UNIVERSITY OF SOUTHAMPTON

SCHOOL OF ECONOMIC, SOCIAL & POLITICAL SCIENCES

USING BEHAVIOURAL RISK ATTITUDES TO ANALYSE HOW HOUSEHOLDS’ PAST EXPERIENCE INFLUENCES THEIR SUBSEQUENT

INVESTMENT DECISIONS IN THE STOCK MARKET

Isobelle Heidi Ager

Presented for B.Sc. (Social Sciences) in Economics and Finance April 2019

I declare that this dissertation is my own work, and that where material is obtained from published or unpublished work, this has

been fully acknowledged in the references.

Signed: Isobelle Heidi Ager, 30/04/19

1

Abstract

In this paper use is made of data of the DNB Household Survey, to model the

relationship between the House Money Effect and shareholding of household

investors. The focus is on proving its existence within the sample, and in

doing so, motivating economic policy to help reduce inefficient biases that

impact investment behaviour.

2

Acknowledgements

I would like to thank Matt (Ortiz De Zarate Rodriguez M. F.) for devoting his

own time to helping both transform my dataset and better my understanding

in using Stata. His knowledge, patience and generosity allowed me to conduct

a thorough piece of statistical analysis that I can be proud of.

3

Contents

1. Introduction ..................................................................................................................... 4

1.1 Research Motivation ..................................................................................................... 4

1.2 Literature Review .......................................................................................................... 6

1.3 Research Question Explained ............................................................................... 10

2. Data and Methodology ............................................................................................. 12

2.1 Sources .............................................................................................................................. 12

2.2 Model for Analysis ...................................................................................................... 13

3. Results .............................................................................................................................. 16

3.1 Intuition and Assumptions ..................................................................................... 16

3.2 Regressions and Interpretations ........................................................................ 21

4. Discussion and Conclusion .................................................................................... 25

4.1 Overview and Limitations....................................................................................... 25

4.2 Relation to Existing Literature ............................................................................. 25

4.3 Policy Implications ................................................................................................ 26

5. Bibliography.................................................................................................................. 27

6. Appendix ......................................................................................................................... 32

4

1. Introduction

1.1 Research Motivation

The stock market puzzle (SMP) has been researched at great length. With

regard to attempted explanations, as discussed during the literature review

which accompanies this report, Gardini and Magi (2006) point to

participation costs (both financial and non-financial) as the dominant reason

for so few households investing. Consequently, one could view other

contributing factors1 as either positive or negative influences on participation

costs as a reason for non-participation. There was also an extensive focus on

proving the importance of behavioural finance theory (BFT) for bettering our

understanding in this research area.

If Campbell (2006) and many others use standard finance theory

(SFT) to outline the irrationality behind not investing given the positive

equity premium, how can we explain the irrational behaviour of individual

investors once they have entered the stock market?

Khorunzhina (2012) demonstrates the argument of stock market

participation costs by using data from the Panel Study of Income Dynamics to

estimate their magnitude. They provide confirmation of fixed entry costs by

explaining that whilst they appear greater for first time investors, costs

decrease relative to continued participation. With this hypothesis of

1 Research by Guiso et al. (2002) surrounding income, wealth and age; subsequently Cocco et al. (2005) who relate this element of growing older with one’s employment position via the life cycle; and as addressed by Garcia and Tessada (2013) and Rooij et al. (2011), the influence of education and financial literacy respectively.

5

homogeneity being widely accepted, research focus is further shifted to

questioning the activity of individuals who already participate in the stock

market. Barber et al. (2008) assess, by comparison, the subpar nature of

individual investors’ trading strategies, which puts them at a disadvantage

within the stock market where more sophisticated investors and institutions

also operate. They computed losses that seemed to result in institutional

gains and consequently, Barber et al. (2008) advise these investors to avoid

their own weaker trading strategies and instead invest in a diversified, low-

cost mutual fund.

This report turns to the house money effect (HME); the pattern of

investment behaviour described by Wen et al. (2014) that means individual

investor’s stock market losses and gains will influence their reference point

for subsequent investment decisions, due to changing risk preferences.

Specifically; the common perception that profits in period t-1 encourage

traders to increase risk-taking in period t, because the gains they have made

will act as a cushion for absorbing any losses, thus, gambling more with the

house’s money. By using panel survey data spanning 1994-2018, this analysis

demonstrates the relationship between a risk measure that mirrors the HME,

and the number of shares one chooses to hold. Figure 1 outlines the presence

of characteristics among the sample participants2.

2 All those noted with † are dummy variables; their average is to be interpreted as the percentage of the sample displaying the characteristic (i.e. their id code coincided with the variable taking a value of 1, noting that for gender this represents the percentage of males). The remaining non-binary variables can be interpreted as follows; numshares captures the amount of companies a participant holds shares with and only those holding shares are included in the sample so that all answers will be greater than 1 with an average of 3. The average answer to year of birth was 1950 and the calculation of ntot (net total income) is explained in detail within each year’s codebook available at https://www.dhsdata.nl/site/users/login.

