economic developments. In this paper, the relevant economic factor that will be studied regards the developments of the housing market. In line with research by Bolt et al. (2019), Shiller (2015), Gelper et al. (2007) and Kilic & Cankava (2016), the non-fundamental used to study this developed is consumer confidence. Rather than repeating studies looking at the long-term impact of consumer confidence on economic variables, this research will focus on what the impact of the change in consumer confidence is on the growth rate of housing prices.1 The reason for this is two-fold: first, it is a relatively novel approach and literature describing this sort of analysis in this setting is scarce. This enhances the possibility of contributing new insights to the existing literature. Secondly, Shiller (2015) has shown that excessive increases in consumer confidence can lead to an overheated housing market, with possible very large effects on the economy as a whole. The growth rate is therefore a very relevant aspect of consumer confidence, although Shiller has not studied the impact of the specific change in growth rate, as is intended in this research. This research goal culminates in the first and main research question within this study:

1. Does the change in consumer confidence affect the growth rate of housing prices, and to what

extend?

Despite the main focus in this research on the effect of a non-fundamental measure on housing prices, there is entire body of research in which has studied the impact of more traditional, fundamental, economic variables. For this reason, fundamentals will be included in this research as well, and their effects on housing price growth will be tested.

A first fundamental which is theorized to have an effect on housing prices is household debt. For many households, their house is the most valuable asset they own (Lowe, 2017). At the same time, for many households mortgage loans constitute the largest portion of what they owe. Rising household debts are therefore a good indicator of an increased spending on housing, which due to demand effects increases the prices of the housing market (Dombret & Goldback, 2017). Based on this, the following controlling research question can be formulated:

2. Does a change in household debt levels have a positive effect on housing prices growth?

1_{Growth is here meant both in a positive and a negative sense, it can therefore also refer to a negative value}

(8)

7

Another fundamental described in the existing literature as having an effect on housing prices, is disposable income or purchasing power. Research hereby consistently shows that increases in income, lead to increasing housing prices (Liu, 2019; Abelson et al., 2005; Gallin, 2006; Maza & Pages, 2007). Based on this, the research questions with regard to purchasing power within this research is given as:

3. Does a change in purchasing power have a positive effect on housing price growth?

Finally, Sun & Tsang (2019) conclude that changes in interest rates can influence housing prices, whereby lower interest rates lead to higher housing prices due to the possibility of lower cost financing. Moreover, McGurk (2019) concludes that inflation levels too play a role in the price increases of houses. This is in line with the general principle of inflation, for example as given by Eurostat (2019) stating that inflation reflects a general price level increase. Based on these two fundamental effects, the final controlling research question is formulated as:

4. Do interest and inflation have an effect on the development of housing prices?

To summarize the introduction; the research described in this paper aims to contribute to the existing literature regarding the role that non-fundamentals, or ‘soft’ factors, have on real economic developments. More specifically, it aims to contribute to the scientific knowledge regarding consumer confidence and developments in the housing market, by assessing the role of a change in consumer confidence as an indicator for the growth rate of housing prices. Although current literature exists which discusses the role of consumer confidence on a number of economic outcomes, primarily consumption, the impact of the degree of change of consumer confidence on these variables remains unclear. This research aims therefore specifically on making a contribution to filling the existing gap in the scientific literature.

To study the effect of consumer confidence changes on housing price growth, a linear OLS regression analysis is conducted in this research, on a data sample that consists of data about the European Union from 2007 through 2017. Five controlling variables are introduced to test whether potentially found effects, regarding the main research question, might be better explained by these external factors.

(9)

8

2. Literature Review

This chapter reviews the available scientific literature regarding the research questions posed in the introduction chapter, specifically in relationship to the factors influencing price developments in the housing market.

As explained in the introduction as well, a lot of research has been conducted studying the housing market in a variety of settings. There is little ambiguity within the scientific literature about the importance of the housing market, for both individual wealth and the overall economy. By virtually all accounts, it is considered to be one of the most important economic sectors, with very real and possible severe, and long-lasting effects on economies as a whole (Lowe, 2017; Reinhart & Rogoff, 2009). One of the reasons the housing market is so important to the economy as a whole, is that price movements within this sector have very real consequences for the perceived wealth of households. Perceived wealth in turn influences a range of economic decisions by households, such as increased spending on for example consumption when their perceived wealth increases, and increasing their saving when they perceive their wealth has diminished (Muellbauer & Murphy, 2008; Lambertini et al., 2013). The potential effects resulting from such economic decisions are very large, and can snowball into a number of economic events, even as extreme as prolonged recession and financial crises (Reinhart & Rogoff, 2009; Shiller, 2015; Leamer, 2008). In fact, the 2008-2009 financial crisis was, among other things, the result of a collective overestimation of the future wealth expected from the housing market, both by individuals and by lenders (De Haan et al., 2015).

2.1 Consumer Confidence

(10)

9

2014). Most important to this decision making based on these narratives, for households that is, are the future asset price expectations of assets. Research finds that positive expectations about future price developments show a strong relationship with actual future, long-term price developments (Lambertini et al., 2013). These authors continue their research with a broader study on the effect of household expectations on future developments on a number of sectors of the economy. They found that these future expectations have a direct effect on the development of a number of economic developments, such as GDP, consumption, investment and real wages (Lambertini et al., 2017). Another research looking at the broader impact of non-fundamentals on the actual economy, is one conducted by Gelper et al. (2007). In this study, the effects of consumer sentiment on a broad range of future economic activity was studied, pre-crisis. The authors find that consumer sentiment has predictive power with regard to future economic activity, most specifically consumption: an increasing level of consumer confidence has a strong and significantly positive effect on the amount of consumption in an economy. This consumption effect is reproduced in the literature more broadly, for example by Dees & Brinca (2013), in a study concerning the euro area specifically. These consumption effects of consumer confidence are most pronounced in the short-term, but tend to have an impact in the longer term as well, although these effects decrease over time (Segers et al., 2017; Dees & Brinca, 2013; Gelper et al., 2007; Baur & Heaney, 2017). Segers et al. (2017) continues on this broad perspective with regard to consumer confidence, its relevance and its economic implications. In their research, they describe consumer confidence as ‘a useful variable to measure the current state of the economy as well as to forecast its future states at reasonably short horizons.’

