The impact of homeownership on risk taking in household portfolio choice: Evidence from the Netherlands

The impact of homeownership on risk taking in household portfolio

choice: Evidence from the Netherlands

By:

Martin Haakma

Student number: S3797392

E-Mail:

m.p.haakma@student.rug.nl

University of Groningen

Faculty of Economics and Business

MSc. Finance

Supervisor: Carolina Laureti

Date: 11-02-2021

Abstract

This paper examines whether homeownership has an effect on household portfolio choices. The panel dataset for this research is acquired from the Dutch Central Bank (DNB) Household Survey – DHS over the years 2009-2019. According to the literature, homeownership should be associated with lower financial investments. We performed multiple Probit and Tobit regressions whilst controlling for economic characteristics, behavioral characteristics and socio-demographic characteristics to examine the relationship of interest. In line with other empirical studies, we have not found a systemic relationship whether homeownership affects the holding of risky assets nor the share of risky assets in household portfolios. The results of our sub-samples home tenure and size of the house are somewhat similar and do not show a significant relationship either. In addition to our subsamples, a Logit model has been used for robustness; it leads to the same conclusion again.

JEL classifications: D14, G11

2 1. Introduction

Households are central to our understanding of finance. They are the pivotal decision making agents in financial markets (Gomes et al., 2020). Households take savings decisions which lead in turn to direct or indirect investments in financial markets. Furthermore, they also have to make an increasing number of decisions on their own. Since those households control a large part of wealth in the economy, participation from households in the stock market can have a significant impact on the equity premium (Campbell, 1993; Heaton and Lucas, 2000). Data from Eurostat confirms that the financial assets of households control a large part of wealth in the economy. In the EU-28, financial assets rose on average by 3.9% each year in the period 2008-20181. These household financial assets are defined by the OECD as assets, such as saving deposits, investments in equity, shares and bonds. Moreover, due to the documented deviations from financial theories, studying portfolio composition in the household context is becoming increasingly relevant. Household financial decisions are complex, interdependent, and heterogeneous, and central to the functioning of the financial system (Campbell, 2006; Gomes et al., 2020). Consequently, small improvements in financial decisions of individual households have the potential to generate large economic gains for society (Bhamra et al., 2019).

Conventional financial theory holds that rational economic men should diversify their investment across all risky assets with a certain proportion, according to their different risk preferences (Markowitz, 1952). Households' motivation to invest in risky financial assets has received a lot of academic interest. Researchers are interested in finding out why the majority of households forgo direct or indirect ownership of risky assets such as equity funds, stocks and long-term bonds, given the substantial equity premium. This phenomenon is known as the stock market participation puzzle (Haliassos and Bertaut, 1995). As will be discussed in the next chapter, there are several reasons why individuals and households have or do not have investments. Previous literature has thoroughly analysed the effect of individuals' characteristics on portfolio allocation and participation decisions in risky asset markets. Identifying the drivers of stock market participation could help to explain the puzzle.

The aim of this paper is to understand that an interesting and relevant factor in households' investment decisions might be the degree of homeownership. Chetty and Szeidl (2010) stated this nicely: ‘’Houses are the largest assets owned by most households, but the impact of housing on financial markets remains unclear’’. Homeownership might have an

1

3 explanatory role in portfolio choice, since households may adjust their exposure to financial risks once they are homeowner. Related literature about the effects of homeownership is both theoretical and empirical. Theoretical literature highlights three important mechanisms through which investment decisions in housing affect portfolio choice: illiquidity of the house, hedging of housing price risk and diversification. The illiquid nature of housing and hedging of house price risk generally reduce the demand for risky assets of households (Grossman and Laroque, 1990; Flavin and Yamashita, 2002; Yao and Zhang, 2005). By contrast, empirical studies find no systematic relationship between homeownership and portfolios in practice (Fratantoni, 1998; Heaton and Lucas, 2000; Yamashita, 2003; Cocco, 2005).

This paper builds on the relationship introduced by Chetty and Szeidl (2010) by disentangling the effects of home equity and outstanding mortgage debt, as components of homeownership, on portfolio choices. The wealth effect of housing could improve the probability of financial market participation, while housing mortgages could reduce the probability of financial market participation. Chetty and Szeidl (2010) used data from the Survey of Income and Program Participation (SIPP) in the U.S and found that the stock share of liquid wealth would rise by 1 percentage point – 6% of the mean stock share – if a household were to spend 10% less on its house, holding fixed wealth. This paper uses longitudinal data of 1.986 Dutch individuals over eleven waves from the Dutch Household Survey (DHS) to examine the effect of homeownership on investment decisions made by households in the Netherlands. Consequently, the research question in our paper is as follows: does homeownership influences risk taking in household portfolio choices?

4 disruption can encourage people to buy bigger houses than they otherwise would and this can lead to different results for stock market participation and allocation decisions. On the other hand, home tenure can be a determining factor. As households build up liquid wealth over time after a house purchase, we expect long home tenure to be accompanied by more available capital to invest. These distinctions are therefore a valuable addition to the existing literature.

Our dependent variables of interest are the holding of risky assets and the share of risky assets in household portfolios, as part of their total assets. Risky assets are defined as stocks (direct participation) and mutual funds (indirect participation). This relationship has been tested using a binary Probit model for the participation decision and a censored Tobit model for the allocation decision. Both specifications use random effects and time-specific effects.

This paper finds no significant opposite effect of home equity and mortgage among Dutch households. Our results show that homeownership alone does not affect the participation is risky assets nor the allocation to risky assets. In contrast, households do consider their home equity when making participation and allocation decisions. Mortgage debt appears to only affect allocation decisions. Although the coefficients on home equity and mortgage are statistically significant and positive, our results do not have a valid economic interpretation. The coefficients are very small, making the economic interpretation negligible. Our results could be explained by the investment behaviour and the institutional setting in the Netherlands. However, we argue that the results of our paper are still relevant for households and policy makers. Understanding the investment behaviour of households provides the government with important information which improves the quality of future decisions. Our study contributes to solving the non-participation puzzle and gives implications needed for policy makers to increase stock market participation. As a final point, further research on the topic is needed for more representative results and explanations.

5 2. Literature review

The aim of this chapter is to discuss the synthesis in the existing literature on homeownership and investment behaviour. In the first section we start with properties of the Dutch housing market, the country on which our research data is based. Subsequently, section two will discuss how homeownership might influence investment behaviour of households using theoretical predictions. Our third section elaborates on heterogeneity in investment behaviour and the associated stock market participation puzzle, because actual investment behaviour deviates from conventional investment theory. Section four provides an overview of empirical papers examining the relationship between homeownership and portfolio choice. Finally, the hypotheses of our paper will be expounded in part five.