6

Figure 1

N: 4,943

Variable

Summary statistics

Value =

Average

Skewness

Kurtosis

𝝈

numshares {>0} 3 7.1751 101.8322 4.2315 housemoney† {0, 1} 5.91% 3.7404 14.9909 0.2358

birthyr {>0} 1950 0.1098 2.6332 15.2323 gender† {0, 1} 75.89% -1.2102 2.4646 0.4278

degree† {0, 1} 32% 0.7715 1.5952 0.4665

employed† {0, 1} 50.07% -0.0028 1.0000 0.5001

selfemployed† {0, 1} 3.66% 4.9343 25.3474 0.1878

retired† {0, 1} 27.01% 1.0357 2.0726 0.4440 ntot {≥0} €13,315.84 3.7043 37.0373 €23,459.31

married† {0, 1} 38.5% 0.4727 1.2235 0.4866

children† {0, 1} 31.28% 0.8077 1.6524 0.4637

financialadmin† {0, 1} 82.58% -1.7181 3.9519 0.3793

insurance† {0, 1} 4.25% 4.5368 21.5825 0.2017

homeowner† {0, 1} 27.37% 1.0150 2.0302 0.4459

Prospect Theory3 would explain this trading pattern by the

phenomena that individuals are risk loving with gains and risk averse with

losses. The latter is a view shared by many; Barberis et al. (2001) explain the

HME by a degree of loss aversion in the current period to determine risky

choices, being dependent on prior investment outcomes, because the

magnitude of the premium that is required to compensate for stock volatility

will change accordingly. Finally, Mattos and Garcia (2009) conclude the

disparity among their participants’ trading strategies proves prior profits are

interpreted as different reference points depending on the risk level in one’s

portfolio, acting as a useful explanation for the type of behaviour we witness

in the stock market.

1.2 Literature Review

3 As developed by Kahneman and Tversky (1979).

7

Upon further reading, a few key elements for successful interpretation of this

analysis have emerged. Hsu and Chow (2013)4 had access to individual level

trade data and therefore followed intraday transactions of securities. Overall,

this allows for a particularly advanced measure of the HME; it is preferable to

analyse with daily trade data because closer periods of buying and selling

actions provide a more accurate interpretation of their relationship, rather

than relying on a longer period between realised gains and the subsequent

investment decisions. In doing so, they assigned codes to each trader to

categorise and track their individual trading movements.

In particular: to measure their reference points (of which there are

many variations; Hsu and Chow (2013) use most recent, average and/or

initial purchase price of a share, as well as the highest closing price over a 20-

day period before the share is sold or the highest market price achieved prior

to any realised gain, finding the latter to give the strongest HME). Brown et al.

(2006) focus on IPOs because this confirms the price at which subscribers

make an initial purchase, also choosing to use averages because realised

gains and losses were felt to be enough to reflect the sentiment of the market

as either bullish or bearish, but not particularly relative to investors’

opportunities. They then calculated gain and loss proportion ratios, where a

value greater than one is interpreted as the account holder either being more

likely to hold winning than losing stocks, or losing than winning ones

respectively, to test for the HME with these changing ratios.

4 Like many other papers discussed throughout this report, namely; Brown et al. (2006) who use the Australian Stock Exchange (ASE) for daily share registry data, Kaniel et al. (2004) who use the New York Stock Exchange and Mattos and Garcia (2009) who use the agricultural futures and options market.

8

Gains or losses made on share trades5 relative to a reference point

should determine the individual’s risk preference. Again, throughout the

literature there exist multiple methods for calculating gains and losses; Hsu

and Chow (2013) take the product of share quantity and the difference

between sales price and purchase price, then vary between the law of first-in-

first-out to follow the physical flow of inventories, and last-in-first-out to

account for the possibility of the availability heuristic6. Likewise for risk, they

use standard deviation of return of the stock in the transaction, computed as

a weighted average because research indicates that the majority of such

traders fail to account for return correlations7.

In most cases, a form of regression is run to test the null hypothesis

that an investor will take greater risk following a period of gains (relative to

their reference point); Hsu and Chow (2013) confirm the HME with a positive

coefficient of 1.743.

Barberis et al. (2001) propose an alternative measure for HM

strategies by considering the impact of financial wealth fluctuations on

deciding how much to invest in the stock market. They assume investors

maximise an objective function of utility with input factors of consumption

and value of asset prices. They also mention a consumption-based model

proposed by Campbell and Cochrane (1999); using a habit level of

consumption as the reference point, moving closer to or further away from

this point generates time-varying risk aversion. These alternatives still find

5 Hsu and Chow (2013) also adjust for dividend profits due to the nature of certain stocks. 6 As outlined by Tversky and Kahneman (1974). 7 Odean (1998) and Hsu and Chow (2013) discuss mental accounting and narrow framing; investors tend to think of each stock (and their gains/losses) as a separate account.

9

fluctuations in financial wealth due to prior investment performance to result

in loss aversion behaviour.

Both Hsu and Chow (2013) and Brown et al. (2006) analysed the

influence of sale period duration to account for how price pressure impacts

demand. A decreasing slope coefficient demonstrates how the HME is

stronger over shorter and more instantaneous periods, as does it depreciate

over time. Mattos and Garcia (2009) explain the weakening of this effect is

due to investors becoming more accepting of the cost; the magnitude of

which relies on the endowment effect8.

As previously mentioned, it is popular to categorise investors in order

to see whether the HME prevails among different types (such as differing

levels of sophistication, captured by the size and quantity of their trades).

Hsu and Chow (2013) show that under the largest gains category (deemed

the most sophisticated within the sample), the HME exists via a coefficient of

just 0.973 compared with the smallest category which displays a much larger

result of 5.897. Whilst there is a notable difference in susceptibility, even the

more knowledgeable investors demonstrate the psychological bias. Through

additional testing, Brown et al. (2006) revealed the observed behaviour is not

attributable to diversification techniques, therefore more likely to be a

behavioural bias.

