Precautionary saving in the Netherlands during the financial crisis.
Abstract: this study quantifies the precautionary saving motive of Dutch households. The source of uncertainty used is a subjective earnings uncertainty measure, provided by the LISS panel. Existing literature shows that the impact of precautionary saving is negligible if the study makes use of a subjective earnings measure. Previous studies for the Netherlands all estimated the precautionary saving behavior by using wealth accumulation data. This study differs in the empirical approach chosen, as precautionary saving is estimated using annual household non-durable consumption data. No significant effect of precautionary saving on annual household non-durable consumption is found. Therefore, the main finding of this study is that Dutch households do not engage in self-insuring behavior, but are also poorly incentivized to do so. By law, every employed individual in the Netherlands is insured against unemployment. This might lead to moral hazard behavior of Dutch households.
Rijksuniversiteit Groningen Faculty of economics and business MSc. Economics
Master’s Thesis Economics EBM877A20
Date: June 23, 2017
Supervisor: Prof. dr. R.J.M. Alessie Co-Assessor: V. Angelini
Mike Hendriks S1885464
Keywords: Life-cycle model, consumption behavior, precautionary saving, survey data
2
1. Introduction
The period between 2008 and 2015 is known as the financial crisis, during which individuals had to cope with increasing job uncertainty. The number of Dutch households that received unemployment benefits increased from 318.000 to 614.000. In the same period the Dutch government engaged in policy reforms concerning the public unemployment insurance (UI) scheme. As of January 1, 2016, the unemployment benefits in the Netherlands have become less beneficial to the unemployed. The duration of unemployment benefits used to be from a minimum of three months to a maximum of 38 months, depending on job history. After January 1, 2016 the maximum duration is shortened by 1 month per quarter, to a maximum of 24 months in 2019. Individuals are entitled to one month of unemployment benefits for every year they worked in their first 10 working years. After these 10 years, every additional year of work entitles them to another half month of unemployment benefits. Before January 1, 2016, this relationship between years of work and the amount of months you were entitled to unemployment benefits used to be one for one. All these changes made the unemployment benefits scheme a lot less generous. Rational individuals would like to optimize consumption over their lifetime. Therefore they should take current period information into account together with expectations concerning their future income when optimizing consumption. Especially in times when the probability of losing your job is high and unemployment benefits schemes are becoming less generous should households incorporate income uncertainty. Income uncertainty might incentivize households to self-insure against future income shocks by consuming less and saving more. This study is set out to investigate self-insurance behavior of households.
The most famous economic model that deals with consumption decisions is the life-cycle model, first developed by Franco Modigliani in the 1950s. Modigliani and Brumberg (1954) state that increases in life-time resources lead to proportionate increases in consumption in all periods of life, depending on certain assumptions formed about the preferences of consumers. This means that total consumption is equal to the life-time resources and is smoothed out over each period. The growth in unemployment and job uncertainty during the financial crisis might influence the expectations households form about future income. The original life-cycle model does not deal with uncertainties, but the model was extended by Caballero (1990) to incorporate income uncertainty. The extended model by Caballero forms a starting point for this study.
3
precautionary reasons can be the result of households facing many types of risk. However, in this study income uncertainty is the only source of risk considered. Rational, risk averse households do not like risk and would like to take actions to mitigate risks. Therefore, households accumulate precautionary savings to overcome potential income drops. Mastrogiacomo and Alessie (2014) mention that precautionary saving has a negative effect on consumption. This relationship is at the core of the empirical analysis in this study. The data is derived from the LISS panel conducted by the CentERdata research center of the Dutch University of Tilburg. CentERdata conducted surveys in the period from 2007 to 2016. This timeframe suits this study nicely as the source of (increased) income uncertainty used is the financial crisis. The financial crisis started in the summer of 2007, reached its peak at the end of 2008 and the aftermath was still felt by 2015. To summarize, this study is trying to answer the following research question:
‘Did Dutch households engage in precautionary saving during the financial crisis?’ The answer to this question has social policy implications. An ongoing debate over social policies is one about the ability of individuals to self-insure. In this study, self-insurance would imply that by precautionary saving households are able to overcome potential incomeless periods, without the need of any social insurance scheme. The findings from this study are especially important for the UI scheme offered by the Dutch government. The value of UI is mitigated by the ability of households to self-insure. Government intervention may even lead to crowding out of precautionary saving if the existence of social insurance programs leads to moral hazard complications. Knowing that you are (fully) insured against the adverse event of losing your job, could entice you to engage in more risky behavior: you might not save (enough) for precautionary reasons. The following quote is from Feldstein, based on his study in 2005:
‘Existing programs have substantial undesirable effects on incentives and therefore on economic performance. Unemployment insurance programs raise unemployment. Retirement pensions induce earlier retirement and depress saving. And health insurance programs increase medical costs’.
4
offering a social insurance program is that households are protected against adverse income shocks. Households value insurance schemes because this allows them to smooth their consumption over time and across various states the world. The distortion due to social insurance programs is moral hazard behavior by households. Indeed, Engen and Gruber (2001) show that more generous UI crowds-out other sources of income support. Their findings show that households tend to save less if unemployment benefits get more generous: for every extra dollar of unemployment benefits, household saving drops by seventy cents. Gruber (1997) also finds imperfect self-insurance by households. Gruber states that unemployment benefits help smooth consumption, which is something households apparently were unable to do themselves. Gruber states that every one dollar of unemployment benefits reduces the drop in consumption by thirty dollar cents.
A different motivation for government intervention is paternalism. Governments may simply feel that individuals will not appropriately insure themselves against risks if the government does not force them to do so. The intervention arises due to the fact that households seem unable, or unwilling to maximize their own utility by smoothing consumption. In the Netherlands, paying for UI is obliged by law. UI is financed through a payroll tax on employers and workers. This means that every individual is insured against unemployment by law. This might increase moral hazard behavior. The very existence of social insurance schemes influences consumption behavior by individuals. Today, social insurance is a key issue on political agendas and although it plays a little role in the original formulation of the life-cycle model by Modigliani and Brumberg (1954), the framework should be extended to include social insurance. This helps policy makers in predicting consequences of alternative policies. Until the abovementioned uncertainties are perfectly incorporated into the life-cycle model, studies like this one help in explaining consumer behavior in response to such social policies.
5
against unemployment. The duration of these unemployment benefits is not indefinite and the income replacement rate of these unemployment benefits is not one for one. However, unemployment benefits seem to be high and long enough to crowd out precautionary saving due to income uncertainty. This means that lowering unemployment benefits in the Netherlands could have detrimental consequences for any unemployed individual.
This study contributes to the existing literature in the following ways. Previous estimates for the precautionary saving effect of Dutch Households have all been estimated using wealth accumulation data. This study differs in the empirical approach chosen as the precautionary saving motive of Dutch households is measured using annual non-durable consumption data. This empirical approach has not been used before for Dutch households. Secondly, the data from the LISS panel provides the possibility to investigate the precautionary saving motive during the financial crisis. If Dutch households were ever to engage in precautionary saving behavior one would expect that to be in periods of economic downturn. Thirdly, the LISS panel provides explicit data on household risk preferences. From this data, a household specific risk aversion parameter can be constructed. This approach offers unique insights into the precautionary saving motive of households when we combine household risk preferences with their subjective income uncertainty.
