Health Deterioration and Portfolio Risk: Evidence From the Netherlands1

(1)

Health Deterioration and Portfolio Risk: Evidence

From the Netherlands

1

T.G. (Ties) Busschers

2

10-1-2019

MSc. Thesis

Abstract

This thesis investigates the positive relationship between the holdings of risky assets and an individual’s health. Specifically, it is tested whether a decrease in an individual’s self-expected likelihood of reaching a certain age, leads to lower holdings of risky assets. Thereby, this thesis is the first to employ this novel health indicator in the health-portfolio nexus. The results are compared with a health indicator that is well embedded within the literature, self-assessed health. However, the new health indicator has the edge that it explicitly takes the time horizon of the investor into account. By employing the LISS dataset, evidence is found that a one standard deviation decrease in self-assessed health and in the new health indicator, reduces the probability of holding risky assets with 1.58 and 1.59 percentage points, respectively. A similar consistent positive relationship between both health indicators and the share of risky assets in financial wealth is found. Risky assets are transformed into riskless financial products after a health deterioration.

Keywords: Portfolio Allocation, Health Economics, Background Risk JEL Classification: D14, G11 and I12

Supervisors: dr. J. O. (Jochen) Mierau & R. D. (Roel) Freriks, MSc. Course Code: EBM866B20

1_{I would like to express my gratitude to Roel Freriks and Jochen Mierau for their excellent} guidance and advice. Without their help this thesis would not have existed.

(2)

1

1. Introduction

Portfolio allocation is a straightforward and uncomplicated matter, according to the stylized classical models of portfolio allocation. As pointed out by Markowitz (1952), all individuals should hold the same diversified portfolio. In turn, the level of risk averseness, embedded within the individual, determines the proportion of wealth allocated to risky assets. This proportion is constant and orthogonal to an individual’s health, time horizon and age. This notion, however, is disputed. For example, an old financial adage, 100 minus your age, proposes that an individual should allocate a 100 percent minus your age as a proportion of wealth to risky assets. Moreover, this notion contrasts to many empirical findings that relate financial risk-taking behavior to age (Ameriks and Zeldes, 2004).

Recently, the notion that the share of risky assets in total wealth is orthogonal to an individual’s health condition, is challenged by researchers (Rosen and Wu, 2004; Bressan, Pace and Pelizzon, 2014). These researchers, among others, took an interest into the empirical relationship between an individual’s health and portfolio allocation. These researchers provide empirical results on how portfolio decisions are affected by an individual’s health. Most frequently, they find that a deterioration of an individual’s health leads to a lower proportion of financial wealth allocated to risky assets. Furthermore, this leads to disaffection from the equity market. This thesis confirms the positive relationship between health and the holdings of risky assets. An addition to the current literature is provided, by employing a new health indicator that explicitly takes the time horizon into account.

The literature voices several ways in which an individual’s health affects the portfolio allocation. Some argue that health directly enters the utility function, where an enhanced mortality risk lowers the maximum financial risk one can take (Edwards, 2008). Others propose that a deterioration of one’s health is associated with substantial medical expenditures. In order to finance these medical expenditures, financial assets are liquidated. Resulting in lower holdings of risky assets (Ayyageri and He, 2017). Lastly, some consider bequest motives and time horizon as factors driving the relationship between health and portfolio choice (Feinstein and Lin, 2006; Coile and Milligan, 2009).

(3)

2

Often this is done with the inclusion of an individual’s age in the analysis. Dow (2009) notes that age is often unrelated to the time horizon of the investor. For this reason, a new health indicator is developed. This indicator takes the time horizon directly into account. The health indicator measures a change in the self-expected likelihood of reaching a certain (distant) age. A lower likelihood is recognized as a health deterioration. Furthermore, the health indicator dominating the current literature, self-assessed health, is flawed. This thesis combines the two pressing issues in the health-risky assets literature. An additional health indicator is provided, while simultaneously taking the time horizon properly and directly into account. The main question of this thesis is whether both health indicators yields results that are consistent with each other.

The analysis is conducted on the Longitudinal Internet studies for the Social Sciences (LISS). LISS gathers information on approximately 7000 individuals, this sample is representative for the Dutch national population. The existing literature tends to employ datasets that are narrowed down to the oldest groups of society. Therefore, conclusions drawn by other researchers in the literature, are limited to the elderly. By employing the LISS datasets, conclusions can be drawn that are population-wide.

This thesis adds to the literature by being the first that employs a dataset that is a representative sample of a population. The most innovative part of this thesis is the inclusion of the new health indicator, change in self-expected likelihood of reaching a certain age, in the health-risky assets nexus. This variable takes the time horizon explicitly into account, which has not been done properly yet. The two health indicators yield consistent results with each other, in both sign and magnitude.

The sections below are, in sequence, the following: literature review, data section, empirical section and conclusion and discussion.

2. Literature review

(4)

3

The use of self-assessed health as a health indicator has risen over the last few decades, and the use of this variable is popular (Doiron et al., 2015). This is remarkably noticeable in the health-risky assets literature. Rosen and Wu (2004) are the first to explicitly investigate the relationship between health and portfolio allocation. Rosen and Wu (2004) construct a binary variable that equals one if the person has a self-assessed poor (or worse) health status, and zero otherwise. Rosen and Wu (2004) find that a poor self-assessed health reduces the likelihood of holding risky assets with a statistically significant amount of 1.7 percentage points. Further, Rosen and Wu (2004) find that a poor self-assessed health increases the share of safe assets in financial wealth with approximately 5 percentage points. Whereas the ratio of risky assets to financial wealth is decreased with approximately 9 percentage points. No evidence is found that a potential third variable problem dilutes the results 3. Cardak and Wilkins (2009) and Cai et al. (2013) find results consistent with Rosen and Wu (2004) on, respectively, an Australian and Chinese dataset.

Love and Smith (2010) are unable to find the statistically significant effects of self-assessed health on portfolio characteristics. Love and Smith (2010) control for unobservable heterogeneity of the respondents with a fixed effects estimator, this analysis largely nullifies the results.

