The effect of financial literacy and family size on
voluntary retirement planning
Master thesis in Finance
University of Groningen
Faculty of Economics & Business
January 11, 2018
Stefan Janssen
S2492547
[2]
1. Introduction
This paper studies the effect of financial literacy and family size (i.e. the amount of children in the family) on voluntary retirement planning. Classic life-cycle theory predicts individuals to save part of their current income for later stages in life where they do not expect to generate income anymore, to be able to smooth consumption levels over lifetime. However, where nearly all classic economic theories assume individuals to be perfect rational economic agents, in real life this is not always the case. Modern economic theories more often implement human behavioural characteristics into their respective models, as it becomes clearer nowadays that human behaviour differs substantially from that of rational economic agents.
Therefore, this paper studies the effect of two of these human behavioural characteristics on voluntary retirement planning. The first characteristic is financial literacy, which can be described as a combination of an individuals’ knowledge about economic theories, knowledge about financial topics, such as the ability to solve (basic) economic calculations, and the ability to manage current consumption levels with respect to future savings. Over the last couple of years, an increasing number of papers studying the relation between financial literacy and financial planning have been published, of which some will be discussed in the literature section of this paper. The most important findings of these papers are that more financial literate individuals are better able to plan for retirement and are better able to stick to their plans (Lusardi and Mitchell, 2011; Van Rooij et al., 2012).
Besides the effect of financial literacy on voluntary retirement planning, this paper takes the factor family size into account. It thereby mainly focusses on the role of children on the ability to save for retirement. Where the effect of financial literacy on (retirement) savings has been studied intensively the last couple of years, there is relatively little information about the role of family size on individual retirement savings. This paper takes this family factor into account by focussing on modern-day households in the Netherlands.
The role of children in the family and the motive of having children in a developed economy such as in the Netherlands differs largely between nowadays and, say, 150 years ago.
Furthermore, this difference can still be observed between developed- and developing economies. In the past, the role of children was mainly related to help providing food and to take care of their parents in later years of life. In that respect, it was desired to have a
[3]
one household has decreased substantially, with an average household size of 2.17 individuals in 2015.1 This drop can be explained by many different factors, such as increased availability of food and other primary sources of life, the development of pension provisions, and the shift towards a more knowledge based economy, leading to more career opportunities for both males and females (Fokkema and Esveldt, 2008). Another possible factor is the cost of having children, which is not only related to the cost of providing food and other materials, but also to the fact that in most cases one of the parents might have to partially give up their career to take care of the child. In this respect, one could say that the opportunity costs of having children have increased over time (Butz and Ward, 1979; Galor and Weil, 1993; Adsera, 2005). Having children might therefore lead to a decrease in lifetime income, which affects the ability to save. In this respect, it can be argued that having children is also a behavioural characteristic that leads to deviating saving behaviour compared to rational economic agents. Moreover, next to the motive of having children, the composition of retirement income has changed substantially over the years. Where there were almost no pension arrangements in the Netherlands in the past, nowadays every resident in the Netherlands automatically builds up a basic pension income during their life. Furthermore, most companies have certain pension arrangements for their employees, which together with the basic pension ensures at least some level of income during retirement. Besides these ‘non-voluntary’ pension provisions, for some individuals it is still desired to have more income during years of retirement. This income can be generated through ‘voluntary’ retirement planning during working years, which will be the variable of interest in this paper.
To study the effect of financial literacy and family size on voluntary retirement planning, the following research question is being tested:
How do financial literacy and family size relate to voluntary retirement planning?
Following the earlier mentioned findings and motivations in the current literature with respect to the role of financial literacy, and the changing role of children in modern day families related to voluntary retirement savings, the hypotheses being tested in this paper are: H1: Financial literacy positively relates to voluntary retirement planning.
1 Statistic provided by the Dutch Bureau of Statistics at
[4]
H2: Family size negatively relates to voluntary retirement planning.
Next to these hypotheses, a third hypothesis is added to test for the interaction effect between financial literacy and children on voluntary retirement savings. This hypothesis is stated in two ways:
H3a: The interaction of financial literacy and family size is negative in relation to voluntary retirement planning.
H3b: The interaction of financial literacy and family size is positive in relation to voluntary retirement planning.
The reason that the third hypothesis is stated in two ways is that both hypotheses can be expected based on the first two hypotheses. More precisely, the expectation of the third hypothesis is that the effect of financial literacy on voluntary retirement savings is different for different family sizes, and that the effect of children on voluntary retirement savings differs between families with a financially literate household head compared to families with a financially illiterate household head. With respect to the first statement of the third
hypothesis (H3a), the motivation for the hypothesis is the expectation that the effect of financial literacy on voluntary retirement planning is strong for small families, but that this effect weakens when family size increases. More precisely, the increasing cost of additional children is expected to outweigh the effect of financial literacy, leading to a negative
interaction term. With respect to the second statement of the third hypothesis (H3b), the motivation of the hypothesis is the expectation that there is a difference between financially literate respondents and financially illiterate respondents with respect to maintaining a certain savings level after the birth of children. More precisely, it can be expected that financially literate respondents are better able to recalculate their saving needs, and adjust their spending patterns accordingly, after the birth of children. In that case, an increasing amount of children would only have a negative effect on financially illiterate respondents, leading to a positive coefficient of the interaction term.
[5]
favour nor against one of the arguments mentioned earlier. With respect to the third hypothesis, the findings of this paper indicate that the interaction term between financial literacy and children is positive, indicating that financially literate respondents are better able to keep increasing their savings after the birth of children compared to financially illiterate respondents. However, as the results of the interaction term are mostly insignificant, there is little evidence to support this statement.
Further motivation and discussion of this research and its results will be discussed in the remainder of this paper, which is organized as follows: section 2 provides an overview of the relevant literature with respect to the Dutch pension system, financial literacy and family size, section 3 discusses the collection and transformation of the data used in this study, section 4 provides an explanation of the methodology used in this study, section 5 shows and discusses the most important results, section 6 discusses the limitations of this research and section 7 concludes.
2. Literature review
In this section, a description of the Dutch pension system will be given, and the current literature about the variables financial literacy and family size with respect to economic studies will be discussed.
[6]
out benefits in the future. These premiums are partly paid by employers and partly by employees, with the larger part being paid by the employer. If people want to save more for their retirement, for example because they want to retire early or simply fill gaps in their savings from the first and second pillar, they can have private voluntary pension savings. Typically, annuity insurance contracts, such as pension insurance contracts and single premium insurance contracts, are being offered. These saving contracts enjoy favourable tax conditions, which makes them more attractive for retirement savings compared to regular saving accounts.
