Is there a relationship between risk aversion, impatience and
household savings behaviour?
A micro empirical case study of the savings behaviour of Dutch households
Gorka Bemer 5838398 Master’s thesis
University of Amsterdam
Faculty of Economics and Business (FEB) Msc Economics; monetary policy and banking
Supervisors: dr. Michiel Bijlsma (CPB) and Nancy van Beers (CPB) Amsterdam, november 2013
First reader:
prof. Casper van Ewijk (CPB and UvA) Second reader:
Abstract
Recent interest in the determinants of savings has been sparked by the financial crisis and its ongoing influence on the financial situation of households worldwide. A study at the individual household level by using micro data can give varying insights on household savings behaviour. This study uses a micro dataset from the CentER data-panel to analyze the relation between risk attitudes, impatience and the savings behaviour of Dutch households. Then we assume the preferences are constant and look for a change in savings since the crisis, that is a break in the dataset. Also, this research analyzes if the connection between these preferences has been affected by the financial crisis. We find that there is clear link between risk attitudes, time preferences and the savings of Dutch households. We can also conclude there has been a significant in savings since the crisis. However, we can not conclude that the relation between savings and risk attitudes and time preferences has changed since the financial crisis.
Table of contents
1. Introduction 4
2. Theoretical framework 6
2.1. Life cycle permanent income hypothesis 6
2.2. Precautionary savings 7
2.3. Buffer-stock savings model 8
2.4. Preferences, savings and idiosyncratic shocks 10
2.5. Hypotheses 11
3. Methodology 12
3.1. Reduced form regression equation 12
3.2. The dataset 13
3.3. Sample selection 13
3.4. Basic demographic characteristics 14
3.5. Variables 14 3.5.1. Wealth measure 14 3.5.2. Permanent income 15 3.5.3. Income risk 16 3.5.4. Risk aversion 17 3.5.5. Impatience 18 3.5.6. Crisis dummy 18 3.5.7. Controls 19 4. Empirical analysis 19
4.1. Impatience, risk aversion and savings 20
4.2. Change in savings behaviour 24
4.3. Change in the relation between preferences and savings 25
4.4. Checking for robustness 27
5. Conclusion 30
References 31
Appendix A 36
1. Introduction
Recent interest in the determinants of savings has been sparked by the financial crisis and its ongoing influence on the financial situation of households worldwide (Carrol et al. 2012, Slacalek and Sommer 2011). Consumption and therewith saving decisions are one of the most important decisions households constantly have to face. The decision is however complex due to the intertemporal horizon, as well as the different motives to save and the interaction among these motives (Ziegelmeyer 2009). Economic theory on savings has a long tradition which started early in the twentieth century. Altough, economic theory gives some structure to the understanding of savings behaviour the complexity of the phenomenon requires additional empirical observations to get more clear insights.
Rising unemployment, pension reforms and declining housing prices have directly affected the financial wealth of Dutch households for the last few years. As a consequence, we could expect that Dutch households have adjusted their savings to face these rising uncertainties. This could be the result of a change in the perception of risks. This study will focus on the effect of an idiosyncratic shock, the financial crisis, on the savings behaviour of Dutch households’ trough the intermediary of their preferences. The focus will especially be on risk attitudes and time preferences. An important thing to keep in mind is that preferences might have stayed constant since the crisis but the
perception of households of risks and uncertainties might have changed. This could possibly explain a change in the savings of households.
For this study, panel data from the DNB Household survey will be used. These yearly surveys started in 1993, and on average 2000 households participate in the survey. The survey contains questions about the financial position of Dutch households and contains psychological aspects that can be used to proxy for risk attitudes and time preferences.
This study is closely related to the papers from Lusardi (1998), Ziegelmeyer (2009) and Kolerus et al. (2012). These three papers use household survey data to estimate some aspects of savings behaviour. The first analyzes the U.S. households. The second and third analyze German households. The contribution of our research is that household data will be used to analyze
preferences that originates from the DNB household survey. The dataset is assembled by the CentER research institution which is closely linked to the Tilburg University and sponsored by the Dutch central bank (DNB). Specializing in internet surveys, CentERdata annually questions approximately 1,500 households (over 2,500 persons) about their financial characteristics and behaviour (e.g., their savings and investment behaviour, their housing wealth, mortgage and other debts, accrued pension rights etc.).
Browning and Lusardi (1996) concluded that the puzzling results of savings and consumption studies called for more sophisticated and reliable data sources. The authors made several recommendations for existent and future surveys e.g. the inclusion of questions that could be used within the
behavioural economics framework. The data that are used for this research originates from the DNB household survey (DHS) and includes many of these recommendations. In that way, the survey enhances the possibilities of research on the financial behaviour of Dutch households. By using the DNB household survey a different dataset is used than similar studies used.
This studies contributes to the existing line of research by testing the findings with a new dataset. Additionally, this research investigates the effect of the financial crisis. Previous research could not account for the crisis because of the simple fact that it had not occurred yet or that there were not enough years of post-crisis data available yet. The one study that does account for the crisis is Kolerus et al. (2012), but as stated before the study uses a German dataset.
The empirical analysis is carried out as follow: a form of wealth is used as the dependent variable. The right-hand side of the equation will be formed by permanent income, household characteristics and will be complemented by risk attitude and time preference. Which are the explanatory variables of interest. Results will show that there is a significant relation between savings and these preferences. Although a test will indicates that the savings behaviour of Dutch households has significantly changed since the crisis, there are no robust results that argue in favor of a change in the relationship between savings and preferences.
The remainder of the paper is structured as follows: section 2 reviews the main savings theories, the existing literature and discusses the hypotheses to be tested. Section three will give a description of the dataset, the model and the variables that will be used. The empirical analysis begins in section 4 with the estimation of multiple OLS and GLS regressions. Various specifications of the regressions are presented. Furthermore, the section contains a test for a break in the dataset and checks the robustness of the results. Finally, section 5 will conclude.
2. Theoretical framework
2.1 Life cycle permanent income hypothesis
The basic theoretical framework for analyses of saving behaviour is the standard life-cycle permanent income model. This model is a combination of the permanent income hypothesis of Friedman (1957) and the life-cycle model of Modigliani and Brumberg (1954). The life cycle model makes some simplifying assumptions in order to characterize a well-defined optimization problem (Shefrin and Thaler, 1988). The permanent income hypothesis states that choices made by consumers regarding their consumption patterns should be based on their permanent income, rather than changes in temporary income. The life-cycle permanent income model combines the assumption that savings is a function of permanent income1 and agents are rational. The rational agents are forward looking and will plan their consumption and saving over their entire lifetime. This results in an agent that will maximize his intertemporal utility by choosing the optimal amount of consumption (and saving) in each period, that is smooth his lifetime consumption.