Kilic & Cankava (2016), who state that consumer confidence effectively captures the effects of uncertainty and other psychological factors, conclude too that consumer confidence has explanatory power on specific forms of economic activity. Among the most pronounced effects of consumer confidence, they find, are manufacturing-related variables and housing market variables. They conclude that higher levels of consumer confidence lead to higher levels of economic activity, and potentially higher housing prices. Clayton et al. (2009) come to a similar conclusion with regard to the housing prices, stating that ‘private real estate markets [are] highly susceptible to sentiment-induced mispricing.’ Investor sentiment has therefore a role in the determination of asset prices, independent of underlying market fundamentals (Shiller, 2015; Baker & Wurgler, 2007). These conclusions that households’ future expectations are related to house price developments have been studied in case studies as well. Research by Rouwendal & Longhi (2008) found for example that consumer confidence is correlated with housing prices in the Netherlands, and Rapach & Strauss (2007) concluded that consumer confidence seems to influence housing prices in the US.

(11)

10

This academic consensus about the role of increasing consumer confidence, on economic outcomes such as housing prices, leads to the following hypothesis in this research:

H1: Increases in consumer confidence have a positive effect on housing price growth.

2.2 Household Debt

Besides the recent trend to look at non-fundamentals and its influence on price developments in the housing market, there is an entire body of research looking at other, fundamental, explanations for changes in housing prices. For a substantial, and majority, amount of economic researches this is still the main mode of looking at price developments.

Anderson et al. (2016) for example identify that the increases in household debt leading up to the financial crisis, led to excessive growth in housing prices. They identify this growth in household debt pre-crisis as well as one of the reasons why the recession in the aftermath of the crisis was so severe. However, they do explain how in non-crisis time an increase in household debt spurs housing prices: when housing prices rise, the value of a home as collateral for an existing loan increases, thereby increasing the borrower’s potential total loan size. Higher collateral makes the borrower a better client to the bank, and therefore more qualified to take on additional loans. As long as household debt levels can be serviced, an increase will lead to an increasing demand for a number of economic goods, including housing, thereby increasing their prices (Dombret & Goldback, 2017). The same effect is found by Cecchetti et al. (2011), who however warns that this is valid until a certain level of leverage. The positive relationship put forward in the literature regarding the relationship between increasing household debt levels and housing prices leads to the following hypothesis to be tested in this research:

H2: Increasing household debt contributes to higher housing price growth.

2.3 Housing Cost Overburden

(12)

11

H3: Increasing housing cost overburden rates contributes to lower housing price growth.

2.4 Disposable Income

Another variable identified in the scientific literature as having a potential effect on housing prices, is the previous ratio underlying variable, disposable income. This is defined by the Cambridge Dictionary (2019) as ‘the money that you can spend as you want, and not the money that you spend on taxes, food and other basic needs’.

Many empirical studies have researched the relationship between this disposable income and housing prices. Although different authors have studied a variety of geographical areas,

timeframes and economic systems, the findings from these studies are relatively similar in their conclusions: there is a long-run equilibrium relationship between the disposable income of individuals and their expenses on housing (Black et al., 2006). What is noticeable within the literature, is over and over the result is found that rising incomes lead to rising prices. As people have more money to spend, more of them will qualify income wise for the purchase of a home. As the pool of potential homebuyers’ increases, an increase in demand for the existing houses occurs, thereby increasing the prices of these houses. Furthermore, because people generally have more money available, it also means that there is an increase in the proportion of the population that can pay more for a house. Some authors go based on this as far as to describe the relationship between income and house prices as ‘co-integrated’ (Gregoriou, 2013; Malpezzi, 1999; Meen, 2002). Andrews (2010) reinforces this conclusion, and provides evidence in his research that the relationship between income and housing prices is almost fully elastic. Based on these findings, the income effect on housing prices will be tested by means of the following hypothesis:

H4: Increasing disposable income has a positive effect on the price development of housing.

2.5 Interest Rates

(13)

12

Hintermaier & Koeniger (2018) supports this claim, but they warn that low interest rates provide a breeding ground for confidence-driven crises.

The above described effect of interest rates on housing prices leads to the following hypothesis:

H5: Increasing interest rates have a negative effect on the price development of housing.

2.6 Inflation

A final economic fundamental assumed to increase prices is inflation. Inflation is best defined as ‘a broad increase in the prices of goods and services, not just of individual items’ (European Central Bank, 2019). This means that when there is inflation in an economy, the value of a currency loses value over time; the higher the rate of inflation, the faster the currency loses value. Per definition, when there is inflation in an economy over time, a single unit of currency is capable of purchasing less at any moment in the future than it is today. Due to the deteriorating value of currency over time, people will start to demand more of the same currency when selling their goods, services and houses. This notion of inflation is such a given, that it is barely discussed or defined in the literature. Because it is such a core economic concept, it is included in this study however.

