«Annamaria Lusardi Dartmouth College and NBER Peter Tufano Harvard Business School and NBER December 22, 2008 We analyze a national sample of ...»
Of course, all of these univariate measures are likely correlated, and therefore we must consider all of the demographic variables simultaneously. Furthermore, we must use a multivariate approach if we hope to understand the marginal relationship between debt literacy and behavior. Since the dependent variable is an indicator for the four clusters we have identified in the data, we use a multinomial logit analysis.
We have four correlated measures of financial literacy: the self-reported measure of literacy and objective measures resulting from the answers to the three questions discussed above. We further organize these data when performing the empirical work in order to characterize the types of errors individuals make. For example, persons with incorrect answers to the question about interest compounding are divided into two groups: those who made ―underestimates‖ and those who made ―over-estimates‖ of how quickly debt can double. Moreover, we add a dummy for those who do not know the answer to this question as this is a sizable and also distinct group of respondents. As we argued earlier, prior research shows that this group tends to characterize those with the lowest level of knowledge. We also include a dummy for those who refuse to answer the literacy questions.17 All incorrect responses to the second literacy question were underestimates of how many years it would take to eliminate credit card debt. We aggregate the responses into those who make large underestimates (answer it would take less than 5 years and between 5 and 10 years to eliminate credit card debt) versus those who chose a longer yet incorrect time period (between 10 and 15 years). The erroneous answers to the third question characterize two distinct types of respondents: those who fail to realize that the implicit interest rate out of a stream of payment is higher than 20%, and those who fail to recognize that the stream of payments has a higher present value and incorrectly state that the two payment options are the same. We keep these two groups separate. For the second and third measure of literacy we again add dummies for those who do not know the answer or refuse to answer.
Among the demographic variables, we include age and age squared to capture the potential non-linear impact of age. We also include dummies for gender, race and marital status.
We add dummies for larger household sizes, characterizing those with four members and those 17 This is a small but rather heterogeneous group of respondents. For some questions, there is a high prevalence of African-Americans who refused to answer the literacy questions.
17 with five or more members, and a dummy for those who are not employed; these families may be more vulnerable to shocks. Finally, we add dummies for household income and wealth. We include these dummies to proxy for both the resources that respondents have available for their consumption and also to buffer themselves against shocks. Moreover, income and wealth can control for skills and ability (in addition to education) as well as control for individual preferences, such as patience and thriftiness.
Table 9 reports the marginal effect of each variable in the multinomial logit across the four clusters. Rather than reporting the estimates with respect to a reference group, we calculate the marginal effects in comparison to all the other clusters. We first consider the self-reported measure of literacy, which is the most comprehensive measure of knowledge. Those who display higher levels of literacy are more likely to locate in the first cluster (in control). Levels of literacy above the mean score (score higher than 4) are associated with higher chances of being among those in control, and chances become higher at top levels of knowledge (scores of 6 and 7). In other words, those who report higher levels of financial knowledge are more likely to pay credit cards on time. Note that African-Americans and Hispanics and those with large families are less likely to locate in the in-control cluster. Individuals in this cluster are also those with high incomes (income greater than $75,000) and high wealth; individuals in cluster 1 are less likely to report financial assets in the three lowest brackets, and particularly below $50,000.
Self-reported financial knowledge is not related to the behavior of those in cluster 2, the borrower/savers.18 Those individuals have relatively high income, as noted before, and they do not display characteristics that are usually associated with debt problems (e.g., large families, not employed, split families). Income and race (specifically, not being Hispanic) are the only variables that characterize those in cluster 2. However, the borrower/savers do carry balances and tend to pay finance charges. The behavior in this group may simply be due to ―inattention‖ as pointed out in other papers that look at credit card mistakes.19 The dummies for self-reported high financial literacy turn negative when considering cluster 3, the overextended. Even after controlling for many demographic traits, respondents in this cluster are much less likely to report high levels of literacy, and estimates are larger and more negative for those who chose the highest score. These respondents are also those more 18 Note that this finding goes against the argument of ―learning by experience.‖ Respondents in cluster 2 have the highest experience with saving and borrowing. They own the highest percentage of assets and have used borrowing the most. Nevertheless they carry balances on their credit cards and pay fees and finance charges.
19 See Scholnick, Massoud and Saunders (2008).
18 likely to have lower levels of wealth, to be African-American, and to have large families.
However, even after accounting for demographics, income and wealth, literacy remains an important and significant predictor for being over-extended.
Low levels of financial literacy also characterize those in cluster 4, the fringe group.
These respondents are much less likely to report high levels of literacy. Respondents in this cluster also have low levels of income; for example they are disproportionately more likely to have income less than $30,000. They are also more likely to report they are not employed.
Employment status, income, and self-reported literacy are the most important predictors for the respondents in this cluster.