The literature review that accompanies this report mentions also

testing for the disposition effect9. Further reading indicated the more

prominent nature of the HME by comparison, and Kaniel et al. (2004)

8 Investors being less willing to give up their shares than they are willing to pay for new ones. 9 As described by Wen et al. (2014).

10

suggests the HME tempers the disposition effect. For this reason and due to

data availability, this report focuses solely on the HME.

1.3 Research Question Explained

Daily trade data was extremely limited and restricted10, leaving survey data

as the most individual-level and suitable option. Seeing as it was not possible

to witness exactly how much an individual made/lost per trade11, therefore

ruling out measuring the impact this had on their subsequent risk preference,

the reverse relationship is captured by regressing numshares against the

HME variable. For the housemoney dummy to be equal to a value of 1, the

individual will have indicated that they agree more than they disagree with a

statement taken from the panel survey about them being prepared to risk

losing money when there is also a chance to gain12. Therefore, it is assumed

that the gains/losses an individual makes in trading stocks is what moves

their reference point to make each new investment decision under changing

risk preferences, such that the sign and size of the housemoney coefficient

reflects the impact this has on the amount of shares they choose to hold.

In attempt to remain following the logic of the literature by capturing

the time-varying essence of risk preferences, this analysis first used the lag of

housemoney, only to reveal that the year-long period between the collection

of surveys meant the gap between variables made them too unrelated to

10 For example, the Taiwan and the London Stock Exchange grant access to participant trading data only to member firms, otherwise just aggregate-level data is viewable. 11 Even questions on interest and/or dividends received were not consistent enough throughout the sample period to be used as a measurable variable. 12 Please refer to the Economic and Psychological Concepts section of each codebook available at https://www.dhsdata.nl/site/users/login.

11

capture a significant relationship. It is likely that if participants thought

specifically about the interrelatedness of the questions and answered each

year based on any changes since the last, these lagged results would have

been more plausible.

12

2. Data and Methodology

2.1 Sources

All data is from De Nederlandsche Bank (DNB) Household Survey (DHS)13, as

their unique provision of yearly panel surveys consisting of over 2,000

households (over 16s only) was most consistent and accommodating, despite

the final dataset being unbalanced14. This use was inspired by Rooij et al.

(2011) whose analysis contains the DHS for studying the relationship

between financial literacy and stock market participation, highlighting

suitable demographic variables within its modules and encouraging this over

the Health and Retirement Study because of the sample bias towards over

50s.

Combined with fine-tooth combing of the DHS codebooks, a process of

elimination resulted in a compilation of variables that were definitely present

and available in each year of the sample period to avoid large gaps in the

data. Overall, the model should be an improvement on Hsu and Chow (2013)

whose analysis only includes one explanatory variable and consequently is

more likely to capture a positive bias in the estimated coefficient. Similarly,

the numshares15 variable covers participants’ wider portfolio by asking a

13 A project by CentERdata; Institute for data collection and research. 14 Given the assumptions hold, Wooldridge states that the unobserved effect within FE controls for unbalanced data. 15 Noting that originally this model used a variable that captured the total market value of one’s shares, which appeared to reflect market sentiment more so than one’s choice to invest given their HME.

13

more general question about stock holding, as opposed to Brown et al. (2006)

who are limited to asking questions surrounding stockholding in the ASE.

Both Rooij et al. (2011) and Mattos and Garcia (2009) report the

importance of including variables to account for age, gender and degree. The

former suggest self-employment and retirement may have a significant

impact on one’s choice to invest because of the exposure to risk one would

already be facing in the labour market and due to one’s decumulation

position in the life-cycle, respectively. Based on their testing, this model also

attempts to capture a measure similar to financial literacy under

financialadmin; an indicator as to whether the participant is involved in the

financial conduct of the household.

The insurance dummy takes a value of 1 when a participant holds at

least one saving/endowment policy; Garcia and Tessada (2013) suggest the

effect being well-insured may have on individuals’ willingness to risk capital

on the stock market. The final three dummies (married, children and

homeowner) are additional factors that could shift one’s portfolio away from

stockholding, recognising that it may not just be the psychological bias of

HME that results in changing investment decisions but also having family

responsibilities and/or as highlighted by Khorunzhina (2012), another

financial asset that acts as a substitute.

2.2 Model for Analysis

Rooij et al. (2011) acknowledge the potential problems with simply relying

on Ordinary Least Squares (OLS) regression; there is substantial error when

measuring estimates due to omitted variable bias and endogeneity and a

14

need for robust standard errors. To generate more significant results, both a

Fixed Effects (FE) and Random Effects (RE) transformation is applied to this

panel data. The model is outlined in Figure 2 (the latter being a variation

including the natural logarithm of non-binary variables in order to curb

outliers and allow a more convenient interpretation of results, noting the

[parenthesis] is to denote what Greene (2012) explains as the additional

random error term necessary for a RE transformation, as well as both the

group specific 𝑢𝑖𝑡 element used in FE and the unknown entity-specific

intercept 𝛼𝑖):