This paper is organized as follows: the theoretical model is presented in section 2, section 3 introduces the data, section 4 discusses the empirical results, section 5 concludes the paper and section 6 provides discussion points of this study and pointers for future research.
2. Theory
‘The Life-Cycle Hypothesis (LCH) posits that the main motivation for saving is to accumulate resources for later expenditure and in particular to support consumption at the habitual standard during retirement. According to the model, saving should be positive for households in their working span and negative for the retired ones, and wealth therefore should be hump-shaped’ (Modigliani, 1986).
6
Modigliani argues, then consumption-smoothing generates wealth accumulation until the age of retirement. After retirement, wealth would decline to keep the consumption path constant and equal to the habitual standard. Therefore, the amount of and the motive for saving are deeply rooted in the expected path of income (Caballero, 1991). Note that in the abovementioned LCH uncertainty has not been taken into account.
When uncertainty is introduced into the model, the motive for and amount of wealth accumulation might change. Uncertainty is inextricably connected to risk and consequently, individual characteristics decide how people deal with these risks. The actions taken to mitigate risks decide the individuals’ level of risk-aversion and the level of prudence. Risk aversion is defined as the extent to which individuals are willing to bear risk, whereas prudence is defined as the actions taken to mitigate risks. Vergara (2016) argues that if consumers are risk averse, uncertainty about income causes disutility. Vergara shows that there exists a clear positive relationship between increasing labor-income risk and saving when the utility function exhibits prudence. However, when the source of uncertainty comes from other types of risk, this condition is no longer sufficient to guarantee the precautionary effect. The precautionary saving effect is described by Vergara as the extra saving generated by uncertainty regarding income. Eeckhoudt and Schlesinger (2009) show that the disutility from labor-income uncertainty is reduced by precautionary saving.
2.1 The life-cycle model
This study incorporates perceived labor-income risk into the basic life-cycle model described previously. When making choices under uncertainty, individuals with rational preferences maximize their expected utility. This means that individuals incorporate (uninsurable) labor-income risk when they choose consumption (𝑐𝑖𝜏) and wealth (𝐴𝑖𝜏) in their utility maximization problem. The utility maximization problem individuals face takes the following form:
7
Where 𝐸𝑡 is the expectation operator conditional on information available at time t, 𝛿 denotes the intertemporal discount factor equal to (1+𝜌1 ), 𝜌 is the rate of time preference of consumption, 𝑈(𝑐𝑖𝜏) is the within period utility function, 𝑟 is the real interest rate and 𝑦𝑖𝜏 is labor-income in period 𝜏. The subscript 𝑖 indicates an individual and thus the panel structure of the data and T is the final period. By assumption, 𝐴𝑖0= 0 (people enter the model with zero initial assets) and in this model the only source of uncertainty is (uninsurable) non-capital income. The above model further assumes perfect capital markets, such that individuals can borrow without collateral. This means that individuals do not face liquidity constraints, such that there is no need to include liquidity constraints in the model. Due to the perfect capital market assumption the real interest rate is equal to the rate of time preference (i.e. 𝑟 = 𝜌).
In the model by Caballero (1990), the time horizon is infinite, income follows a Autoregressive Moving Average (ARMA) process and the return on assets is certain. These assumptions allow Caballero to derive closed form solutions for consumption and wealth. In deriving closed form solutions, a random walk assumption for income could also be used. Under such assumptions next year income uncertainty is informative for life cycle income risk (Mastrogiacomo & Alessie, 2014). However, Guvenen (2007) argues that income shocks tend to be less persistent than suggested by a random walk model. To solve this complication, Guiso, Jappelli and Terlizzese, (1992) and Angelini (2009) derive closed form solutions for wealth in the case of an AR(1) income process without drift. Mastrogiacomo and Alessie (2014) extend the studies by Guiso, Japelli and Terlizzese (1992) and Angelini (2009) and show that an AR(1) income process can be generalized to also include predictable changes to income over the life cycle and still produce similar results in terms of precautionary saving. Mastrogiacomo and Alessie assume the following AR(1) stochastic process for non-capital income of household i in year t:
𝑦𝑖𝑡 = 𝑥𝑖𝑡′𝜉 + 𝜀
𝑖𝑡 (4)
Where 𝑥𝑖𝑡 denotes a vector of explanatory variables, from which the future values are known to the consumer and 𝜀𝑖𝑡 is assumed to follow an AR(1) process, i.e.:
𝜀𝑖𝑡 = 𝜆𝜀𝑖𝑡−1+ 𝑣𝑖𝑡 ; 𝑣𝑖𝑡~𝐼𝐼𝐷(0, 𝜎𝑖2) (5)
8
Mastrogiacomo and Alessie (2014). Both studies make use of a subjective earnings variance measure (i.e. the proxy used to estimate the unobservable variance of income 𝜎𝑖2). Therefore, the theoretical model by Mastrogiacomo and Alessie will be used to estimate the effect income uncertainty has on current period consumption.
2.2 Consumer utility function
This study is set out to isolate the effect precautionary saving has on consumption. As mentioned before, precautionary saving arises when individuals are confronted with income uncertainty. Both Caballero (1990) and Mastrogiacomo and Alessie (2014) make use of an exponential utility function with Constant degree of Absolute Risk Aversion (CARA) to isolate the precautionary savings term in their closed form solutions. The exponential utility function that will be used throughout this study is stated in equation (6), where 𝜃 is the coefficient of risk aversion and differs by person. 𝜃 is assumed to be non-negative. Exponential utility functions have several preferred characteristics. Equation (7) shows that utility is an increasing function in consumption whereas equation (8) shows that there are diminishing returns to consumption: individuals are indeed risk averse.
𝑈(𝑐𝑡) = − (1 𝜃) 𝑒−𝜃𝑐𝑡 (6) 𝜕𝑈(𝑐𝑡) 𝜕𝑐𝑡 = 𝑒−𝜃𝑐𝑡 > 0 (7) 𝜕2𝑈(𝑐 𝑡) 𝜕2𝑐 𝑡 = −𝜃𝑒 −𝜃𝑐𝑡< 0 (8)
However, using exponential utility functions also has drawbacks. A possible drawback would be the possibility of negative consumption. Zeldes (1989) imposes non-negativity constraints and shows that adding such constraints impedes finding a tractable solution. The biggest downside to using exponential utility functions is the fact that absolute risk aversion is constant for every individual. This is shown in equation (9), which gives the Arrow-Prat measure of risk aversion (i.e. the coefficient of absolute risk aversion).