2.1 Properties of the Dutch housing market.

As indicated in the introduction, our paper is based on Dutch household survey data. Due to historical figures and publications, it can be concluded that the Netherlands is a unique country to study the impact of homeownership on investment behaviour. We will successively discuss that this argument is justified by figures on mortgages and homeownership, which are set out in a report from the European Mortgage Federation (EMF Hypostat, 2020). Subsequently, we elaborate on the development of investment behaviour in the Netherlands.

6 the default probability ranges between 1.5 percent and 5 percent using all outstanding mortgage loans in the Dutch market; a significantly low default probability.

The Netherlands is also a unique country when looking at the housing market. For instance, the Netherlands is almost at the bottom of the list in terms of homeownership in Europe. However, Butrica and Mudrazija (2017) argue that reasons for differences in homeownership rates across countries are difficult to substantiate. Each country has its own culture, demographics, policies, housing finance systems, and, in some cases, a past history of political instability that favours homeownership. Figures on the housing market in 2018 show that the owner occupation rate in the Netherlands is 68.9% compared to the EU27 average of 66.2%. In addition, this housing market is characterized by high demand, lowering supply and rising prices. Houses are in 2019 on average 6.9% more expensive than a year earlier. The fact that prices have risen more than expected is explained by a further fall in mortgage interest rates, making buying a more attractive option than renting. Furthermore, several scholars indicate that the Dutch housing market is characterized by low taxation of homeownership. In particular, the very generous deductibility of mortgage interest artificially increases house prices. Moreover, the mortgage interest deductibility disproportionately favour high-income taxpayers and have ambiguous effects on home tenure and spending options (Vandevyvere and Zenthöfer, 2012; Haffner, 2002; Wolswijk, 2010).

7 According to them, the Netherlands form an interesting research country due to the institutional setting, because financial markets are well developed compared to surrounding countries. In addition, there are also extensive information channels through which the common household can learn about investment possibilities. Stock market capitalization by households in the Netherlands fluctuates around 35.7% according to Guiso et al. (2003). Furthermore, increases in overall financial risk taking has been accompanied by a reduction of traditional savings accounts and more investments in housing. That households appear to be more risk tolerant than before can be explained by competitive forces in the banking and insurance industry, a period of stable economic growth and a booming stock market (Alessie and Hochguertel, 2003). Our next section examines homeownership as a potential driver or barrier to invest in risky assets.

2.2 The influence of homeownership on investment behaviour.

In the current literature, there is an ongoing debate about how homeownership might affect household portfolio choices. Traditionally, homeownership is recognised as a major indicator of economic well-being at both the household and market levels (Megbolugbe and Linneman, 1993). In the U.S., the importance of homeownership is intertwined with the set of beliefs and expectations identified as the ‘American Dream’ (Rossi, 1980). As a result, owning a home may be seen as evidence that a person is a competent and worthy individual (Rohe and Basolo, 1997). Others argue that homeownership is a valuable institution because it allows families to build wealth and serves as a measure of financial security (Goodman and Mayer, 2018; Arundel and Doling, 2017). However, unlike financial assets, housing is illiquid, subject to fluctuations in house value, and is costly to adjust (Chetty and Szeidl, 2007). Pelizzon and Weber (2008) argue that efficiency of household portfolios strongly depends on homeownership, also in the case of nonzero correlation between housing and stock returns.

8 with a negative effect on stock market participation. The rationale is that households with a high mortgage may be more conservative in their financial portfolio decisions to guard against a simultaneous drop in stock and house prices (Michielsen et al., 2016). As a result, households with high mortgage debt are more likely to fall into negative net wealth and this reduces their risk tolerance. Furthermore, the effect of housing on portfolios is likely to be reinforced by more risky housing markets with price fluctuations (Chetty and Szeidl, 2016). In addition to these two important components of homeownership, several papers suggest that there are three main mechanisms through which investment decisions in housing affect portfolio choice (see e.g., Chetty and Szeidl, 2010; Chetty et al., 2017; Lyng and Zhou, 2019; Michielsen et al., 2016; Fougere and Poulhes, 2012). As a result, the following mechanisms will be discussed in turn: hedging of housing price risk, illiquidity of the house and diversification.

The most obvious form of uncertainty associated with homeownership is housing price risk (Fratantoni, 2001). Fratantoni (2001) believes that housing prices at the local level can be volatile. As a result, agents do not know with certainty the future sales price of their houses. Furthermore, there is committed expenditure risk. Homeowners are committed to fixed payments over a long horizon even though they are subject to labour income uncertainty. Chetty and Szeidl (2007) document that more than 50% of the average U.S. household budget remains fixed when the household faces moderate income shocks such as unemployment. Sinai and Souleles (2005) indicate that owning a house provides a good hedge against fluctuations in housing costs, but in turn introduces asset price risk. The authors argue that hedging in the form of wealth accumulation is especially relevant for households who want to make the switch from cheaper to more expensive housing. In related work, Flavin and Yamashita (2002) demonstrate that liquidity constraints caused by a house purchase can have an enormous effect on the risk-return trade-off available to the household. In contrast to Sinai and Souleles (2005), Flavin and Yamashita (2002) emphasize that hedging is interesting for highly leveraged homeowners. Accordingly, they should hedge their house price fluctuations risks by holding fewer risky financial assets. Their argument is consistent with Fratantoni (1998) who believes that housing price risk will push the household to reduce their risky asset holdings. Arrondel and Savignac (2009) elaborated further on this. The authors argue that households should moderate their global exposure to housing price risk by limiting the share of their financial wealth invested in stocks. In other words, an increase in the housing to-net-wealth ratio is associated with lower stock market investments for a given total financial wealth.

9 will drop before the purchase of a house (see e.g., Paxson, 1990; Grossman and Vila, 1992; Tepla, 2000). These papers argue that households become more risk averse if there are borrowing constraints, such as with mortgage. Their intuition is that additional risk aversion is caused by the possibility that the constraint might be binding in the near future. The model of Grossman and Laroque (1990) confirms this assumption. Consequently, it predicts that if it is costly to adjust the level of an illiquid durable consumption good, households are more risk-averse after the purchase of a durable consumption good such as a house. In conclusion, the second mechanism implies that higher property holding may lead to reduced holdings of risky assets.