The effects described that increasing inflation leads to upward pressure on housing prices, is captured in the following hypothesis:

(14)

13

3. Data and Methods

3.1 Data Collection and Sources

In order to conduct a proper, thorough analysis of any given quantitative phenomenon, it is important to make use of consistent, high quality data. For this purpose, a plurality of data sources have been considered as input for the analyses in this research, many of which meet an independency threshold and report on one or more of the considered variables. These include organizations such as the United Nations Statistics Division, the OECD and The Worldbank. There were however two identifiable issues casting doubt on the data quality, with regard to the before named sources: (1) they did not, or not comprehensively, report on all of the necessary variables and (2) for certain variables it was unclear what their data was actually based upon. Especially the former posed a threat to the consistency and comparability of the data, as the direct consequence was that a comparison of data from different sources would be necessary, featuring possible differing measuring methods and/or standards.

As a consequence of the above mentioned considerations, to ensure data consistency and quality, the choice was made to make exclusive use of data from Eurostat as input for the analyses conducted in this research. Eurostat not only meets an independency threshold, it also reports consistently on all of the studied variables, for all of the EU-28.2 Moreover, it serves as the official statistical office of the European Union, ensuring that their data on European statistics are consistent, reliable and of high quality.3 The fact that Eurostat reports extensively on all of the studied variables, eliminates the potential risk of having to compare cross-source analyses, thereby reducing the risk of noise in both the data and the output of the final analyses.

Eurostat makes data available about a broad range of economic, demographic, geographic and other topics. This data has a spatially comprehensive scope, a broad time horizon, usually features many subcategories and covers a plurality of measuring methods; this often includes reporting the same data in absolute, indexed and relative ways. In this research has been chosen to collect, where available, the data in such form, that the fewest manual manipulations of the data are necessary, to reduce the risk of human error in the process. Therefore where available, data in the form of yearly or quarterly percentual changes have been collected, which lend themselves best for a comparison of yearly growth rate changes. When this data was not available, the data has been collected in absolute terms, because this form of data is most suitable to manually establish year-on-year change percentages. This outcome closely approximates the other, percentual, form of data used and therefore ensures data consistency across variables, eliminating interpretation or any other subjective form of influence on the data.

Another virtue about Eurostat is that it does not guess, nor estimate, missing datapoints in the past. It only reports values for those periods it has actual, factual input to back up their claims. A slight downside to this approach is that it leads to certain missing datapoints in the final dataset,

2_{Except Greece, which does not provide housing statistics.}

3_{Eurostat’s official mission includes this principle as well: ‘To provide high quality statistics for Europe.’}

(15)

14

as not all countries have always reported on all variables in all periods. These missing datapoints are however relatively limited, and considered acceptable – perhaps even desirable – to ensure factual, high quality data.

The data collected from Eurostat encompasses 27 of the current 28 countries within the European Union; Greece has been left out, as it does not report on the development of housing prices in its territory. It furthermore covers the period from 2006 through 2017 consistently for most of the variables, a period that has been chosen due to the availability of the most important data: most countries report from 2007 yearly on the changes in house prices, and at the moment of writing, not all countries had reported on their 2018 figures.4 The collected data, according to the above mentioned criteria, entails changes in the following variables: house prices, consumer confidence, housing cost overburden, household debt, purchasing power, inflation and interest. As stated, all of the collected data is collected either in percentual or absolute form, which is then used in the analyses as a numerical, continuous form of data.

3.2 Specification of the final data sample

To study the effects of consumer confidence on housing prices, a final sample of data consisting of 27 out of 28 European Union member states (excluding Greece), spanning a time period from 2006 through 2017, has been selected. For each mentioned country and time period, data on a number of variables is collected, whereby the variables are in line with the hypotheses from chapter 2; for each hypothesis, there is one corresponding independent variable.

1. Housing prices changes serves in this research as the dependent variable, which is theorized to be determined by a number of independent variables. Housing price as a variable within the context of this research defined by the definition of Eurostat (2019): ‘the price changes of residential properties purchased by households, both newly-built and existing ones, independently of their final use and independently of their previous owners.’ The variable therefore encompasses the entirety of the price changes in the residential housing market. The data on housing prices has been retrieved from Eurostat, and is on a quarterly basis describing the year-on-year price changes. For example, the 2008-Q1 data describes the price change of the first quarter of 2008 compared to the prices of the first quarter of 2007. The data on housing prices is numerical and continuous, as price changes can theoretically take on any positive or negative value. 2. Consumer confidence is the main independent variable within this research and is

theorized to be an important determinant of the changes in housing prices. Consumer confidence itself is best defined as: the amount of confidence consumers feel about their current and future economic and financial situation (Gelper et al., 2007). The variable is based on monthly EU survey data and reports consequently on monthly consumer confidence statistics for the entire EU-28, and takes the form of a positive number when consumers answer on average more positive than negative, and takes on a negative

4_{The discrepancy between 2006 and 2007 is due to the fact that 2006 data is necessary to calculate year-on-year}

(16)

15

number when consumers answer on average more negative than positive. This data, too, takes on the form of numerical, continuous data.

3. Housing cost overburden is an independent control variable, theorized to have a potential negative effect on the growth of housing prices. Housing cost overburden is as a variable is best defined by Eurostat (2019) as: ‘the percentage of the population living in a household where total housing costs represent more than 40% of the total disposable household income’. It is therefore the part of the population that spends such a significant amount of their income on housing needs, that the burden of this expenditure is considered excessive and unsustainable (European Index of Housing Exclusion, 2017). This variable is numerical and continuous.