Some scholars voice their concerns with the use of self-assessed health as a health measure. Rice and Yardley (1991) find a strong significant effect of mood on subjective well-being. Joiner, Schmidt and Telch (1996) find that mental issues, unrelated to physical health, affect how people perceive their physical health. Further, Joiner, Schmidt and Telch (1996) find that individuals suffering from panic disorders, perceive that they have a lower physical health. This is unrelated to any illness. Moreover, Joiner et al. (1996) note that an individual’s subjective health is strongly associated with short-term mood and emotions. Furthermore, many discrepancies between self-assessed health and more objective health measures are found. (Doiron et al., 2015). For these reasons other researchers employ other health indicators. This approach is, for instance, taken by Fan and Zhao (2009), and Bressan, Pace and Pelizzon (2014). Fan and Zhao (2009) construct four health indices, covering various dimensions of health. The inability of fulfilling daily tasks, and whether the respondent has experienced a stroke or heart attack, decreases the share of risky assets in total assets with approximately 2

(5)

4

and 1 percentage points, respectively. Another attempt at covering various dimensions of health by employing different health measures, is Bressan, Pace and Pelizzon (2014). They employ self-assessed health, the presence of chronic diseases, and mental health as the health indicators. Only self-assessed health has a statistical positive association with the risk characteristics of the portfolio.

As noted above, there are some attempts to address the various dimensions of health in the health-risky assets nexus. Some researchers voiced their concerns on self-assessed health as a health indicator (Joiner et al., 1996). A different issue within this literature, and portfolio choice in general, is the inclusion of the time horizon of the investors. Researchers, such as Rosen and Wu (2004) and Edwards (2008), remark that the time horizon of an individual is important to include in the analysis of portfolio choice. All researchers in the health-risky assets nexus, except Edwards (2008), include the time horizon of the investor with the inclusion of age in the analysis. Rosen and Wu (2004) argue that age varies with the time horizon of the investor. However, Dow (2009) finds that age is not closely related to the reported time horizon of the investor. Thus, the time horizon of the investor is unsuccessfully and unsatisfactorily accounted for.

Edwards (2008) takes the time horizon of the investor into account by the inclusion of a covariate that measures expected remaining years. This variable hinges on some unlikely assumptions 4, and is added next to the health variable. In this thesis time horizon is directly taken into account via the health indicator. A shortened time horizon could be the result of a deterioration of the health condition. Thereby, it can act as the sole health indicator (instead of acting as a control variable, as done by Edwards, 2008), while simultaneously taking the time horizon explicitly into account. Which has not been taken into account sufficiently so. The main question of this thesis is whether both health indicator yield results that are consistent with each other.

In conclusion, the literature finds a positive relationship between health and the holdings of risky assets on non-representative datasets. In this thesis a national representative dataset will be employed. Most scholars employ self-assessed health as a health indicator, however this health indicator is flawed. Further, the time horizon is unsuccessfully accounted for, by many researchers (Rosen and Wu, 2004; Edwards, 2008). In short, this thesis adds on the existing literature by taking the time horizon explicitly into account, via the health indicator.

(6)

5

Simultaneously, this thesis provides an additional health indicator, that can be used next to, or even instead of, self-assessed health.

3. Data

3.1 Data Description

The data to be used in this thesis is the Longitudinal Internet Studies for the Social Sciences (LISS) dataset. It consists of approximately 7000 individuals. The LISS panel is administrated by CentERdata (Tilburg University, the Netherlands). The panel is based on a true probability sample of households. The LISS panel consists of different cross-sectional datasets, covering a wide range of topics. For this thesis the following cross-sectional LISS panel datasets are employed: Background Variables; Health; Economic Situation: Assets;

Economic Situation: Income; Economic Situation: Housing.

The questions on self-expected likelihood of reaching a certain age were first asked to the respondents in 2009. A change of the likelihood of reaching a certain age is used as the main independent variable. Therefore, the wave of 2009 is dropped from the analysis. The

Health dataset is released annually. The last Health wave that can be used, is conducted in 2015.

Respondents are asked to value financial assets at the last day of the previous year. Assets is released only biennially. Therefore, Assets of the waves 2012, 2014 and 2016 are used 5, covering data on the years 2011, 2013 and 2015. Thereby, the analysis is reduced to the years 2011, 2013 and 2015. Further, the remaining surveys cover data of the years 2011, 2013 and 2015.

Unfortunately, a large number of respondents were unable to complete the survey properly, resulting in large amounts of observations being dropped from the analysis 6_. Improper responds include, for instance, illogical responses. Only respondents that are responsible for the financial matters in the household are taken into account.

5_{Assets 2010 cannot be used, because it covers data on 2009. Data on the year 2009 is dropped,} as mentioned above.

(7)

6

3.2 Variables

3.2.1 Dependent Variables

The dependent variables are: the probability of holding risky assets; the share of risky assets to total wealth and financial wealth; and the amount of risky assets. Respondents are asked whether they were in the possession of risky assets on 31 December in the year, prior to the year the survey is conducted in. Risky assets include growth funds, share funds, bonds, debentures, stocks, options, warrants and other non-riskless financial assets. Unfortunately, the LISS panel dataset does not provide the possibility to differentiate between stocks, options, warrants on one hand, and growth funds, share funds, bonds and debentures on the other hand. For this reason, risky assets should be interpreted as assets that are more risky than completely riskless assets.

Respondents are required to fill out the exact amount of risky assets they possess. If a respondent does not know the exact amount of the risky assets, they are asked to appoint the total amount of risky assets to certain pre-determined value ranges. In this follow-up question they are allowed to state, again, that they do not know the exact value. Observations unwilling or unable to provide the exact amount, are dropped from the analysis. For those that assign their amount of risky assets to a value range, the median value of the range is chosen as the amount of risky assets 7.

For the dependent variables, a measure of total wealth is required. Total wealth is defined as liabilities deducted from assets. Financial wealth is the sum of savings and risky assets minus current liabilities. Outliers are deleted. The two dependent variables that are ratios, are not necessarily in the range of zero and one. For example, a negative ratio implies that the person is short in risky assets. Values of this ratio that are smaller than -0.1, or larger than 1.1, are deemed an outlier, and therefore excluded from the analysis 8_.

7_{The median value is determined on those that do know their risky assets, and possess risky} assets within that value range.

(8)

7

3.2.2 Independent Variables 3.2.2.1 Self-Assessed Health

As a starting point of the empirical part of this thesis, self-assessed health is employed as the health indicator. This health indicator is well embedded within the literature. This variable ranges from one to five, where one corresponds to poor health and five to excellent health.