9
Absolute risk aversion measures the rate of decay for marginal utility: the rate at which marginal utility decreases when wealth is increased by one euro. Constant absolute risk aversion implies that risk aversion is independent of individual income and wealth characteristics. This goes against our intuition, as absolute risk aversion should be decreasing in income. A lottery where a person can win or lose 100 euro’s is potentially life-threatening for an individual with initial assets equal to 101 euro’s, whereas it is essentially trivial for an individual with initial wealth equal to 1.000.000 euro’s. Therefore, individuals with higher incomes are expected to have lower absolute risk aversion. To solve the issue of constant absolute risk aversion one could use a different type of utility function. A function that has Decreasing Absolute Risk Aversion (DARA) is a Constant Relative Risk Aversion (CRRA) utility function. However, when using a CRRA utility function no closed form solutions for consumption have been found in the literature, whereas closed form solutions have been found for the CARA utility function. Angelini (2009) and Alessie and Lusardi (1997) argue that the main advantage of using closed-form solutions is that it is possible to gain important insights on the effect of precautionary saving on the level of consumption and savings over the life-cycle. Using a CRRA utility function would mean that intertemporal consumption decisions need to be solved using the Euler equation approach. The Euler equation approach focuses on a specific first-order condition implied by the optimization problem faced by a consumer, allowing the estimation of preference parameters (Alessie & Teppa, 2010). The Euler equation is given by equation (10).
𝜕𝑈(𝑐𝑡) 𝜕𝑐𝑡 = ( 1 + 𝑟 1 + 𝜌) 𝐸𝑡( 𝜕𝑈(𝑐𝑡+1) 𝜕𝑐𝑡+1 ) (10)
By assumption 𝑟 = 𝜌, such that the Euler equation shows that the marginal utility of consumption is equal over time: individuals smooth consumption over the life-cycle. The Euler equation only takes two periods into account. In this perspective, making use of closed form solutions allows for a more powerful test of the validity of the precautionary saving model (Alessie & Teppa, 2010).
10
However, since this study’s main focus is to find the precautionary savings motive of wealth accumulation and not habit formation, using a time separable utility function seems justified.
The important thing to note is that individuals need to engage in prudent behavior for precautionary savings to occur. To show that individuals indeed engage in prudent behavior, the measure of absolute prudence by Kimball (1990) will be used. Kimball shows convincingly that prudence, like risk aversion, should decline with wealth. However, as equation 5 shows, using a CARA utility function means that absolute prudence is constant and equal to 𝜃. However, a positive 𝑃(𝑐𝑡) is needed for people to engage in prudent behavior. Since by assumption 𝜃 > 0 (i.e. rational individuals with a CARA utility function) people indeed engage in prudent behavior. But since the absolute prudence measure is constant and independent of individual characteristics, income has no influence on the measure of prudent behavior.
𝑃(𝑐𝑡) = − 𝜕3𝑈(𝑐 𝑡) 𝜕3𝑐 𝑡 𝜕2𝑈(𝑐 𝑡) 𝜕2𝑐 𝑡 = 𝜃 > 0. (11)
As equation (11) shows, for positive prudence we need a positive third derivative of the utility function (since the second derivative of the abovementioned CARA utility function is negative: shown by equation (8)). The necessity of a positive third derivative for a precautionary saving motive to exist was first shown by Leland (1968) and Sandmo (1970). As can be seen from equation (12), a preferential feature of using a CARA utility function is that indeed the third derivative is positive.
𝜕3𝑈(𝑐 𝑡) 𝜕3𝑐
𝑡 = 𝜃
2𝑒−𝜃𝑐𝑡 > 0 (12)
Since this study assumes people to be rational and risk averse, prudent behavior explains the actions taken by consumers to insure themselves against possible income shocks. Therefore, the above stated model predicts that consumers who face income uncertainty will take actions to mitigate labor income risk.
2.3 Closed form solution
Mastrogiacomo and Alessie (2014) solve1 the maximization problem stated in equation (1) and derive the closed form solution for consumption. This closed form solution for consumption is given in
1
11
equation (13). Using a CARA utility function makes it possible to isolate the precautionary savings motive in the closed form consumption function:
𝐶𝑖𝑡 =(𝑡 − 𝑇) ln 𝜙 2𝜃 − 𝜃𝜎𝑖2 2(𝑇 − 𝑡 + 1)∑ (1 − 𝜆𝑙)2 (1 − 𝜆)2𝑙 𝑇−𝑡 𝑙=1 + 𝑌𝑝𝑖𝑡 (13)
Where 𝜎𝑖2 is the one-period-ahead household income uncertainty. Income shocks are given by 𝜆. Just as in the studies by Mastrogiacomo and Alessie (2014), Guiso et al., (1992) and Lusardi (1997) this study will assume the degree of persistence as measured by 𝜆 to be same for all households. Mastrogiacomo and Alessie (2014) mention that although this is a strong assumption to make, results vary not much across households with different levels of persistence when they control for different levels of education of the household head. 𝑌𝑝𝑖𝑡 denotes permanent income and is equal to:
𝑌𝑝𝑖𝑡 = (∑(1 + 𝑟)𝑡−𝜏 𝑇 𝜏=𝑡 ) −1 ((1 + 𝑟)𝐴𝑡−1+ ∑(1 + 𝑟)𝑡−𝜏𝐸 𝑖𝑡𝑦𝑖𝜏 𝑇 𝜏=𝑡 ) (14)
The only unknown variable is 𝐸𝑖𝑡𝑦𝑖𝜏 for which Mastrogiacomo and Alessie make use of the method by Kapteyn, Alessie and Lusardi (2005). In their study, Kapteyn, Alessie and Lusardi first estimate capital income over the life cycle and use the estimated income equations to predict future non-capital incomes of all households in the sample. The income model used by Kepteyn, Alessie and Lusardi (2005) to describe household income is the following:
ln 𝑦𝑖𝑡 = 𝜕0+ ∑ 𝜕1 5 𝑗=1 𝑎𝑔𝑒𝑖𝑡𝑗 + ∑ 𝛾𝜏𝑇𝐷𝜏𝑡 1989 𝜏=1985 + ∑ ∆𝑗𝐿𝐷𝑗𝑖𝑡 6 𝑗=2 + ∑ 𝜑𝑗𝑆𝐷𝑗𝑖𝑡 5 𝑗=2 + 𝑋𝑖𝑡′𝛽 + 𝑢𝑖+ 𝜀𝑖𝑡 (15)
12
The household income equation used by Kapteyn, Alessie and Lusardi can be summarized by the following equation:
ln 𝑦𝜏= 𝑋𝑡+ 𝑓(𝜏)+𝜖𝜏 (16)
Where 𝑋𝑡 is a vector of demographic and socio-economic characteristics at time t. Included variables are: number of adults in the household, number of children in different age groups (6 or younger, between 7 and 12, between 13 and 17, 18 years or older), dummy variables indicating the education level of the head of the household (primary, lower secondary, higher secondary, and university education). 𝑓(𝜏) is a function of age and 𝜖𝜏 is the error term ( 𝜖𝜏~𝐼𝐼𝐷(0, 𝜎𝜀2)). The regression results are shown in appendix A.