Figure 2

𝑛𝑢𝑚𝑠ℎ𝑎𝑟𝑒𝑠𝑖𝑡 = 𝛽0 + 𝛽1ℎ𝑜𝑢𝑠𝑒𝑚𝑜𝑛𝑒𝑦𝑖𝑡 + 𝛽2𝑏𝑖𝑟𝑡ℎ𝑦𝑟𝑖 + 𝛽3𝑔𝑒𝑛𝑑𝑒𝑟𝑖

+ 𝛽4𝑑𝑒𝑔𝑟𝑒𝑒𝑖𝑡 + 𝛽5𝑒𝑚𝑝𝑙𝑜𝑦𝑒𝑑𝑖𝑡 + 𝛽6𝑠𝑒𝑙𝑓𝑒𝑚𝑝𝑙𝑜𝑦𝑒𝑑𝑖𝑡

+ 𝛽7𝑟𝑒𝑡𝑖𝑟𝑒𝑑𝑖𝑡 + 𝛽8𝑛𝑡𝑜𝑡𝑖𝑡 + 𝛽9𝑚𝑎𝑟𝑟𝑖𝑒𝑑𝑖𝑡 + 𝛽10𝑐ℎ𝑖𝑙𝑑𝑟𝑒𝑛𝑖𝑡

+ 𝛽11𝑓𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙𝑎𝑑𝑚𝑖𝑛𝑖𝑡 + 𝛽12𝑖𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒𝑖𝑡 + 𝛽13ℎ𝑜𝑚𝑒𝑜𝑤𝑛𝑒𝑟𝑖𝑡

+ 𝛼𝑖 + 𝑢𝑖𝑡 [+𝜀𝑖𝑡]

𝑙𝑛𝑛𝑢𝑚𝑠ℎ𝑎𝑟𝑒𝑠𝑖𝑡

= 𝛽0 + 𝛽1ℎ𝑜𝑢𝑠𝑒𝑚𝑜𝑛𝑒𝑦𝑖𝑡 + 𝛽2𝑙𝑛𝑏𝑖𝑟𝑡ℎ𝑦𝑟𝑖 + 𝛽3𝑔𝑒𝑛𝑑𝑒𝑟𝑖

+ 𝛽4𝑑𝑒𝑔𝑟𝑒𝑒𝑖𝑡 + 𝛽5𝑒𝑚𝑝𝑙𝑜𝑦𝑒𝑑𝑖𝑡 + 𝛽6𝑠𝑒𝑙𝑓𝑒𝑚𝑝𝑙𝑜𝑦𝑒𝑑𝑖𝑡

+ 𝛽7𝑟𝑒𝑡𝑖𝑟𝑒𝑑𝑖𝑡 + 𝛽8𝑙𝑛𝑛𝑡𝑜𝑡𝑖𝑡 + 𝛽9𝑚𝑎𝑟𝑟𝑖𝑒𝑑𝑖𝑡 + 𝛽10𝑐ℎ𝑖𝑙𝑑𝑟𝑒𝑛𝑖𝑡

+ 𝛽11𝑓𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙𝑎𝑑𝑚𝑖𝑛𝑖𝑡 + 𝛽12𝑖𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒𝑖𝑡 + 𝛽13ℎ𝑜𝑚𝑒𝑜𝑤𝑛𝑒𝑟𝑖𝑡

+ 𝛼𝑖 + 𝑢𝑖𝑡 [+𝜀𝑖𝑡]

FE are useful for analysing the time-varying impact of an event on an

outcome variable ((ln)𝑛𝑢𝑚𝑠ℎ𝑎𝑟𝑒𝑠𝑖𝑡), under the important assumption that

time-invariant characteristics of individuals (𝑏𝑖𝑟𝑡ℎ𝑦𝑟𝑖, 𝑔𝑒𝑛𝑑𝑒𝑟𝑖), which either

15

influence the predictor variable (ℎ𝑜𝑢𝑠𝑒𝑚𝑜𝑛𝑒𝑦𝑖𝑡) or not, are uncorrelated

with any other individual characteristics such that neither the constant nor

error term capture them. Both Greene (2012) and Carter Hill et al. (2011)

speak favourably of how this helps to reduce bias in the estimated coefficient,

and how the creation of time-demeaned data loses the unobserved effect

without sacrificing any years. Bartels (2008) also implies that using dummies

allows for coefficients that are mediated by cross-country differences in

absorbing anything that may lead to unobserved heterogeneity, therefore

capturing the pure effect.

They also describe RE as a random intercept and partial pooling

model; the variation across entities is instead assumed uncorrelated with the

predictor such that it influences the outcome rather than being adopted by

the intercept, leading to what Greene (2012) refers to as a much stronger

assumption of zero correlation between the group-specific error and

predictor. Whilst this could lead to omitted variable bias, it allows estimation

of time-invariant variables (Wooldridge explains the process of Generalised

Least Squares (GLS) resulting in quasi-demeaned data) and the

generalisation of influences beyond sample use.

16

3. Results

3.1 Intuition and Assumptions

Figure 3 achieves very few statistically significant results; combined with the

conclusion arrived at from the Chow Test16, this confirms the need for panel

transformations.

Figure 3

numshares

Pooled OLS

*p<0.01; **p<0.05; ***p<0.10

(robust standard errors)

P>F=

0.0000

housemoney† 0.2399 (0.2735)

N 4,943

birthyr -0.0254* (0.0051)

gender† 0.7578* (0.1240)

𝑅2

0.0377

degree† 0.4953* (0.1323)

employed† -0.3596** (0.1501)

selfemployed† 0.4266 (0.3405)

retired† 0.2082 (0.2103)

ntot 4.68e-06 (3.32e-06)

married† -0.1169 (0.1203)

children† -0.2302 (0.1421)

financialadmin† 0.1738 (0.1840)

insurance† 0.7043*** (0.4266)

homeowner† -0.2537*** (0.1381)

16 See 3.2

17

Both Carter Hill et al. (2011) and Wooldridge relay the importance of

satisfying regression assumptions, to avoid misconceptions in the form of

reverse causality, omitted variable bias and measurement error from

recollection bias causing estimates to tend towards zero. First, we prove the

linearity assumption holds for all variables in the raw data; Figure 4 shows

that whilst the sign/size may currently appear unintuitive without necessary

modelling transformations, the fitted values demonstrate weakly

positive/negative though linear relationships.