Yao and Zhang (2005) emphasized that diversification is an important mechanism as well. They examined optimal dynamic portfolio decisions for investors who acquire housing services from either renting or owning a house. They argue that investors who own a house will choose a substantially different portfolio allocation. Furthermore, such investors should hold a higher equity proportion in their liquid financial portfolio to take advantage of the diversification benefit. This benefit is afforded by the low correlation between stock returns and housing returns. Moreover, these investors should partly substitute home equity for risky stocks. As stated earlier, homeowners with mortgage debt face expenditure risk due to committed mortgage payments over a long horizon. Fratantoni (2001) and Hu (2005) emphasize that liquidity demand is important in the presence of mortgage debt. Subsequently, Becker and Shabani (2010) invented a debt retirement channel which dominates the concern of liquidity demand. According to their debt retirement channel, households should increase the equity share of their liquid wealth if they hold mortgage debt after a house purchase, conditional on stock market participation (Becker and Shabani, 2010; Lyng and Zhou, 2019). For instance, if households invest their liquid wealth in stocks, this allows them to earn a higher expected return than the return on safe assets. On the other hand, if they don't, they may be better off by using the liquid wealth to pay back the mortgage debt. Repayment of mortgage debt offers households a return equal to the interest rate on the mortgage loan. Noteworthy, that return is almost always higher than the return on safe assets.

2.3 Heterogeneity in investment behaviour of households.

10 individual’s risk aversion parameter (Breuer et al., 2014). Investors should care only about the mean and variance of their portfolio. Moreover, the standard portfolio choice framework predicts that households' wealth is fully diversified and does not vary during the life cycle (Merton, 1971; Arrondel and Savignac, 2009). Economic theory usually assumes risk aversion as a stable personal trait and it is defined as a preference for a sure outcome over a prospect with an equal or greater expected value (Tversky and Kahneman, 1981). In addition, precautionary saving theory predicts that households will respond to an unavoidable risk both by saving more, and by shifting their portfolio toward safe assets (Kimball, 1991; Fratantoni, 2001). Consequently, households that are more risk averse should be less inclined to invest in risky assets (Guiso et al., 2003; Breuer et al., 2014).

Homeownership as a driver or barrier in investment decisions of households can be explained by behavioral finance theory. Behavioral finance is concerned with psychological influences on individual investor behaviour (Charness and Gneezy, 2010). As pointed out by Shiller (2003), behavioral finance theory is now one of the most vital research programs, and it stands in sharp contradiction to much of efficient markets theory. It was created because researchers had seen too many anomalies. As a result, there was too little inspiration that conventional theoretical models captured important fluctuations in household decisions (Shiller, 2003; Thaler, 1980; Kahneman and Tversky, 1979). In fact, many consumers act in a manner that is inconsistent with conventional economic theory. Bhamra and Uppal (2019) argue that financial markets are not a mere sideshow to the real economy and that improving the financial decisions of households can lead to large benefits, not just for individual households, but also for society. However, the study of household finance is challenging because household behaviour is difficult to measure (Campbell, 2006).

11 even if households with very low holdings of financial assets are excluded, only 12% to 18% of households invested between 20% and 80% of their financial assets in stock holdings (Shum and Faig, 2006). Consequently, the low participation rate cannot be explained by wealth alone.

Because limited stock market participation has a direct impact on the level of the equity premium, proper identification of the determinants of stock market participation decisions is necessary for a better understanding of the puzzle (Bonaparte and Kumar, 2013). Several scholars indicate that even small costs are enough to keep many households out of the stock market, especially since the marginal investor wants to invest limited amounts in the stock market (see e.g., Haliassos and Michaelides, 1999; Polkovnichenko, 2000; Paiella, 2001; Vissing-Jorgensen, 2002). Other main drivers of limited stock market participation are short sale constraints and income risk, inertia, and nonstandard preferences (Haliassos and Bertaut, 1995; Linnainmaa, 2005; Benartzi, 2001). The aim of our paper is to link homeownership to an important economic outcome: participation in the stock market. Clear is that homeownership alone has no exclusive power to explain the stock market participation puzzle. Consequently, other barriers to enter the stock market must exist. We will successively discuss other important papers intending to account for the stock market participation puzzle.

The effect of age seems to be somewhat ambiguous in the literature. Fagereng et al. (2017) indicate that participation in the stock market is not only limited at all ages, but it tends to follow a life‐cycle pattern. This life-cycle pattern is in line with Dohmen et al. (2011) and Cappeletti et al. (2014) who emphasize that the relation between stock market participation and age is nonlinear, being much stronger in the last part of one’s career. More specifically, households rebalance their portfolio away from stocks before they reach retirement and exit the stock market after retirement (Fagereng et al., 2017). Indeed, some studies have found that younger respondents are more likely to invest in stocks than the elderly (Campbell and Viceira, 2002; Veld-Merkoulova, 2011). However, other papers argue that individuals do not gradually decrease their risky asset shares as they age (Ameriks and Zeldes, 2004; Gomes, 2020).

12 confirmed by other studies who suggest that many families shy away from the stock market because they have little knowledge of stocks, the workings of the stock market, and asset pricing (Van Rooij et al., 2011; Almenberg and Dreber, 2015).

Besides demographic factors, several studies investigated psychological, behavioral and social determinants. For instance, Coval and Thakor (2005) argue that investors who are more optimistic about future returns should allocate more money to risky financial assets. Furthermore, Brown et al. (2008) and Hong et al. (2004) present evidence supporting a positive association between social interaction and stock market participation. They established a causal link between an individual’s decision to own stocks and the average stock market participation of the individual’s community.

As a final point, risk aversion is an important psychological determinant. The paper by Laakso (2010) makes clear that risk aversion stands out as the single most significant driver of stock market participation. Furthermore, risk aversion seems to be an important channel through which also other previous discussed drivers of stock market participation operate. In addition, literature shows us that risk aversion does not only affect portfolio composition, but the overall decision of becoming a stockholder, due to the fixed cost of investing (Bertaut, 1998; Paiella, 2001; Guiso et al., 2003; Vissing-Jorgensen, 2002; Bucciol and Miniaci, 2011). To sum up, all the discussed scholars provide important insights in the determinants of household portfolio choices. However, a considerable part of participation and allocation decisions still remains unexplained (Laakso, 2010).

2.4 Empirical evidence on homeownership and household portfolio choices.

13 Several papers studied the causal effects of homeownership decisions on the level and composition of household savings accumulation. For instance, Krumm and Kelly (1989) present evidence of saving rises prior to the house purchase and after the house purchase, to rebuild liquid assets. Savings prior to the house purchase are due to down payments in order to qualify for a mortgage loan (Engelhardt, 1996). Beverly and Sherraden (1999) show that household members’ decisions to save are affected by past life course experiences. They found evidence that many households reduce consumption and increase savings in order to accumulate the necessary down payments. Furthermore, once individuals get used to this disciplined lifestyle of saving, it continues after having entered homeownership. Other papers have found evidence that saving behaviour is influenced by house price fluctuations. Using 1984-1989 PSID Data, Hoynes and McFadden (1994) found that a 1% increase in regional house prices will lead to a household savings rate increase of 0.26% point. In contrast, Skinner (1993) found that for households of age above 45 years, savings declined by 2.8 cents for one dollar of housing wealth increase.