The idea of the whole pension system is to let people save for their retirement, in order to prevent a shortage of assets during these years where income is low. The role and importance of saving in general within economic context has been subject of numerous studies. Perhaps one of the most famous theories with respect to saving is the theory of Keynes (1936), who was among one of the first to discuss different motives for saving. One of these saving motives is ‘foresight’ (also called the life cycle motive), which can be described as saving for expected future differences in income. In light of his study, saving for retirement fits well within this ‘foresight’ motive, as individuals expect that they cannot work and receive income from this work forever, so they need to save some of their current income to ensure cash reserves in the future. Another famous theory about saving is the lifecycle theory, described by Modigliani and Brumberg (1954). This theory describes how individuals save part of their current income for later stages in life where they do not expect to generate income anymore. Furthermore, they argue that the level of retirement savings must be proportional to its basic earnings capacity, and that the timespan in which these savings are made is independent of the level of income. Moreover, they disagree with Keynes’ statement that when real income increases, the proportion of income saved should also increase, by claiming that, in essence, the proportion of income saved is independent of the level of income.
[7]
example that, when taking human aspects into consideration, the earlier mentioned life cycle theory no longer holds.
In a related study, Thaler and Benartzi (2004) use these behavioural economic aspects in their analysis of the SMarT (Save More Tomorrow) program in the US. They mention, among other things, that real household behaviour differs from the optimal behaviour due to a lack of self-control to save part of current income for retirement years. This is an indication that individuals do not save enough for retirement according to theory, as a consequence of behavioural phenomena. The SMarT plan was designed to overcome some of the most common behavioural phenomena, such as lack of self-control, loss-aversion and
procrastination. Moreover, they find that the plan was successful in most of the cases, which provides evidence that behavioural aspects do play an important role in the real world economy.
Relating to the findings of Thaler and Benartzi, Choi et al. (2005) also find evidence that behavioural phenomena play an important role in saving behaviour. One of their key findings is that individuals choose options that require the least amount of effort, which they call passive decision making. This finding is also linked to the deviation from saving behaviour theory, as individuals save less than theory suggests because it takes effort to save for retirement.
As mentioned earlier, Mullainathan and Thaler mention the capacity to perform complex calculations as one of the requirements for retirement saving. Classic economic theories suggest that every individual is rational, and would therefore be able to compute these complex calculations. However, it seems obvious that in the real world this is not the case. Therefore, the analysis of such problems should include a measure to control for the fact that not every individual might be able to perform these calculations. One could describe the extent to which individuals understand (complex) economic instruments as financial literacy. Lusardi and Mitchell (2014) describe financial literacy as “ peoples’ ability to process
economic information and make informed decisions about financial planning, wealth accumulation, debt, and pensions” (Lusardi and Mitchell, 2014, pp 6). Moreover, they find that the overall financial knowledge of people is (too) low and they argue that these gaps have to be filled by, among other things, improving financial education.
[8]
significantly less likely to participate in the stock market. The main argument for this finding is that stocks can be seen as complex assets, and financial illiterate people may not know how to trade on these assets. Furthermore, Behrman et al. (2012) find a positive relation for the interaction of financial literacy and schooling on household wealth accumulation and argue that investments in financial literacy could lead to large wealth payoffs. In addition, Von Gaudecker (2015) finds a positive association between financial literacy and portfolio diversification, which he attributes to the phenomenon of overconfidence. This means that illiterate people have a false believe of their own knowledge and therefore have under diversified portfolios.
A number of authors have linked financial literacy to (retirement) savings in their study. For instance, Clark, d’Ambrosio, McDermed and Sawant (2003) investigate the role of financial literacy on retirement savings allocation by measuring the impact of a financial education program. They measure the impact of the educational event by comparing the results of a pre-event survey with the results of the after-pre-event survey and find that after the educational event, participants change their behaviour towards retirement savings. More specifically, they find a significantly positive effect of the educational event on revising retirement goals. Lusardi and Mitchell (2007) study the relationship between financial literacy and retirement preparedness using data from the 2004 Health and Retirement Survey in the US. They find that, according to literacy surveys, the financial knowledge in many developed countries is low. They find this result particularly worrying for the fact that low financial literacy relates to inability to save for retirement, as people simply do not know that they need to save for later. In a related study, Lusardi and Mitchell (2011) find a clear relation between financial knowledge and planning. Moreover, they find that those with more knowledge that are more likely to plan, more often keep to that planning and rely on formal planning methods.
[9]
argument of Mullainathan and Thaler about the need to be able to perform complex calculations.
Where the role of financial literacy in economic decision making is becoming more and more important and has been investigated a number of times, the role of family size (and
composition) in economic decision making is a relatively less studied subject. However, when keeping behavioural aspects as a deviation from the ‘optimal economic behaviour’ in mind, one could argue that family size has an impact on the way individuals make decisions. Angrist and Evans (1996) find a drop in female labour supply after the birth of the child, but do not find this drop for husbands, using 1980 and 1990 Census Public Use Micro Samples in the US. Furthermore, they find that these effects of children on labour supply to be smaller for women with college education and for women whose husbands earn relatively high wages. Moreover, they find no significant raise in labour hours or income by the husband during the period around and after the childbirth. Therefore, this drop in female labour supply implies a drop in household income due to the birth and caretaking of children.
An increase in household size does not only affect the income of the household, but can also affect the proportion of income it is able to save. This effect can either be positive or negative, as one can think of the increased costs which reduce the ability to save, but for families with high incomes in the first place, child birth can lead to a reduction in wasteful investments and increase the proportion of excess income saved. Freyland (2004) argues that modelling
households as a single decider with respect to household saving behaviour could involve some large mistakes. He argues that other factors, such as difference of age between the partners and the length of time children stay at home could also have an impact on the household savings rate. Furthermore, he argues that the time path of children within the household is a dimension of heterogeneity among households. Households with children in general face higher expenses and therefore lower savings during the time children are in the household, compared to childless households. In a related article, Freyland (2005) provides evidence of this assumptions using German SOEP Data. He finds, among other results, a positive effect of children on savings for couples with young children but a negative effect for older couples. However, he fails to identify positive saving effects because of parents saving for their children’s education.
[10]
mentioned motives is the bequest motive (referred to by Keynes as the ‘pride’ motive) which basically states that individuals want to save to leave resources to their heirs (Canova and Webley, 2005). For instance, Love (2005) finds that changes in family composition can lead to large adjustments in portfolio choice and household savings. Furthermore, he finds a significant effect of the number and age of children on these adjustments. This implies that, based on the current literature, the birth of children can have an effect on total household income, proportion of income saved and allocation of savings.
Overall, these effects all seem to look negative with respect to retirement savings. However, having children can also be seen as a substitute for cash reserves during retirement years, as children can take care of their parents during this time. In other words, children can also be seen as a form of retirement saving. Galasso, Gatti and Profeta (2009) find, especially in countries with low-developed capital markets, a significantly positive effect for children as a substitute for pension savings. The reason for this, according to Galasso et al. (2009, pp 19) is that ‘This effect is larger in countries with less developed capital markets, because when the saving instrument is more costly, individuals rely more heavily on the fertility choice to ensure an adequate level of resources for their second period of life.’ This finding then implies that in countries with more developed capital markets, individuals should rely less on children to ensure future resources. However, their paper does not provide evidence for this intuition, leaving a potential gap in the literature.