18

Figure 4

19

Next, we turn to the random sampling assumption. Wooldridge

explains that whilst it is unlikely to hold for this type of data, it can be

considered covered by its relation to independence. Carter Hill et al. (2011)

tell us that under FE and when there is a large N relative to small T (as is our

case), relying on asymptotic approximations satisfies another assumption of

𝑢𝑖𝑡 being independently and identically distributed. Therefore, seeing as

independence implies random sampling (but not vice versa), to assume no

violations is perfectly reasonable.

Their work moves us on to endogeneity, essentially requiring

satisfaction of the zero unconditional mean. Figure 5 shows that since P is not

< 0.05, we cannot reject the null hypothesis (𝐻0) of no omitted variables.

Logic suggests that having no omitted variables to exist in the error term

means there is no correlation between explanatory variables and the error.

Further benefits here include no upward bias of estimated coefficients and

consistency with the Hausman17 conclusion that under satisfaction of this

assumption, the RE model is also suitable.

Figure 5

Ramsey RESET test

F(3, 4903) 1.35 P>F 0.2577

Whilst Stata would omit any variables suffering from collinearity,

Carter Hill et al. (2011) emphasise its importance as another assumption, so

this report has tested for inflationary variables. Figure 6 reveals no variance

inflation factor (VIF) to be greater than 10 and no degree of collinearity

17 See 3.2

20

(1/VIF, also known as tolerance) to be equal to 0.1, therefore none are a

linear combination of the others.

Figure 6

Variable

VIF

1/VIF

housemoney† 1.04 0.9651 birthyr 2.19 0.4570

gender† 1.12 0.8912

degree† 1.11 0.9026

employed† 2.31 0.4333

selfemployed† 1.25 0.8027

retired† 2.46 0.4070 ntot 1.20 0.8314

married† 1.30 0.7670

children† 1.28 0.7832

financialadmin† 1.05 0.9556

insurance† 1.04 0.9584

homeowner† 1.19 0.8426 Mean 1.53

Further to this is Figure 7 for heteroscedasticity; with P < 0.05, we

reject 𝐻0: constant variance and conclude the need for a White estimator of

standard errors in regressions.

Figure 7

Breusch-Pagan / Cook-Weisberg test

chi2(1) 649.98 P>chi2 0.0000

Finally, seeing as the dataset is what Hoechle refer to as micro-panel

in that the T is relatively small by comparison to N, we can safely assume

there exists no autocorrelation, therefore accurate standard errors of

coefficients and 𝑅2.18

18 This rules out the use of First Differencing (FD) for two reasons; Wooldridge explains how both a small T with respect to N and serially uncorrelated errors makes FE preferable to FD.

21

3.2 Regressions and Interpretations

Figure 8 demonstrates the Chow Test conclusion; as explained by Carter Hill

et al. (2011), the FE F-value being < 0.05 means we reject 𝐻0: sufficient joint

model, and instead require a panel model due to interaction terms.

Figure 8

numshares

Corr(𝑢𝑖, 𝑋𝑏)= -0.0954

FE

Corr(𝑢𝑖, 𝑋) assumed 0

RE GLS

*p<0.01; **p<0.05; ***p<0.10

(standard errors)

P>F= 0.0000

FE

P>chi2= 0.0000

RE

housemoney† 0.3640*** (0.2150)

0.3299 (0.2052)

N 4,943

birthyr omitted -0.0323* (0.0068)

No. of groups 1,599

gender† omitted 0.6798* (0.1987)

𝑅2 within

0.0251 0.0238

degree† 0.5975** (0.3045)

0.4235** (0.1647)

𝑅2 between

0.0094 0.0431

employed† 0.0697 (0.2801)

-0.1784 (0.1912)

𝑅2 overall 0.0080 0.0342

selfemployed† -0.1879 (0.4951)

-0.0287 (0.3772)

𝜎𝑢

3.5398 2.5583

retired† 0.1867 (0.2526)

0.1105 (0.2053)

𝜎𝑒 3.0045 3.0045

ntot 3.12e-06 (2.59e-06)

3.45e-06 (2.31e-06)

rho 0.5812 0.4203

married† -0.1272 (0.1285)

-0.0762 (0.1167)

children† -0.2261*** (0.1339)

-0.1700 (0.1217)

financialadmin† 0.0932 (0.3961)

0.1795 (0.1928)

insurance† 0.5502** (0.2536)

0.6061** (0.2419)

homeowner† -0.0286 (0.1268)

-0.0435 (0.1100)

constant 1.5462** (0.4930)

64.4869* (13.2377)

Greene (2012) and many others turn to the Hausman test to decide

the most efficient transformation given the data. Figure 9 shows P is not <

0.05, thus we cannot reject 𝐻0: no systematic difference in coefficients. This is

favourable because now either transformation is plausible and we can make

22

greater use of the statistically significant coefficients. Nonetheless, FE is

generally preferred for policy evaluation and should therefore be given more

weight for interpretation.