14 Most other empirical papers do not find a systematic relationship between housing and portfolio choice. A seminal contributor in the field is the work of Cocco (2005). His focus was on the homeowners' portfolio choice problem. He found evidence that investment in housing plays a crucial role in explaining the patterns of cross-sectional variation in the composition of wealth and the level of stockholdings. And together with Heaton and Lucas (2000) they found evidence that the effect of housing property on stockholdings is positive and risky asset shares are positively correlated to mortgage debt. Furthermore, the results of Cocco (2005) imply that due to investment in housing, younger and poorer investors have limited financial wealth to invest in stocks. Consequently, house price risk crowds out stockholdings, and this crowding out effect is larger for low financial net-worth. According to Yamashita (2003) and Fougere and Poulhes (2012), there is a hump-shaped pattern for the proportion of stocks in financial portfolios. They found evidence that when housing represents either a very small part or an important part of the total household wealth, household’s stockholding is low. Unlike using US data, Saarimaa (2008) used a theoretical simulation model with Finnish asset return data. His main result indicates that the more valuable house a homeowner resides in, at a given level of net wealth, the less likely it is to own stocks. Consistent with Yamashita (2003), there is a negative correlation between stockholding and housing wealth. The same argument persists in the work of Wu and Qi (2007). They provide evidence that housing investments in China crowds out holding proportions of risky assets such as stocks, funds and foreign exchange.

15 households have weaker buffers to withstand unexpected shocks to their incomes or to interest rates (Meen, 2011).

Several other papers further highlight the role of housing wealth in determining optimal portfolio choice and life-cycle consumption allocation. For instance, Cardak and Wilkins (2009) and Chetty and Szeidl (2016) provide evidence that wealth improvements from housing capital gains enable households to be less risk averse. Consequently, households are able to engage in riskier investments. Similar results emerge in the study by Chen, Shi and Quan (2015). Using CHFS data from 2011, they show that a rising housing market will increase the value of household housing wealth and associated holdings of stocks and other financial assets. Other empirical evidence considers the housing value-to-wealth ratio (Stokey, 2009; Flavin and Yamashita, 2011). For instance, Flavin and Yamashita (2002) studied the impact of the portfolio constraint, imposed by the consumption demand for housing, on the household's optimal holdings of financial assets. They found evidence that investments in housing by younger and less wealthy investors reduces the investment in risky financial assets, which is consistent with previously reported findings by Cocco (2005). They also prove that if a low-wealth household purchases a home, it would be exposed to relatively high risk because its portfolio would not be diversified enough (Flavin and Yamashita, 2002). In later work, Flavin and Yamashita (2011) made use of the Survey of Consumer Finances and construct a model of optimal portfolio allocation incorporating the role of housing as collateral. Their results imply that holding risk aversion constant, the percentage of the portfolio held in stocks is decreasing in the ratio of house value-to-net-wealth. As a result, older households with a lower ratio of house value-to-net-wealth will generally hold more of their portfolio in stocks than younger households. In addition, Cardak and Wilkins (2009) found that housing as collateral can create opportunities for households to invest in risky assets such as stocks. However, this only applies to countries and regions with a well-established credit system.

16 decisions for investors who acquire housing services from either renting or owning a house. Their results show that when indifferent between owning and renting, investors owning a house hold a lower equity proportion in their portfolios.

Finally, some evidence from the Netherlands also suggests that demand for financial wealth is systematically different for renters and homeowners (Hochguertel and van Soest, 2001). Using Dutch collective Bank Study data, they found evidence that investing in financial wealth involves a positive threshold for homeowners. Subsequently, this threshold has to be overcome before positive financial asset holdings are possible. This threshold is magnified by the level of housing investment. In addition, as homeowners are affected by decreases in their property value, homeowners reduce their conditional demand for financial assets when houses fall in value.

2.5 Hypothesis development

Economic literature on homeownership and household portfolio choice argues that housing investment in general crowds out financial investments. The effect of this relationship in the Netherlands is expected to be similar to those found in the papers of Chetty and Szeidl (2010) and Fougere and Poulhes (2012). Our first hypothesis is therefore formulated as follows:

H1: There is an (negative) association between homeownership and financial risk-taking behaviour.

This hypothesis enables us to test whether the effect of homeownership is limited to the participation decision or also influences the share invested in the stock market by using different models. Relying on the discussed literature, we could argue that historic practices, culture and most importantly tax legislation, especially those pertaining to the deductibility of mortgage interest, have shaped the Dutch residential mortgage market in quite a unique way. In particular, this makes the Netherlands an interesting domain for our research. In addition, our prediction is that the substantial incentives for homeownership (from tax breaks to low mortgage-interest rates) in the Netherlands distort demand, encouraging people to buy bigger houses than they otherwise would and that this has a negative effect on stock market participation. Hence, the following second hypothesis is added to our paper:

17 Our last hypothesis is in line with the theoretical framework of Alessie et al. (2000) and Lyng and Zhou (2019). Both authors pointed out that households tend to save more if they plan to buy a house. In order to purchase a house, stock market participation will decrease as households will reallocate their liquid wealth between risky investments and safer investments, leaving them with cash to spare for uncertain times. Once the house has been bought, the liquid assets will be low due to investments in the house. We then expect capital invested in the stock market to remain low. But, as households build up liquid wealth over time after a house purchase, we expect that the equity market participation rate will gradually increase if the housing market remains somewhat normal. Therefore, a positive effect on risk taking is expected if respondents are homeowners for a long time. Our third and final hypothesis is therefore formulated as follows:

H3: There is an (positive) association between long home tenure and financial risk-taking behaviour.

3. Data

In this chapter, the first section provides an overview of the data source. Section two is about the variable definitions. In section three we elaborate on the sample construction. We close this chapter with section four, which describes the descriptive statistics of our pooled sample and our subsamples house size and home tenure.

3.1 Data source DHS

The data for this research is acquired from the Dutch Central Bank (DNB) Household Survey – DHS2. It is a panel survey supported by CentERdata, which is part of Tilburg University, the Netherlands. The panel is representative of the Dutch population with respect to a number of important demographic characteristics. In other words, the average panel member has the same experiences and knowledge as the average person living in the Netherlands. The DHS data are collected by means of an internet panel and people without connectivity participate through a TV set top box or using an obtained computer and internet connection. The DNB Household Survey is a unique data set allowing people to study both psychological and economic aspects of financial behaviour. This panel survey among 2.000 Dutch households was launched in 1993. The reason why DHS is used for this paper is that this

2 _{More information about the DHS survey can be found at:}

18 panel contains a large amount of information about employment, pensions, housing, mortgages, income, possessions, loans, health, economic and psychological concepts and personal characteristics. The data is divided in different modules, which are bundles of related questions. It is collected annually among participants in the panel and each year, the survey adds new participants to refresh the sample and to deal with attrition (Teppa and Vis, 2012). Teppa and Vis (2012) define attrition as a reduction or decrease in size, number, or strength of the sample. It occurs because of a failure to relocate or recontact a respondent after the initial wave of data collection, and because of non-cooperation. Winkels and Withers (2000) describe attrition as ‘the panel researcher’s nightmare’. According to them, members who drop out of a panel may differ systematically from those who stay in. As a result, the dataset of continuing members may no longer be representative of the original population. Earlier research by Falaris and Peters (1998) provided evidence that survey attrition either has no effect on the regression estimates, or only affects the estimates of the intercept. Furthermore, it does not affect estimates of background (family) slope coefficients. Nevertheless, even if the attrition rate is small, it can introduce bias to survey estimates and therefore attrition is important to take into account (Fitzgerald et al., 1998).