This paper tries to close this gap by taking into account the variable of financial literacy and family size using recent data for a country with a developed capital market (The Netherlands, 2016). Therefore, it relates to the study of Lusardi and Mitchell in a way that it also studies the relation between financial literacy and retirement planning. Furthermore, it relates to Van Rooij et al. in a way that it also uses data from the DNB Household Survey. It complements the study of Van Rooij et al. by using more recent data of almost 10 years later. This is
particularly complementary, since technological innovation has grown significantly in the last couple of years, making it much easier nowadays to acquire information. In other words, financial innovation has reduced information costs, which was one of the most important arguments of Van Rooij et al. with respect to the importance of financial literacy.
[11]
3. Data collection and transformation
The data used to test the hypotheses comes from the DNB Household Survey (DHS) and is collected by CentERdata. The DNB Household survey consists of several modules containing data about various financial and psychological concepts. The data is collected every year for around 2000 households and around 5000 individual respondents. For this thesis, the data wave of 2016 is being used. The variables of interest are collected from the various modules and combined into a sub sample. The final sample contains data about age, gender, household size, income, (retirement) savings, and financial knowledge for a total of 1970 respondents. Table 1 (in the appendix) provides an overview of the exact variables from the survey used for this thesis. It shows the module from which the variable was collected, the original variable name, the (new) name used in this study, and a general description of the variable.
< Insert table 1 here>
Some variables are coded, or measured on a certain scale. The numbers on these scales correspond to certain responses. Table 2 (in the appendix) provides an overview of the variables that are measured on these scales and shows the meaning of the corresponding answers.
<Insert table 2 here> Retirement planning
[12]
in the third pillar if they believe their benefits from the first and second pillar are not
sufficient enough. Therefore, it seems more logical to add the benefits from the second pillar as a control variable in the regression.
The variable used to measure the amount of pillar 1 pension benefits is the AOW variable. This variable directly measures the yearly pillar 1 income when respondents have reached the age of 65. It is important to notice that this is the amount as of 2016, meaning that this is most likely to increase over the years until respondents have reached the age of 65. To measure the amount of pillar 2 savings, the variable pillar 2 is used. This variable is the total amount of savings respondents have built up through their pension funds and is measured as the total yearly income from pillar 2 when respondents have reached the age of 65 or decide to retire. The most important variable to measure the amount of the savings in the third pillar is the retirement savings variable, together with the retirement account variable. The retirement account variable measures how many saving accounts respondents have that are specifically labelled for retirement (annuity insurance accounts). The retirement savings variable than measures what the total amount of savings on these saving accounts is. To make matters more clear, from now on pillar 3 retirement savings will be referred to as voluntary retirement savings (VRS), whereas pillar 1 and pillar 2 savings will be referred to as non-voluntary retirement savings.
Financial literacy
The DNB Household Survey contains only one question about financial literacy. This
question asks respondents how they judge their own knowledge regarding financial problems. Respondents only have four answer possibilities, see table 2 (perception variable). This is not a very in-depth measure of financial literacy, and it is subject to, among other things,
[13]
better able to judge their own knowledge, making the response to the perception question more reliable.
Family size
Besides the effect of financial literacy on voluntary retirement savings, this paper studies the effect of family size, and in particular the effect of having children on voluntary retirement savings. The DNB Household Survey does not contain a question about the total amount of children respondents have, but they do ask how many children are living in the household, and how many children have left the household already. Therefore, to measure the effect of children on voluntary retirement savings, these variables are combined to measure the total amount of children. Furthermore, the household size itself is also measured in the survey. The sample contains mostly Dutch respondents, so therefore most households consist of two parents and their children, since polygamy is forbidden by law in the Netherlands. However, there might still be variation in household size, for example because of a divorce or one of the grandparents is living in the household. To control for this variation, the number of household members, additional to the number of children, is being used as a control variable in the regression.
Control variables
Of course, the amount saved for retirement cannot completely be explained by financial literacy and the amount of children alone. In fact, one can think of numerous variables that can have an impact on retirement planning. In order to make the analysis as complete as possible, other important variables that can have an impact on retirement planning are added to the regressions as control variables. The most important control variables are age,
household income and non-voluntary retirement savings.
3.1. Descriptive statistics and frequencies
Descriptive statistics of the raw data sample are presented in table 3a. When looking at the age of the population, the distribution seems relatively normal, with a mean age of 53.17, a minimum of 1 and a maximum of 94. By further inspecting the data, however, there is a small number of respondents for which the age is below 16. Since the description of the Household Survey states that all household members above the age of 16 are asked to fill in the
[14]
Furthermore, when looking at total household income, voluntary pension savings, total general savings and non-voluntary pension savings, one can notice that the mean amount is far closer to the minimum amount than to the maximum amount, indicating the presence of outliers. However, in contrast to the age variable, it is not clear whether these are due to errors in processing the data. Furthermore, it could be that the maximum values for these variables are specific outliers due to one-time events, such as an in heritage, and are therefore not related to the variables of interest, financial literacy and family size. Figures 1a-d in the appendix show the boxplots for the variables total household income, voluntary retirement savings, general savings and non-voluntary retirement savings, respectively. Based on the results of these boxplots, respondents 4377601, 6884401, 6904601, 337401, 3569201, 4193001, 4894601, 1560702, 3235701 and 1922012 are marked as outliers and are therefore excluded from the sample. As the drivers for these outliers are not known, the exclusion of the outliers is purely on a statistical basis. To increase transparency of the research, in the results section there will be a reference to results including the outliers.
Another value to notice is the minimum value of income, which is minus 4154.75. The codebook describes the total income after tax as the total income before tax, minus the taxes, plus a couple of other variables including profit and alimony. Since these variables can have negative values, the total income after tax can indeed be negative. Therefore, it seems plausible that the minimum value for income is negative.
Tables 3a and 3b show the descriptive statistics of the data sample before and after the removal of earlier mentioned outliers, respectively. The data of table 3b, after the removal of outliers, is the data sample used in the remainder of this paper.
Table 4 shows the frequencies of the variables Gender, Degree, Retirement savings accounts and General savings accounts. When looking at the frequencies of table 4, one can observe that a small majority of the sample is male (57.3%), and that the minority of the sample has either a vocational college degree (HBO) or a university degree (WO) (41.6%). Furthermore, one can observe that 74.9% of the sample does not have a savings account specifically
labelled for retirement, leading to a largely skewed distribution when taking the whole sample into account. This can be related to the findings of Thaler and Benartzi, and Choi et al., with respect to deviating from saving theory (e.g. saving less than theory suggests) due to
behavioural characteristics. When looking at general, unlabelled, saving accounts, however,
[15]
the data shows that a large part of the respondents (72.4%) does have a saving account. This is an early indication that, although a large part of the sample is able to save, only a small part chooses (or is able) to save specifically for retirement.