3.2.2.2 Change in Expected Likelihood of Reaching a Certain Age

The health measure to be used in this thesis is an explicit change in time horizon. Edwards (2008) includes a similar, although significantly different, variable. This variable is included along a self-assessed health variable in the analysis. In this thesis a change in time horizon of an investor will act as the health measure, rather than a control variable.

Data availability problems restrict Edwards (2008) to make some improbable assumptions. These assumptions are not required in the health measure in this thesis. The variable to be used in this thesis equals one if the person has experienced a lower self-assessed likelihood of reaching a certain age, compared to previous wave, and a zero otherwise. Thus, the variable measures a shortening in the life horizon of the person through a health deterioration 9. This variable is called Health Depreciation.

3.2.3 Control Variables 3.2.3.1 Assets

Some researchers control for total wealth (Cardak and Wilkins, 2009), while others decompose total wealth (Bressan, Pace and Pelizzon, 2014). This thesis will decompose total wealth into assets and liabilities. This allows for different effects of assets and liabilities on the holdings of risky assets. Love and Smith (2010) note that it is imperative to control for total wealth and income. For the reason that many unobserved factors that simultaneously affect stock holdings and health, can also affect wealth and income 10_{. Moreover, the level of assets} can proxy for socio-economic status, the relationship between socio-economic status and health is well embedded within the literature (Adler and Ostrove, 1999).

9_{Appendix A contains a detailed description of the construction of this variable.}

(9)

8

Assets is calculated as the sum of current and non-current assets. Current assets comprise of the amount of risky assets, savings and money lent out to friends and family, and a remaining part 11. Non-current assets consist of real estate; vehicles 12; the share in the equity of a private limited companies, financial partnerships and one-man companies and the value of the house(s) the respondent owns. All assets are valued at the last day of the year, prior to the year the wave is conducted in. Except houses, these are valued as the expected market value at the moment the survey is conducted 13_{. The questions on the value of assets follow the same} method as the determination of the value of risky assets. The respondents unwilling or unable to answer the question on house value are inquired about the WOZ-value of their house 14_.

Berkowitz and Qiu (2006) find an asymmetric effect of liquid and illiquid assets on the share of risky assets to (financial) wealth. Therefore, current assets (liquid) and non-current assets (illiquid) enter the empirical analysis separately. To detect non-linear effects of current and non-current assets on the dependent variable of interest, the squares of the respective variables is added. This replicates the procedure of Rosen and Wu (2004) and Edwards (2008), they include the square of every wealth variable they include.

3.2.3.2 Liabilities

High levels of liabilities might deter individuals to obtain risky assets, as the individual is unable to possess risky assets, due to a substantial financial burden. Further, high levels of liabilities can cause a health deterioration (Adam et al., 2013). Thus, liabilities is added as a control variable, since it simultaneously affects health and risky assets.

Liabilities is the sum of current and non-current liabilities. Current liabilities constitute to short-term debt; credit card debt and short-term credit. Non-current liabilities comprise of mortgages on real estate and houses, and study grants provided by the government. The

11_{The remaining part consists of cash money, musical instruments, art works, jewelry,} collections and so on.

12_{Cars, motorcycles, boats and (static) caravans.}

13_{All house values are discounted with a general Dutch house price index (HPI) to the correct} date (Centraal Bureau voor de Statistiek, Kadaster, 2018).

(10)

9

determination of debt values follows the same pattern as the method employed for the value of assets.

For a potential asymmetric effect of liabilities on the holdings of risky assets, the analysis differentiates between current and non-current liabilities. Again, the squares of current and non-current liabilities are added to detect any non-linear effects of these variables on the dependent variables. This follows the procedure in Rosen and Wu (2004) and Edwards (2008).

3.2.3.3 Remaining Control Variables

The remaining control variables are mainly demographic variables. These variables are conventional in the literature on explaining portfolio behavior, on US and European data (Bertaut and Starr-McCluer, 2002; Carroll, 2002). The inclusion of these demographic variables can be justified by their association with preferences for risk, and thus the holdings of risky assets, and an individual’s health condition. Issues of a third variable bias are mitigated by the addition of control variables that affect simultaneously health and the holdings of risky assets.

Although the direction of the relationship between age and risk appetite is disputed, there is a consensus that a relation between age and risk appetite exists (Morin and Suarez, 1983). It is evident that health and age are related, generally when an individual ages, his or her health deteriorates (Deeks et al., 2009). The square of age is included to detect and non-linear effects of age on the holdings of risky assets.

Gender is included, which equals one for men and zero for women. The relation between gender and risk appetite is well embedded in the literature (Böheim and Lackner, 2015). Although, the direction of gender and health is unclear, it is evident that health and gender are related (McDonough and Walters, 2001). For this reason, gender is controlled for.

Whether a respondent lives together with a partner is included, as a significant effect of relationship status on the risk level of the portfolio is found in Rosen and Wu (2004). A large amount of literature on relationship status, health and health behavior indicates that marriage conveys health benefits (Fuller, 2010).

(11)

10

to the respondent only having finished primary education, and six corresponds to a university degree.

The number of children living in the household is included, as this may affect portfolio choice through a bequest motive (Edwards, 2008). Moreover, parenthood is associated with health. Parenthood may induce individuals to have health attitudes that reduce their sensitivity to diseases (Verbrugge, 1983).

Next, the level of urbanization in the area the respondent lives, is included. This number ranges from one to five, where one corresponds to at least a population density of 2500 citizens per square kilometer, and a five corresponds to less than 500 citizens per square kilometer. Although the literature on the effect of urbanization on risk appetite is limited. Shi and Yan (2017) find some evidence that living in an urban area enhances risk appetite. At the same time, urbanization does affect health, for example through the large availability of unhealthy food in the city (World Health Organization, 2010).

(12)

11

Table 1

Descriptive Statistics (N=1729)

Name Min. Max.