Figure 9

Hausman

FE RE

Difference FE coefficient – RE coefficient

(b-B)

Standard

error housemoney† 0.0340 0.0642

degree† 0.1739 0.2562

employed† 0.2481 0.2048

selfemployed† -0.1592 0.3208

retired† 0.0762 0.1471 ntot -3.34e-07 1.17e-06

married† -0.0510 0.0537

children† -0.0561 0.0559

financialadmin† -0.0863 0.3460

insurance† -0.0559 0.0761

homeowner† 0.0149 0.0465 chi2(33) 43.99 P>chi2 0.0957

The final interpretation can be found in Figure 10:

23

Figure 10

lnnumshares

Corr(𝑢𝑖, 𝑋𝑏)= -0.2087

FE

Corr(𝑢𝑖, 𝑋) assumed 0

RE GLS

*p<0.01; **p<0.05; ***p<0.10

(robust standard errors)

P>F= 0.0002

FE

P>chi2= 0.0000

RE housemoney† 0.0990**

(0.0436) 0.0858** (0.0408)

N 2,031

lnbirthyr omitted -23.0746* (4.8835)

No. of groups 708

gender† omitted 0.2054* (0.0635)

𝑅2 within

0.0722 0.0673

degree† 0.0675 (0.1102)

0.0898 (0.0611)

𝑅2 between

0.0022 0.0684

employed† 0.0454 (0.0775)

0.0195 (0.0584)

𝑅2 overall 0.0033 0.0613

selfemployed† -0.0488 (0.1653)

0.0555 (0.1387)

𝜎𝑢

0.8057 0.6712

retired† 0.0288 (0.0843)

0.0174 (0.0730)

𝜎𝑒 0.4570 0.4570

lnntot -0.0061 (0.0134)

-0.0064 (0.0125)

rho 0.7566 0.6833

married† -0.0041 (0.0263)

-0.0052 (0.0254)

children† -0.0155 (0.0268)

0.0019 (0.0254)

financialadmin† 0.0820 (0.1001)

0.1347** (0.0572)

insurance† 0.0912 (0.0610)

0.0814 (0.0571)

homeowner† -0.0082 (0.0238)

-0.0012 (0.0226)

constant 0.3395*** (0.1895)

175.0255* (36.9994)

With P-values < 0.05 we infer both models as suitable. All else being

equal, on average and given statistical significance; under FE those displaying

the HME are 9.9% more likely to increase the number of companies they hold

shares with. By comparison, this figure under RE is 8.58%. In order to make

some further inference we continue with RE and state that given statistical

significance, seeing as a higher value for birthyr indicates a younger

participant, the negative sign in front of the coefficient implies that being a

year younger decreases one’s likelihood of participating by 23.0746%, the

24

direction of which is consistent with the literature. Similarly, being male

increases this likelihood by 20.54%, and being involved in the financial

administration of one’s household too increases the likelihood by 13.47%.

As presented during the literature review that accompanies this report,

Garcia and Tessada (2013) highlight increased probability of having health

and life insurance given an individual has high school education. Here we

briefly analyse the relationship this education may have with one’s

willingness to risk their capital seeing as they are better insured. The sample

reveals 3.9% of those with a degree also hold insurance, and 6.7% of this sub-

sample displays the HME. Rooij et al. (2011) share similar results in their

positive correlation between financial literacy and having hired a financial

advisor, thus proving how better understanding from education encourages

one to engage in a policy/agreement that puts them in a better position for

financial activities such as stock trading.

25

4. Discussion and Conclusion

4.1 Overview and Limitations

As already discussed19, data availability20 meant the analysis took a different

course to most existing literature, and by comparison, both financialadmin

and insurance are likely to be inefficient measures due to the subjective

nature of survey questions and the limited categorisation of

saving/endowment policies.

Nonetheless, in line with initial aims of the project, findings reveal a

positive relationship between the behavioural bias that is the HME and

shareholding within the sample.

4.2 Relation to Existing Literature

Given that the main predictor variable is statistically significant (and larger

under the policy focused FE), these interpretations are most reasonable. The

majority of those that remain insignificant still appear to follow an intuitive

sign given the direction in existing literature.

Kahneman (2011) uses BFT to support the Efficient Markets

Hypothesis (EMH) by suggesting that those subject to biases are the source of

inefficiency in creating over/under reactions to market news. Barber and

19 See 1.3 20 Perhaps with greater access to daily trade data, a Differences-in-Differences regression would have been appropriate for policy evaluation as Carter Hill et al. (2011) highlight its benefits in better explaining causal effects via the average treatment effect.

26

Odean (2013)21 infer that studying the EMH is what has led to more research

on superior than household investors, despite our most profound result

being the disparity of HME among shareholders, where subjective experience

with gains/losses on the stock market cause trading strategies to vary from

SFT.

4.3 Policy Implications

Kaniel et al. (2004) confirm the bias as diversifiable and Rooij et al. (2011)

feel targeting specific groups, such as retirement, would be most effective.

Despite portfolio choice models assuming so, they find not even basic

financial knowledge to be widespread, concluding this to be a result of more

recent market liberalisation and structural reforms to social security that

have shifted decision-making responsibility toward private individuals.

The existence of the HME across all investor types provides justified

economic argument for the implementation of financial education

programmes, to address both the SMP in curbing participation costs and to

ensure greater awareness of available investment tools, in turn allowing

greater exploitation of stock market gains. This should create a more level

playing field in the market; Odean (1998) talks of participants’ trading

activities (especially in large volume) affecting overall market prices because

of the positive relationship, thus transmitting to market stability as current

trading prices become the future reference point.