Table 3a: Descriptive statistics of raw data sample
Table showing the number of observations, minimum value, maximum value, mean value and standard deviations of key variables before removal of outliers.
N Minimum Maximum Mean Std. Dev
Age 2031 1.00 94.00 53.17 18.67
Household size 2031 1.00 8.00 2.36 1.24
Financial literacy 1938 0.00 1.00 0.42 0.24
Children 2031 0.00 12.00 0.93 1.46
Total household income 1376 -3408.00 561793.84 20344.75 25993.898
Voluntary retirement savings 1769 0.00 2685188.50 4854.07 69346.00
Total general savings 1769 0.00 2871270.75 25347.72 86589.26
Non-voluntary retirement savings
1698 0.00 796369.00 9244.77 24967.13
Valid N (list wise) 1243
Table 3b: Descriptive statistics of data sample after removing outliers
Table showing the number of observations, minimum value, maximum value, mean value and standard deviations of key variables after removal of outliers.
N Minimum Maximum Mean Std. Dev
Age 1970 16.00 94.00 54.32 17.42
Household size 1970 1.00 8.00 2.33 1.23
Financial literacy 1882 0.00 1.00 0.42 0.24
Children 1970 0.00 12.00 0.94 1.47
Total household income 1333 -3408.00 174391.00 29093.25 18585.39
Voluntary retirement savings 1723 0.00 462817.00 2911.94 19811.23
Total general savings 1723 0.00 579344.00 23073.88 47412.87
Non-voluntary retirement savings
1646 0.00 171000.00 8771.56 14502.18
[16] Table 4: Frequencies of data sample
Frequency table showing frequencies of responses, after removing outliers, for variables Gender, Degree, Retirement saving accounts, and general saving accounts. Percentages of respondents is given between brackets. Gender is split between males (response 0) and females (response 1). Responses of 0 for Degree, retirement saving accounts and general saving accounts indicate that respondents do not have a degree, retirement saving account or general saving account, whereas a response of 1 indicates they do.
Response Gender Degree Retirement saving
accounts General saving accounts 0 1129 (57.3) 1150 (58.4) 1476 (74.9) 296 (15.0) 1 841 (42.7) 820 (41.6) 247 (12.5) 1427 (72.4) Total 1970 2031 1723 (87.5) 1723 (87.5) Missing 247 (12.5) 247 (12.5) Total 1970 1970 1970 1970 4. Methodology
After collecting the data from the different modules, a new dataset needs to be created to be able to perform the regressions. The first step in this process is to match the responses from different modules to their respondents. Therefore, every respondent has a unique identifying number that is being calculated from the variables nohhold and nomem using:
𝑅𝑒𝑠𝑝𝑜𝑛𝑑𝑒𝑛𝑡 𝐼𝐷 = (𝑛𝑜ℎℎ𝑜𝑙𝑑 ∗ 100) + 𝑛𝑜𝑚𝑒𝑚 (1)
After all the respondents and their responses have been matched, the income of all household members is cumulated into one variable labelled ‘total household income’. Thereafter, all household members excluding the household head are removed from the sample to prevent double counting. In this process, the household member with the highest individual income is treated as the head of the household. Furthermore, the financial literacy variable that is measured on a scale from 1 to 4 is transformed to a 0-1 range by subtracting 1 from the original score and dividing this new score by 3, and the total household income and non-voluntary retirement savings variables are divided by 1000 to test the effect of income and non-voluntary retirement savings in thousands of euros instead of in single euros.
In addition to the variables collected from the DHS, some extra variables are created and used in the study. First of all, the variable age is calculated as 2016 – Birthyear. Furthermore, the variable non-voluntary pension savings is calculated by adding the pillar 1 and pillar 2
[17]
the perception variable for measuring financial literacy. Every respondent with a degree in vocational college (HBO) or university (WO) receives a ‘1’ for degree, whereas all other responses will be treated as ‘0’.
Now that the data is being transformed, the first two hypotheses can be tested. The most simplistic way to test the hypothesis is a simple OLS regression containing the dependent variable voluntary retirement savings (VRS) and the independent variables of interest, financial literacy and the amount of children. A regression of this sort looks as follows: 𝑉𝑅𝑆 = 𝐶 + 𝛽1𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽2𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝜇 (2)
Where C is a constant and µ is the error term. Furthermore, the variable children measures the total amount of children as the sum of the children living inside the household and children living outside the household.
However, as mentioned earlier, table 4 shows an indication for the presence of skewed data for the dependent variable voluntary retirement savings. To test the indication of a non-normal distribution, the Jarque-Bera test is performed to test for a non-normal distribution. Figure 2a in the appendix shows the histogram and corresponding results of the Jarque-Bera
normality test for the variable voluntary retirement savings. Since the p-value of this statistic is zero, the null-hypothesis of normally distributed residuals can be rejected at the 1% level, therefore a non-normal distribution can be assumed. It is important to notice that the
distribution would be approximately normal if the excess of zero values would not be in the data sample. This will be dealt with later on.
As a normal distribution of error terms is one of the key assumptions for BLUE (best linear unbiased estimators) coefficients, in this case it can be assumed that the coefficients are biased. One way to solve this problem is to take the log-transformed value of the voluntary retirement savings variable as the dependent variable. However, since a large part of the responses in this case is zero, this part of the responses will be left out of the analysis after a simple log-transformation, leading to a large decrease in number of observations, and
therefore to a large decrease in statistical power of the analysis. As this effect is not desirable, the inverse hyperbolic sine of the voluntary retirement savings (IHSVRS) variable is taken as the dependent variable in the regression, using
[18]
Compared to a normal log transformation, the advantage of using the IHS log transformation is that the zero values will not be left out of the analysis. Figure 2b shows the histogram and the results of the Jarque-Bera normality test for the inverse hyperbolic sine of the voluntary retirement savings variable. Although the p-value is still zero, and therefore normality cannot be assumed, this transformation is closer to a normal distribution than the standard voluntary retirement savings variable.
After replacing the dependent variable VRS with the inverse hyperbolic sine transformed value of voluntary retirement savings, the hypotheses can be tested again. Moreover, in order to be able to test the third hypothesis, namely the interaction effect of both explanatory variables, the interaction effect of financial literacy and children on VRS needs to be embedded into the regression. The regression than look as follows:
𝐼𝐻𝑆𝑉𝑅𝑆 = 𝐶 + 𝛽1𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽2𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝛽3𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑙𝑖𝑡𝑒𝑟𝑎𝑐𝑦 ∗
𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝜇 (4)
Where IHSVRS is again the inverse hyperbolic sine transformed value of the variable voluntary retirement savings, C is a constant, and µ is the error term.