Attrition is supposed not to be a major problem if it is random (Nunan et al., 2018). A Logit regression was performed to test whether respondents randomly enter or leave our panel during the years. The Pseudo R-square from this attrition Logit can be interpreted as the proportion of attrition that is non-random (Outes-Leon and Dercon, 2008). A dummy variable is created where a score of 1 is assigned to households which drop out of the sample in our timespan, and a score of 0 if they remain. Next, it is regressed on different observable characteristics. The results can be found in appendix D and the results confirm that attrition is supposed to be random. The corresponding p-values show us that only 2 out of 16 variables are statistically different from zero at the 5% level. In addition, there are 3 variables significant at the 10% level. The resulting Wald chi-square statistic indicates that these five variables are jointly statistically different from zero at the highest level of significance in predicting attrition3_.

Apparently, the DNB Household Survey was subjected to attrition between 2009 and 2019. Only a relatively small number of households were observed in all the waves, but non-random attrition in our panel is only 12.8%. Consequently, we believe that the effects of attrition are not a major issue in our panel. As indicated earlier by Teppa and Vis (2012), the administrators of CentERdata are doing their best to deal with attrition properly. In addition, even if attrition follows some non-random patterns, attrition remains overwhelmingly a random phenomenon (Outes-Leon and Dercon, 2008; Nunan et al., 2018).

19 The dataset tracks the same household over a number of years, which allows us to obtain a representative picture of respondents’ housing, mortgages and demographic properties. In addition, the questions asked to measure the required variables have proved to be relatively consistent throughout the years of data collection (Teppa and Vis, 2012). This enables us to assess an extended time interval. This study uses all waves of data from years 2009 till 2019 as data relevant to our variables were particularly sparse in earlier years of the survey, as we will explain later. The DHS consists of six questionnaires and considering the empirical part of this thesis, the required variables can be found into all the sections of data in the DHS panel. The categories that are used are: general information on the household, household and work, accommodation and mortgages, health and income, assets and liabilities, economic and psychological concepts and aggregated data on income, assets, liabilities and mortgages. In the next section we elaborate on the construction of our variables.

3.2 Variable construction

The following section describes successively the dependent, independent and control variables included in our analysis. The exact composition and description of each variable is included in appendix A.

The first dependent variable relates to the holding of risky assets, which we will define as stocks (direct participation) and mutual funds (indirect participation). It is based on a question in the questionnaire regarding assets. Participants were asked whether they possess growth funds, share funds, debentures, stocks, options or warrants. The holding of risky assets is a binary variable. A dummy variable is created where a score of 1 is assigned to a household if they possess either stocks or mutual funds, and a score of 0 when a household possesses neither of them. Consequently, total stockholdings, the direct and indirect stock market participation together, is taken into consideration for determining the dependent variable ownership of risky assets.

20 As mentioned in the literature review, the most important independent variable to include in our analysis is homeownership. Homeownership is defined as a household that owns a home (Letkiewicz and Heckman, 2018). From the Housing and Mortgages questionnaire we obtain one main question regarding ownership and one main question regarding mortgage. Homeownership is measured as a binary variable that takes the value of one if the household owns a home and zero otherwise. Related to the findings in the paper of Chetty and Szeidl (2010), we add mortgage and home equity as independent variables. For mortgage, we used a question from the same questionnaire whether households have a mortgage and how much the mortgage is. Home equity is defined as house value minus outstanding mortgage debt.

To test our second hypothesis, whether there is an association with house size and financial risk-taking, we use a variable that distinguishes between type of house. Respondents were asked in what type of home they live, ranging from a small flat to a large farm or gardener's home, which is therefore considerably larger. We define a small house as: semi-detached houses, terraced houses, a flat, apartment, upstairs or downstairs apartment. A large house is defined as a detached house, a farm, or gardener's house. For the third hypothesis, a distinction must be made between years of home tenure. Subsequently, home tenure is defined using a question where respondents had to state the purchase year of their current home. We define long home tenure as houses that were bought more than 10 years ago. Short home tenure is defined as houses bought by their owner in the past 10 years.

To accurately measure the moderating effect of homeownership on stock market participation and allocation decisions of households, controlling for additional variables affecting the variance of our dependent variables is a prerequisite (Wooldridge, 2010). As a result, several relevant control variables are used in the regression to take into account differences in households’ economic, behavioral and socio-demographic characteristics. Otherwise, it may cause omitted variable bias. The term omitted variable bias refers to any variable not included as a variable in the regression that might influence our dependent variables (Brooks, 2019). These controls aim to address the confounding variables in the investigated relationship. These confounding variables are extraneous variables whose presence affects the variables being studied. In that case, our results do not reflect the actual relationship between the variables under study (Brooks, 2019). In line with multiple scholars in this field, there will be controlled for an extensive set of variables.

21 sources of income and wage-replacing transfers, then mortgage interest payments and taxes are subtracted. This calculation is already done by the DNB Household Survey. The log of net income is calculated to remove potential outliers and make the distribution of income look more normal. We also account for wealth, as several papers show a positive relation between wealth levels and stock market participation (Moskowitz and Vissing, 2002; Campbell, 2006; Guiso et al., 2003). Wealth is the total amount of all financial and non-financial assets held by the individual. These asset groups are indicated in appendix A. We use the logarithmic form, because both variables are positively skewed. Secondly, it is of relevance to include liabilities. Most studies mainly investigate the asset side of the balance sheet in the context of household portfolio choice, but households also consider their outstanding liabilities when making investment decisions (Becker and Shabani, 2010). The variable liabilities is calculated by adding different debt components available in the questionnaire. Since there may be a nonlinear relationship with the dependent variables, we also use the quadratic form. Relying on empirical evidence, a positive relation is expected for income and wealth. We expect the opposite for liabilities.