In panels b through d of Table 9 we have replaced the self-reported measure of literacy with the three measures of debt literacy. Since the estimates of the demographic variables are not affected by the measure of literacy we use, this discussion focuses on coefficients related to debt literacy. Those who over-estimate how long it takes for debt to double may be lulled into borrowing more or not paying on time. Indeed, those who are less likely to be knowledgeable about interest compounding, both because they over-estimate the number of years it takes for debt to double or because they do not know the answer to this question, are less likely to be incontrol (cluster 1) and more likely to belong to the fringe (cluster 4). As mentioned above, these two clusters characterize very different types of borrowing behavior and debt literacy remains a predictor of these two groups even after accounting for a rich set of characteristics, including income and wealth. Being unable to answer the question about interest compounding also characterizes those who belong to cluster 3, the overextended who tend to carry balances and pay finance charges and penalty fees. On the other hand, those who do not know the answer to the question about interest compounding are less likely to belong to cluster 2, our borrower/savers who are likely to carry balances and not pay on time.
Turning to the question about minimum credit card payments, we find that those who make mistakes in answering this question, both small and large, are significantly more likely to belong to the fringe group (cluster 4). Those who display the lowest level of debt literacy, i.e., respond that they do not know the answer to this question, are also more likely to belong to this group. Conversely, those who make small mistakes or do not know the answer to the question are less likely to belong to the in-control or borrower/saver clusters.
Estimates for the third question of financial literacy, which was answered correctly only by a small fraction of respondents, show similar findings; those who answered this question 19 incorrectly (i.e., chose option (a) or thought the two options are the same) or do not know the answer to the question are much less likely to belong to the in-control cluster. On the other hand, those who make mistakes in answering this question are more likely to belong to the overextended cluster. As with other literacy questions, those who are less knowledgeable are also less likely to belong to cluster 2, again emphasizing the differences between this cluster and clusters 3 and 4.20 In summary, for each measure of financial literacy, there is a strong relationship between literacy and debt behavior. The more financially knowledgeable who grasp basic concepts about debt are much more likely to be in control of their finances, while those less literate are more likely to be over-extended or be fringe borrowers. Thus, the relationship illustrated in Table 8 continues to hold even after accounting for demographic traits. The curious group is those in cluster 2, our borrower/savers who are rather knowledgeable and have high incomes, yet tend to carry credit card balances and pay finance charges. One may argue that these charges are not sizable and are not of much consequence for borrowers. In the next section we try to address this issue by examining self-reports of debt loads.
8. Difficulties paying off debt According to intertemporal models, consumers borrow to smooth consumption over the life-cycle. Variations in debt over time and across individuals would not necessarily indicate that anyone was ―overlevered‖ or ―underlevered.‖ Yet imperfections in financial markets and shocks might lead individuals to conclude that their debt level was suboptimal. Some may suffer from credit constraints and are unable to borrow as much as they would like. Others may be hit by unexpected negative shocks and carry higher debt loads than they might otherwise have had.
In the survey, we sought to understand whether people have difficulties paying off their debt. While we recognize the potential problems with self-reported measures of debt levels, these reports give information about credit constraints and consumers’ interest in additional
borrowing. To gauge debt levels, we asked individuals the following question:
Which of the following best describes your current debt position?
a. I have too much debt right now and I have or may have difficulty paying it off.
b. I have about the right amount of debt right now and I face no problems with it.
c. I have too little debt right now. I wish I could get more.
20 If debt literacy is measured with error and the errors are random (the classical measurement error problem), then our estimates of debt literacy underestimate the true effect.
In aggregate, in November 2007, before the financial crisis hit the economy, already 26.4% of respondents in our representative sample of Americans said they have or may have difficulties paying off debt (overburdened). Another group, 11.1% ―just didn’t know‖ (unsure) their debt position. Thus, close to 40% of Americans could experience some problems with debt, even when the economy was not in a recession. We focus primarily on these two groups.
Paralleling our analysis in the last section, we first report on the traits of these different groups in univariate terms (Table 10) and then provide a multinomial logit analysis of debt loads (Table 11).
Looking at Table 10, one can see that relative to those comfortable with their level of debt, the overburdened are younger, and have lower financial assets and incomes. Note that they are disproportionately drawn from the overextended cluster, while almost none are part of the in-charge segment. In terms of debt literacy, the overburdened rank themselves the lowest of the four groups, although their actual level of debt literacy (as measured by percentage correct) was only somewhat lower than those who considered their debt levels to be about right.
The ―unsure,‖ the 11% who were unable to judge whether they had too much or too little debt, tended to be disproportionately female (nearly 70%), African-American (18%), and unmarried (60%), the same characteristics displayed by those with low debt literacy. With respect to income, they are disproportionately drawn from the lowest income group (59% making under $30,000 per year), and have considerably less wealth than the 60% who categorized their debt load as ―about right.‖ With respect to financial literacy, their debt literacy is considerably weaker than those who judged their debt to be either about right or even too high.
People in this group also were more likely to select ―do not know‖ as the answer to the debt literacy questions than were the other two groups. This group is disproportionately drawn from the ―fringe‖ segment.
We perform a multinomial logit analysis of the three groups mentioned above: the overburdened, the unsure, and those with the right amount of debt. As predictors for these debt outcomes, we use demographic variables including age and age squared, and dummies for gender, marital status, race, family size, employment status and income and wealth. Moreover, we add dummies for the different measures of financial literacy.