Mean (All) Mean (HRA=0) Mean (HRA=1) Mean (HD=0) Mean (HD=1) Demographic Variables Age ₂₁ 85 58.540 58.124 59.840 57.020 59.544*** Gender ₀ 1 0.613 0.584 0.704*** 0.602 0.621

Living Together with a Partner ₀ 1 0.648 0.657 0.621 0.640 0.654 Educational Attainment ₁ 6 3.942 3.786 4.427*** 3.955 3.933 Number of Children at Home ₀ 5 0.501 0.515 0.458 0.520 0.488 Urbanization of Area of Living ₁ 5 3.086 3.134 2.936*** 3.078 3.090 Health Variables Self-Assessed Health 1 5 3.092 3.062 3.186*** 3.058 3.114 Health Depreciation (HD) ₀ 1 0.602 0.614 0.566* 0 1*** Financial Variables (x1000’s) Total Wealth _-49 2022 236 193 370*** 237 235 Total Assets _0.0 2322 356 313 491*** 363 352 Current Assets _-9 1300 61 35 141*** 60 61 Non-Current Assets ₀ 1431 296 278 351*** 303 291 Total Liabilities ₀ 809 121 121 121 126 117* Current Liabilities ₀ 222 2.758 3.018 1.947 2.816 2.720 Non-Current Liabilities ₀ 809 118 117 119 123 114*

Monthly Income Net of Taxes ₀ 11.000 1.972 1.840 2.383*** 1.976 1.968 Dependent variables

(13)

12

We begin our exploration of the relationship of health and risky assets by showing the (control) variables and how they relate to a variation in health and risky assets. Table 1 displays the descriptive statistics of the variables to be used in the empirical analysis. Further, table 1 depicts the mean differences for those that do hold risky assets and those that do not possess risky assets. Due to the non-normality characteristics of income and wealth variables, percentile scores are presented in table 2. The probability of being a male, and educational attainment are higher for those holding risky assets, compared to those that do not possess risky assets. Whereas, a lower level of urbanization in the area of living is associated with a lower probability of holding risky assets.

Moreover, the mean differences of those that experience a depreciation in their health and those that do not experience this, are displayed in table 1. Those experiencing a lower likelihood of reaching a certain age, seem to have the same demographic characteristics, except age, as those that do not. The subsample 𝐻𝐷 = 1 have a lower likelihood of holding risky assets, lower amount of risky assets and lower risky assets ratios than the subsample 𝐻𝐷 = 0. However, only the probability of holding risky assets is statistically significant. Nonetheless, some evidence is shown that a decreased likelihood of reaching a certain age is associated with lower holdings of risky assets.

Table 2

Percentile Scores of Financial Variables (N=1729) in Thousands (x1000’s)

Percentiles Name 10% 25% 50% (Median) 75% 90% Total Wealth 19 78 180 322 506 Total Assets 170 217 291 416 623 Current Assets 2 9 26 60 153 Risky Assets 0 0 0 0 30

Risky Assets if HRA=1 3 9 24 60 180

Non-Current Assets 149 194 249 349 503

Total Liabilities 0 30 102 182 260

Current Liabilities 0 0 0 0 1.800

Non-Current Liabilities 0 27 100 175 255

(14)

13

4. Empirical Results

4.1 Self-Assessed Health 4.1.1 Empirical strategy

The primary objective of this thesis is to assess the relationship between health and the appetite for risky assets. This assessment will be executed with the aid of Health Depreciation. However, the starting point of the empirical analysis follows the typical approach taken by the existing literature, such as Rosen and Wu (2004). That is, self-assessed health is the initial health indicator. This is done in order to compare the results in this thesis with others employing non-representative datasets. Further, the results of this health indicator will be compared to the results with Health Depreciation as the health indicator.

Common practice in the literature, is a regression analysis with random effects. However, the assumptions of the random effects estimator are unlikely to hold 15. The polar opposite of the random effects estimator, is the fixed effects approach. The fixed effects estimator does not have this restrictive assumption. However, all time invariant observations are deleted by the fixed effects estimator. Due to the time invariant characteristics of the dependent variables, random effects will mostly be employed.

Formally, the Hausman test decides whether the random or fixed effects estimator is appropriate. In order to employ the Hausman test, one should be able to execute the fixed effects estimator. For this reason, the Hausman test can only be completed for the analysis with the dependent variable Amount of Risky Assets. This variable is sufficiently time variant. The Hausman test does not reject the null hypothesis of using the random effects estimator 16_. However, since the fixed effects estimator controls for time invariant unobserved heterogeneity, a fixed effects estimator is executed as well.

A tobit model will be employed to manage the majority of respondents possessing zero euros of risky assets 17_{. The tobit model will have truncation at zero, to deal with the fact that} most respondents have zero euros in risky assets.

15_{For example, respondents that have an occupation with a high social status are more likely to} experience moments of stress. This in turn, can affect health. Thus, correlation between the error term and the explanatory variables exists.

16_{The Hausman test statistic is 𝜒}2_{(7) = 9.09, resulting in 𝑝[𝜒}2_{(7)] = 0.246.}

(15)

14

Table 3 specification 1 is a logistic random effects regression with the dependent variable that measures whether a respondent possesses risky assets or not. Specification 2 is the random effects model with a dependent variable that measures the amount of risky assets one possesses. Specification 3 is the tobit model. Specification 4 is the fixed effects model. In table 4 specification 1, the share of risky assets to total wealth is regressed on self-assessed health and the list of covariates. In specification 2 the dependent variable is interchanged for the share of risky assets in financial wealth. Robust standard errors, if the statistical software allows it, are employed to take heteroskedasticity into account. Robust standard errors are unavailable for tobit models, therefore standard errors are bootstrapped with 100 iterations 18_{. No alarmingly} high correlations of the independent variables are found, collinearity is of no concern.