However, this model is highly likely to suffer from omitted variable bias, as one can think of other important explanatory variables with respect to voluntary retirement savings that are left out of the analysis. To reduce this omitted variable bias, some of these explanatory variables can be included into the regression as control variables. These control variables include age, additional household members, degree (dummy), household income and non-voluntary retirement savings. By keeping ordinary least squares as the regression method and by adding these control variables, the regression looks as follows:
𝐼𝐻𝑆𝑉𝑅𝑆 = 𝐶 + 𝛽1𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽2𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝛽3𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑙𝑖𝑡𝑒𝑟𝑎𝑐𝑦 ∗ 𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝛽4𝐴𝑔𝑒 + 𝛽5𝐴𝑔𝑒2+ 𝛽6𝐴𝑑𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙 ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑚𝑒𝑚𝑏𝑒𝑟𝑠 + 𝛽7𝐷𝑒𝑔𝑟𝑒𝑒 + 𝛽8𝐻𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖𝑛𝑐𝑜𝑚𝑒 + 𝛽9𝐻𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖𝑛𝑐𝑜𝑚𝑒2 + 𝛽
10𝑁𝑜𝑛 −
𝑣𝑜𝑙𝑢𝑛𝑡𝑎𝑟𝑦 𝑟𝑒𝑡𝑖𝑟𝑚𝑒𝑛𝑡 𝑠𝑎𝑣𝑖𝑛𝑔𝑠 + 𝜇 (5)
Where the first terms are the same compared to (4), the variable additional household
[19]
One can think of several non-linearity’s in the effect of age, for example when people that are close to retirement are not working that hard anymore and already start consuming part of their pension savings instead of increasing them. It is also possible to think of a non-linear effect of household income, for example when a household has (more than) sufficient assets and starts making wasteful investments instead of saving extra for their retirement.
To check if the adding of the control variables can be justified, a redundant variables test is being performed. This test tests if the added control variables are jointly insignificant. Table 5 in the appendix shows the results of this test. From table 5 it can be concluded that the null-hypothesis of jointly insignificant control variables can be rejected at the 1% level. Therefore it can be assumed that the control variables are indeed useful in the regression. However, with the control variables added into the regression, there might be a new problem, which is
multicollinearity. It could be that there is some form of correlation between the explanatory variables themselves. The consequence of multicollinearity is that the individual estimates of coefficients measuring the effect on the dependent variable will be less precise compared to the case where all explanatory variables are uncorrelated. To test if this is really a problem in the sample, the correlation matrix for the explanatory variables is being constructed and shown in table 6 in the appendix. From table 6 it can be seen that overall the correlations between explanatory variables is low, with the highest correlation between the variables age and children (0.51). When using a correlation of 0.7 as a rule of thumb, one does not need to worry about these values of correlation. Therefore, multicollinearity does not seem to be a problem in the regression.
[20] 𝐼𝐻𝑆𝑉𝑅𝑆∗ = 𝐶 + 𝛽
1𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽2𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝛽3𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑙𝑖𝑡𝑒𝑟𝑎𝑐𝑦 ∗
𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝜇 (6)
And the Tobit regression with control variables looks as follows:
𝐼𝐻𝑆𝑉𝑅𝑆∗ = 𝐶 + 𝛽1𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝐿𝑖𝑡𝑒𝑟𝑎𝑐𝑦 + 𝛽2𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝛽3𝐹𝑖𝑛𝑎𝑛𝑐𝑖𝑎𝑙 𝑙𝑖𝑡𝑒𝑟𝑎𝑐𝑦 ∗ 𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝛽4𝐴𝑔𝑒 + 𝛽5𝐴𝑔𝑒2+ 𝛽6𝐴𝑑𝑑𝑖𝑡𝑖𝑜𝑛𝑎𝑙 ℎ𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑚𝑒𝑚𝑏𝑒𝑟𝑠 + 𝛽7 𝐷𝑒𝑔𝑟𝑒𝑒 + 𝛽8𝐻𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖𝑛𝑐𝑜𝑚𝑒 + 𝛽9𝐻𝑜𝑢𝑠𝑒ℎ𝑜𝑙𝑑 𝑖𝑛𝑐𝑜𝑚𝑒2 + 𝛽 10𝑁𝑜𝑛 − 𝑣𝑜𝑙𝑢𝑛𝑡𝑎𝑟𝑦 𝑟𝑒𝑡𝑖𝑟𝑚𝑒𝑛𝑡 𝑠𝑎𝑣𝑖𝑛𝑔𝑠 + 𝜇 (7) Where 𝐼𝐻𝑆𝑉𝑅𝑆 {𝐼𝐻𝑆𝑉𝑅𝑆 ∗𝑖𝑓 𝐼𝐻𝑆𝑉𝑅𝑆∗> 0 0 𝑖𝑓 𝐼𝐻𝑆𝑉𝑅𝑆∗ ≤ 0
In this way, the Tobit regression allows to test for the relation between VRS, Financial
[21] 5. Results
Results of specific regressions are referred to by their number as described in the
methodology section (for example, (4) refers to regression 4 described in the methodology section). Table 7 shows the results of regression (2). The results indicate that VRS is positively influenced by financial literacy and children. More specifically, in this simplistic regression, voluntary retirement savings are about 4741 euros higher for financially literate respondents compared to financially illiterate respondents, and seem to increase by about 562 euros for every additional childbirth. Furthermore, it can be noticed that the coefficient for the variable financial literacy is statistically significant at the 1% level, whereas the coefficient for the variable children is not statistically significant. However, significance in this case is based on the t-statistic and the corresponding p-value. Since a normal distribution cannot be assumed, one needs to keep in mind that this measure of significance is not always reliable. Table 7: Results of regression (2)
Results of regression (2). Dependent variable: voluntary retirement savings. Standard errors are reported within brackets. F-statistic corresponds to a Wald test of the hypothesis that all coefficients excluding the constant are equal to 0. Regression using heteroskedasticity and autocorrelation (HAC) robust standard errors. P-values smaller than 0.01, 0.05 and 0.1 are denoted by ***, ** and * respectively. Variable Coefficient Financial literacy 4740.950*** (1751.707) Children 409.692 (269.984) Constant 561.726 (853.891) Observations 1686 R-squared 0.004 Adjusted R-squared 0.003 F-statistic 3.470**
Regression (2), as mentioned before, is likely to be biased and therefore serves as the baseline, from which improvements need to be made. Furthermore, to be able to test
hypothesis 3, regression (4) is performed. The results of this regression, along with the results of (5), (6) and (7), are presented in table 8. Compared to table 7, the results in table 8 show coefficients for the dependent variable IHSVRS instead of VRS. Keeping in mind that
[22]
be interpreted as approximately a (Coefficient *100) % change in VRS for every unit increase in the explanatory variable3.
The results of (4) show an interesting change in coefficients compared to (2). After adding the interaction term Financial literacy * Children to the regression, the coefficients change
substantially. Furthermore, by including the interaction term, the interpretation of the
coefficients changes. Where the coefficients in (2) estimate the sole effect of financial literacy and children on VRS, in (4) (and further regressions) the effect is measured by a combination of the estimates of the coefficients in (3) and the coefficient of the interaction term.