Our behavioral control variables say something about the characteristics of a household (personality traits). We include risk tolerance, because various scholars suggest that actively participating in the stock market is correlated with an individual’s risk preference (see, e.g., Vissing-Jørgensen and Attanasio, 2003; Van Rooij et al., 2011). Individuals willing and prepared to take on larger risks are more likely to invest in stocks. Contrarily, risk averse individuals are less likely to invest in risky assets. To measure the degree of risk aversion, we use the question “What would you say was the risk factor that you have taken with investments over the past few years?” from the DHS. The participants were able to answer this question on the basis of five possible answers, ranging from 1 - “I have taken no risk at all” to 5 - “I have often taken great risks”. Since a higher score corresponds to a higher risk tolerance, a positive relation is expected.

We also control for planning horizon. Planning horizon tends to influence investments in financial risky assets, such as stocks, options, and mutual funds. A longer planning horizon leads to an increasing share of risky financial investments (Lee et al., 2015). In the questionnaire, respondents were asked to indicate the time range they consider when planning expenditures and savings. Again, the answer choice includes five options: 1 - “the next couple of months”, 2 - “the next year”, 3 - “the next couple of years”, 4 - “the next 5 to 10 years”, and 5 - “more than 10 years from now”. Since a higher score corresponds to a longer planning horizon, a positive relation is expected.

22 their lifecycle (Alessie et al., 1997). Age is included in linear and quadratic form to account for these life-cycle effects (Atella et al., 2012). We calculate age as the year of the questionnaire minus the birth year. Furthermore, retirement is examined as a dummy variable where a score of 1 is assigned to a household if their age is above 67, and zero otherwise. The effect of age on risky assets holding is not convincing in the existing literature. Hence, the effect can be both positive or negative.

Secondly, we add health. Rosen and Wu (2004) find evidence for holding a larger part of wealth in safe assets instead of risky assets, caused by an individual’s poor health condition. Therefore, controlling for health is relevant in the context of stock market participation. To measure the self-perceived health, we use the question where respondents had to assess their own health. The answer choice ranges from 1 to 5, where 1 corresponds to ''excellent'' and 5 corresponds to ''poor''. Since a higher score corresponds to a poorer health, a negative relation is expected.

Thirdly, scholars do agree on education influencing stock market participation rates as well (see, e.g., Haliassos and Bertaut, 1995; Van Rooij et al., 2011). They argue that higher educated participants are more likely to take risk. Consequently, a higher level of education might influence portfolio composition due to the improvement of financial knowledge. A respondent’s level of education will be determined by a question where respondents had to fill in their highest level of education attended. The answer choice ranges from 1 to 7, where 1 corresponds to ''(continued) special education'' and 7 corresponds to ''university''. Since a higher score corresponds to a higher education level, a positive relation is expected.

Fourthly, we control for children. Households with more children are less likely to invest in the capital market, because they are not supposed to make mistakes in their investment decisions (Dimmock et al., 2015). Households were asked to indicate how many children the family consists of. Since a higher score corresponds to more children, a negative relation is expected.

23 1 if the respondent is male or married, and 0 otherwise. Since a score of 1 corresponds to being a male or married, a positive relation is expected.

3.3 Sample construction

As mentioned before, the DHS covers seven questionnaires. These questionnaires are provided in separate datasets and have to be matched for each person within a single year and have to be appended to a dataset covering all years. As a first step, the different questionnaires for each individual within the same year are merged based on a unique indicator. Each observation in the dataset contains a unique household identification number and person in household number. Combining these using the following formula yields a unique person identification number that can be used to identify individuals across years.

𝑃𝑒𝑟𝑠𝑜𝑛 𝑖𝑑 = (ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖𝑑 ∗ 100) + 𝑝𝑒𝑟𝑠𝑜𝑛 𝑖𝑛 ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖𝑑 (1) Due to significant changes in the accommodation and mortgages questionnaire after 2008, we only use data from 2009 onwards. Afterwards, datasets from 2009 till 2019 are appended on this person id, keeping only unique observations. Ultimately, we have a dataset of eleven years which we will use to test our hypotheses. Since the dataset is carefully constructed by an institutional research agency, it does not suffer from implausible values for some of the variables. All household members over the age of 16 within a participating household are interviewed. However, only data is retained from the household member that has filled out the accommodation and mortgages questionnaire. This is done for two important reasons. Firstly, the accommodation and mortgages questionnaire is always completed by only one member of the household, often the head of the household. According to Nyhus and Webley (2001), the head of household has the most influence on households’ financial decision-making. Therefore, only answers from heads of households are considered in this study. Throwing away data from household members who have not completed this questionnaire ensures that in each wave, all households only enter the dataset once. Secondly, the most important independent variable for our analysis, homeownership, comes from the accommodation and mortgages questionnaire.

24 question on homeownership or respond to one of the dependent variables, which makes it impossible to investigate the research question.

The original dataset retrieved from the DHS contained 14.947 observations. After all of the above considerations are taken into account, the result of our panel data is a pooled sample of 6.059 observations in eleven periods, 2009-2019. These observations come from 1.986 individuals over time. Respondents are always head of the household, so the 6.059 observations are from 1.986 unique households. Our resulting panel dataset consists of roughly 3.05 observations on average per individual over 11 different years. The panel is not balanced since the participations are not necessarily consecutive. Consequently, one could neglect to participate during a specific time wave and households that fulfil the inclusion criteria can enter during the sample period. Also in our sample, multiple households drop and join the survey for unspecified reasons. This dataset allows us to test whether homeownership influences the holding of risky assets and the risky asset share of households, while controlling for a broad set of household characteristics.

3.4 Descriptive statistics pooled sample

We start with an overview of our pooled sample with 6.059 observations. We have 4.990 observations for homeowners in the pooled sample. Table 1 shows a large difference in the rate of homeownership in the presence of mortgage. At least, 1.630 households are homeowners without having mortgage and 3.360 households own a home and have mortgage. This table indicates a difference between the presence of mortgage among homeowners of approximately 29% (1.730). There are of course no households with mortgage without owning a home. Finally, 1.069 households are neither homeowners nor mortgage holders.

Table 1. Cross-tabulation mortgage by homeownership.

* This table is derived from our pooled sample with 6.059 observations from 1.986 households. We have 4.990 observations for homeowners in the pooled sample.

Table 2 presents summary statistics of our pooled sample. 44% of the respondents own risky assets and this is about 13% of their total assets. This is in line with previous research on stock market participation in the Netherlands (Guiso et al., 2008). We see that most respondents own a house, because 82.4% of the respondents in the sample own a home with a value of €241.559. We see that home tenure is on average 19 years, but it varies from 0 to 103 years.

Homeownership

Mortgage No Yes Total

No 1.069 1.630 2.699

Yes 0.00 3.360 3.360

25 The houses are burdened with a mortgage of €85.355 on average and a semi-detached, single-family-house is the most common.