(16)

15

Table 3 Dependent Variable:

Risky Assets

Ownership Amount of Risky Assets Demographic Variables (1) (2) (3) (4) Self-Assessed Health 0.022* 2077 8352* 2272 (0.012) (1712) (4611) (2109) Age 3.20e-04 404.1 840.1 6082** (0.005) (463) (1394) (2530) Age2 1.21e-06 -2.304 -3.708*** -15.39 (4.31e-05) (4.565) (12.41) (10.38) Gender 0.022 2848 3504 Omitted (0.021) (2200) (7115) Omitted Partner -0.048** -778.3 -9365 -2386 (0.020) (2045) (7340) (1764) Educational Attainment 0.021*** 2089*** 9150*** 3537 (0.008) (852.7) (2543) (2600) Number of Children in Household 0.007 -445.9 1096 -1094

(0.011) (1040) (4007) (1674)

Urbanization -0.011 -1478 -5052* 4532

(0.007) (1257) (2896) (6102) Financial Variables

Current Assets _2.03e-06*** 0.240*** 0.700*** 0.169*** (2.06e-07) (0.057) (0.092) (0.059) Current Assets2/103 -1.36-09*** 2.21e-04*** -1.15e-04 3.33e-05

(1.68e-10) (6.24e-05) (8.58e-05) (7.52e-05)

Non-Current Assets 7.57e-08 0.012 0.038 4.17e-04

(1.45e-07) (0.020) (0.065) (0.015) Non-Current Assets2_/103

-2.64e-11*** 1.17e-05 -3.15e-05 -1.62e-06 (4.64-12) (1.93e-05) (5.65e-05) (1.47e-05)

Current Liabilities 1.81e-06 0.154* 0.720 0.160

(1.31e-06) (0.088) (0.762) (0.103) Current Liabilities2_/103 _-1.75e-08 _-0.001** _-0.006 _-0.002* (1.37e-08) (0.001) (0.169) (0.001) Non-Current Liabilities -3.08e-08 0.026 0.062 0.021

(1.98e-07) (0.030) (0.076) (0.029) Non-Current Liabilities2/103 -8.49e-11 -1.61e-05 -8.64e-05 -6.04e-05

(1.90e-10) (6.00e-05) (1.38e-04) (5.76e-05) Monthly Income Net of Taxes 1.69e-05 0.609 0.460 6.149*

(1.11e-05) (1.571) (3.096) (3.359)

Control for Year Effects Yes Yes Yes Yes

Number of observations 1729 1729 1729 1729

R2_(within) _- _0.236 _- _0.271

(17)

16

Notes: */**/*** indicate significance at the 10%/5%/1%. Specification 1 and 2 are a tobit model with truncation at zero. Average marginal effects are displayed. For both specifications the standard errors are bootstrapped with 100 iterations.

Table 4 Dependent variable: Risky Assets-Total Wealth Ratio Risky Assets- Financial Wealth Ratio Demographic Variables (1) (2) Self-Assessed health 0.024 0.049* (0.019) (0.028) Age -0.003 0.001 (0.006) (0.011)

Age2 _3.60e-05 _1.99e-05

(4.82e-05) (9.84e-05) Gender 0.038** 0.066 (0.019) (0.054) Partner -0.051** -0.089 (0.022) (0.054) Educational Attainment 0.027*** 0.071*** (0.007) (0.020) Number of Children in Household 0.014 0.020

(0.012) (0.028)

Urbanization -0.009 -0.025

(0.007) (0.018) Financial Variables

Current Assets 1.95e-06*** 3.26e-06***

(2.19e-07) (4.40e-07) Current Assets2/103 -1.27e-09*** -2.32e-09***

(2.18e-10) (5.90e-10)

Non-Current Assets -1.74e-07 3.53e-07

(1.76e-07) (3.49e-07) Non-Current Assets2/103 -3.02e-12 -2.34e-10

(1.35e-10) (2.42e-10)

Current Liabilities 1.28e-06 4.48e-06

(2.59e-06) (8.65e-06) Current Liabilities2/103 -1.22e-08 -3.88e-08

(6.11e-08) (2.16e-07)

Non-Current Liabilities 3.59e-07* 4.491e-08

(1.94e-07) (4.46e-07) Non-Current Liabilities2_/103 _-1.80e-10 _1.57e-10

(3.36e-10) (9.07e-09) Monthly Income Net of Taxes 3.79e-06 1.32e-05

(9.17e-06) (1.69e-05)

Control for Year Effects Yes Yes

(18)

17

4.1.2 Results

Table 3 and 4 depict the results of the regression with self-assessed health as the health indicator. In table 1 it is shown that the self-assessed health variable ranges between one and five. Therefore, the result of specification 1 can be interpreted as follows: if self-assessed health status decreases with one point, on average, the probability of stock ownership diminishes with 2.2 percentage points. However, Rosen and Wu (2004) constructed a dichotomous variable that indicates whether a respondent has a poor self-assessed health or not. They find that experiencing poor health decreases the probability of holding risky assets with 1.7 percentage points. Following this procedure, an average marginal effect of 3.5 percentage points is found, this result is not formally presented. Thus, this thesis finds a moderately more pronounced effect of self-assessed health on the probability of holding risky assets. Limited evidence is found that self-assessed health is statistically associated with the amount of risky assets. Only one of the three specifications is statistically significant.

In table 4 we find that a one-point increase of self-assessed health raises the share of risky assets in financial wealth with 4.9 percentage points. For reasons of comparability a binary variable is constructed, as done in Rosen and Wu (2004) and Cardak and Wilkins (2009). This approach finds that poor self-assessed health lowers the share of risky assets in financial wealth with 7.2 percentage points. The 7.2 percentage points is between the 9.9 found in Rosen and Wu (2004) and the 5.0 percentage points in Cardak and Wilkins (2009). Self-assessed health has no statistical effect on the share of risky assets in total wealth. Although, the results are not presented, riskless savings rise with a lower self-assessed health. So, health has an opposite effect on riskless savings. A lower self-assessed health decreases the holdings of risky assets, and raises riskless savings. This implies that risky products are transformed to riskless financial products, as a consequence of a health deterioration.

(19)

18

have a larger coefficient than their non-current counterparts, this is consistent with Berkowitz and Qiu (2006).

In short, the results in the literature align, although some deviations exist, with those found in this thesis on a Dutch representative dataset.

4.2 Health Depreciation 4.2.1 Empirical Strategy

In section 4.1 a statistically significant relationship between self-assessed health on one hand, and the probability of holding risky assets and the share of risky assets in financial wealth, on the other hand, has been established. The relationship is established on a Dutch representative dataset. This contrasts to the non-representative datasets employed in the existing literature. As discussed above, this thesis adds to the existing literature by combining the two pressing issues in the literature. A new health indicator is employed that measures a change in the self-expected likelihood of reaching a certain age. This health variable provides an additional health indicator that takes the time horizon properly and explicitly into account.