4. Household debt is an independent variable, which is theorized to have a positive effect on the growth of housing prices (Dombret & Goldback, 2017). The definition of household debt used in this research is same as given by Eurostat (2019), namely: ‘The Household debt is the stock of liabilities held by the sector Households and Non-Profit institutions serving households.’ In the data this is expressed as a percentage of GDP, in numerical and continuous form.

5. Purchasing power is also an independent variable, which reflects the amount of relative spending power a household holds. The Cambridge Dictionary (2019) defines it as ‘the value of money considered as the amount of goods it can buy’. Eurostat uses a specific measure of purchasing power in its data with regard to this topic, namely the Purchasing Power Parity (PPP). This PPP is a measure of an artificial currency unit in euros, that compares the real final expenditures of the countries in the EU with one another and has the same purchasing power throughout the entire EU. It is a weighted average of the national currencies of the EU Member States and reflects the weighted average price levels in Member States (OECD, 2007). This data is numerical and continuous in form. 6. Inflation is used as another independent control variable, and is described as: ‘a broad

increase in the prices of goods and services, not just of individual items. As a result, you can buy less for €1’ (European Central Bank, 2019). Inflation is therefore best defined as: the loss in general value of a currency. Inflation is reported in the form of a percentage per year and comes in the form of numerical, continuous data.

7. Interest is the final of the independent control variables used in this research. It is defined by the Oxford Dictionary as ‘money paid regularly at a particular rate for the use of money lent, or for delaying the repayment of a debt.’ It is therefore considered as the price of money. Interest data is given in the form of a monthly average percentage based on the official central bank rate in that given period; in this case the 3-month rate has been used to represent the interest value.5

As stated, all of the data has been collected in numerical, continuous form. For this purpose, the data is collected exclusively in the form of percentual changes, or in the form of absolute numbers, based upon which percentual changes can be calculated. All data has been gathered in

5_{The 3-month rate was most readily available for all of the countries for the given period. For analysis purposes the}

(17)

16

annual form where available, and where not available – for example with consumer confidence and housing prices – the average year-on-year growth rate per month or quarter has been collected, in order to enable calculating the yearly growth rate. As stated previously, all the used data is collected exclusively from Eurostat, from its official database portal.

3.3 Data transformation

To test the extend of the relationship between consumer confidence and housing prices, while controlling for a number of variables, at first a number of data transformations had to be conducted, to ensure that the data would be showing the same year-on-year differences, and be suitable for comparison. In order to arrive at average year-on-year price changes for housing prices, the average has been calculated of the four quarters of each year, as given by Eurostat. Since the objective of this study is to test the effect of the changes in house price developments, rather than the actual percentual developments, for each year the year-on-year growth change has been calculated. As the data is quarterly data, the data is for every year transformed to an average of the four yearly quarters, by means of:

4

( ∑ Housing Price Growth q ) / 4 q=1

This is then followed by the simple manipulation of: Housing Price Growth t – Housing Price Growth t-1 for every year of t, which gives the growth of the housing price developments in percentage points on an annual basis. This is a numerical, continuous variable.

Consumer confidence is a measure that is calculated by means of surveys, conducted by Eurostat throughout all of the EU-28 countries on a monthly basis. In this surveys consumers are asked a number of questions, in their native language, about their current and expected future economic and financial situation. Questions include: ‘How do you expect the financial position of your household to change over the next 12 months?’ All answers are recorded on a one to five scale. Based on the answers given in a certain month by consumers in a certain country, the final consumer confidence score is calculated. The consumer confidence score of a country is the average of the responses given, whereby the neutral answer (‘Stay the same’) has a value of zero, positive answers are given a positive score and negative answers are given a negative score. The more positive consumers’ answers, the more positive the country’s score is, and vice versa. A positive number in a given period for a country means that people were on average positive, and the more positive the number is, the more positive their responses were. The first manipulation necessary on the data is the transformation from monthly to yearly data, by means of:

12

( ∑ Consumer Confidence m ) / 12 m=1

This gives the average annual level of consumer confidence, after which a simple manipulation has been conducted here too in order to obtain the year-on-year change in consumer confidence:

(18)

17

With regard to the control variables, similar manipulations of data have been conducted. Housing Cost Overburden numbers by Eurostat are given as a percentage of the population. In order to deduct the yearly growth numbers, the following manipulation has been conducted: Housing

Cost Overburden t – Housing Cost Overburden t-1. For the variable Purchasing Power the

Eurostat data is given in an absolute number, which means that certain countries have higher Purchasing Power than others. To make this data usable for comparison, the growth rate for each country has been calculated year-on-year by means of: (Purchasing Power t – Purchasing Power t-1 ) / Purchasing Power t-1. This outcome gives the percentual change per year, which makes it

suitable for comparison. The numbers on Household Debt are given by Eurostat on a yearly basis, as a percentage of national GDP. In order to extract the yearly growth data in percentage points, the following simple data transformation has been conducted: Household Debt t –

Household Debt t-1. For the interest and inflation variables, the same year-on-year calculations

have been executed, in order to extract the year-on-year changes. All of the control variables encompass numerical, continuous variables.