According to the classic model the agent will borrow if his current income is below his permanent income, and the agent will save, if his income is above his permanent income. As a consequence the agent moves resources from one period to another until his marginal utility is stable over time i.e. smoothing of consumption over the cycle (Romer, 2006). The resulting life-cycle profile is illustrated by figure 1.
Figure 1: Life-cycle saving, income and consumption
1 Permanent income is defined as the total lifetime recourses of the rational agent divided by the number periods the agent will expect to live (Romer 2006, pp. 347-348).
The implications of this profile are well known. As figure 1 shows agents, households or individuals should borrow money to finance their consumption at the beginning of their career (phase A). When earnings start to increase they start to pay off debt and accumulate wealth (phase B), up to the point that they own a sufficient amount of wealth to smooth their consumption during retirement (phase C). The model assumes that the agents are hedged against idiosyncratic shocks and have no capital constraints.
The life cycle permanent income model, hereafter life-cycle model, has been the basis for many empirical studies. Browning and Crossley (2001) showed that predictions of the life-cycle framework indeed explain some of the observed patterns of household savings behaviour.
In the last decades, however, empirical studies have shown that the framework has its limitations. Recent empirical evidence showes that American workers, for example, save less than the life-cycle model would predict. This means that they save less than what would be needed to support their consumption after retirement. Hence, they experience a decline in their standard of living after retirement (Lusardi 1999, Bernheim et al. 2003). Another contradiction is the savings behaviour of German households. Despite a generous health and a general pension system the German elderly continue to save after retirement (Börsch-Supan et al. 2003). A similar contradictory trend has been found by using Italian data (Ando et al. 1993).
2.2 Precautionary savings
An extension of the life-cycle model was proposed by Leland (1968) by introducing the effect of uncertainty on savings. This so-called precautionary savings theory argues that savings are not only a tool to smooth consumption over the life-cycle but also function as an insurance against income shocks. The precautionary savings model extends the traditional model by introducing imperfect capital markets where future labor income is uncertain. In this model savings depends not only on expected income but also on the variance of income (Arrondel 2002, p.188).
These assumptions suppose a certain class of utility functions. In a simple two period setup with known labour income in period 1 and uncertain labour income in periode 2, consumption in periode 1 should be chosen in a way that the expected utility in period 2 is maximized. Leland (1968) described in his article that risk aversion alone could not capture higher savings by increasing wage uncertainty. Leland (1968) and later Sandmo (1970) concluded that precautionary demand for savings where the results of a concave utility function with a positive third derivative. Logically, a quadratic utility function cannot represent the precautionary savings motive since it has a third derivative of zero. Kimball (1990) extended this analysis by introducing the concepts of relative and
absolute prudence which resemble the Arrow-Pratt measure of relative and absolute risk aversion. While risk aversion is based on a negative second derivative, the third derivative needs to be positive in the presence of prudence (Ziegelmeyer 2009). These findings have led to the use of constant relative risk aversion (CCRA) utility functions in the theoretical literature of precautionary savings (Kimball 1990, Guiso et al. 1992, Arrondel 2002). In this paper we assume that prudence is highly related to risk aversion.
The driving force of the precautionary savings motive is twofold. First insurance markets are imperfect or nonexistent. Second, macroeconomic shocks make insurance impossible, even if markets are perfect. As a consequence households cannot insure themselves against income risk and face income uncertainty. To face this uncertainty households consume less and accumulate wealth i.e. save a part of their current income. The addition of precautionary savings to the life-cycle model makes it less restrictive and makes it possible to model more complex savings behaviour (Browning and Lusardi, 1996, pp.1798, 1808).
An addition to the precautionary savings hypothesis is the introduction of liquidity constraints. If the possibility of a household to get credit is restricted it may want to hold a larger amount of precautionary savings as insurance to wealth shocks. On the other hand, if a household can easily get credit it may want to hold smaller amounts of precautionary savings. Carrol et al. (2012) showed that the long structural decline of the savings rate in the United States since the 1960s can be attributed to increased credit availability.
2.3 Buffer-stock savings model
The new approaches of behavioural economics question the life-cycle framework. Some have tried to add behavioral aspects to the classic model. For example, Shefrin and Thaler (1988) formulated a behavioral life-cycle hypothesis. This line of research was further explored by a series of laboratory experiments initially started by Kahneman and Tversky (1979) and Thaler (1994, pp.186-188). They showed that households are not that successful in building rational expectations. Households appear to be more short sighted and less able to process information than their counterpart, the rational economic agent, is assumed to be. These sorts of empirical findings and new approaches opened the road to the development of extensions of the life-cycle model.
A recent theoretical addition to the savings theory was developed by Carroll (1992, 1997) and Deaton (1991). The “buffer-stock” model of saving, hereafter buffer-stock model, is an intertemporal model of consumption behaviour under uncertainty. Within the buffer-stock model the household is assumed to be impatient and prudent. Impatience is the preference to consume
today compared to consuming tomorrow. In other words, the household would like to borrow against future income if the income is constant over time and there is no uncertainty (Ziegelmeyer 2009). As in the precautionary saving model a CCRA utility function with a positive third derivative is assumed, so the “prudent” household will increase its savings with increasing (income) uncertainty.
A household is both facing impatience; it wants to increase its consumption in the present, and prudent, which will lead to higher future consumption. The symbiotic relationship of both characteristics results in a target wealth to income ratio. This ratio determines the buffer stock of optimal wealth a household will hold to insure itself against income uncertainty. If wealth is below its optimal target, prudence will dominate impatience and the household will save more. On the contrary if wealth is above its optimal target, impatience will dominate prudence and the household will dissave until the ratio is restored. Carrol (1997, p. 46) argues that this feature of the buffer-stock model aligns well with the prediction of financial planners, who advise people to hold a certain ratio of wealth for a rainy day.
The buffer-stock model has appeal because as Carrol (1997, pp.32-38) describes in his article it provides answers to three empirical savings puzzles that other models fail to explain. Still the model has its limitations. The model assumes that the household faces the risk that income will go to zero. This assumption is not realistic for countries which have a social security system like the most European countries. Also the model does not provide explanations for the savings behaviour of very wealthy households. Nevertheless, the model provides a sound framework to investigate savings behaviour by assuming a positive relationship between buffer-stock holdings and risk. While assuming a pivotal role for preference parameters, such as the degree of impatience (Kennickell and Lusardi, 2006).