In the derivation of their estimate for 𝐸𝑖𝑡𝑦𝑖𝜏, Kapteyn, Alessie and Lusardi argue that it is important to note that over time the demographic and socio-economic characteristics (variables in the vector 𝑋𝑡) that are used to estimate their income equation might change in predictable ways. This leads to the following estimate for their auxiliary equation for 𝐸𝑖𝑡𝑦𝑖𝜏:
𝐸𝑖𝑡𝑦𝑖𝜏 = 𝑒[𝑋𝑡+𝑓(𝜏)+12𝜎𝑖𝜀2] (17)
and this means that permanent income used in this study will have the following form:
𝑌𝑝𝑖𝑡 = (∑(1 + 𝑟)𝑡−𝜏 𝑇 𝜏=𝑡 ) −1 (𝐴𝑖𝑡−1+ ∑(1 + 𝑟)𝑡−𝜏(𝑒[𝑋𝑡+𝑓(𝜏)+12𝜎𝑖𝜀 2] ) 𝑇 𝜏=𝑡 ) (18)
which in turn changes the closed form solution for consumption:
𝐶𝑖𝑡 =(𝑡 − 𝑇) ln 𝜙 2𝜃 ⏟ 1 − 𝜃𝜎𝑖 2 2(𝑇 − 𝑡 + 1)∑ (1 − 𝜆𝑙)2 (1 − 𝜆)2𝑙 𝑇−𝑡 𝑙=1 ⏟ 2 + 𝑌⏟𝑝𝑖𝑡 3 (19)
13
indicates that patience (𝜙 > 1) has a negative effect on current period consumption, as it is easier for patient people to postpone consumption to future periods. The second term (underlined with the number 2) is the precautionary savings term. In this model, people consume or save their income and consumption and saving are substitutes. For saving to increase, consumption needs to decline. This theory dates all the way back to Keynes’ General Theory of Employment, Interest and Money, where Keynes defined saving as the excess of income over consumption (Keynes, 1936. p56). Since precautionary saving is a component of total savings, the expected effect of precautionary saving on current period consumption is negative. Risk aversion influences the amount of precautionary saving. Higher values of 𝜃 increase the precautionary saving component. Also the interaction between 𝜃 and the one-period-ahead household labor-income uncertainty (𝜎𝑖2) is an important explanation for the amount of precautionary saving. More risk averse individuals (higher 𝜃) respond more heavily to higher labor-income uncertainties (higher 𝜎𝑖2), whereas by itself the uncertainty concerning labor-income increases the amount of precautionary saving. The precautionary saving motive also depends on age. This effect is shown by 𝑡, and enters the precautionary saving term in two positions. 𝑇 is equal to 98 and represents the final period of the planning horizon. The effect of age on the precautionary saving term is twofold. The first part of the precautionary saving term is increasing in age, since a person’s remaining timeframe declines (i.e. 𝑇 − 𝑡 gets smaller). However, the second part of the precautionary saving term is declining in age. Overall, the precautionary saving term increases with age. An explanation might be that finding a new job is harder at older ages, such that older working individuals face income uncertainty. Consequently, older individuals might feel more pressure to engage in precautionary saving. Note that the precautionary saving term is increasing in the persistence of income shocks (i.e. 𝜆). This effect is shown in table 1, which shows simulation results for different values of 𝜆, for given values of 𝜃, 𝜎𝑖2 and 𝑡. In these simulations individuals are assumed to be risk averse. The findings show that for higher values of 𝜆 the precautionary savings term goes up.
𝝀 Prec. Saving 0.2 0.050903 0.4 0.080906 0.6 0.155720 0.8 0.464493
14
The third term (underlined with the number 3) in equation (19) is the certainty equivalence term, which is defined in equation (18). This term states that lifetime consumption is equal to lifetime income when the state of the world is one without labor-income uncertainty.
2.4 Literature review
Mastrogiacomo and Alessie (2014) quantify the relative importance of the precautionary motive for wealth accumulation over time. They note that empirical evidence suggests that the impact of precautionary savings is small if one uses a subjective measure of uncertainty about next year income. An alternative would be to use an objective method to proxy for income uncertainty by exploiting life-cycle income variation. Such models find that precautionary saving motives explain large parts of total wealth accumulation. The method chosen to measure the precautionary savings motive seems to give contradictory results.
15
According to Kennickell and Lusardi (2005) the broad range of findings in the literature is driven in large by a variety of conceptual and empirical issues in estimating household total wealth and the consumption function. Kennickel and Lusardi mention eight issues. These issues will be briefly discussed below as most have important implications for the empirical model used in this study.
2.4.1 The measure of wealth and risk
It is important to note that the differing findings for the precautionary savings motive might come from differing wealth measures. Some papers use total net wealth as a measure for total wealth. Where total net wealth is defined as total assets minus total liabilities. Browning and Lusardi (1996) show that the problem with this specification is that because of differing risk and liquidity characteristics of the underlying assets and liabilities, it is inappropriate to aggregate them. Others, like Engen and Gruber (2001), look at liquid wealth. But the problem with this method is that it might be too restrictive. The largest assets households normally hold are housing wealth and designated retirement accounts. These should be taken into account when calculating total wealth, but these assets are hard to liquidate or are not easily accessible. The literature is unclear over a perfect measure of total household wealth.
Other complications arise when one would like to measure income risk. In the existing literature usually one of two methods is applied. Some papers (like Caballero (1991) and Browning and Lusardi (1996)) model a household-specific stochastic process for income that they estimate using panel data. From which the resulting variance of income is used as a proxy for income risk. Caballero (1991) and Browning and Lusardi (1996) both show that the estimated variability in income is not without flaws. Both studies argue that it is hard to distinguish empirically between transitory income and measurement error. This problem could be solved by using subjective measures of income risk. However, such subjective measures also suffer from measurement error. When using a subjective measure for income risk, empirical results for the effects of precautionary saving tend to be implausibly low (Guiso et al. (1992), Lusardi (1997),1998)).
16
2.4.2 Preferences
Preferences regarding risk also influence precautionary saving. Of particular importance is the coefficient of risk aversion (Caballero (1990, 1991)), which might differ substantially between households. Individuals who are very risk averse might self-select into jobs that offer less income risk. Lusardi (1997) shows that if risk aversion and prudence are positively correlated, then these individuals should also save more.
2.4.3 Liquidity constraints
The existence of liquidity constraints means that there is presence of some imperfection in the financial market, that prevents people from borrowing. This could influence consumer behavior, as unconstrained households might not need precautionary savings to shield themselves against negative income shocks. The ability to borrow reduces the need to precautionary save. The problem is that this (dis)ability to borrow is usually unobservable in micro-economic data sets. This means that for models that use closed form solutions to estimate the precautionary effect, it is hard to take liquidity constraints into account. A direct way to detect liquidity constraints is to ask consumers directly whether they applied for and were denied credit (Jappelli, 1990)). However, a problem with this type of question is that consumers may have been denied credit for good reasons, or may have decided not to apply for credit on the assumption that this would be refused to them (i.e. the discouraged borrower effect). Less directs tests rely on the fact that liquidity constraints are more likely to be binding for young people and for those whose liquid assets are particularly low (Zeldes, 1989). Neither of these proxies are flawless.
The focus of this paper is on the effect income uncertainty has on current period consumption levels. Although there is empirical evidence that liquidity constraints are indeed important in explaining household consumption behavior (see for instance Zeldes (1989a)), applying liquidity constraints to the closed form solutions above is very hard, if not impossible. See the discussion about the Euler equation approach above. However unfortunate, this means that liquidity constraints will not be included in this model.