In the case of (4), VRS is about 84.2% higher for financially literate people compared to financially illiterate people. Furthermore, VRS decreases by about 6.7% for every additional childbirth (instead of increasing in (2)). Moreover, the interaction term between financial literacy and the amount of children is positive. This can be interpreted in two ways, namely that the amount of VRS increases by an additional 29.1% (on top of the 84.2%) for every additional child in the family of a financially literate respondent, or that the amount of VRS is about 29.1% per child higher for financially literate respondents with children compared to financially illiterate respondents with children.
As mentioned before, control variable are added to regression (4) to decrease the omitted variables problem. After these variables have been added, the results of (5) are obtained. From table 8 it can be observed that the inclusion of control variables increases both the R-squared and adjusted R-squared, indicating a stronger regression compared to (4). One can observe that the inclusion of control variables changes the coefficients of the variables of interest again. With the inclusion of control variables, the results still indicate a significantly positive effect of financial literacy on VRS (although one still needs to be cautious about the reliability of the significance measure in this case). Compared to (4), the results of (5) show that the coefficient of the variable children now becomes positive (again). Furthermore, the coefficient of the interaction term is positive. However, these coefficients are highly insignificant and therefore do not seem to provide much information about the effect of children on VRS (at this stage).
3 For relatively large values of VRS, IHSVRS approximates log(VRS) for which the coefficients can be interpreted
as (coefficient * 100)% change in VRS for every unit increase of the explanatory variable. In reality, an additional factor equal to √𝑉𝑅𝑆2+ 1 needs to be added to the derivative of the ihsvrs function, but as the
[23] Table 8: Results of regressions (4), (5), (6) and (7)
Results of regressions (4), (5), (6) and (7). Dependent variable: ihsvrs, which is the inverse hyperbolic sine-transformed value of voluntary retirement savings. (4) And (5) are results of an OLS regression, whereas (6) and (7) are results of a Tobit regression, left censored at zero. Standard errors are reported within brackets. F-statistic corresponds to a Wald test of the hypothesis that all coefficients excluding the constant are equal to 0. Regression using heteroskedasticity and autocorrelation (HAC) robust standard errors. P-values smaller than 0.01, 0.05 and 0.1 are denoted by ***, ** and * respectively. Regression (4) (5) (6) (7) Variable Coefficient Sdsdffd Financial literacy 0.842** (0.382) 0.920** (0.465) 5.291** (2.648) 4.066 (3.057) Children -0.070 (0.089) 0.049 (0.120) -0.660 (0.758) 0.236 (0.873) Financial literacy*Children 0.291 (0.222) 0.136 (0.273) 1.901 (1.624) 0.989 (1.769) Degree 0.001 (0.209) -0.576 (1.246) Age 0.260*** (0.038) 1.706*** (0.266) Age2 -0.002*** (0.000) -0.016*** (0.003)
Additional household members -0.253*
(0.134) -1.325 (1.013) Household income 0.015 (0.011) 0.085 (0.068) Household income2 -0.000 (0.000) -0.000 (0.000)
Non-voluntary retirement savings 0.002
(0.008) 0.030 (0.041) Constant 0.867*** (0.180) -5.480*** (1.089) -14.521*** (1.662) -53.371*** (7.344) Observations 1686 1163 1686 1163 R-squared 0.008 0.056 - Adjusted R-squared 0.007 0.048 - F-statistic 4.715*** 6.857*** -
[24]
result is not significant, there is little evidence to support this statement based on the current data. Furthermore, one can observe the positive coefficient for the age variable, and the negative coefficient for the quadratic term of the age variable, which are both statistically significant at the 1% level. This indicates that VRS increases for every year of age, but that this increase becomes smaller with every additional year increase in age, and might even decrease after a certain point. A possible explanation for the presence of this non-linear effect of age is that in the beginning, most people earn more and are therefore able to save more every year, but at a certain point, for example when buying a house or when having children, expenses increase, decreasing the amount of money saved. Furthermore, one can think that in later years, when approaching retirement, individuals might actually start consuming their savings, in which case VRS can even decrease when age increases further (but still below 65). Another possible explanation might simply be a generational effect, as current retired
respondents faced different economic conditions during their working period compared to current working respondents. It might be that because of lower costs of life in the past, the need to save for retirement was simply less than it is nowadays.
Moreover, the results of (5) indicate that VRS decreases by about 25.3% for every additional household member excluding the household head and children. Besides that, VRS seems to increase by about 1.5% for every 1000 euro increase in household income, but the results do not indicate the presence of non-linear effects of household income (such as wasteful
investments for the very rich). Rather surprisingly is the positive coefficient for non-voluntary retirement savings which indicates that VRS increases by about 0.2% for every 1000 euro increase in non-voluntary retirement savings. This might be an indication that VRS acts as a complement to non-voluntary retirement savings, rather than as a substitute. However, apart from the coefficient for additional household members it is again important to notice that the results are highly insignificant, so that it might be an indication, but there is little evidence for it.
[25]
the effect of financial literacy on VRS is positive, given that VRS is larger than zero.
Furthermore, the interaction term between financial literacy and children is also positive in (6) and (7). When looking at the variable children, one can observe that also for the Tobit
regression, the coefficient is negative in the case without control variables, but becomes positive when the control variables are added. Therefore, in terms of signs, both regression types generate similar results with respect to the variables of interest. This is particularly interesting, since in the case of the OLS regression one cannot be sure about the reliability of the significance measure. Since the Tobit regression only looks at the cases that are larger than zero for which the distribution is approximately normal, the similar results indicate more reliability with respect to the OLS results.
An interesting result in (7), however, is that the argument that respondents with a higher IQ are better able to estimate their financial knowledge does not seem to hold in this regression, as the coefficient of the degree variable is negative. Besides this, in terms of significance and coefficient signs, the results of the Tobit regression (7) are comparable to the results of the OLS regression (5).
As mentioned in the data section, some respondents were marked as outliers and were therefore left out of the analysis. If these responses were not left out of the analysis, results would be slightly different. Table 9 in the appendix therefore shows the results for regressions (4), (5), (6) and (7) including the respondents that were originally marked as outliers. From table 9 it can be seen that with respect to the variables of interest to test the hypothesis, there are no significant deviations compared to table 8.