Essentially this section replicates section 4.1 with a different health indicator. The results are shown in table 5 and 6. Only the dependent variable that measures the amount of risky assets has sufficient variation to execute a fixed effects estimator. The Hausman test does not reject the use of the random effects estimator 19. In order to control for unobserved time invariant heterogeneity, the fixed effects estimator is executed.

4.2.2 Results

The results in table 5 and 6 show that a shortened time horizon lowers the probability of holding risky assets and lowers the share of risky assets in financial wealth, respectively, with a statistically significant 3.2 and 4.8 percentage points. Consistent with the results of self-assessed health in section 4.1, the health indicator on the remaining two dependent variables are insignificant. Both health indicators have an entirely different scale. Therefore, the results are best compared by standardization of the health indicators. A one standard deviation decrease of the health condition, with respect to self-assessed health, constitutes to a lower probability of owning risky assets of 1.58 percentage points. For a lower likelihood of reaching a certain age this number constitutes to 1.59 percentage points. Further, a one standard deviation decrease

(20)

19

of the health condition, measured by self-assessed health, decreases the share of risky assets in financial wealth by 3.60 percentage points. The corresponding number for the other health indicator is 2.37 percentage points. Thus, the results employing both health indicators do not only reveal a consistent pattern in terms of the direction of the relationship. They do show effects that are truly similar in magnitude 20. So, the results are robust for two entirely different health indicators, one that measures self-assessed health and the other that directly incorporates the time horizon 21_{. The signs of the covariates in table 5 and 6 paint a similar picture as in table} 3 and 4.

The amount of savings one has and the share of savings in financial wealth are regressed on Health Depreciation and the list of covariates. These results are not presented. However, this analysis reveals that savings and the share of savings in financial wealth rise in response to a shortened time horizon. Thus, the analysis suggests that those experiencing a shortened time horizon, convert their risky assets into a riskless product. Again, this result is consistent with the results found in section 4.1.

20_{The second health indicator, whether the individual experienced a lower likelihood of} reaching a certain age is not a continuous variable, but a binary variable. Therefore, technically, a one standard deviation increase is not observed in practice. This issue can, for example, be addressed by creating a binary variable for self-assessed health. Thereby, the scale of both health indicators is identical. If an individual has a moderate or poor self-assessed health, the probability of holding risky assets and the share of risky assets in total wealth is decreased by, respectively, 3.5 and 7.2 percentage points. These numbers are, in sign and magnitude, similar to the results in table 5 and 6.

(21)

20

Table 5 Dependent variable:

Risky Assets

Ownership Amount of Risky Demographic Variables (1) (2) (3) (4) Health Depreciation -0.032** -670.3 -3766 -666.2 (0.013) (960.9) (3531) (897.3) Age -2.35e-04 356.1 625.2** 6388** (0.004) (468.3) (1392) (2529) Age2 5.15e-06 -2.066 -2.507** -16.84 (4.12e-05) (4.552) (12.42) (10.67) Gender 0.023 2755 3654 Omitted (0.020) (2185) (7170) Omitted Partner -0.045** -635.4 -8777 -2216 (0.019) (2060) (7517) (1748) Educational Attainment 0.021*** 2184** 9506*** 3449 (0.007) (851.0) (2570) (2602) Number of Children in Household 0.007 -473.8 1189 -1180

(0.010) (1035) (4049) (1636)

Urbanization -0.011 -1495 -5153 4464

Current Assets _2.01e-06*** 0.240*** 0.694*** 0.168*** (2.27e-07) (0.056) (0.0925) (0.058) Current Assets2/103 _-1.35e-09*** 2.22e-04*** -1.10e-04 3.62e-05

(1.72e-10) (6.29e-05) (8.68e-05) (7.39e-05)

Non-Current Assets 6.00e-08 0.012 0.042 -1.78-04

(1.37e-07) (0.020) (0.066) (0.015) Non-Current Assets2/103 _-1.27e-11*** 1.20e-05 -3.57e-05 -1.54e-06

(4.01e-12) (1.91e-05) (5.65e-05) (1.53e-05)

Current Liabilities 1.75e-06 0.153* 0.716 0.155

(1.27e-06) (0.087) (0.738) (0.103) Current Liabilities2/103 -1.72-08 -0.001** -0.006 -0.002* (1.34e-08) (6.14e-04) (0.016) (0.001) Non-Current Liabilities -2.52e-08 0.026 0.063 0.024

(1.90e-07) (0.030) (0.079) (0.030) Non-Current Liabilities2/103 -8.70e-11 1.59e-05 -9.39e-05 -6.34e-05

(1.89e-10) (5.96e-05) (1.47e-04) (5.97-05) Monthly Income Net of Taxes 1.66e-05 0.656 0.806 6.155*

(1.06e-05) (1.569) (3.194) (3.332)

Number of Observations 1729 1729 1729 1729

R2 (within) - 0.234 - 0.268

(22)

21

Notes: */**/*** indicate significance at the 10%/5%/1%. Specification 1 and 2 are a tobit model with truncation at zero. Average marginal effects are displayed. For both specifications the standard errors are bootstrapped with 100 iterations.

Table 6 Dependent variable: Risky Assets-Total Wealth Ratio Risky Assets- Financial Wealth Ratio Demographic Variables (1) (2) Health Depreciation -0.012 -0.048* (0.012) (0.027) Age -0.004 5.25e-05 (0.005) (0.012)

Age2 _3.99e-05 _2.75e-05

(4.54e-05) (1.05e-04) Gender 0.039* 0.067 (0.022) (0.047) Partner -0.049* -0.085 (0.025) (0.053) Educational Attainment 0.028*** 0.073*** (0.007) (0.018) Number of Children in Household 0.014 0.021

(0.012) (0.028)

(1.98e-10) (3.89e-10)

(1.78e-07) (3.21e-07) Non-Current Assets2/103 -1.10e-11 -2.48e-10

(1.47e-10) (2.42e-10)

(4.67e-08) (1.75e-07)

(1.81e-07) (4.60e-07) Non-Current Liabilities2_/103 _-1.84e-10 _-1.94e-10

(1.19e-05) (1.66e-06)

(23)

22

5. Conclusion and Discussion

This thesis studied the relationship between health and portfolio allocation. A positive strong, and statistically significant, relationship between an individual’s health condition and the holdings of risky assets exists. Even after controlling for assets, liabilities, monthly income and a list of demographic variables. This evidence is found using two health indicators. The first health indicator is self-assessed health. Self-assessed health and its positive associations with the holdings of risky assets is well embedded within the literature. Both the probability of holding risky assets and the share of risky assets in financial wealth is positively related to subjective well-being. That is, a lower health status is related to lower holdings of risky assets. Consequently, these risky assets are transformed into riskless financial products. However, assessment of the individual’s own health condition is subject to short-term emotions and mood.