2.4.4 Other forms of insurance.
17
2.4.5 Functional forms
Much like other studies, in estimating permanent income this study makes use of the logarithm of net wealth as the functional form. This transformation excludes all households with net wealth of zero or below. But, as Kennickell and Lusardi (2005) argue, these household might reflect features of the welfare system or individual preferences. Perhaps their current wealth is a result of past income uncertainty and dropping these observations might lead to a selectivity bias in the estimate of precautionary saving. Also, dropping households with net wealth of zero or below means that using net wealth as a proxy for liquidity (un)constrained households is no longer an option.
2.4.6 Macro-level and other shocks.
Most of the studies on precautionary saving largely ignore the importance of macro shocks (Kennickell & Lusardi, 2005). Macro shocks influence the amount of household total wealth. The problem is that these shocks make it hard to reliably estimate the precautionary savings motive with a single cross-section of wealth data. The alternative would be to use many years of observations, but these data are extremely hard to derive in. Not accounting for past shocks might lead to biased results about the level of precautionary saving (Carroll, Dynan & Krane, 2003)).
2.4.7 Precautionary saving and (stock) portfolio choice
Kennickell and Lusardi (2005) show that portfolio choices are directly influenced by perceived income uncertainty. People who face high income risk or who own a business are less likely to invest in stocks (Guiso et al. (1992), Haliassos and Bertaut (1995) and Hochguertel (2003)). From this Kennickell and Lusardi argue that households making high returns of the stock market are likely to face lower income risk. Although this argumentation is not backed up with empirical findings, the implications are clear. Regressing wealth on income risk is likely to confound this effect. Adding this to the model below is not within the scope of this study, but is certainly applicable for future research.
2.4.8 Other motives to save
Isolating the precautionary saving motive from other saving motives is difficult. Again, this is positively related to perceived risk. Entrepreneurs tend to save more, due to their higher income risk. Yet, their higher wealth is usually illiquid, as it is mostly held in their business (Gentry & Hubbard, 2004). Other motives to save are a bequest motive or saving for a comfortable retirement (Hurst & Lusardi, 2004). With most of the existing empirical data it is hard, or even impossible, to control for these factors (Kennickell & Lusardi, 2005).
18
the fact that one compares different countries with different institutions needs to be taken into account. Cultural aspects and estimations during differing time periods also influence the differing empirical results. Most of the abovementioned issues by Kennickell and Lusardi (2005) have been taken into account in the theoretical and empirical model used. Unfortunately, some issues are impossible to account for.
2.5 Empirical model
The purpose of this study is to measure the effect that perceived labor-income uncertainty has on household consumption in the Netherlands. The estimation procedure of the consumption function given by equation (19), shows resemblance with the method applied by Teppa (2014). For simplicity, equation (19) is shown again below.
𝐶𝑖𝑡 =(𝑡 − 𝑇) ln 𝜙 2𝜃 ⏟ 𝐼𝑖𝑡 − 𝜃𝜎𝑖 2 2(𝑇 − 𝑡 + 1)∑ (1 − 𝜆𝑙)2 (1 − 𝜆)2𝑙 𝑇−𝑡 𝑙=1 ⏟ 𝑆𝑖𝑡 + 𝑌⏟𝑝𝑖𝑡 𝑃𝑖𝑡 (20)
The specification used is in levels, in which annual non-durable consumption (𝐶𝑖𝑡) is regressed over the three separate terms stated in equation (20): the (im)patience term (𝐼𝑖𝑡), the precautionary savings (𝑆𝑖𝑡) term and the certainty equivalence term (𝑃𝑖𝑡), which is equal to equation (18). The three terms were explicitly discussed in section 2.3 and will be further discussed in section 3 of this study. However, for completeness a brief discussion of these terms will be given here. 𝐼𝑖𝑡 is decreasing in age and increasing in 𝜃 if the household is patient (i.e. 𝜙 > 1). If households are impatient (𝜙 < 1), then 𝐼𝑖𝑡 is increasing in age and decreasing in 𝜃. 𝑆𝑖𝑡 is increasing in 𝜃, 𝜎𝑖2, age (𝑡) and 𝜆. Whereas 𝑌𝑝𝑖𝑡 is the certainty equivalence term. The real interest rate used in estimating 𝑌𝑝𝑖𝑡 (given by 𝑟 in equation (14)) is 0.03.
The model will be extended by allowing for demographic and socio-economic variables and for an unobserved individual effect 𝛼𝑖. The corresponding estimation equation is the following:
𝐶𝑖𝑡 = 𝛽0+ 𝛽1𝐼𝑖𝑡− 𝛽2𝑆𝑖𝑡+ 𝛽3𝑃𝑖𝑡+ 𝛾′𝑍
𝑖𝑡 + 𝛼𝑖+ 𝜀𝑖𝑡 (21)
19
variables are measured at the household level (i) at time t and 𝜀 denotes the error term. The effect of the precautionary savings motive on consumption is given by 𝛽2, for which the following hypothesis will be tested:
𝐻0: 𝛽2= 0 (22)
(23) 𝐻1: 𝛽2< 0
The expected sign of 𝛽2 is negative. Consumers with preferences described by a CARA utility function, facing income uncertainty, will increase their savings for precautionary reasons. This in turn will have a negative effect on 𝐶𝑖𝑡.
3. Data
Although there are many similarities between the above mentioned studies by Caballero (1990), Mastrogiacomo and Alessie (2014) and this study, there is one major difference. The studies by Caballero and Mastrogiacomo and Alessie use wealth accumulation to predict the precautionary saving motive, whereas this study makes use of consumption data provided by the LISS panel of the (Dutch) Tilburg University. Estimating the precautionary savings effect of Dutch households experiencing income uncertainty using consumption data has not been done before. An explanation might be that high quality consumption data is hard to derive.
The data used in this study is derived from the LISS panel. The LISS panel is an online longitudinal survey that consists of about 5.000 households and is representative of the Dutch population. The surveys are conducted by CentERdata at the Tilburg University. CentERdata conducted surveys in the period from 2007 to 2016. To reduce potential measurement error, the LISS panel offers panel members a compensation for every completed questionnaire. Most of the questionnaires are filled out on a regular basis. The fact that participation is compensated induces panel members to fill out the questionnaire with care. This means that the LISS panel delivers high-quality, high-frequency data, at the individual and household level.
20
survey. This survey is conducted on a monthly basis, starting in January 2008 and ending in December 2016. For the purpose of this study, only the background variables surveys between 2009 and 2015 will be used. The LISS panel also offers data on household Assets, provided in the surveys ‘Economic Situation: Assets’ and ‘Economic situation: Housing’. However, this data is of a much lower quality and frequency. Data on household Assets and Housing should be used in the prediction of permanent income, given by 𝐴𝑖𝑡−1 in equation (18), but only if the data is of sufficient quality. This means that the permanent income term will not feature household assets. As equation (18) shows, assets would increase permanent income only slightly. Therefore leaving out assets will probably not lead to a very different permanent income measure. However, adding household assets would give a better estimate of the effect household permanent income has on annual non-durable household consumption. It would be interesting to add the household assets to the permanent income term when more and better data on household assets becomes available.