3.4 Method and Empirical Model

Due to the data transformations, as described in section 3.3, the dataset now takes on the form of a year-on-year change for every variable. As every datapoint is given exclusively in this form, all of the direct over-time developments have been excluded from the final dataset.6

The causal model chosen in this research as best suited to estimate the effects of this confined type of data, is the Ordinary Least Squares (OLS) linear regression model, as the relationship between the variables in the current data form can only be estimated by means of linearity. The OLS-regression model test statistically for the relationship between one or more independent variables on a specific dependent variable (housing prices changes), and requires data to be in a numerical, continuous form. A requirement for the data in a linear regression is that it needs to be multivariate normal, that means that it takes the data follows a normal distribution, and is not random and/or skewed. This requirement will be tested by means of plotting normal distribution histograms of the variables. Furthermore, the data should not include multicollinearity and have a linear relationship between the dependent and independent variables. To test this, the correlation between the main independent variable, Δ consumer confidence, and the control independent variables will be assessed. In case a significant correlation is found, the model will be re-estimated without these variables, to test whether this has an effect on the found relationships with the dependent variable. A final test of the data will assess whether the relationship between dependent and independent variables is in fact linear, and not for example exponential. The testing of these criteria is part of the robustness checks that are conducted in chapter 4.

The first, and main, hypothesis to be tested in this paper concerns the influence of the change in consumer confidence on Δ housing price growth. The expectation is that there is a linear,

6_{However, these can be estimated based on the data, by giving the 2007 datapoints an index-value of 100, and}

(19)

18

positive relationship between consumer confidence changes and housing price changes. This hypothesis is expressed in the following linear bivariate regression line:

dY = c + dCx1 + e (1)

Where dY represents the change in housing price growth, and dC represents the change in consumer confidence.7 Furthermore, c represents a constant.

Next to the main effect, a number of control variables have been introduced in this paper, and of these variables the potential moderating effects will be tested, to ensure that potential results are not innately better explained by the controlling variables than the independent variable of consumer confidence. These control tests are grouped in two sorts of control variables, before being tested in a wholesome model testing the influence of all the control variables simultaneously. The first group of tested variables as controlling for the influence of consumer confidence, concerns the household specific variables, namely Household Debt, Housing Cost Overburden and Purchasing Power. In line with the hypotheses, the expression is given as:

dY = c + dCx1 + dDx2 – dHx3 + dPx4 + e (2) Whereby dD represents the year-on-year change in household debt levels, dH represents the annual change in housing cost overburden levels, and dP gives the annual change in purchasing power for households.

Next to this, the second group of controlling variables tested for its influence on the relationship between a change in consumer confidence levels and household debt levels, concern the traditional economic fundamentals of interest and inflation. The expected expression for this relationship is the following:

dY = c + dCx1 – dIx5 + dGx6 + e (3)

Whereby dI gives the annual change in interest rate levels and dG represents the annual change in inflationary levels.

These tests for the controlling influences of the, potential, controlling variables on the relationship between the change in consumer confidence and the change in housing prices, is then expressed in the linear multivariate regression model of:

dY = c + dCx1 + dDx2 – dHx3 + dPx4 – dIx5 + dCx6 + e (4) In the later regression, the effect of all controlling variables is tested on the causal relationship between consumer confidence and housing prices. Only if a statistically significant relationship is found, given the control variables, is it possible to conclude positively that there is in fact a causal relationship between these main variables. Another possible result however is that the analysis shows that one or multiple of the control variables provide better, more significant, explanations for the found variance in the growth of housing prices.

(20)

19

4. Empirical Results

4.1 Data description

The collected data, as described in chapter 3, has for each variable an upper limit of 27 (countries) * 11 (years) = 297 datapoints. The dependent and main variable, Δ housing price growth, has a total of 266 datapoints in the data within a range of -54.4 and 43.3 and a standard deviation of 9.686. The fact that the variable does not have all possible 297 datapoints, is attributable to a number of countries not reporting on housing prices in the first year(s) of the studied time period. The main independent variable, consumer confidence change, has data available for each period and country in the studied time period. The range of data and standard deviation is smaller than for the dependent variable, with a range of -27.8 to 26.2 and a standard deviation of 7.182.

Of the control variables Δ housing cost overburden rate and Δ household debt both have a full dataset of 297 observations. The ranges and standard deviation of these variables are smaller than those of the two main variables, as described above. The three remaining control variables – Δ interest change, Δ purchasing power change and Δ inflation change – have a number of missing values, respectively 18, 27 and 22. The amount of missing values is relatively small, and with 266+ datapoints for each variable in the dataset, the data consists of sufficient datapoints for to enable reliable testing of the hypotheses. The range and standard deviation is relatively small compared to the other variables.

The summary statistics for the data sample used in this research are given in table 1 below. These statistics show e.g. the amount of datapoints, mean and standard deviation for each variable. One feat of the data that stands out, is that the spread between the minimum and maximum value of Δ Housing Prices Growth is substantially larger than the spread of the other variables.

Variable Obs. Mean Std. Dev. Min Max

Δ Housing Prices Growth 266 -.50338 9.68535 -54.4 43.4

Δ Consumer Confidence 297 .34882 7.18215 -27.8 26.2

Δ Housing Cost Overburden 297 -.13603 1.95696 -12.2 7.9

Δ Household Debt 297 .22256 3.48329 -24.4 11.8

Δ Purchasing Power 281 2.31495 3.83361 -14.6 19.8

Δ Interest 275 -.31382 1.18023 -4.9 5.0

Δ Inflation 270 -.15444 2.16030 -12.0 5.4

(21)

20

4.2 Results and model estimation

Multiple regression analyses have been conducted in order to test for hypotheses H1 through H6. These analyses are conducted both in a bivariate regression – that is, without considering the impact of multiple independent variables simultaneously – and in a multiple regression analysis to estimate the model, thereby considering the different independent variables simultaneously. At first the impact of all the independent variables separately is analyzed, after which the outcome of the entire model is studied.

4.2.1 Bivariate regression analysis

At first for every independent variable identified in the research – Δ consumer confidence, Δ housing cost overburden, Δ purchasing power, Δ household debt, Δ inflation and Δ interest – a linear regression is executed. These bivariate regression analysis show whether the variables have a direct, significant effect on the dependent variable, and whether the hypotheses based on the literature hold true in a one-on-one regression.