Households have different preferences and motivations that drive their savings behaviour. The literature has provided possible theoretical factors that explain why households increase or decrease the amount they save. Numerous motivations have been proposed to explain specific savings behaviour e.g. old-age provision, bequest and insurance. Still, initial household preferences are the fundamental building blocks of the newer theoretical savings models. In this current paper preferences, especially risk aversion and impatience, will be examined. Notice that this paper focuses on risk aversion and not prudence. In other words, a utility function with a negative second derivative is assumed for the individual household. However, it is not assumed that the third derivative is positive which would make it a prudent household. This choice is made due to the boundaries of the dataset. By making this choice this paper is more closely related to Lusardi (1998), Ziegelmeyer (2009) and Kolerus et al (2012) than to the theoretical literature of precautionary
savings like Kimball (1990) Guiso et al. (1992) and Arrondel (2002) that assume relative risk aversion (CCRA) utility functions. We will analyze the relationship between impatience, risk aversion and the savings behaviour of Dutch households.
2.4 Preferences, savings and idiosyncratic shocks
Since the peak of the business cycle in 2007 the savings rate of U.S. households has shown a remarkable rise. The savings rate has risen more in the aftermath of the recent financial crisis than after all other recessions since the Second World War. According to Carrol et al. (2012) and Slacalek and Sommer (2011) this sharp rise in the savings rate can be attributed to three different channels;- precautionary, wealth and credit channels. These three channels passed the revue in the previous sections. Carrol et al. (2012) quantify these three channels and base them on theoretical considerations as risk aversion, time preference and expected income growth. In their model these household preferences are determinants of the separate channels that affect the savings rate. One of their conclusions is that the recent rise in the savings rate was driven by tightening credit conditions which forced households to increase their savings. A result that was also found by Mian et al. (2013).
Households save in response to uncertainty about employment, earnings and numerous other factors. Logically, the way in which an individual household copes with these kind of uncertainties depends on individual (time-) preferences and attitudes towards risk. These household characteristics shape decisions about consumption and savings in a fundamental way and are captured in the utility function of the individual household. How households adjust to an economic shock could be the result of these same preferences and attitudes towards risks. One could assume that an idiosyncratic shock, as the recent financial crisis, has a significant influence on the way households perceive their current and future wealth and that they will adjust their consumption and savings behaviour accordingly. However, an idiosyncratic shock will change the circumstances to which they will adjust their savings behaviour but not the underlying utility function. In other words, the preferences of households might stay constant while the perception of risk and uncertainties changes. Risk attitudes and time preferences have been used in some empirical studies about savings behaviour of households before. Overall, risk aversion measures are used more often than time preferences in empirical savings studies.
By using a dataset of the Health and Retirement Study (HRS), Lusardi (1998) found that the amount of wealth U.S. households hold for precautionary motives is partly based on their preferences. With an empirical estimate of a regression of wealth divided by permanent income on a
set of household characteristics and an income risk measure she found that more risk averse households accumulate more wealth. Also, based on her results she concluded that respondents with long planning horizon, i.e. patient households, have higher wealth stocks than households with shorter planning horizons. Kennickell and Lusardi (2006) on the contrary could not find results that show a robust connection between time preference and household saving behaviour based on two waves of the U.S. survey of Consumer Finances. Finally, Kolerus et al. (2012) found significant results in one of their estimations using the German SAVE study. Their result supports the view that more risk loving individuals tends to have lower savings rates.
2.5 Hypotheses
In this study the relation between the preferences of Dutch households and their savings behaviour is examined, in particular if this relation has changed since the financial crisis2. This will be approached by determining which household characteristics contribute to the degree of savings. On the basis of several hypotheses the savings behaviour will be analyzed.
Three hypotheses will cover the main aspects of this paper. First two individual hypotheses will be analyzed seperately and the third hypothesis is a combination of the first two. The hypotheses are based on the intuition that the financial crisis affected the savings behaviour of Dutch households trough a change in the preferences of the households.
H1: Prudence and impatience have a significant impact on the savings behaviour of Dutch households.
Two subjective measures will be used in the regression analysis to analyze the relationship between prudence, impatience and savings behaviour. As stated before we assume that risk aversion and prudence are highly related. According to different savings theories one would expect that prudence and impatience significantly affects savings behaviour. One would expect that a relatively more risk averse household would save more than a relatively more impatient household and vice versa.
H2: The savings of Dutch households has significantly changed since the financial crisis.
The second hypothesis will be tested by testing for a break in the regression coefficients. As stated before, we can expect that there has been a significant change in the savings of Dutch households since the financial crisis by the precautionary, wealth and credit channels. By introducing
2 The bankruptcy of Lehman Brothers is taken as the start of the financial crisis. Section 3.5.6 will elaborate on the exact “break” that is used to test if the savings behaviour has changed since the financial crisis.
a binary interaction variable for every independent variable the Chow-test can be used to test for the null hypothesis that there is no “break” in the regression coefficients. In other words, test for the null hypothesis that savings has not significantly changed since the financial crisis.
H3: The relationship of risk aversion and impatience with the savings behaviour of Dutch households has significantly changed since the financial crisis.
This hypothesis combines the previous two. According to the buffer-stock theory households will compensate a shock e.g. a decline in wealth, by saving more or less to restore their target wealth ratio. One could argue that the financial crisis has not only affected the amount households save. The financial crisis could be a shock with such an idiosyncratic nature that it also affects the uncertainty households have about their future wealth situation. In other words, preferences could be unchanged after a macroeconomic shock but perceptions of risks and uncertainties might change.
3. Methodology
3.1 Reduced form regression equation
Before giving a description of the dataset and used variables it is useful to introduce the regression equation that will be used:
( ) ( )3
In both equations is a measure of household wealth held by household i. Permanent income is reflected in the equation by and is some kind of income risk measure household i faces. Finally, is a vector of control variables e.g. age, number of children and marital status. The second equation will be used in this research to allow for non-homothetic preferences. The logarithm of permanent income is added to the right hand side as can been seen in equation B. The functional form of the logarithm and the permanent income variable on the right hand side has an advantage. As a consequence, the influence of explanatory variables on the ratio of precautionary savings on permanent income must not be constant and is allowed to change e.g. with permanent income (Ziegelmeyer, 2009, pp. 62-63). The regression equation will be elaborated after the description of the dataset, demographic characteristics, justification of different explanatory and control variables.
3
Almost all micro-empirical studies3 analyzing precautionary savings behaviour uses this equation. Some examples; Guiso et al. (1992, p. 324), Lusardi (1997, p. 323; 1998, p. 449), Arrondel (2002 p. 188), Essig (2005, p. 5), Kennickell and Lusardi (2006, p. 6), Bartzsch (2006, p.4), Ziegelmeyer (2009, p.92), Kolerus et al. (2012, p. 11).