Hurst, Kennickell, Lusardi and Torralba (2005) show that it is important to make a distinction in the data between two different groups in the population: business owners and other households. Business owners tend to face higher income uncertainty than other households. Business owners hold large amounts of wealth for reasons unrelated to precautionary saving. If one would include business owners in the estimation then this might lead to a high correlation between wealth and income uncertainty, regardless of whether or not the precautionary saving, due to income uncertainty, is important (Kennickell & Lusardi, 2005). To control for this bias, business owners are not included in the estimations.
21
safe option is at a sure payoff level of 30 euro’s. When individuals chose the sure payoff for the first time at the level of 35 euro’s they receive a risk aversion level of 1. Although the sure payoff in this case is equal to the expected value of the lottery, and thus is certainty equivalent, these individuals still receive a positive risk aversion parameter equal to 1. This is due to the fact that these individuals also could have chosen to participate in the lottery with an equal expected payoff. However, by choosing the safe option these individuals apparently fear the risk of receiving only 5 euro’s and choose the safe option of the sure payoff, equal to 35 euro’s. Individuals choosing the sure payoff for the first time at higher levels than 35 euro’s, or never choosing the sure payoff, get a risk aversion score of 0. An example of such a trial can be found in appendix B. A potential downside to this survey is that it is conducted only once, in 2009. However as was shown in section 2, assuming that an individual’s utility function is characterized by CARA preferences leads to constant absolute risk aversion (see equation (9)). This means that in the model used by this study, it is safe to apply the measure of risk aversion from the year 2009 to all other years as well. The other variables used in estimating the empirical specification given in equation (20) are discussed below.
Annual household non-durable consumption – just as Teppa (2014), this variable is constructed as the self-reported monthly household consumption multiplied by twelve. Considering the scope of this study, only the non-durable consumption questions are relevant. Appendix C shows the consumption expenditures included in the annual household non-durable consumption parameter. Annual household net income – this variable is computed as the household net monthly income (labor and non-labor income) multiplied by twelve.
Permanent income (𝑷𝒊𝒕) – the permanent income parameter is constructed as shown in equation (18) above and is represented in equation (21) by 𝑃𝑖𝑡. In estimating the permanent income variable, significant positive autocorrelation was found when estimating equation (5) above. The autocorrelation coefficient was equal to 0.385 and was significant at the 1% level. This suggests that the model specification for permanent income should be adjusted. However, the model specifications in this study were not properly adjusted. Future empirical models should adjust their model to deal with this autocorrelation.
22
deals with Kennickell and Lusardi’s preference issue mentioned in section 2.4.2. Table 2 gives the distribution of the risk aversion parameter used to predict the precautionary savings effect.
𝜽 Percentage 0 26.65% 1 8.53% 2 9.59% 3 9.63% 4 45.61%
Table 2. Distribution of the risk-aversion parameter. The total number of risk aversion observations is 2.462.
(Im)patience (𝑰𝒊𝒕) – consumer (im)patience is given by 𝐼𝑖𝑡 equation (21) and is equal to: 𝐼𝑖𝑡 =(𝑎𝑔𝑒 − 𝑇) ln 𝜙
2𝜃
(23) 𝐼𝑖𝑡 gives the relationship between an individual’s age and risk aversion, and his patience preferences. 𝑇 is equal to the final time period on an individual’s planning horizon, and is equal to 98. It is important to note that (𝑎𝑔𝑒 − 𝑇) < 0, for any level of age included in any of the estimations of section 4. Also, as by assumption in section 2.2, 𝜃 ≥ 0, which leads to 2𝜃 ≥ 0. Therefore, an individual is patient (𝐼𝑖𝑡 < 0) if (𝑎𝑔𝑒 − 𝑇) ln 𝜙 < 0, which holds if ln 𝜙 > 0. This means that 𝜙 > 1. An individual is impatient if 𝐼𝑖𝑡 > 0, which implies that 0 < 𝜙 < 1. The specifications in section 4 assume households to be patient. Consumer patience should have a negative effect on annual household non-durable consumption, since 𝐼𝑖𝑡 < 0. When consumers are patient, they have a lower intertemporal discount rate and value consumption in the future highly. This means patience has a negative effect on annual household non-durable consumption. Since no data on 𝜙 was available in the LISS panel, 𝐼𝑖𝑡 was estimated using different values of 𝜙. These values range from 0.5 to 8. Household income uncertainty (𝝈𝒊𝟐) – since this variable is unobservable, the proxy this study will use comes from a question out of the ‘Economics Situation: Income’ survey. Participants were asked to answer the following question concerning their expectations about their job security in twelve months:
‘Do you think there is any chance that you might lose your job in the coming 12 months? You indicate this in terms of a percentage. 0% means that you are sure you will not lose your job, and 100% means that you are sure that you will use your job.’
23
given as the answer to the above mentioned question. This transformation deals (at least partly) with the income risk measurement error issued by Kennickell and Lusardi (2005). Kennickell and Lusardi also note that one should include other types of risk as well, when estimating the precautionary savings effect. However, since this is the first time the precautionary saving effect is estimated for Dutch households using consumption data, only non-insurable labor-income risk is added to the model. Nevertheless, adding multiple types of risk might offer interesting insights and offers opportunities for future research.
Together the risk aversion parameter and the income uncertainty measure largely define an individual’s precautionary saving preferences, given by 𝑆𝑖𝑡 in equation (21). Table 3 shows an auxiliary regression of annual non-durable consumption on risk aversion (𝜃)and income uncertainty (𝜎𝑖2). Although no significant effects have been found, the sign of both coefficients point in the right direction. This is as expected from the theoretical model described in section 2.
Annual household non-durable consumption Pooled OLS
𝜃 -148.57
(178.566)
𝜎𝑖2 -2951.9
(2918.12)
Table 3. Auxiliary pooled Ordinary Least Squares (OLS) regression of annual non-durable consumption on the risk aversion and income uncertainty parameters. Standard errors are clustered by household and are given in parentheses.
Precautionary saving term (𝑺𝒊𝒕) – the precautionary saving term is given by 𝑆𝑖𝑡 in equation (21) and is equal to: 𝑆𝑖𝑡 = 𝜃𝜎𝑖2 2(𝑇 − 𝑎𝑔𝑒 + 1) ∑ (1 − 𝜆𝑙)2 (1 − 𝜆)2𝑙 𝑇−𝑎𝑔𝑒 𝑙=1 (24)
The precautionary saving term is increasing in 𝜃, 𝜎𝑖2 and 𝜆2. The precautionary saving term is decreasing in age. The last time period (𝑇) used in the estimation is equal to 98. For a more extensive discussion of this term, see also section 2.3.
Vector of demographic and socio economic characteristics – the vector of demographic and socio-economic characteristics exists of the number of adults in the household, the number of children
2
24
living in the household, the gender of the household head and a variable indicating the educational level of the household head. Gender is measured using a dummy variable which is equal to zero if the household head is a female and is equal to one if the household head is male. The education variable runs from one to three, where one indicates the household head as no or a low level of education, two means the household head has a college degree and households heads with a university degree get a score of three. It is important to control for the household characteristics, as consumption patterns may vary across differing households characteristics.