The main hypothesis in this research relates to the effect on Δ consumer confidence on Δ housing price growth. The results of the bivariate analysis with regard to these two main variables, are shown in table 2 and illustrated by figure 1. The results show that in bivariate regression, not considering any control variables, the effects of Δ consumer confidence on Δ housing price growth are strongly significant and positive. This shows that when Δ consumer confidence is positive, Δ housing price growth rates tend to increase. The coefficient of this relationship gives a rounded value of .74479, meaning that a 1-point increase in Δ consumer confidence, leads to roughly three quarters of one percentage point in additional house price growth. The explanatory power of this relationship is given by a R-squared value of .2909, which shows that consumer confidence explains a substantial amount of

the variance in house price changes.

Variable _{Δ Consumer} Confidence Δ Housing Price Growth .74479*** (.07156) Constant -1.13870*** (.50470) Observations 266 R-squared .2909

* Significant at the level of p < .10. ** Significant at the level of p < .05. *** Significant at the level of p < .01 Table 2: bivariate regression of Δ consumer confidence on Δ housing price growth

(22)

21

Furthermore, the same form of bivariate regression analyses has been conducted with regard to the control variables, of which the results are shown below in table 3. These results show that all five variables independently of one another have a statistically significant relationship with the dependent variable change in housing prices, with a p-value of less than .05. Of these, only housing cost overburden and interest do not have a statistically significant influence on the level of p < .01 on the change in housing prices. Δ housing cost overburden shows a positive and significant effect on Δ housing price growth, with however a low R-squared of only .0209. For Δ purchasing power, the results show that an increase in purchasing power has a positive and significant (p < .01) effect on the change in housing prices as well. The direct explanatory power of this variable is given by R-squared as .0367. Δ household debt is found, in a one-on-one linear regression, to have a significant (p < .01) and negative effect on the changes in housing prices. That is to say, when household debt growth increases, housing prices growth tends to become negative. The explanatory of this variable, with an R-squared of .1342, is somewhat higher than that of purchasing power and housing cost overburden. Furthermore, both the control variables Δ inflation and Δ interest show a statistically significant and positive effect on the Δ housing price growth, with respectively p < .01 and p < .05. This result shows that both an increasing level of inflation and an increasing level of interest rates have a positive effect on the Δ housing price growth. The explanatory power of Δ interest is however rather limited, with a R-squared value of only .0159, while Δ inflation has a R-squared of .0663.

Variables (1) Δ Housing Cost Overburden (2) Δ Purchasing Power (3) Δ Household Debt (4) Δ Inflation Δ Interest (5) Δ Housing Price Growth .70657** (.29750) .52370*** (.17035) -1.00023*** (.15640) 1.13942*** (.27162) 1.04802** (.52720) Constant -.41756 (.58982) -1.66370** (.72286) -.52971 (.55364) -.38273 (.57828) .00054 (.57302) Observations 266 250 266 250 246 R-squared .0209 .0367 .1342 .0663 .0159

Table 3: Bivariate regression of the control variables on the dependent

* Significant at the level of p < .10 ** Significant at the level of p < .0 *** Significant at the level of p < .01

4.2.2 Initial multiple regression analysis

(23)

22

purchasing power and Δ interest now lose their statistically significantly impact on Δ housing price growth, given the influence of the other independent variables.

Variables Δ Housing Price Growth Standard Error t-value

Δ Consumer Confidence .53583*** 0.07987 6.71

Δ Housing Cost Overburden .33691 0.22887 1.47

Δ Purchasing Power .02400 0.16619 .14 Δ Household Debt -.47683*** 0.14852 -3.21 Δ Inflation 1.49212*** .62554 4.67 Δ Interest -.87832 .62554 -1.40 Constant -1.09754* .63739 -1.72 Observations 215 R-squared .3506 Prob > F .000

Table 4: Multiple regression analysis with all variables included * Significant at the level of p < .10 ** Significant at the level of p < .05. *** Significant at the level of p < .01

4.2.3 Outliers and omission

After the execution of the full regression analysis, as given above in table 4 with all the independent variables included, the residuals of this outcome have been analyzed. This is important, as it enables testing for outliers in the data. In order to get an initial insight of this, the scatterplot in figure 2 was constructed, which shows the distribution of the residuals. Based on a first glance observation, it appears that there are a couple of observations deviating relatively far from the regression line of Δ consumer confidence on Δ housing price growth.

However, based on a scatterplot alone it is hard to quantify whether and which observations might be considered outliers, and it serves more to get an impression of the data. To actually analyze which values might be considered outliers, a boxplot is graphed. As shown in figure 3, a number of observations are considered outliers based on the statistical rule that observations 1.5 times the interquartile range (lower quartile - 1.5 IQR, upper quartile + 1.5 IQR) can be considered as such. This method of determining outliers, gives 9 potential outliers in the data, 2 on the higher end of the observations and 7 on the lower end. When taking a closer look at these outliers, it turns out that two thirds of these outliers concern observations of the crisis years (2008-2009) or the direct aftermath. The observations identified as outliers, by means of the boxplot, are given in table 5.

(24)

23 Year Country 2008 Estonia 2008 Poland 2009 Bulgaria 2009 Slovakia 2010 Estonia 2010 Hungary 2013 Ireland 2015 Ireland 2017 Hungary

Table 5: Observations considered outliers based on the boxplot

As the in table 5 mentioned values fall outside of the 1.5 IQR, they may be considered as outliers and be omitted. These outlier values, in this mainly rare crisis values with extreme values for Δ house price growth, potentially influence the outcomes of the analysis to such a degree that they can been excluded from the data.