3.2 The dataset
Browning and Lusardi (1996) stated that the puzzling results of savings and consumption studies called for more sophisticated and reliable data sources. The author made several recommendations for existent and future surveys e.g. the inclusion of questions that could be used within the
behavioural economics framework. The data that are used for this research originates from the DNB household survey (DHS) and includes a lot of these recommendations. The panel started in 1993 and contains information on economic and psychological aspects of the financial behaviour of a sample of Dutch households. The dataset is assembled by the CentER research institution which is closely linked to the Tilburg University and sponsored by the Dutch central bank (DNB). Specializing in internet surveys, CentERdata annually questions approximately 1,500 households (over 2,500 persons) about their financial characteristics and behaviour (e.g., their savings and investment behaviour, their housing wealth, mortgage and other debts, accrued pension rights etc.). The survey contributes significantly to the possibilities of research on the financial behaviour of Dutch
households.
Households yearly flow in, and drop out of the panel. This has two consequences. First, households are part of the panel for a couple of years on average. Second, the number of households that participate differs per year. As was stated by van Beers and Bijlsma (2013) the structure of the panel has implications for representativeness of the panel. Differences in
recruitment, approach and willingness to participate to the panel creates differences between the sample and the population. This makes the data less useful for the computation of the savings behaviour of the average Dutch household. For instance, urban and young households are under-represented while wealthy households are over-under-represented. Nevertheless, the survey still is a useful source to examine developments in savings behaviour and differences between households.
3.3 Sample selection
The micro panel dataset consists of thirteen yearly rounds of the DHS between the years 2000 and 2012. The sample is constructed by merging all thirteen rounds. For each year one observation is kept per household. The sample then numbers a total of 23,592 observations.
Unfortunately not all 23,592 observations can be used in this analysis. First, households don’t necessarily fill in all the variables of interest. This so called item non-response creates a partial lack of information. Some respondents agree to participate in the survey but do not answer certain questions. As a consequence for some observations their is a lack of data on some items. This significantly reduces the number of observations when the regression analysis is performed. For
instance, the question that is used to construct a measure of risk aversion has only 15,111 observations (including 88 observations that state don’t know)4. Logically, by including more variables with non-response items the sample size is reduced with an associated loss of statistical efficiency.
Second, we restrict our sample by dropping some outliers. The handling of outliers is critical to get unbiased estimation results. By using the logarithm as functional form for the explained variable, possible problems with outliers may be reduced. The sample is additionally restricted by dropping the 2nd percentile and the top 2% of the explained variable.
3.4 Basic demographic characteristics
Table 1 in appendix A shows the basic demographic characteristics of the sample for the thirteen rounds between 2000 and 2012. Statistics concerning age, sex, marital status, average income, education and region are tabulated. The statistics in table 1 are weighted. This means that weights are added at the household level to increase the representativeness of the sample.5 By using this method, data on income becomes more representative for the population. Nevertheless, this method has no influence on characteristics as age and education, so differences between the sample and the population still exist.
3.5 Variables
This section will give a description of the dependable, explanatory and control variables that will be used in the regression analysis. First the variables that are at the focus of this research will be described before there will be given a description of the necessary control variables.
3.5.1 Wealth measure
According to the precautionary savings framework an increase in uncertainty will increase savings. This means that asset accumulation will increase as well. In theory, savings is related to wealth trough the intertemporal budget constraint. Following this relationship the impact of uncertainty on savings should be equal to the impact on net wealth. In practice, savings and net wealth differ (Guiso et al, 1993). This makes it difficult to define a measure of wealth that satisfies the needs of precautionary savings. Most micro-empirical studies use some form of financial wealth or total net wealth (Ziegelmeyer, 2009).
4 For the subjective risk aversion measure I drop all observations which states don’t know as response. 5 See Van Beers and Bijlsma (2013) and Alessie, Lusardi and Aldershof (1997) for further explanation of this weighting method.
But the financial and total wealth measures are quite different. This is the consequence of the combination of accessibility and liquidity of the underlying assets and liabilities of the wealth measures and savings motives of the household. The time required for accessibility of money plays a major role and depends on the risks and uncertainties a household has to face. The main explained variable that will be used in this study is liquid assets and liabilities. Also the wealth measure does not contain liquid, but risky, assets as stocks. This financial wealth measure is constructed as follow:
Financial wealth = (checking accounts + savings + bonds + savings insurance + life insurances) -
(credit cards + debet checking accounts)
But this financial wealth measure could be overly restrictive. So a total net wealth measure is constructed to test for robustness of the financial wealth results. In line with the literature housing wealth is left out of the equation.
Total net wealth = (cars + motorbikes + boars + caravans + lent money + savings + investments + checking accounts + bonds + stocks + savings insurance + life insurance) -
(credit cards + debet checking accounts + personal loans + consumptive credit + loans family and friends + student loans + other loans)
By using these two measures we use an indirect approach of determining the amount households save. This is contrary to some studies that use a direct measure for determining the amount households save or want to save. For example, some studies that use the German SAVE study which is comparable to the DHS uses this approach (Schunk 2007, Kolerus et al 2012). The difference is that the SAVE survey contains direct questions about savings. This provides the researchers a direct measure of savings. Even better, a question about the desired amount of precautionary savings makes it possible to make comparisons between financial wealth, total net wealth and the amount of wealth households’ whish to hold. Unfortunately, the DHS does not contain an adequate question in the survey to use this method. In all regressions the wealth measure is restricted to drop outliers. The second percentile and the top 2% are left out of the econometric analyses.
3.5.2 Permanent income
Permanent income represents the income a household would earn in the absence of idiosyncratic shocks. According to the permanent income hypothesis households should consume part of this permanent income and not a part of current income. Because households face risks and uncertainty a household should hold an amount of precautionary savings to smooth consumption around their
permanent income. In this way households have the possibility of maintaining a living standard while facing risks and uncertainty. Because households strive to have a smooth standard of living, precautionary savings should be set into relation to permanent income (Engen and Gruber, 2001).
The permanent income measure is constructed by regressing the logarithm of household income6 on a group of household characteristics. I use age, sex, region, marital status, occupational dummies, education, drinking behaviour, smoking and health condition. The results of this prediction are used as a proxy of the permanent component of income.
3.5.3 Income risk
The computation of an income risk measure seems to be the most difficult aspect of precautionary savings research. The literature provides different approaches which all have some limitations. A commonly used proxy for income risk is the variance of total income (Carrol and Samwick, 1998). An objection against this measure is that it may be possible that a household knows its income variation and therefore can insure against it. Another method to proxy for income risk is by introducing occupational dummies in the regression analysis. The intuition behind this approach is that groups with a risky occupation will tend to save more because they face higher income risk. Earlier studies showed that risky occupational groups in fact save less. This could indicate that people self-select themselves in different types of occupations according to their risk aversion (Skinner 1988). Some studies have used a regional unemployment figure to proxy for the risk that someone becomes unemployed and loses (part of) his income. This source of income risk is exogenous and only reflects the general probability of becoming unemployed and not the individual’s probability. Lusardi (1998) used a subjective measure to proxy for income risk. Respondents where asked the probability they thought they had of losing their job in the next year.