[26]
Regarding the difference between male and female respondents, there are a couple of interesting results. For both male and female respondents, the coefficient for the financial literacy variable is positive, which is in line with the findings of the overall sample presented in table 8. Table 10, however, shows that for male respondents, the effect of children on VRS is positive, whereas for females this is negative. Not controlling for statistical significance, this difference can be partly explained by the findings of Angrist and Evans (1996), as they found a drop in female labour supply after childbirth, but they did not find this drop in male labour supply. This drop in female labour supply is likely to lead to a decrease in female income, therefore decreasing the female saving opportunities in case of childbirth, compared to male saving opportunities. When looking at the interaction term between financial literacy and children, one can observe that the coefficient is negative for male respondents, whereas this is positive for female respondents. Also, for male respondents the coefficient of the degree variable is positive, whereas for female respondents this is negative. However, given that these results are highly insignificant, one cannot draw a strong conclusion from these findings. Moreover, the results show that for both male and female respondents there is
significant evidence for the presence of non-linear saving effects of age, which was also found for the overall sample in table 8. Furthermore, for both genders there seems to be a positive relation between household income and VRS, and a negative relation between additional household members and VRS. However, these results are again highly insignificant. Lastly, the results show a significantly positive effect of non-voluntary retirement savings on VRS for female respondents, compared to an insignificantly negative effect for male respondents. This might be an indication that non-voluntary retirement savings act as a complement to VRS for females, whereas it acts as a substitute for VRS for males.
[27]
Table 10: Results of regression (5) for male, female, ‘poor’ and ‘rich’ respondents
Results of regression (5) for male, female, poor and rich respondents using OLS regression method. Dependent variable: ihsvrs, which is the inverse hyperbolic sine-transformed value of voluntary retirement savings. Standard errors are reported within brackets. F-statistic corresponds to a Wald test of the hypothesis that all coefficients excluding the constant are equal to 0. ‘Poor’ respondents correspond to the 25% of respondents with the lowest household income in the sample, whereas ‘rich’ respondents correspond to the 25% of respondents with the highest household income in the sample. Regression using heteroskedasticity and autocorrelation (HAC) robust standard errors. P-values smaller than 0.01, 0.05 and 0.1 are denoted by ***, ** and * respectively.
Subgroup
male female poor rich
Variable Coefficient Sdsdffd Financial literacy 1.172* (0.670) 0.343 (0.654) -0.045 (0.576) 0.857 (0.874) Children 0.145 (0.185) -0.123 (0.194) -0.316** (0.126) 0.303 (0.320) Financial literacy*Children -0.066 (0.346) 0.687 (0.496) 0.710* (0.401) -0.239 (0.520) Degree 0.368 (0.277) -0.519 (0.324) -0.246 (0.341) 0.009 (0.362) Age 0.309*** (0.053) 0.192*** (0.060) 0.202*** (0.046) 0.340*** (0.091) Age2 -0.003*** (0.000) -0.002*** (0.001) -0.002*** (0.000) -0.003*** (0.001)
Additional household members -0.276
(0.219) -0.236 (0.195) 0.094 (0.221) -0.768*** (0.249) Household income 0.007 (0.016) 0.023 (0.017) -0.107 (0.087) 0.008 (0.030) Household income2 0.000 (0.000) -0.000 (0.000) 0.005 (0.003) 0.000 (0.000)
Non-voluntary retirement savings -0.012
(0.009) 0.031** (0.015) 0.025 (0.017) -0.007 (0.009) Constant -6.695*** (1.508) -3.486** (1.652) -3.325** (1.374) -6.512** (3.026) Observations 705 458 427 434 R-squared 0.070 0.065 0.071 0.065 Adjusted R-squared 0.057 0.044 0.048 0.043 F-statistic 4.762*** 3.010*** 3.163*** 2.950***
[28]
This difference in findings for ‘poor’ and ‘rich’ respondents might be due to the fact that for larger incomes, the distinction between saving and spending can be stronger. From another point of view, with low income and low saving ability in the standard situation, the changing effect of increasing financial literacy does not have a large impact if the saving ability stays low. For larger incomes, there can be a transfer from wasteful investments to voluntary retirement savings in the case of an increase in financial literacy or birth of a child.
When looking at table 11, some interesting results for different age groups can be observed. For all age groups, the positive effect of financial literacy on VRS seems to hold. When looking at the effect of children, from table 11 one can observe that again all results are highly insignificant, but that the coefficient for respondents at the age of 46-65 is negative, whereas for other age categories this is positive. Not controlling for significance, a possible
explanation for this finding is that at the age category of 46-65 years old, the cost of children are highest (costs of education, pocket money, healthcare, and so on), leading to a decrease in saving ability. In that way, it could be linked to the findings of Freyland (2005), who found a negative effect on savings for couples with older children, compared to a positive effect for couples with younger children. This might also be a possible explanation for the negative coefficient of the interaction term between financial literacy and children for older working respondents, as well as for retired respondents. Another interesting result from table 11 is that for both young working respondents and retired respondents, the effect of age seems to be opposite of the findings for the overall sample. For young working respondents, a possible explanation might be that at a young age, individuals are building their careers, and face large expenses (i.e. buying a house, car, and so on) while income is not yet at a steady state.
[29]
Table 11: Results of regression (5) for different age groups
Results of regression (5) for different age groups using OLS regression method. Dependent variable: ihsvrs, which is the inverse hyperbolic sine-transformed value of voluntary retirement savings. Standard errors are reported within brackets. F-statistic corresponds to a Wald test of the hypothesis that all coefficients excluding the constant are equal to 0. Regression using heteroskedasticity and autocorrelation (HAC) robust standard errors. P-values smaller than 0.01, 0.05 and 0.1 are denoted by ***, ** and * respectively.
Age group
16-45 years 46-65 years >65 years
Variable Coefficient Financial literacy 0.061 (0.454) 2.240** (1.060) 1.727* (0.885) Children 0.120 (1.366) -0.001 (0.245) 0.218 (0.166) Financial literacy*Children 2.560 (2.475) -0.200 (0.549) -0.492 (0.347) Degree 0.072 (0.266) 0.533 (0.494) 0.019 (0.293) Age -0.451** (0.201) 0.425 (0.682) -1.652*** (0.563) Age2 0.007** (0.003) -0.003 (0.006) 0.010*** (0.004)
Additional household members -0.224
(0.172) -0.397 (0.270) -0.011 (0.224) Household income -0.005 (0.020) 0.030 (0.018) 0.036** (0.012) Household income2 0.000 (0.000) -0.000 (0.000) -0.000*** (0.000)
Non-voluntary retirement savings 0.009
(0.010) 0.006 (0.018) -0.003 (0.012) Constant 7.295** (3.128) -12.231 (18.755) 65.764*** (21.862) Observations 354 374 435 R-squared 0.149 0.048 0.084 Adjusted R-squared 0.125 0.022 0.063 F-statistic 6.019*** 1.832* 3.899***
[30]
Table 12: Results of regressions (5) and (7) with general savings as dependent variable
Results of regression (5s) and (7s), where the s is an indication of a regression with general savings as dependent variable instead of VRS. Dependent variable: ihssavings, which is the inverse hyperbolic sine-transformed value of general, unlabelled savings. (5s) Uses OLS regression method, whereas (7s) uses Tobit regression method, left censored at 0. Standard errors are reported within brackets. F-statistic corresponds to a Wald test of the hypothesis that all coefficients excluding the constant are equal to 0. Regression using heteroskedasticity and autocorrelation (HAC) robust standard errors. P-values smaller than 0.01, 0.05 and 0.1 are denoted by ***, ** and * respectively.