21 Their research also justifies HM behaviour with overconfidence, sensation seeking, familiarity and like Kaniel et al. (2004), evidence of household investors behaving like contrarians as their market share is the minority by comparison to institutional investors who trade more aggregately.

27

5. Bibliography

Barber, B. M., Lee, Y., Liu, Y., and Odean, T. (2008) ‘Just How Much Do

Individual Investors Lose by Trading?’, The Review of Financial Studies,

22(2), pp. 609-632. doi: 10.1093/rfs/hhn046.

Barber, B. M. and Odean, T. (2013) ‘The Behavior of Individual Investors’,

Handbook of the Economics of Finance, pp.1533-1570. Available at:

http://dx.doi.org/10.1016/B978-0-44-459406-8.00022-6 (Accessed 2

December 2018).

Barberis, N., Huang, M. and Santos, T. (2001) ‘Prospect Theory and Asset

Prices’, The Quarterly Journal of Economics, 116(1), pp. 1-53. Available

at: http://faculty.som.yale.edu/nicholasbarberis/bhs_jnl.pdf

(Accessed 5 February 2019).

Bartels, B. L. (2008) ‘Beyond “Fixed Versus Random Effects”: A Framework

for Improving Substantive and Statistical Analysis of Panel, Time-

Series Cross-Sectional, and Multilevel Data’, pp. 1-42. Available at:

https://cpb-us-

e1.wpmucdn.com/blogs.gwu.edu/dist/1/170/files/2016/10/bartels_

cluster_confounding-1s24lhf.pdf (Accessed 13 April 2019).

Brown, P., Chappel, N., Da Silva Rosa, R. and Walter, T. (2006) ‘The Reach of

the Disposition Effect: Large Sample Evidence Across Investor

28

Classes’, International Review of Finance, 6(1-2), pp. 43-78. Available

at: https://onlinelibrary.wiley.com/doi/epdf/10.1111/j.1468-

2443.2007.00059.x (Accessed 4 February 2019).

Campbell, J. Y. (2006) ‘Household Finance’, The Journal of Finance, 61(4), pp.

1153-1604. Available at: https://doi.org/10.1111/j.1540-

6261.2006.00883.x (Accessed: 1 December 2018).

Campbell, J. Y. and Cochrane, J. H. (1999) ‘By Force of Habit: A Consumption-

Based Explanation of Aggregate Stock Market Behavior’, Journal of

Political Economy, 107(2), pp. 205–251. Available at:

http://schwert.simon.rochester.edu/f532/jpe99_cc.pdf (Accessed 5

February 2019).

Cocco, J. F., Gomes, F. J. and Maenhout, P. J. (2005) ‘Consumption and Portfolio

Choice over the Life Cycle’, The Review of Financial Studies, 18(2), pp.

491-533. doi: 10.1093/rfs/hhi017.

Carter Hill, R., Griffiths, W. E. and Lim, G. C. (2011) Principles of Econometrics

Fourth Edition. United States of America: John Wiley & Sons, Inc.

Garcia, R. and Tessada, J. (2013) ‘The Effect of Education on Financial Market

Participation: Evidence from Chile’, Finance UC SOC1102, pp. 1-40.

(Accessed 6 December 2018).

29

Gardini, A. and Magi, A. (2006) ‘Stock Market Participation: New Empirical

Evidence from Italian Households’ Behavior’, Department of Statistical

Sciences University of Bologna, pp. 1-22. Available at:

https://pdfs.semanticscholar.org/c3c3/ba8bf670e8d608db174e775b

e5915288d98.pdf (Accessed: 6 November 2018).

Greene, W. H. (2012) Econometric Analysis Seventh Edition. Harlow:

Pearson Education Limited.

Guiso, L., Haliassos, M. and Jappelli, T. (2002) Household Stockholding in

Europe: Where Do We Stand and Where Do We Go? Available at:

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=346542

(Accessed: 6 November 2018).

Hoechle, D. Robust Standard Errors for Panel Regressions with Cross-Sectional

Dependence. Available at:

http://fmwww.bc.edu/repec/bocode/x/xtscc_paper (Accessed 13

April 2019).

Hsu, Y. L. and Chow, E. H. (2013) ‘The house money effect on investment risk

taking: Evidence from Taiwan’, Pacific-Basin Finance Journal, 21(1),

pp. 1102-1115. Available at:

https://doi.org/10.1016/j.pacfin.2012.08.005 (Accessed 5 December

2018).

30

Kahneman, D. (2011) Thinking Fast and Slow. New York: Farrar, Straus

and Giroux.

Kahneman, D. and Tversky, A. (1979) ‘Prospect Theory: An Analysis of

Decision under Risk’, Econometrica, 47(2), pp. 263-292. Available at:

http://links.jstor.org/sici?sici=0012-

9682%28197903%2947%3A2%3C263%3APTAAOD%3E2.0.CO%3B-

3 (Accessed: 28 October 2018).

Kaniel, R., Saar, G. and Titman, S. (2004) Individual Investor Sentiment

and Stock Returns. Available at:

https://archive.nyu.edu/bitstream/2451/26545/2/FIN-04-023.pdf

(Accessed 6 February 2019).

Khorunzhina, N. (2012) Dynamic Stock Market Participation of Households

with Heterogeneous Participation Costs. Available at:

http://www.efa2012.org/papers/f3g3.pdf (Accessed: 31 January

2019).