For this reason, an additional health indicator is developed. This thesis adds to the literature by the inclusion of a novel health indicator in the health-risky assets nexus. The respondents are required to fill in their expected likelihood of reaching a certain (distant) age. A lower likelihood of reaching a certain age is considered a health deterioration. This variable explicitly incorporates the time horizon. Several researchers note the importance of the time horizon in the interplay between portfolio selection and health, such as Rosen and Wu (2004) and Edwards (2008). Most frequently, this is incorporated with the inclusion of an age variable in the analysis. However, Dow (2009) notes that age is often unrelated to the time horizon of the investor. In Edwards (2008) this is incorporated with the inclusion of a control variable that measures the remaining years of life of the investor. That control variable hinges on some unlikely assumptions. However, in this thesis the time horizon is directly incorporated through the novel health indicator. Thereby, this thesis combines the two pressing issues in the health-risky assets literature. An additional health indicator next to self-assessed health is provided, while simultaneously taking the time horizon properly and directly into account.

(24)

23

reaching a certain age, is 1.59 and 2.37 percentage points. Thereby, the results are consistent and robust, in both sign and magnitude, for entirely different health measures.

Further, this thesis adds to the existing literature by employing a dataset that is a reflection of the entire Dutch population. The existing literature tend to focus on the elderly. The dataset employed in this thesis, LISS, enables that the scope can be enlarged to a sample representative of the entire Dutch population. Thus, conclusions that are drawn, hold population-wide.

An interesting hypothesis that sheds some light on the mechanism that might drive the results, is presented in Bodie, Merton and Samuelson (1992). Bodie, Merton and Samuelson (1992) suggest a theoretical model in which individuals vary their labor supply to compensate for variability in investment returns. The ability to compensate for low investment returns, decreases with age. More frequently, older people tend to experience the inability to work. Thereby, their ability to compensate for low returns is decreased. Bodie, Merton and Samuelson (1992) view this as an explanation for the fact that older people tend to hold safer portfolios. A similar line of thought can be applied for people of poorer health or those that experience a decrease in expected time horizon due to falling ill. These people are less able to compensate for low returns, thereby reducing the riskiness of their portfolio. This induces a movement towards safer investment options. Indeed, this is observed in the data.

(25)

24

References

Adam, E. K., McDade, T. W., Nandi, A. and Sweet, E. (2013). The high price of debt: Household financial debt and its impact on mental and physical health. Social Science & Medicine, 91, 94-100.

Adler, N. and Ostrove, J. (1999). Socioeconomic Status and Health: What We Know and What We Don’t. Annals of the New York Academy of Sciences, 3-15.

Ameriks, J. and Zeldes, S. P. (2004). How Do Household Portfolio Shares Vary With Age? working paper, Columbia University, Graduate School of Business.

Ayyagari, P. and He, D. (2017). The role of medical expenditure risk in portfolio allocation decisions. Health Economics, 26, 1447-1458.

Berkowitz, M. K. and Qiu, J. (2006). A further look at household portfolio choice and health status. Journal of Banking and Finance, 30(4), 1201-1217.

Bertaut, C. C. and Starr-McCluer, M. (2002). Household portfolios in the United States. In: Guiso, L., Haliassos, M., Jappelli, T. (Eds.), Household Portfolios. MIT Press, Cambridge, MA, 181–218.

Bodie, Z., Merton, R. C. and Samuelson, W. F. (1992). Labor supply flexibility and portfolio choice in a life cycle model. Journal of Economics Dynamics and Control, 16, 427-449. Böheim, R. and Lackner, M., (2015). Gender and Risk Taking: evidence from jumping competitions. Journal of the Royal Statistical Society, 178(4), 883-902.

Bressan, S., Pace, N. and Pelizzon, L. (2014). Health status and portfolio choice: Is their relationship economically relevant? International Review of Financial Analysis, 32, 109-122. Cai, M., Ni, Y., Pan, R. and Shi, J. (2013). Do Heterogeneous Background Risks Matter to Household Asset Allocation? Asia-pacific Journal of Financial Studies, 42, 563-589.

Cardak, B. A. and Wilkins R. (2009). The determinants of household risky asset holdings: Australian evidence on background risk and other factors. Journal of Banking and Finance, 33(5), 850-860.

(26)

25

Centraal Bureau voor de Statistiek, Kadaster (2018). House prices continue to rise. [online] Available at: https://www.cbs.nl/en-gb/news/2018/29/house-prices-continue-to-rise [Accessed 11-10-2018].

Cinciripini, P. M., Cofta-Gunn, L., Daza, P., Fouladi, R. T., Gritz, E. R., Irvin, J. E., Mazas, C. Wetter, D. W. and Wright, K. (2005). What accounts for the association of education and smoking cessation? Preventive Medicine, 40(4), 452-460.

Coile, C. and Milligan, K. (2009). How household portfolios evolve after retirement: the effect of aging and health shocks. Review of Income and Wealth, 55, 226–248.

Deeks, A., Lombard, C., Michelmore, J. and Teede, H. (2009). The effects of gender and age on health related behaviors. BMC Public Health, 9(213).

Doiron, D., Fiebig, D. G., Johar, M. and Suziedelyte, A. (2015). Does self-assessed health measure health? Applied Economics, 47(2), 180-194.

Dow, J. P. (2009). Age, investing horizon and asset allocation. Journal of Economics and Finance, 33(4), 422-436.

Edwards, R. D. (2008). Health Risk and Portfolio Choice. Journal of Business and Economic Statistics, 26(4), 472-485.

Fan, E. and Zhao, R. (2009). Health status and portfolio choice: Causality or heterogeneity? Journal of Banking & Finance, 33, 1079-1088.