Functions of age – much like Mastrogiacomo and Alessie (2014) functions of age are added in the prediction of permanent income for two reasons. One, to control for age effects. Two, adding functions of age help to better pick up the business cycle effect. Especially when Deaton and Paxson (1994) time dummies are added.
(Deaton and Paxson) Time dummies – in the estimation of permanent income Deaton and Paxson (1994) time dummies will be used. Due to the fact that in the estimation process of permanent income there is already controlled for cohort and age effects, adding simple time dummies is not possible. This is due to the direct relationship between the current year and the sum of a person’s age and the year of birth. The Deaton and Paxson transformation of the time dummies means that all time effects add up to zero and are orthogonal to a time trend. This relatively simple transformation of the time dummies implies that all transitory time effects are assumed to be business cycle shocks, and not for instance due to changes in preferences (Mastrogiacomo and Alessie, 2014). These dummies, together with the abovementioned cohort dummies and the functions of age are added to the model to deal with the macro-level and other shocks issue of Kenickell and Lusardi (2005).
25
Overall, clear outliers are deleted from the model such that a total of 2085 observations are left over the years 2009, 2010, 2012 and 2015. Descriptive statistics of the above mentioned variables can be found in table 4.
Variables Mean sd Min. Max.
Annual household non-durable consumption 20832 11306 0 96945
Annual household net income 35880 16176 0 138000
Permanent income 41393 18532 2042 148130
Household risk aversion paramater 2.39 1.709 0 4
Household income uncertainty 0.082 0.092 0 0.25
Age of the household head 46.62 10.23 21 65
Number of adults per household 1.699 0.459 1 2
Number of children per household 0.941 1.107 0 6
Gender of the household head (1=male) 0.765 0.424 0 1 Education low (1=low or no education) 0.589 0.492 0 1 Education college (1=college degree) 0.282 0.45 0 1 Education university (1=university degree) 0.129 0.335 0 1
Table 4. Descriptive statistics of the variables used in the empirical analysis. The total number of observations is 2085. The minimum and maximum values of the variables are shown by Min. and Max. respectively.
The descriptive statistics show that on average the households tend to be risk averse. The household characteristics further show that overall the household head is a male and is roughly 47 years old. Households tend to have less than 1 child living at home and household heads tend to have no or a low level of education. More precisely, from the total sample a total of 58.86% of the households has no or a low level of education.
4. Results
26
Correlations Permanent income Education
Permanent income 1.000 .
Education 0.3569*** 1.000
Annual non-durable consumption 0.3208*** 0.1668***
Table 5. Correlations between permanent income, education and annual non-durable consumption. ***=1% statistically significant.
In estimating any of the models mentioned below it is assumed (as described in section 2), that consumers are patient. Thus, by assumption 𝜙 > 1. The models were estimated using a value of 𝜙 = 8, indicating that individuals are very patient. Different values of 𝜙 > 1 have been used to look at the effects patience has on annual non-durable consumption. However, in none of the models estimated in the tables below did differing values of 𝜙 have very different findings. This means that the estimates using 𝜙 = 8 give a good picture of the effects of 𝐼𝑖𝑡 on 𝐶𝑖𝑡.
Table 6 shows the estimation results for the different coefficients of equation (21). The results are shown for two alternative specifications: pooled Ordinary Least Squares (POLS) and Fixed Effects (FE). Due to the panel structure of the data, estimating the model with Ordinary Least Squares (OLS) would lead to biased results and severe autocorrelation. Therefore the first specification in table 6 is the pooled OLS regression. The second specification in table 6 gives the FE estimation results. FE are used to control for unobserved heterogeneity and omitted variable bias.
The pooled OLS estimate shows some significant findings. Keeping all else fixed, permanent income has a significant positive effect on annual non-durable consumption. This is in line with the expectations of the sign of 𝛽3 in equation (21). If permanent income goes up with one euro, annual non-durable consumption goes up with 0.18 euro (significant at the 1% level). Other significant findings are that household annual non-durable consumption drops by €790.41 euro’s for every child living in the household. This finding is significant at the 1% level. Also, households in which the household head has a university degree consume more than households in which the household head is poorly educated or has no education at all. If the household head has a university degree the household consume on average €1859.05 euro more on an annual basis (significant at the 10% level). The level of 𝜙 used in the estimation of the (im)patience term was equal to 8. This means that the model assumes households to be patient. The pooled OLS specification finds a positive significant effect for the patience term. This means that, keeping risk aversion and age constant, higher values of 𝜙 lead to higher values of consumption. We would expect this relationship to be negative. However, the (im)patience term is significant only at the 10% level.
27
increase of €0.09 of annual non-durable consumption (significant at the 5% level). The FE estimation also finds a significantly negative effect for the number of children living at home. It should be stressed that the FE specification has a poor fit. The F-test shows that this model specification would be rejected, even at the 10% level.
The most important estimate for this study is the estimate for 𝛽2 in equation (21), the estimate for the precautionary savings term. Although the sign is negative in the pooled OLS regression, neither model is able to find a significant effect of precautionary saving on annual non-durable consumption. One explanation might be measurement error. As mentioned in section 2.4.1, Kennickell and Lusardi argue that using subjective questions on personal income changes, that occur in the next twelfth months, might not incorporate enough variation to represent life cycle income risk (In Mastrogiacomo & Alessie, 2014). Consequently, both estimations fail to reject 𝐻0: 𝛽2= 0 given in equation (22). Variables Model 1 Pooled OLS Model 1 FE Patience 34.850* 239.643 (19.940) (240.941) Precautionary Saving -5131.168 556.002 (8054.193) (11852.09) Permanent Income 0.1843*** 0.0902** (0.0226) (0.0377) Gender of the household head 307.810 .
(714.283) .
Number of children living at home -790.413*** -1461.81** (287.916) (725.132) Cohabiting partner 114.696 539.920 (732.892) (1930.072) Education: college 557.784 4613.520 (661.769) (3337.417) Education: university 1859.048* 9469.337 (1030.093) (6007.307) Constant 13308.01*** 18196.270*** (814.1152) (4346.89) Deaton and Paxson time dummies Yes Yes
N 2085 2085
Household clusters 870 870
R2 0.1052 0.0119
Prob > F 0.0000 0.1067
28
4.1 Sensitivity analysis
The precautionary saving motive depends on a number of household characteristics. To further look into the precautionary motive a couple of submodels will be estimated, all dependent on specific household characteristics. These findings are shown in tables 7 and 8. The model specification is still represented by equation (21), which will be estimated again with pooled OLS and FE. Model 2 shows the estimates for all households who’s household head is younger than 41 years of age and older than 21 years. Model 3 shows the estimates for the households that have a household head with an age between 41 and 65 years. Model 4 uses as a sample only the households that face positive income uncertainty, such that 𝜎𝑖2> 0. Model 5 shows the findings if the risk aversion parameter (𝜃) is equal to one, for every household. Model 5 corrects for potential measurement error in 𝜃, but still assumes households are risk averse. Models 2 and 3 are represented in table 7, whereas models 4 and 5 are shown in table 8. Each model will be briefly discussed below.