The age of respondents varies between 21 and 94. The average age of respondents is 58 and approximately 32% of the individuals are retired. Furthermore, we see that men are over-represented in this paper since only 18.4% of the respondents are women. This can be explained by the fact that men are often head of the household. The marital status dummy shows that 62% of the respondents in the sample have a registered partnership or are married. The average educational level is 5, which corresponds to senior vocational training or training through an apprentice system. The participants assess themselves as relatively healthy. The average self-reported health is 2.1, which denotes a “good” health condition.

In addition, the respondents can be seen as risk-averse with a score of 2. This score indicates that they sometimes took small risks, but certainly no major risks. Planning horizon also scores a 2, which means that respondents use the next couple of years to decide about what part of their income to spend, and what part to save. We could argue that they have a long investment horizon. The respondents earns roughly €34.000 net of tax. This average value plus a standard deviation indicates that the household income in the sample is mostly within €56.000. On average, they have €77.000 in financial assets and €25.000 in non-financial assets. Noteworthy is the low value of liabilities, only €3.000. Nevertheless, large outliers are possible up to almost two million. Net income and financial assets are expressed in real terms. It is important to note that the distribution of financial assets is strongly non-normal. This is evident from the fact that the mean of financial assets is closer to the 75th percentile than the median. Therefore, we use the logarithm of financial assets in the estimations. The logarithm transforms the distribution into a more normal one.

26 The summary statistics of our subsample home tenure are shown in appendix C. As expected, we see that long home tenure leads to 4.4% more ownership of risky assets and 3.6% more risky assets in household portfolios. Furthermore, the majority of households (3.567 vs 1.567) have owned a house for more than 10 years. The other statistics are in line with our expectations. Firstly, recent buyers have €35.875 less home equity and €100.745 more mortgage debt, as expected for recent home buyers. Secondly, they are slightly less wealthy, consistent with being younger on average. Thirdly, they have a slightly higher risk tolerance and a slightly shorter planning horizon. In terms of income and wealth, it is curious that recent homeowners earn €6.820 more on average. Although these results are not a substitute for the more comprehensive analysis that follows, they indicate the basic patterns. It suggests that longer home tenure leads to more risk taking in household portfolio choices.

A high or perfect relationship between explanatory variables precludes estimates of the final regression coefficients from being reliable and thus leads to incorrect conclusions. This phenomenon is called multicollinearity (Brooks, 2019). To detect whether there is multicollinearity between the independent variables, a correlation matrix has been constructed. This matrix has been added to appendix E. A coefficient of 0 would indicate no relation at all. Rules of thumb from basic statistical textbooks together with empirical work from Taylor (1990), indicate that correlation coefficients ranging from 0.500 up till 0.700 are identified as highly correlated.

27 Table 2. Summary statistics of the pooled sample.

Variable Obs Mean Std. Dev. Min Max

Dependent variables

Ownership of risky assets 6.059 0.444 0.497 0 1

Share of risky assets 5.565 0.134 0.233 0 1

Independent variables

Owner of a house 6.059 0.824 0.381 0 1

Value of the house 6.043 241.559 184.341 0 5.450.000 Mortgage amount 6.043 85.355 115.222 0 1.100.000 Home equity 6.043 156.203 174.095 -650.000 5.449.000

Size of the house 6.059 3.292 1.472 1 6

Home tenure 4.989 18.77 12.40 0 103 Demographic variables Age 6.059 58.299 14.46 21 94 Retired (age >67) 6.059 0.323 0.468 0 1 Number of children 6.059 0.485 0.935 0 6 Male 6.059 0.816 0.387 0 1 Married / partnership 6.059 0.619 0.486 0 1 Education level 6.059 5.176 1.462 1 7

Self-perceived health status 6.059 2.132 0.699 1 5 Behavioral variables Risk preference 6.059 2.152 1.018 1 5 Planning horizon 6.058 2.711 1.088 1 5 Economic variables Net income 4.808 33.642 22.080 -2.760 460.315 Financial assets 5.774 76.972 168.976 0 3.410.230 Nonfinancial assets 5.774 25.089 106.160 0 4.461.500 Liabilities 5.774 3.032 30.232 0 1.812.500

Important note: Homeownership refers to the purchase of owner occupied housing, i.e., purchases of

28 4. Methodology

This section contains the methodology and several conditions needed to process the DHS data. As stated before, the data consists of unbalanced data, which is common with panel data (Torres-Reyna, 2007). Our first section deals with the choice between random effects and fixed effects models. Subsequently, we discuss in section two the methodology regarding the participation decision of households. Our third section discusses the methodology for the allocation decision of household’ assets to the stock market. We conclude this chapter with a discussion of methodological limitations.

4.1 Random effects versus Fixed effects.

Our sample consists of panel or longitudinal data and has observations of individuals over time. Panel data modelling is a method to estimate data which is both time-series and cross-sectional (Brooks, 2019). In addition, it accounts for individual-specific heterogeneity. This makes a simple pooled OLS approach not suitable for our analysis. Pooled OLS is suitable when using different samples for each period of the panel data (Wooldridge, 2010). However, we are going to observe the same sample of individuals over time and pooled OLS does not recognizes this kind of panel structure. Instead, there are broadly two classes of panel data models that can be used in financial research: fixed effects and random effects. The use of random effects in relation to pooled OLS is supported by the result of the Breusch and Pagan Lagrangian multiplier test. This test yielded a significant result in favour of random effects4. Consequently, we are able to reject the null hypothesis that there is no evidence of significant differences across entities i.e., no panel effects.

To control for possible unobserved heterogeneity, we could include fixed effects. The purpose of fixed effects is to capture the time variation in our variables. Furthermore, it treats the unobserved individual heterogeneity (αi) for each respondent to be correlated with our explanatory variables. Fixed effects estimation involves a transformation to remove the unobserved individual heterogeneity prior to the estimations. The fundamental difference between fixed effects and random effects is that the latter assumes that the individual-specific effects (αi) are independent of the regressors (Brooks, 2019). This individual-specific effect is included as the error term with random effects.

29 The Hausman test provides information about which estimator is more appropriate for our analysis. By rejecting the null hypothesis that the error term is uncorrelated with the independent variables at 5 percent level (0.05 <p), we conclude that fixed effects is more appropriate than a random effects model5_{. This is the main disadvantage of the random effects}

model. Individual characteristics, which are not observed, are rarely uncorrelated with the independent variables. Nevertheless, we opt for random effects. If we use fixed effects, we lose the ability to determine the influence of time-invariant variables that affect our dependent variable, but do not vary over time. Consequently, individuals whose characteristics do not change during the sample period will not contribute to the estimation. Moreover, the table in appendix F shows that the within-variance in the dependent variables is not enough for a fixed-effects approach. Some researchers argue, and demonstrate, that when within-group variation is small, the widely used Hausman test to decide between estimator type (FE vs. RE) is problematic, as it relies on both the between- and within-variations to be large enough (Hahn et al., 2009). Within variation for our dependent variable ownership of risky assets is 21.3% and for the share of risky assets it is only 10.2%. In addition, numerous of our control variables are time-invariant where the between variation is also obviously larger than their within variation.