Feinstein, J. S. and Lin, C. (2006). Elderly asset management. SSRN Working Paper 956399. Fullet, T. D. (2010). Relationship Status, Health and Health Behavior: An Examination of Cohabiters and Commuters. Sociological Perspectives, 53(2), 221-246

Heaton, J. and Lucas, D. (1997). Savings Behavior and Portfolio Choice: Evidence from the Tax Data. Working Paper, Northwestern University, 1997.

Joiner, T. E., Schmidt, N. B. and Telch, M. J. (1996). Factors influencing health perceptions in patients with panic disorder. Comprehensive Psychiatry, 37, 253-260.

Levy, H. (2000). The Financial Impact of Health Insurance. UC-Berkeley. Mimeo.

(27)

26

McDonough, P. and Walters, V. (2001). Gender and health: reassessing patterns and explanations. Social Science & Medicine, 52(4), 547-559.

Markowitz, H. (1952). Portfolio Section. The Journal of Finance, 7(1), 77-91.

Morin, R. A. and Suarez, A. F. (1983). Risk Aversion Revisited. The Journal of Finance, 38(4), 1201-1216.

Rice, R. W. and Yardley, J. K. (1991). The Relationship between Mood and Subjective Well-Being. Social Indicators Research, 24(1), 101-111.

Rosen, H.S. and Wu, S. (2004). Portfolio choice and health status. Journal of Financial Economics, 72(3), 457-484.

Shi, X. and Yan, Z. (2017). Urbanization and risk preference in the people’s republic of China: A decomposition of the self-selection and assimilation effects, ADBI Working Paper Series, No 710.

Verbrugge, L. M. (1983). Multiple Roles and Physical Health of Women and Men. Journal of Health and Social Behavior, 24, 16-30.

World Health Organization. (2010). Urbanization and health. Bulletin World Health Organization, 88, 245-246.

(28)

27

Appendix A

Construction of the variable Health Depreciation

Respondents are inquired about the likelihood of them reaching the ages of 75 to 95 in steps of five years. The answer possibilities range from zero to ten. The first two self-assessed life expectancy questions are asked to those of at least age 16 and those that are at least ten year younger than the age they are inquired about. The remaining three life expectancy questions are for those that are at maximum twenty years younger and at minimum ten years younger than the age they are inquired about. Subsequently, a variable is constructed that measures the change per wave of self-assessed life expectancy. If a respondent experiences a lower likelihood of reaching a certain age, compared to previous wave, Health Depreciation equals one. Health

Depreciation equals zero in all other cases.

The questions on the younger ages, have priority over later ages. That is, the question on the likelihood of the age is only considered if the other age questions are left blank. Health

Depreciation equals zero in all other cases. Questions on self-assessed life expectancy are not

(29)

28

Appendix B

Table B1 Dependent Variable:

Risky Assets

Ownership Amount of Risky

Demographic Variables (1) (2) (3) (4)

New Severe Health Condition -0.058* -183.7 -8855 -5095 (0.030) (3535) (10427) (4013) Age -0.001 349.8 484.6 6744*** (0.005) (458.7) (1388) (2546) Age2 1.52e-05 -2.028 -0.961** -15.15** (4.29e-05) (4.560) (12.35) (10.50) Gender 0.024 2745 3882 Omitted (0.020) (2196) (7242) Omitted Partner -0.047** -648.5 -8923 -2516 (0.020) (2073) (7520) (1842) Educational Attainment 0.020** 2195** 9468*** 3806 (0.008) (856.7) (2616) (2824) Number of Children in Household -0.007 -508.9 1104 -1456

(0.010) (1044) (4036) (2824)

Urbanization -0.010 -1482.7 -5024* -4772

Current Assets _2.04e-06*** 0.239*** 0.700*** 0.170*** (1.93e-07) (0.056) (0.093) (0.058) Current Assets2/103 _-1.36e-09*** 2.22e-04*** -1.16e-04 3.19e-05

(1.75e-10) (6.26e-05) (8.82e-05) (7.21e-05)

Non-Current Assets 8.26e-08 0.0131 0.041 2.50e-04

(1.46e-07) (0.020) (0.066) (0.014) Non-Current Assets2/103 _-3.36e-11*** -1.24e-05 -3.39e-04 -1.87e-06

(5.28e-12) (1.91e-05) (5.79e-04) (1.41e-05) Current Liabilities _1.77e-06*** 0.151* 0.706 0.155

(1.29e-06) (0.086) (0.755) (0.102) Current Liabilities2_/103 _-1.72e-08** _-0.001** _-0.006 _-0.002* (1.33e-08) (0.001) (0.017) (0.001) Non-Current Liabilities -2.68e-08 0.026 0.053 0.023

(1.97e-07) (0.030) (0.079) (0.029) Non-Current Liabilities2_/103 _-7.66e-11 _-1.52e-05 _-9.08e-05 _-6.13e-05

(1.89e-10) (5.96e-05) (1.46e-04) (5.90e-05) Monthly Income Net of Taxes 1.81e-05* 0.656 0.715 6.118*

(1.08e-05) (1.571) (3.192) (3.366)

Number of Observations 1729 1729 1729 1729

R2_(within) _- _0.234 _- _0.272

(30)

29

Notes: */**/** indicate significance at the 10%/5%/1%. Specification 1 and 2 are a tobit model with truncation at zero. Average marginal effects are displayed. For both specifications the standard errors are bootstrapped with 100 iterations.

Table B2 Dependent Variable: Risky Assets-Total Wealth Ratio Risky Assets- Financial Wealth Ratio Demographic Variables (1) (2)

New Severe Health Condition -0.040 -0.097*

(0.032) (0.057)

Age -0.004 -0.001**

(0.005) (0.011)

Age2 _4.63e-05 _4.35e-05

(4.23e-05) (9.80e-05) Gender 0.039* 0.068 (0.021) (0.047) Partner -0.050** -0.088* (0.023) (0.051) Educational Attainment 0.027*** 0.071*** (0.008) (0.019) Number of Children in Household 0.014 0.020

(0.011) (0.026)

(2.51e-10) (3.72e-10)

(1.75e-07) (3.60e-07) Non-Current Assets2/103 6.99e-12 -2.40e-10

(1.22e-10) (2.76e-10)

(7.38e-08) (5.06e-07)

(1.78e-07) (5.06e-07) Non-Current Liabilities2_/103 _-1.74e-10 _-1.58e-10

(1.06e-05) (1.76e-05)