4.1.1 Effect of different age samples on the precautionary saving motive
29
The effect of higher education in the FE specification in model 2 on annual non-durable consumption is significantly positive. Young households in which the household head has a university degree spend on average €20498.89 more on annual non-durable consumption than poorly or not educated households. This is equal to €1708.24 more non-durable consumption per month. As can be seen from table 5, permanent income is positively correlated with education. Table 5 also shows that permanent income is significantly positively correlated with annual non-durable consumption. This could explain the significantly higher annual non-durable consumption levels by higher educated households found in the FE specification in model 2.
Variables Model 2 21 < age < 41 Pooled OLS Model 2 21 < age < 41 FE Model 3 41 < age < 65 Pooled OLS Model 3 41 < age < 65 FE Patience 25.853 379.197 15.137 -303.217 (32.586) (542.617) (22.285) (289.226) Precautionary saving 1640.939 18488.230 -6392.085 -6838.167 (22318.09) (28444.41) (8685.209) (13125.01) Permanent income 0.2134*** -0.083 0.2151*** 0.1836*** (0.0487) (0.0683) (0.0257) (0.0446)
Gender of the household head -2051.560* . 865.352* .
(1202.559) . (870.829) .
Number of children living at home 2085.234*** 3367.239** -1680.315*** -3779.683***
(570.427) (1483.076) (314.940) (853.917) Cohabiting partner -3229.441* 9266.561** 21.105* -2643.204 (1723.659) (3809.593) (824.362) (2312.758) Education: college 64.855 13276.07** 1044.163 2466.497 (1223.857) (5794.854) (734.983) (3898.667) Education: university 888.804 20498.89*** 2353.824* -2316.011 (1598.496) (7782.417) (1276.055) (10488.93) Constant 12645.25*** 13449.45 12438.46*** 13712.13*** (2033.565) (11790.19) (850.308) (4777.211)
Deaton and Paxson time dummies Yes Yes Yes Yes
N 525 525 1560 1560
Household clusters 261 261 686 686
Within R2 0.1356 0.0966 0.1303 0.0370
Prob > F 0.0000 0.0019 0.0000 0.0001
Table 7. Pooled OLS and FE estimation results for equation (21), where the dependent variable is household annual non-durable consumption. Model 2 includes only households with a household head younger than 41 years of age and older than 21 years. Model 3 contains only households in which the household head is older than 41 years of age and younger than 65 years. Reported R2 for FE specification is the within R2. ***=1%, **=5% and *=10 statistically significant. Standard errors are clustered by household and are given in parentheses.
30
4.1.2 Effects of income uncertainty selections and risk aversion adaption on the precautionary saving motive
The theoretical model in section 2 shows that the precautionary saving motive is increasing in income risk. This means that for given values of risk aversion and age, higher probabilities of households losing their job should lead to higher precautionary savings. Model 4 uses as a sample only the households that face positive income uncertainty, such that 𝜎𝑖2 > 0. The estimates are shown in table 8. Just as in the full specification in table 6 above, both the pooled OLS and the FE specifications in model 4 find significant positive effects of permanent income on annual non-durable consumption. However, neither specification is able to find a significant effect for the precautionary savings term on household annual non-durable consumption.
31 Variables Model 4 𝝈𝒊𝟐> 𝟎 Pooled OLS Model 4 𝝈𝒊𝟐 > 𝟎 FE Model 5 𝜽 = 𝟏 Pooled OLS Model 5 𝜽 = 𝟏 FE Patience 46.487** 387.048 47.269* 210.341** (24.349) (326.053) (28.352) (95.011) Precautionary saving -2335.358 2528.228 2043.678 -429.509 (9855.754) (21475.12) (27886.44) (35986.03) Permanent income 0.1698*** 0.1501** 0.1889*** 0.083** (0.0265) (0.0638) (0.0231) (0.0378)
Gender of the household head 810.961 . 154.591 .
912.363 . (712.189) .
Number of children living at home -169.166 -1096.645 -761.497*** -1310.294*
(356.762) (1080.249) (288.422) (1931.04) Cohabiting partner 193.097 33.72 34.559 771.379 (877.052) (2431.649) (741.554) (1931.04) Education: college 1252.218 6288.63 675.683 4371.893 (841.711) (3932.925) (667.180) (3328.167) Education: university 1667.524 11801.68* 1883.466* 9193.324 (1148.851) (6728.137) (1032.786) (5980.662) Constant 12754.10*** 17162.34*** 15074.27*** 25762.35*** (1148.829) (6247.404) (1529.255) (5619.137)
Deaton and Paxson time dummies Yes Yes Yes Yes
N 1269 1269 2085 2085
Household clusters 621 621 870 870
R2 0.0939 0.0208 0.1047 0.0151
Prob > F 0.0000 0.1422 0.0000 0.0305
Table 8. Pooled OLS and FE estimation results for equation (21), where the dependent variable is household annual non-durable consumption. Model 4 includes only households that face positive income risk. Model 5 uses a different risk aversion measure of households to correct for potential measurement error. Model 5 assumes 𝜃 =1 for every household in the sample. Reported R2 for FE specification is the within R2. ***=1%, **=5% and *=10 statistically significant. Standard errors are clustered by household and are given in parentheses.
32
5. Conclusion
The purpose of this study was to find the effect precautionary saving motives have on consumption choices by households in the Dutch population. Financial wealth accumulated for precautionary reasons can be the result of individuals facing many types of risk. However, in this study income uncertainty during the financial crisis was the only source of risk considered. Rational, risk averse individuals do not like risk and would like to take actions to mitigate risk. Therefore, individuals accumulate precautionary savings to overcome potential income drops. This could have a negative effect on annual non-durable consumption. Therefore the research question of this study was:
‘Did Dutch households engage in precautionary saving during the financial crisis?’ Previous studies have found negligible or very relevant findings for precautionary saving, depending on the way income uncertainty is measured. Studies that use a subjective measure of income uncertainty found negligible effects for the precautionary saving motive. Studies that use life-cycle earnings variance as a proxy for income uncertainty found more relevant findings for precautionary saving. Almost all of the previously described studies estimated the precautionary saving motive using data on wealth accumulation whereas this study made use of consumption data to estimate the precautionary saving motive.
33
6. Discussion
A possible explanation for the findings of this study might be the existence of UI provided by the Dutch government. UI offered by the government could give rise to moral hazard behavior of Dutch households. The very existence of public insurance schemes could crowd out precautionary saving motives. In the debate between protection versus distortion of public insurance schemes, the findings of this study show that protecting Dutch households against adverse income shocks due to labor-income uncertainty, probably distorts the precautionary saving motive. This means that any policy reform regarding public UI schemes could have large effects. A reduction, or even abolishment of such a scheme implies that a huge source of income for the unemployed will disappear, whereas the unemployed also fail to accumulate precautionary wealth themselves. Iincreasing the benefits received from UI might only increase the scope for moral hazard behavior of households. Consequently, governments should have clear policy goals for their social programs. This study argues that if the goal is to increase self-insurance behavior of Dutch households, than lowering unemployment benefits might entice households to do so. This study also shows that increasing the generosity of the public unemployment program might lead to households engaging even more in moral hazard behavior. However, it is very difficult to fully incorporate social insurances in economic models like the one described in section 2. Therefore, the direct effect of social insurance policies is hard to measure, but might give very interesting insights on precautionary saving behavior of households.