Albeit the different approaches do not contain a perfect measure for income risk this research will still borrow two of these measures. This is done because income risk is a fundamental part of different precautionary savings theories. First, we will proxy income risk by introducing occupational dummies7. Because the analysis also contains measures of prudence and impatience I think part of the self-selection is in fact neutralized. Second, the subjective measure is constructed as Lusardi (1998) and Kolerus et al. (2012) have done. The DHS contains a question that asks respondents the probability that they will lose their job in the following year. The proxy is then constructed as follows:
6 Construction of the permanent income measure is described in appendix B. 7
( )
UNP is the likeliness households give to becoming unemployed in the following year. Households can attach a probability between 0 percent and 100 percent to this eventuality. This second measure may be favorable to the first one but as we will see this measure significantly decreases the sample size because of frequent non-response. This is due to the fact that the question about the probability of job loss was only introduced in 2005 so this question is missing for five waves of our dataset. 3.5.4 Risk aversion
The savings theories, and especially the buffer-stock theory, have shown that precautionary savings are mainly driven by a combination of income uncertainty, prudence and impatience. The more prudent a household, the greater the amount of precautionary savings it should be willing to hold. Just as with the income risk measure a critical point of micro-empirical studies is to find a appropriate measure for risk aversion.
Again, two kinds of measures are mainly used in the savings literature. The first one is a direct measure of risk tolerance. Respondents are asked several questions about their willingness to take a bet to increase income in the next year. The risk and payoff increase with each question and the answers are used to create a measure for risk aversion8. This kind of measure has the firmest grounding in economic theory but comes with two problems. First of all there is the practical problem that the DHS survey does not have such questions. This, of course, makes it quite difficult to introduce such a risk measure. Nevertheless, Alessie and Teppa (2002) have empirically shown that this risk tolerance variable has little explanatory power.
Another approach is to use a subjective measure for risk aversion. Fortunately, the DHS survey provides data that can be used to construct such a subjective measure. The starting point for this measure is the following question that is asked the respondents of the DHS survey;
I find it more important to invest safely and to get a guaranteed return than to take risks in order to possibly get a higher return.9
The question has an ordinal structure which means that respondents can choose within a scale ranging form completely disagree (=1) to completely agree (=7). Respondents who strongly agree with this statement are not willing to take risks and are thus considered to be very risk averse. The answers are used to create three dummies in the regression analysis by setting low risk aversion (= 1
8 See Barsky et al. (1997) for theory and construction of such a measure.
9 This kind of survey question was previously used to construct a risk aversion measure in Alessie et al. (2000) , Ziegelmeyer (2009) and Kolerus et al. (2012).
and 2), medium risk aversion (= 3, 4, and 5) and high risk aversion (6 and 7). This subjective measure
needs some intuition to make it a risk aversion measure and could be discredited as being too simple. Still, Alessie and Teppa (2002) have shown that these simple intuitive measures in fact appear to be more powerful predictors of risk tolerance than more sophisticated measures.
3.5.5 Impatience
Impatience is a time preference. Impatience could be described as the preference for immediate utility over delayed utility. The importance of this kind of measure was described in the section about the buffer-stock model. In the literature a smoking dummy is often used as proxy for a measure of time preference. The underlying idea is that smokers prefer the current utility of smoking over the negative influence it could have on their health on the long term. This proxy has been used by Lusardi (2003), Kennickell and Lusardi (2005) and Ziegelmeyer (2009). Nevertheless, this study makes the choice for another kind of impatience proxy. As Frederick et al. (2002) described in their article someone may be a smoker but still make thoughtful decisions about his retirement packages. This puts the adequacy of smoking behaviour as proxy in perspective.
The basis for the impatience measure is a question about the expenditure profile of the household in the DHS survey. The respondents are asked to answer the following question:
Some people spend the money they receive immediately. Others save some part of the money they receive to have some reserves. Could you indicate what you do with your money after you have spend the money on food, rent or mortgage and other basic needs.
The question has an ordinal structure and the respondent can choose within as scale ranging from 1 (=immediately spend) to 7 (=save biggest part). The answers are used to create three dummies in the regression analysis by setting high impatience(= 1 and 2), medium impatience (= 3, 4, and 5) and
low impatience (6 and 7).
3.5.6 Crisis dummy
A crisis dummy is created which serves several purposes. It gives the possibility to detect if there has been a change in the savings behaviour since the financial crisis. Also, by interacting the crisis variable with other variables of interest it is possible to analyze if some specific household characteristics have changed in the way they influence precautionary savings. The crisis dummy contains the value 1 if the year of response is 2009, 2010, 2011 or 2012. A reasonable question that follows is to ask why 2008 is not included in the crisis dummy. This research takes the fall of Lehman Brothers in 2008 as the start of the crisis and at that time the 2008 wave of the DHS survey was not
completed yet. The survey was conducted between February and September 2008, which means that the biggest part of the respondents could not take into account the start of the financial crisis. 3.5.7 Controls
In this paragraph the control variables that will be used in the regression analysis will briefly be discussed. Even if the focus of this research is on the risk attitude and impatience variables it is necessary to have control variables to get unbiased estimates. Common controls that do not need explanation are basic characteristics as gender, age, unemployment, education and marital status. The use such an extensive amount of controls has several reasons. First, the objective is to give a complete description of the data and to do this by using the richness of the DNB household survey. Second, I would like to show that the preference variables do not capture another omitted effect (Lusardi and Kennickell, 2006).
Two variables are introduced to control for household composition. One variable states the number of children in the household, and another one for the number of people in the household. The effect of children on precautionary savings is not clear. Having children may increase the willingness to increase precautionary savings but children may also limit to possibilities to do so. Additionally, extra people in the household that are not children will in most cases mean that there is an older person, probably a parent, living in the household. Older people often have their own source of income, but this may not be sufficient to offset the care and medicines that are needed. One would expect that precautionary savings increase to insure against an older person needing a bigger part of the household income.
Another important control is the expectation the about future financial situation of a household. According to the buffer-stock model if a household expects their wealth to decrease in the future they should increase their wealth to income target (Carrol 1997). In the DHS survey household are asked a question about their expectation on the development of their financial situation in the next year, in comparison to their current financial situation. This question is used to control for expectations about the future financial situation.
Finally, I include one dummy for home ownership and one for people who have a chronicle disease. Both should have an important effect on precautionary savings.