4.2.4 Regression diagnostics

In this research, the OLS linear regression model is used to test the causal relationship between variables. An OLS regression model, and the corresponding input data, however needs to meet certain criteria, in order for the results to be considered reliable. For this purpose, a number of relevant regression diagnostics are conducted to test whether the input meets the requirements for the conducting of a linear regression.

4.2.4.1 The relationships between the independent and dependent variables are linear The first necessity of a linear regression, is that the relationship between the dependent and independent variables is in fact linear and not for example exponential. The relationship of the independent variables with the dependent variable have therefore been plotted on a scatterplot, in order to test whether there are clear signs that a relationship is non-linear. In case one of the independent variables has a non-linear relationship with the dependent variable, one would expect to see a curving pattern of datapoints. As the output in figure 4 shows, non-such curvature appears to be present in the data. Although precise linear relationships are hard to assess based on the scatterplot, it does become clear that there are not clear curved patters visible. The input data for the regression data therefore meets the requirement that the relationship between independent and dependent variable is linear.

(25)

24 4.2.4.2 Normality of the variables

In order to test for the normality of variables, scatterplots with an overlaying normal

distribution curve have been constructed to gain insight in the distribution of the variables. Scatterplots representing data with a normal distribution should center around a central datapoint, with around the same amount of datapoints to either side of the center, taking the form of a in downward sloping curve (bell curve). In the histograms below, figures 5 and 6, it is clear that the main variables, Δ consumer confidence and Δ housing price growth are

relatively normally distributed variables, despite a higher peak. Based on the histograms, the

assumption can be made that both variables are more or less normally distributed.

Figure 5: The distribution of Δ consumer confidence Figure 6: The distribution of Δ house price growth

When assessing the normality of distribution with regard to the control variables, the scatterplots of housing cost overburden, purchasing power, inflation and household debt in figures 7-9 show that these variables follow the properties of a normal distribution relatively well. For these variables we conclude therefore that the data is sufficiently normally distributed.

(26)

25

Figure 7: The distribution of Δ housing cost overburden Figure 8: The distribution of Δ purchasing power

Figure 9: The distribution of Δ inflation Figure 10: The distribution of Δ interest

The data with regard to the changes in interest, as shown in figure 10, however seems not to be normally distributed. This effect is explainable, because the main central bank within the EU, the European Central Bank, has steadily decreased the interest rates in the aftermath of the final crisis, starting in 2008 and continuing all the way through 2017 (ECB, 2019). As a result, a large proportion of the datapoints falls in the category of a slightly decreasing interest rate and as a result violates the conditions of normal distribution.

Despite not all the variables adhering completely to the standards of a normal distribution, it is clear that except for the interest changes, the variables all very much center around a particular value and have a similar number of observations on both sides of the center. The absence of skewed distributions within variables, makes the variables suitable for a linear regression analysis, as conducted in this research.

(27)

26 4.2.4.3 Multicollinearity

Multicollinearity occurs in data when the independent variables are significantly correlated with one another. The problem with such multicollinearity is that it becomes harder to distinguish between the effect of each independent variable on the dependent variable. Part of the explanatory power attributed to one variable, might in fact to some degree be caused by another correlated variable. It is therefore important to ensure that there is no multicollinearity between the independent variables, to ensure that the outcomes are reliable. The

main concern in this research is the potential multicollinearity of the main independent variable, consumer confidence. In order to test whether this variable is truly independent in the research, a correlation analysis has been conducted between consumer confidence and the independent control variables.

Purchasing Power

Household Debt

Inflation Interest Housing Cost

Overburden Consumer

Confidence

.1319** -.3349*** -.1857*** -.0642 .0383

Table 6: Correlation analysis between Δ consumer confidence and the other independent variables

* Significant at the level of p < .10, ** Significant at the level of p < .05, *** Significant at the level of p < .01

The outcomes, given in table 6, show that a change in consumer confidence is significantly (p < .05) correlated with both changes in purchasing power, inflation (positively) and household debt (negatively). These correlations suggests that consumers are more optimistic about their current and future economic situation, when their purchasing power increases and their household debt decreases. Both of these effects seem logical, as both an increasing income and decreasing liabilities are indicators of a stronger financial position of a household. Moreover, higher inflation might be related to higher incomes, which could explain higher consumer confidence. That these variables are significantly correlated, indicates however that in the model the found effects of one variable might in fact be explained by means of another variable. As consumer confidence is the main independent variable in this research, the effects of multicollinearity with regard to this particular variable will have to be tested by means of a robustness check.

Another measure suitable for assessing the level of multicollinearity in a regression analysis, is based on the outcome directly, by means of a Variance Inflation Factor (VIF) test. This VIF test shows the degree of multicollinearity between the independent variables in a given output of a regression analysis, and shows a high level of multicollinearity when the given VIF values are greater than 10 (University of California, 2019). The initial regression analysis as reported in chapter 4.2.2, including all five of the control variables, has been tested on this measure and is reported in the below table 7. The output from the table shows that all the independent variables

(28)

27

have a VIF-score that is far below the threshold of 10. This test therefore concludes that the level of multicollinearity between the independent variables is relatively low.