Regression (5s) (7s) Variable Coefficient Financial literacy 1.068** (0.564) 0.353 (0.395) Children -0.256* (0.155) -0.214** (0.106) Financial literacy*Children 0.462 (0.305) 0.378* (0.218) Degree 0.914*** (0.228) 0.457 (0.166) Age -0.111*** (0.039) -0.006 (0.031) Age2 0.001*** (0.000) 0.000 (0.000)
Additional household members 0.130
(0.170) 0.167 (0.129) Household income 0.039*** (0.013) 0.032*** (0.009) Household income2 -0.000** (0.000) -0.000** (0.000)
Non-voluntary retirement savings 0.012
(0.007) 0.005 (0.006) Constant 8.473*** (1.065) 7.745*** (0.837) Observations 1163 1163 R-squared 0.062 - Adjusted R-squared 0.054 - F-statistic 7.607*** -
Figure 3 in the appendix shows the histogram and the result of the Jarque-Bera normality test for the inverse hyperbolic sine-transformed value of the savings variable. Although a normal distribution cannot be assumed from the test statistic, the histogram shows that the
[31]
and children are therefore in line with the results of table 8. Interestingly, with general unlabelled savings, there is significant evidence for a negative effect of children on savings. Given that the earlier findings for the children variable were (highly) insignificant, this finding indicates that there might be evidence for a negative effect of children on general savings, but that further research is necessary to study this effect.
[32]
6. Limitations
This paper uses data of the DNB Household survey to test for the effect of financial literacy and children on voluntary retirement savings. The drawback of using survey data is the risk of biased data, which can be the result of, among other things, rounding, not answering sensitive questions, or simply not answering the truth. Furthermore, there is the risk of errors in the data, which can occur due to wrongly answering questions, or due to errors in processing the survey responses into the database.
Besides the risk of biased data, a limitation in this research is the measurement of financial literacy. Although the measurement power is increased by adding the degree variable, the measurement of the financial literacy variable itself is still measured on respondents own perception of their financial skills. To increase the measurement power of financial literacy, one could follow the example of Van Rooij et al., who added a special module to the DNB household survey to be able to better measure financial literacy. In this specially constructed module, respondents had to answer more detailed questions about financial matters, and had to perform a number of basic-, as well as a number of advanced financial calculations.
Therefore, in order to increase the meaning of financial literacy, modules can be developed to be able to measure financial literacy more precisely. Another possibility is finding a suitable instrument for the variable financial literacy to use in the analysis using the instrumental variable approach. The current dataset of the DHS does not provide a suitable instrument at the moment.
Furthermore, this paper only focusses on annuity insurance accounts as measure of voluntary retirement savings. Of course, one can think of a number of alternatives with respect to saving, such as financial investments in the stock market. This paper does not take these type of assets into account when investing the effect on voluntary retirement savings. Although others (i.e. Van Rooij et al. (2007)) have studied the effects of financial literacy on stock market participation, further research on the effect of family size on other saving options is necessary.
[33]
the data set. Further research can improve the findings of this paper by controlling for more variables with respect to voluntary retirement savings.
7. Conclusion
This paper studies the effect of financial literacy and family size (i.e. the amount of children) on voluntary retirement savings (VRS). It thereby tests the hypotheses that the effect of financial literacy on VRS is positive, the effect of children on VRS is negative, and the
interaction term between financial literacy and children can be either positive or negative. The motivation for the first hypothesis comes from the findings of Van Rooij et al. (2012), and Lusardi and Mitchell (2011), who found that more financially literate people are better able to plan (e.g. for retirement) and are better able to stick to these plans. The motivation for the second hypothesis comes from the findings of Freyland (2005), who argues that saving behaviour depends on household composition, and from the findings of Butz and Ward (1979), Galor and Weil (1993), and Adsera (2005) with respect to the increasing opportunity costs of having children in modern day economies. The motivation for the third hypothesis is that the effect of financial literacy on VRS is expected to differ among different family sizes, and that the effect of (additional) children on VRS differs between financially literate and financially illiterate respondents.
Overall, this paper finds a positive effect of financial literacy on voluntary retirement savings, which is in line with the findings of Van Rooij et al., and Lusardi and Mitchell. In the overall case without control variables, the results is significant at the 5% level for the OLS- and the Tobit regression. In the case with control variables, the result is significant at the 5% level for the OLS regression, but not significant for the Tobit regression. When looking at the results of the heterogeneity analyses, this positive effect seems to be stronger for male respondents compared to female respondents. Furthermore, for poor respondents the results show a (insignificant) negative relation between financial literacy and VRS, whereas this relation for rich respondents seems to be positive. Moreover, the positive relation also seems to be stronger for respondents above the age of 45 compared to respondents below the age of 45 and is also found when testing for the effect of financial literacy on general unlabelled savings. Following the arguments of Van Rooij et al., and Lusardi and Mitchell, as well as the arguments mentioned earlier, the overall positive finding is likely due to the fact that higher financial literacy results in increased planning activity, and therefore higher savings for future retirement years. Moreover, one of the main arguments of van Rooij et al. is that
[34]
perform calculations and plan income over lifetime. Despite the fact that technological innovations have reduced these information costs considerably over the years since the work of van Rooij et al., the argument still seems to hold for data of 2016. The (insignificant) negative finding for poor respondents might be due to the fact that when saving opportunities are low in the first place, the ability to save does not vary between financially literate and illiterate respondents, as this is limited to the level of assets.
The results with respect to the second hypothesis provide little information about the role of family size and the amount of children with respect to saving. In the overall case with the inclusion of control variables, it shows a small insignificant positive effect, which might be somewhat in line with the argument of Canova and Webley that families want to save more to leave resources to their heirs, as the results show that savings increase slightly when family size increases. On the other hand, when the control variables are not taken into account, the results show a small insignificant negative effect of children on VRS, which can be linked in a way to the arguments mentioned earlier with respect to increasing opportunity costs of
children. This negative effect is also found when testing for the effect of children on general unlabelled savings. When looking at the results of the heterogeneity analysis, there also seems to be a negative relation between children and VRS for female and poor respondents as well as for respondents between the age of 46 and 65. For female respondents, this might be due to the fact that in most cases they have to (partially) give up their career, reducing lifetime income. For poor respondents, the increased costs when having a child might have a
significantly larger impact on the ability to save compared to respondents with excess assets. For the respondents aged between 46 and 65, a possible explanation might be that at this age the cost of children are highest (education, healthcare, pocket money and so on). However, as the results are mainly insignificant, this paper provides little evidence of an effect of children on voluntary retirement savings.
Regarding the third hypothesis, the finding of a positive effect of the interaction term between financial literacy and children might indicate that in general, financially literate respondents are better able to maintain their saving levels after the birth of children compared to