4. Empirical analysis
The following section will present the empirical estimates that are the results from the regression analysis. The structure will be based on the hypotheses that were formulated at the end of section 2 and consists of three parts. The first subsection will investigate the relationship between the household preferences, risk aversion and impatience, and the savings behaviour. In the second subsection there will be investigated if the savings behaviour of Dutch households has changed since the financial crisis. Finally, the third subsection will analyze if the influence of these preferences have changed since the financial crisis.
4.1 Impatience, risk aversion and savings
This section will investigate the relationship of the time preference and risk preference measures, respectively impatience and risk aversion, with the savings behaviour of Dutch households. The estimations will be based on the following specification:
( )
Here, is household financial wealth as described in section 3.5.1 and is the predicted log of permanent income which was described in section 3.5.2. For clarity of the specification is used as a vector for household characteristics; age, age², gender, marital status, number of children, number of persons in the household, marital status, education, expectations dummies and home ownership. The variable income risk will be estimated using two approaches. As was described in section 3.5.3 first we will proxy for income risk by using occupational dummies, second we will proxy for income risk by using a subjective income risk formula. All included regressors are described in more detail in the appendix.
According to the savings theories and in line with the hypothesis that was stated in section 2 impatience and risk aversion should have a significant effect on the savings of Dutch households. Both for risk aversion and impatience we use subjective measures which are introduced in the regression by using dummy variables. To avoid perfect multi-collinearity the dummies for high risk aversion and low impatience are dropped from the specification. In other words, high risk aversion and low impatience are the reference for the estimations. For the sake of the argument I will repeat the hypothesis before the results are presented.
H1: Prudence and impatience have a significant impact on the savings behaviour of Dutch households.
Where we would expect that relatively more risk averse households save more than less risk averse households. Impatient household are expected to save less than patient households. Table 1 shows the results of the regression. Regression 1 shows the OLS estimates with the occupational dummies to proxy for income risk. To be more elaborate, the occupational dummies for family business, student, disabled, housekeeping, entrepreneur and retired are used to proxy for income risk10. Each group should represent, in theory, some kind of income risk. The reference is households where the head of the household is employed. Hereafter, this will be referred to as occupational proxy. People who are employed are taken as basis. Regressions 2 shows the GLS estimates with the same occupational proxy. Regression 3 displays the OLS estimates with the constructed income risk measure instead of the occupational proxy. This income risk measure is constructed with the use of a subjective measure which has been introduced in section 3.5.3. Regression 4 uses the same income risk measure while using GLS. As was noted before, both income risk proxies have
10
Further description in section 3.5.3 or Skinner (1988). Dependent variable: Log(Financial wealth)
OLS 1 GLS 2 OLS 3 GLS 4
Variable β Std. Err. β Std. Err. β Std. Err. β Std. Err. Reference
High impatience -0,969 *** 0,071 -0,339 *** 0,060 -1,327 *** 0,112 -0,507 *** 0,095 Low impatience Medium impatience -0,267 *** 0,026 -0,135 *** 0,022 -0,407 *** 0,044 -0,175 *** 0,036 Low impatience Low risk aversion -0,459 *** 0,042 -0,138 *** 0,035 -0,304 *** 0,075 -0,069 0,063 High risk aversion Medium risk aversion -0,102 *** 0,026 -0,039 * 0,022 -0,085 ** 0,043 0,013 0,037 High risk aversion Log(permanent income) 0,659 *** 0,079 0,618 *** 0,068 -0,146 0,198 0,130 0,183 Age 0,023 *** 0,006 0,051 *** 0,008 0,105 *** 0,018 0,092 *** 0,023 Age² -0,007 0,006 -0,031 *** 0,008 -0,099 *** 0,019 -0,08 *** 0,025 Number of children in hh -0,157 *** 0,039 -0,121 *** 0,043 -0,289 *** 0,063 -0,284 *** 0,074 Number of persons in hh 0,126 *** 0,035 0,110 *** 0,037 0,197 *** 0,056 0,236 *** 0,064 Male 0,025 0,036 0,081 0,050 0,149 ** 0,061 0,105 0,083 Female Married -0,076 * 0,042 -0,107 *** 0,038 0,178 ** 0,078 0,030 0,075 Unmarried Home owner 0,604 *** 0,028 0,429 *** 0,038 0,649 *** 0,051 0,525 *** 0,058 No homeownership Primary school -0,238 *** 0,056 0,024 0,046 -1,331 *** 0,204 -1,211 *** 0,295 Acadamic
High school -0,170 ** 0,042 0,090 ** 0,043 -0,76 *** 0,096 -0,651 *** 0,113 Acadamic HBO -0,123 * 0,039 0,157 *** 0,047 -0,453 *** 0,067 -0,388 *** 0,096 Acadamic Financial capable/literate 0,21 *** 0,029 0,096 *** 0,270 0,133 *** 0,045 0,079 * 0,043 Non capable/lit. Employed 0,104 ** 0,046 -0,005 0,038 0,457 *** 0,134 0,132 0,111 No employment Expectations worse fin. sit. -0,119 *** 0,030 -0,025 0,023 0,155 ** 0,065 0,107 ** 0,049 Same expectat. Expectations better fin. sit. 0,324 *** 0,032 0,136 *** 0,025 0,321 *** 0,046 0,149 *** 0,038 Same expectat.
Family business -0,304 0,209 -0,742 *** 0,251 Employed
Student -0,679 * 0,378 -0,301 0,339 Employed
Disabled 0,106 0,067 0,118 ** 0,054 Employed
Housekeeping -0,370 *** 0,111 -0,283 ** 0,114 Employed
Entrepreneur -0,034 0,075 0,095 0,076 Employed
Retired -0,053 0,048 -0,740 * 0,042 Employed
Other occupations -0,214 *** 0,077 -0,069 0,068 Employed
Income risk 4,5E-05 *** 0,00002 6,91E-06 0,000012
Constant -5,718 *** 0,819 -6,208 *** 0,712 0,922 2,071 -1,699 1,948
N 11799 11799 3919 3919
R² 0,19 within 0,034 0,22 within 0,0212
between 0,2265 between 0,2317
overall 0,158 overall 0,1995
Table 2 Relationship between risk aversion, impatience and savings. Estimations (1) and (3) are OLS. (2) and (4) are GLS. Statistical significance is indicated with *,** and *** at the 10%, 5% and 1% level, respectively.
limitations.