Mattos, F. and Garcia, P. (2009) The Effect of Prior Gains and Losses on Current

Risk-Taking Using Quantile Regression. St. Louis, Missouri: Available at:

https://ageconsearch.umn.edu/bitstream/53035/2/confp01-09.pdf

(Accessed: 31 January 2019).

Odean, T. (1998) ‘Are Investors Reluctant to Realize Their Losses?’, The

31

Journal of Finance, 53(5), pp. 1775-1798. Available at:

https://faculty.haas.berkeley.edu/odean/papers%20current%20versi

ons/areinvestorsreluctant.pdf (Accessed 4 February 2019).

Rooij, M. V., Lusardi, A. and Alessie, R. (2011) ‘Financial literacy and stock

market participation’, Journal of Financial Economics, 101, pp. 449-

472. doi: 10.1016/j.jfineco.2011.03.006.

Tversky, A. and Kahneman, D. (1974) ‘Judgement under Uncertainty:

Heuristics and Biases’, Science, 185(4157), pp. 1124-1131. Available

at: http://links.jstor.org/sici?sici=0036-

8075%2819740927%293%3A185%3A4157%3C1124%3AJUUHAB%

3E2.0.CO%3B2-M (Accessed: 29 October 2018).

Wen, F., He, Z. and Chen, X. (2014) ‘Investors’ Risk Preference Characteristics

and Conditional Skewness’, Hindawi Publishing Corporation

Mathematical Problems in Engineering, pp. 1-14. Available at:

http://dx.doi.org/10.1155/2014/814965 (Accessed: 5 December

2018).

Wooldridge, J. M. Econometric Analysis of Cross Section and Panel Data.

Cambridge: The MIT Press.

32

6. Appendix

Please find below the list of Stata commands used to produce the results:

33

Word Count: 3,981.

34

April 2019

Risk Assessment Form for Assessing Ethical and Research

Risks

Please see Guidance Notes at the end of this document.

Students: Please make sure you have discussed this form with your supervisor!

Researcher’s name:

In case of students: Supervisor’s name: Degree course:

Part 1 – Research activities

What do you intend to do? (Please provide a brief description of your study and details of your proposed methods.) This research mainly studies the impact of subjective risk preferences via the house money effect, on the number of companies participants choose to hold shares with, over the time period of 1994-2018 (via panel data regressions in Stata).

Will your research involve collection of information from other people? (If yes, please provide a description of your proposed sample.) No, this research project used information from De Nederlandsche Bank Household Survey (DNB, DHS).

If relevant, what locations are involved? (Please specify which country/region/place you will be working in, and details of where data collection activities will take place (e.g. public or private space).) All the information and data used was collected in the Netherlands and exists privately online unless granted access to the database for research purposes (as I was), but all of my own research and analysis using this data was carried out in the United Kingdom.

Will you be working alone or with others in the data collection process? I will be alone in the data collection process.

Part 2 – Potential risks to YOU as the researcher

Please specify potential safety issues arising from your proposed research activity. (Give consideration to aspects such as lone working, risky locations, risks associated with travel;

Isobelle Heidi Ager

Dr. Jian Tong

B.Sc. in Economics and Finance

35

please assess the likelihood and severity of risks.) If you have already completed a departmental H&S risk assessment, this may be attached to cover these aspects. To my knowledge there are no safety issues.

What precautions will you take to minimise these risks? N/A

Please specify potential distress or harm to YOU arising from your proposed research activity. (Give consideration to the possibility that you may be adversely affected by something your participants share with you. This may include information of a distressing, sensitive or illegal nature.) To my knowledge there is no potential harm to me arising from my proposed research activity.

What precautions will you take to minimise these risks? N/A

Part 3 – Potential risks to YOUR RESEARCH PARTICIPANTS

Please consider potential safety risks to participants from taking part in your proposed research activity? (Give consideration to aspects such as location of the research, risks associated with travel, strain from participation, and assess the likelihood and severity of risks.) If you have already completed a departmental H&S risk assessment, this may be attached to cover these aspects. N/A

What precautions will you take and/or suggest to your participants to minimise these risks? N/A

Please specify potential harm or distress that might affect your participants as a result of taking part in your research. (Give consideration to aspects such as emotional distress, anxiety, unmet expectations, unintentional disclosure of participants’ identity, and assess the likelihood and severity of risks.) N/A

What precautions will you take and/or suggest to your participants to minimise these risks? N/A

Part 4 – Potential wider risks

Does your planned research pose any additional risks as a result of the sensitivity of the research and/or the nature of the population(s) or location(s) being studied? (Give considerations to aspects such as impact on the reputation of your discipline or institution; impact on relations between researchers and participants, or between population sub-groups; social, religious, ethnic, political or other sensitivities; potential misuse of findings for illegal,

36

discriminatory or harmful purposes; potential harm to the environment; impacts on culture or cultural heritage.) No.

What precautions will you take to minimise these risks? N/A

Part 5 – International Travel

If your activity involves international travel you must meet the Faculty’s requirements for Business Travel which are intended to:

1. Inform managers/supervisors of the travel plans of staff and students and identify whether risk assessment is required.

2. Provide contact information to staff and students whilst travelling (insurance contact details, University contact in case of emergency etc.)

Full details are provided in the Faculty H&S Handbook in the Business Travel section. Selecting Business Travel from the Contents list will take you straight to the relevant section.

Departmental H&S risk assessment attached (for Part 2/3)

NO (Delete as applicable)

Business Travel and Risk Filter Form attached (Part 5)

NO (Delete as applicable)