Table 7: VIF with all independent variables

4.2.5 Adjusted multiple regression analysis and model re-estimation

After excluding the outliers, an adjusted multiple regression model is run with all the variables, to test whether the omission of datapoints results in a different outcome than before. The result of this adjusted regression is shown below in table 8. What is most noticeable, is that the main result of the exclusion of outliers is a change in predictive power of the model. The R-squared value, as a consequence of the elimination of outliers, has increased by .1320, from .3506 to .4826. What is furthermore relevant about the results, is that Δ purchasing power now has a significant and positive effect on the Δ housing price growth as well, which it did not have before. Variables Δ Housing Price Growth Standard Error t-value Δ Consumer Confidence .40637*** .04976 8.17

Δ Housing Cost Overburden .19546 .13444 1.45

Δ Purchasing Power .26709*** .09652 2.77 Δ Household Debt -.40014*** .10020 -3.99 Δ Inflation .87880*** .20434 4.30 Δ Interest -.31972 .37716 -0.85 Constant -.67624* .36750 -1.84 Observations 206 R-squared .4826 Prob > F .000

Table 8: Regression analysis post outlier omission * Significant at the level of p < .10 ** Significant at the level of p < .0 *** Significant at the level of p < .01

Variable VIF 1/VIF

(29)

28

Based on the outcome of this adjusted regression analysis, two of the control variables are still not significant: Δ housing cost overburden and Δ interest. As these variables are not significant they do not add explanatory power to the model with regard to the dependent variable.8 To test whether both variables should be omitted from the final model, regressions have been run excluding first one, and then the other of these control variables. Both regressions did not cause any noticeable differences in significance among the two control variables, and as they have no explanatory power on Δ housing price growth, they will both excluded from the final model estimation. The omission of these variables leads to the construction of another multiple regression analysis. The output of this regression analysis is given in table 9.

Based on the output of this analysis, an important results can be identified: three of the four remaining independent variables – Δ consumer confidence, Δ household debt and Δ inflation– now have a statistically very significant effect, at the level of p < .01 or stronger, on the dependent variable Δ housing price growth. The fourth independent variable, Δ purchasing power, however no longer has a significant effect on Δ housing price growth. Next to that, the explanatory power of the model has increased somewhat, compared to the regression model including all of the control variables. As a result of the excluding of the two variables, the R-squared has increased to a value of .4985, compared to a value of .4826 when all variables were included. This increase in R-squared shows that the model now explains more of the variance of the dependent variable, and moreover is better able to explain the variance than before. Based on this the conclusion follows that the omitted control variables only created noise in the outcomes and the estimation of the model.

Δ Consumer Confidence .64186*** .05879 10.92 Δ Household Debt -.33110** .12842 -2.58 Δ Inflation .83547*** .18770 4.45 Δ Purchasing Power .09176 .11361 .81 Constant -.57578 .41480 -1.39 Observations 226 R-squared .4985 Prob > F .000

Table 9: Regression analysis excluding non-significant control variables * Significant at the level of p < .10 ** Significant at the level of p < .0 *** Significant at the level of p < .01

However, table 9 shows that after the exclusion of the variables Δ housing cost overburden and Δ interest, the control variable Δ purchasing power no longer has a significant effect either. For this reason, for the purpose of the final model construction, Δ purchasing power is excluded from the analysis as well, together with Δ interest and Δ housing cost overburden. Although this final model, as given in table 10, has a slightly lower R-squared value than the model including Δ purchasing power, the final omission is justified based on the significance levels of the variables. In the final model, all the remaining independent variables now have a significant effect on Δ housing price growth, while the R-squared value of .4650 shows that the model has slightly less

(30)

29

explanatory power, but still explains a substantial portion of the variance of the dependent variable. The final model estimation, excluding outliers and insignificant variables, is based on this multiple regression analysis, which is given below in table 10.

Δ Consumer Confidence .58551*** .05599 10.46 Δ Household Debt -.41851*** .12748 -3.28 Δ Inflation .84280*** .17939 4.70 Constant -.51316 .34912 -1.47 Observations 241 R-squared .4650 Prob > F .000

Table 10: Final model regression analysis * Significant at the level of p < .10 ** Significant at the level of p < .0 *** Significant at the level of p < .01

Based on the outcomes of the analyses conducted in this chapter, the theorized model as stated in chapter 3 should be readjusted, based upon the empirical findings. After that, the model’s estimated coefficients, based on the outcome of the final regression analysis, can be incorporated. First, the expressions with regard to the variables Δ housing cost overburden, Δ interest and Δ purchasing power need to be excluded from the model, based on the empirical findings that these do not have a significant effect on Δ housing price growth. Moreover, no significant effect of a constant was detected in the analysis, indicating that there is not a normal, baseline, underlying trend in Δ housing price growth. As a consequence, the constant has been eliminated from the model’s estimation as well.

Together, these adjustments lead to the re-estimated, final form of the model as:

dY = dCx1 – dDx2 + dGx6 + e (5)

Note that with the adjustment of the model, the direction of the effect of Δ household debt has been reversed from ‘+’ to ‘–’. This is the case, because the result of the analysis has shown that the variable in fact has a negative effect on Δ housing price growth, instead of the theorized positive effect. Using the results from the regression analysis, we can now estimate the linear model to approximate:

dY = dC(.586) – dD(-.419) + dG(.843) + e (6)

4.3 Robustness checks

In order to test whether the found results in the analysis section hold true when tested against a number of deviating scenarios, robustness checks have been conducted. First, the data underlying the final model estimation is divided into two temporal groups, both containing values of six years.9_{This test is meant to control whether the found effects hold true in two subsamples of the} full period. After this an analysis is conducted including the interaction effects of the independent variables of the final model, to control whether the found effects may potentially be

The Influence of Consumer Confidence on Housing Prices A study of the housing market in the European Union, 2007-2017

Master Thesis International Economics & Business

2018-2019.2