The results in table 2 contain some ambiguity. The impatience dummies are significant and have the same negative sign in all four estimations. Because low impatience is the reference of all four estimations we can conclude that relatively more impatient households tend to save a smaller amount of their wealth. This result is in line with the intuition that impatient households have a preference for immediate utility over delayed utility which will cause them to save a smaller part of their income. The risk aversion dummies, on the other hand, do not show a clear-cut picture. The coefficients are significant and negative in the OLS (1), GLS (2) and OLS (3) estimates. In other words, according to these results a more risk averse household will save more than a relatively less risk averse household. In the fourth estimate with the income risk proxy and the random-effects model, the low and medium risk aversion dummies are not significant. According to these results we can conclude that the coefficients for risk aversion and impatience in estimations (1), (2) and (3) are in line with basic savings theories. Estimation (4) does not allow for such a conclusion. This could be the result of the different income risk measures that have been used and the large drop in the sample size that occurred. As was stated in section 3.5.3 the inclusion of the income risk measure decreases the sample size with two thirds. This of course comes with a loss of statistical power of the third and fourth estimations.
If a household has a permanent income that is 10 percent higher than that of another household it will save 6,59 percent more in the first estimate and 6,18 percent in the second. The effect is not significant in the third and fourth specifications. This is remarkable because, as stated in section 3.5.2, the permanent income theory states that households should save a part of their permanent income. In other words, the fact that there is no relationship between permanent income and financial savings in specification (3) and (4) is counterintuitive.
The age and age² variables are significant in estimates (2), (3) and (4). Age is additionally significant in specification (1). The magnitudes of the estimated coefficients in those three specifications imply a typical life cycle savings pattern and thus we could speak of economic significance. Savings are low at young ages, increase during the working part of life and decrease after retirement.
The variables of household composition show a robust pattern in all four specifications. Previous research show contradictory results. For example, Ziegelmeyer (2009) found that German households with more children or persons save more while Kolerus et al. (2012) showed that bigger households tend to save less. Table 1 shows that every additional child in a household decreases
savings between 12,1 and 28,9 percent. This may be an indication that children limit the possibility for a household to increase financial wealth. Additional persons in a household increase the amount of financial wealth between 11 and 23,6 percent. This could be the result of elderly people living in a household. They tend to bring some kind of income or wealth with them which may increase the possibilities for a household to have financial savings.
The male dummy is only significant in one specification (3). As a consequence, the conclusion can be made that gender does not influence the amount of financial wealth Dutch households in the other three specifications. Respondents indicating that they own their home appear more likely to save more. The coefficients for the home ownership dummy are positive and robust over all four specifications. Home ownership indicates higher financial savings between 12,7 and 40,8 percent than households that do not own their home. This could be attributed to the flows of income which could be the result of owning a home and that can be used to increase financial wealth (Harris et al. 2002). Household that own their home save between 42,9 and 64,9 percent more than households that do not own their home. Finally, the results for the marital status dummy are ambiguous. In the first two specifications the coefficients are significant and negative. In the third estimate the coefficient is positive and significant but in the fourth specification the coefficient is positive but not significant. Thus based on these results previous empirical findings that found a positive effect of being married on savings in the United Stated (Lusardi 1998) and Germany (Schunk 2009, Kolerus et al. 2012) can not be confirmed.
Financial literacy is associated with considerably higher savings. This is in line with the literature. Financially illiterate lack planning in their financial decisions which leads to lower savings . This can be the result of no retirement savings (because of no planning) or “hand-to-mouth behaviour” i.e. setting income equal to consumption (Lusardi 2008). The coefficients show a robust pattern. Overall, lower education and lower savings go hand in hand. The dummy for employed household heads is only positive and significant in the OLS estimations (1) and (3).
The coefficients of the expectations dummies cannot lead to a conclusion of economic significance. The first specification is not in line with theory. Households that expect a worse financial situation in the years to come should increase their savings, according to theory. The coefficients of specification (1) show the contrary. The coefficients of the expectations in the other estimations lack significance or show ambiguous signs. Specifications (3) and (4) state that households that have worse and better expectations for the future save more than households that think their financial situation will stay the same.
The results of the occupational proxy are counterintuitive and in line with earlier research (Skinner 1988). Economic theory predicts that riskier occupations should have a higher amount of precautionary savings but the results do not show this. The basis for the occupational dummies is a dummy for employed people so we would expect for example that entrepreneurs would significantly save more than this group. Estimations (1) and (2) show the dummy for entrepreneurs is not significant. One could argue that working in a family business is safer than the “normal” payroll job but this distinction is quite arbitrary. Unfortunately, the alternative income risk measure also does not provide clear-cut results. The OLS (3) coefficient is significant and small but positive. This is in line with economic theory and previous research (Lusardi 1998). A household facing a higher degree of uncertainty should increase its precautionary savings. Still, the effect is small relative to the other variables.
So which conclusion can be made? Impatience, and thus time preference, seems to have a significant relationship with the savings behaviour of Dutch households. The results for the risk aversion, or risk attitude, variables are significant in specifications (1), (2) and (3) which may be the result from the income risk measure that has been used. Because the estimates for the variables of interest where very similar across the four estimation methods – OLS and GLS with different risk measures – the OLS estimators for these variables are considered as being robust. Also, because the income risk measure could only be constructed with data since 2005 the occupational proxy should be used to make full use of the dataset and increase statistical power. To sum up, the results from the four estimations provide enough evidence to partly support the first hypothesis and the OLS specification with the occupational proxy is the preferred specification.
4.2 Change in savings behaviour
This section investigates whether Dutch households have significantly changed their savings since the financial crisis. As was stated in section 2.4 savings theory predicts that households adjust their wealth structure after the occurrence of a macroeconomic shock. According to research this has significant effects on the savings of households. It has been shown empirically that US households have increased their savings since the financial crisis (Carrol et al. 2012, Mian et al. 2013). The Dutch and US households are different in numerous ways. Nevertheless, we can expect that the financial crisis has affected the risks Dutch household face and therefore their savings has changed since. I
expect that there is a break in the data after the survey year 200811 and will test the following hypothesis that was previously stated in section 2.5;
H2: The savings of Dutch households has significantly changed since the financial crisis.
This hypothesis will be tested by testing for a break in the regression coefficients. This will be done by introducing a binary interaction variable for every independent variable and then using the Chow-test. The dummy that will be introduced and will be interacted with all the other variables is the
crisis dummy that was described in section 3.5.6. The null hypothesis will be tested that there is no
break in the used dataset. In other words, the null hypothesis that savings has not significantly changed since the financial crisis will be tested. The Chow-test will be performed on the four estimations that where discussed in the previous section. Table 3 shows the results.
Clearly, we can reject the null hypothesis that there is no break in the dataset for all four estimations. This result suggests that Dutch households have significantly changed their savings since the financial crisis. A result that is in line with theory and confirms earlier empirical research (Carrol et al. 2012, Slacalek and Sommer 2011).