Another potential motivation for random effects concerns the length of our panel. According to Cameron and Trivedi (2009), a short panel has many entities (large N) but few time periods (small T). A large panel instead has many time periods (large T) but few entities (small N). We believe that our panel falls between the two, but that it has more in common with a short panel. When Torres-Reyna (2007) refers to a large panel, he mentions a period of 20 to 30 years. Our panel contains only eleven years of data and we have many entities. Consequently, a short panel with a large number of variables may cause bias in the fixed effects estimates due to the incidental parameter problem. As a result, the number of nuisance parameters grow quickly as the number of individuals increases (Greene, 2004). It is caused by only having T observations to estimate each αi, As a result, as N grows, the estimate of αi remains random. In linear models this randomness gets averaged out. In nonlinear models like a fixed effects model, it does not. Having said this, fixed effects are generally inconsistent in nonlinear models as N grows with T fixed. Using Random effects is further supported by Torres-Reyna (2007). He argued: “if you have reason to believe that differences across entities have some influence on your dependent variable then you should use random effects”. As a result, random effects estimates will be employed in this paper.

30

4.2 Methodology participation decision

The existing research on the topic proposes various econometric approaches. Some studies have estimated OLS regressions of portfolio shares on property values, mortgage debt and home equity with various control vectors and obtained mixed results (Heaton and Lucas, 2000; Cocco, 2005; Yao and Zhang, 2005). To test the effect of homeownership on stock market participation, we must keep in mind that stock market participation is binary. Since stock market participation in binary, linear regression analysis is unsuitable for examining the effects of home ownership on a household's likelihood to participate in the stock market (Wooldridge, 2010). Dey and Astin (1993) suggest to use either a Logit or Probit model when the dependent variable is binary. The coefficients of these models will predict the conditional probability of a successful outcome that 𝑦𝑖𝑡 = 1| 𝑥′𝑖𝑡. Both models are non-linear models and estimated using

maximum likelihood. According to Brooks (2019), the Logit and Probit models give similar results and the choice to use one of those methods is arbitrary. The models differ in the fact that the Logit model uses the cumulative distribution function of the logistic distribution and the Probit model uses the cumulative distribution function of the standard normal distribution (Dey and Astin, 1993). The Probit model overcomes the limitation of the linear probability model, producing estimated probabilities that are negative or greater than one. In line with other empirical papers in the field of household portfolio choice (see, e.g., Rosen and Wu, 2004; Dimmock and Kouwenberg, 2010), we use a Probit model to test the relationship regarding the participation decision. The estimates and standard errors of the Probit models are obtained using the standard maximum likelihood procedure. Hence, stock market participation is examined by the following random effects Probit model:

𝑦it = α + 𝛽𝑥𝑖𝑡 + ω𝑖t, ω𝑖t = u𝑖 + v𝑖t v𝑖t ~ NID (0, 1), 𝑖 = 1, 2, … , 𝑁, (2)

𝑦𝑖𝑡 = 1 𝑖𝑓 𝑦it > 0, (3)

= 0 𝑖𝑓 𝑦it ≤ 0. (4)

The u𝑖 are assumed independent of the traditional error term (v𝑖t) and explanatory

variables (𝑥𝑖𝑡), which are also independent of each other for all i and t. In addition, the errors

31 and control variables, including year dummies. The last term ω𝑖t is our composite error term, it

captures other sources of heterogeneity in portfolios. In this error term, v𝑖t measures the random

deviation of each entity’s intercept term from the global intercept term α. The other term, υ𝑖t,

represents the individual specific random effect; it has a global constant with a random disturbance around it. After estimating the model, the effects of the variables will be evaluated at the mean with marginal effects, as is conventional when conducting a Probit model. In addition to our Probit model, the Logit model will be used for robustness of our main results in line with Love and Smith (2010).

4.3 Methodology allocation decision (Tobit)

As explained earlier, much recent research into the effect of homeownership on portfolio choices has been done by Messrs Raj Chetty and Adam Szeidl. Our work is closely related to their work, but less in depth given the time and word limitations of our paper. In addition to following Chetty and Szeidl, the work of Love and Smith (2010) has common grounds with our study. They investigated how much of the connection between health and portfolio choice is causal and how much is due to the effects of unobserved heterogeneity. Their strategy entails a censored regression model (Tobit). According to them, a Tobit model is applied in the case of allocation of risky assets; it tackles the issue of clustering at point zero. It is therefore a useful model, because it takes into account selection effects. As explained, not all households own risky assets. This is also the case in our sample, ownership of risky assets is only 44%.

The allocation measure is a limited dependent variable and skewed into one direction. In line with Rosen and Wu (2004), a random effects Tobit model is employed to test the relationship of interest. The Tobit model allows regression of such a dependent variable while it is being censored, so that regression of a dependent variable can occur. It allows us to specify a lower or upper bound to censor the regression, while preserving the linear assumptions required for linear regression (Brooks, 2019). As a result of this natural boundary at zero, the non-response answers on share of risky assets are converted to zero and therefore analysed as no stock and mutual fund holdings. The standard Tobit model can be described by several equations with estimation conditions. This results in the following equations for the random effects Tobit model:

𝑦∗i = 𝛽0 + 𝑥′𝑖𝛽 + 𝜀𝑖, 𝑖 = 1, 2, … , 𝑁, (5)

𝑦𝑖 = 𝑦∗i 𝑖𝑓 𝑦∗i > 0, (6)

32 The construction of the model will be identical to the variables used in the random effects Probit model. However, this time our dependent variable is left-censored at zero and this will be the amount of stocks and mutual funds without missing values. In addition, the observed share of net household wealth invested, is equal to the latent variable if the latent variable is above zero and zero for the other observations. After estimating the model, the effects of the variables will be evaluated at the mean with marginal effects.

4.4 Methodological limitations

Our research method is accompanied by several limitations. We successively discuss interpretation, endogeneity, heteroskedasticity, serial correlation and time-specific effects. Firstly, a drawback of Probit and Tobit models is that the coefficients are difficult to interpret. It has no direct interpretation in terms of the probability that Yi = 1. It only provides information about the sign of the coefficient and its relationship with the probability that Yi = 1 (Dey and Astin, 1993). To be able to interpret the coefficients better, we estimate marginal effects to determine the magnitude. The margins are calculated for each variable at the mean, as is common with Probit and Tobit estimations (Torres-Reyna, 2007). The marginal effects provide the change in the probability of an outcome on risk-taking when an independent variable is increased by 1 unit.