4.3 Change in the relation between preferences and savings
This section will analyze if the relationship between risk aversion, impatience and savings have changed since the financial crisis. The specification that was used in section 4.1 will be extended to investigate the relation. The estimations will be based on the following specification:
( )
( ) ( )
To avoid repetition I refer back to section 4.1 for a description of the variables that where already used and will only elaborate on the newly introduced interaction variables. I expect a change in the relationship between the preferences of Dutch households and their savings behaviour. Significant
11 For explanation of the break date I refer back to the section 3.5.6 which describes the crisis dummy and the choice to exclude the 2008 “wave” of the DHS survey.
OLS 1 Occupational proxy F( 27, 11746) = 2.61 Prob > F = 0.0000 3 Income risk F( 21, 3877) = 1.74 Prob > F = 0.0191 GLS 2 Occupational proxy chi2( 27) = 51.78 Prob > chi2 = 0.0002 4 Income risk chi2( 24) = 74.28 Prob > chi2 = 0.0000 Table 3 Testing for a break in the dataset. Chow-test on four specifications.
interaction coefficients could indicate that there has been such a change in the preferences. For clarity hypothesis three is restated:
H3: The relationship of risk aversion and impatience with the savings behaviour of Dutch households has significantly changed since the financial crisis.
In order to analyze if the crisis has changed the preferences of Dutch households the variables of interest are interacted. As can been seen in the specification the crisis dummy12 is interacted with the impatience and risk aversion dummies. Following the example and argument that was introduced in section 4.3, OLS with the occupational proxy is used. Also, only the variables of interest and their interactions with the crisis dummy will be showed in table 5.13 The results of this specification are used to test hypothesis three.
The coefficients of the impatience dummies are in line with the results that were found in the OLS specification with the occupational proxy in section 4.1. Impatient households tend to accumulate a smaller amount of financial wealth than relatively more patient households. The interactions between the impatience dummies and the crisis dummy do not show a robust pattern. The coefficient of the interaction between highly impatient households and the crisis is negative and significant. This would indicate that the impatient households have additionally decreased their financial holdings. The interaction between medium impatient households and the crisis dummy shows a small negative coefficient but is not statistically significant.
12 The crisis dummy was introduced in section 3.5.6. 13
The complete extended table can be found in the appendix A.
Dependent variable: Log(Financial wealth) OLS
Variable β Std. Err. Reference
High impatience -0,850 *** 0,085 Low impatience
High impatience * Crisis -0,410 *** 0,155 Low impatience, pre crisis Medium impatience -0,251 *** 0,032 Low impatience
Medium impatience * Crisis -0,051 0,056 Low impatience, pre crisis Low risk aversion 0,473 *** 0,053 High risk aversion
Low risk aversion*Crisis 0,034 0,088 High risk aversion, pre crisis Medium risk aversion -0,123 *** 0,033 High risk aversion
Medium risk aversion*Crisis 0,068 0,057 High risk aversion, pre crisis
Crisis 0,059 0,049 pre crisis
Table 4 Relationship between risk aversion, impatience and savings. Estimation with OLS and interaction variables
The results for the risk attitudes dummies and their interactions concur with that of the impatience dummies. Again, the sign and magnitudes of the coefficients resemble the results that were found in section 4.1. The interaction between the dummy for households with low risk aversion and the crisis dummy is positive which would indicate that these households have increased their financial wealth since the crisis. Nevertheless the coefficient is not significant. The interacted medium risk aversion variable shows a better result. It is positive which means that households that have medium risk aversion increased their financial wealth since the financial crisis in comparison to households that are relatively more risk averse. One would expect that households with a relatively high risk aversion would be more sensitive to a financial shock. The argument could however be made that risk averse household already own a sufficient amount of financial wealth to protect themselves from a shock. On the contrary it could be that medium risk averse households felt they had to increase their financial holdings since the crisis to insure against future shocks. Finally, the crisis dummy is relatively small and not significant. Apparently, an after the crisis entry in the dataset does not significantly lead to higher savings.
We can not conclude that the relationship between the preferences of Dutch households and their savings behaviour has significantly changed since the financial crisis. In line with earlier results I found that the preferences do influence the savings behaviour. However results that show a change in the preference variables could only be found in twenty five percent of the variables (one out of four). Overall these results are not robust enough to confirm the third hypothesis. In other words, according to these results the relationship between risk aversion, impatience and the savings behaviour of Dutch households has not changed since the financial crisis of 2008.
4.4 Checking for robustness
As section 4.2 has shown there is a clear break in the dataset around the financial crisis. As a consequence the results for the variables of interest could be influenced by this macroecnomic shock. This section will investigate if the relationship between impatience, risk aversion and financial savings holds when we control for the financial crisis. By introducing year dummies i.e. year intercepts we can control for macroeconomic shocks. This section will test the robustness of the results in section 4.1. First I will try to disentangle the relationship of risk aversion and impatience with the savings behaviour by using OLS and year intercepts. Second, a more broad dependent variable will be introduced. The following specification will be estimated:
( )
This is the specification that was used in section 4.1 with the addition of the year intercepts for the years 2000 to 2011. To avoid perfect multi-collinearity the year 2012 is taken as reference and left out. The results are shown in table 3. To prevent redundant information table 4 only shows the regression results for the variables of interest. The complete results can be found in an extended table in appendix A.
Introducing year intercepts does not produce considerable different results from the estimations in section 4.1 for the impatience and risk attitudes dummies. The signs and the coefficients almost stay the same. We can conclude that even when we control for macroeconomic shocks there exist a relation between impatience, risk attitudes and financial wealth. Thus, the break in the dataset is not solely responsible for the relationship between the dependent variable and the variables of interest. In other words, the relation existed before as well as after the break in the data.
Sections 4.1 to 4.3 used financial wealth as dependent variable. As was stated in section 3.5.1 this kind of wealth measure could be overly restrictive. To prevent that the conclusions of the previous sections were the result of the nature of the underlying assets and liabilities of the financial wealth measure this section replicates some of the previous regressions with a broader wealth measure.
The total wealth measure that was introduced in section 3.5.114 will be used to check the robustness of the regressions that were done with the financial wealth measure. This total wealth measure contains assets and variables that are considerably less liquid. Three sensitivity analyses are reported. The first specification is estimated with OLS and the occupational proxy. The second regression adds the year intercepts to disentangle the relation between the preference variables and
14
Both wealth measures, financial wealth and total wealth, are described in appendix B. Dependent variable: Log(Financial wealth)
OLS
Variable β Std. Err. Reference
High impatience -0,957 *** 0,071 Low impatience
Medium impatience -0,266 *** 0,026 Low impatience
Low risk aversion -0,446 *** 0,042 High risk aversion
Medium risk aversion -0,079 *** 0,026 High risk aversion
Table 5 Relationship between risk aversion, impatience and savings. Estimation with OLS, occupational proxy
and year intercepts.
