The Impact of College Financial Aid Rules on Household Portfolio Choice

Transcription

1 The Impact of College Financial Aid Rules on Household Portfolio Choice The Impact of College Financial Aid Rules on Household Portfolio Choice Abstract Using the 2001 Survey of Consumer Finances, this study investigates the effect of the college financial aid rules on household portfolio choices. The federal algorithm used to compute financial aid does not take into consideration assets accumulated in retirement accounts or as home equity, and households can reduce their implicit marginal financial aid tax rates by moving funds into retirement accounts or by increasing their home equity. Our results show that households who have higher marginal financial aid tax rates have a larger fraction of their portfolios invested in retirement assets and home equity, relative to taxable assets. Patryk Babiarz Department of Consumer Sciences, University of Alabama, Tuscaloosa, AL and Tansel Yilmazer Department of Personal Financial Planning, University of Missouri, Columbia, MO National Tax Journal Vol. LXIl, No. 4 December 2009 I. INTRODUCTION high proportion of students enrolled in post-secondary A institutions receive financial aid from the federal government. The U.S. Department of Education reports that over $4.2 billion were spent on need-based financial aid in the academic year, and need-based financial aid represented about 75 percent of all support dollars spent on post-secondary education (U.S. Department of Education, 2005). A high proportion of U.S. students rely on financial aid. Almost 34 percent of full-time undergraduate students and over 23 percent of part-time undergraduate students received non-returnable grants from the federal government in the academic year. Among the beneficiaries of the financial aid system, full-time undergraduate students received on average more than $7,300 in federal financial support, $3,247 of which was in the form of non-returnable grants (U.S. Department of Education, 2005). Even among full-time students from households with yearly incomes over $100,000, average financial support from the government amounted to $7,263 and $1,659 of this sum were non-returnable grants. The irony of the need-based college financial aid system is that it imposes an implicit tax on household assets and income, and thus penalizes households who save in advance for their child s college expenses. Consequently, the federal algorithm used to compute financial aid eligibility generates incentives for households to reduce their savings in order to increase the amount of financial federal aid that their children receive through federal grants or loans. At the same time, the federal algorithm does not take into 635

2 NATIONAL TAX JOURNAL consideration the assets accumulated in retirement accounts or as home equity, which makes these assets exempt from implicit financial aid taxation. Increasing contributions to retirement accounts and reducing mortgage debt are often advised by Internet resources on how to fund a college education. For example, maximize contributions to your retirement fund or prepay your mortgage are two of the tips offered by the FinAid website. 1 CNNMoney.com also recommends that parents should maximize retirement savings and watch their debt load in order to maximize their student s financial aid eligibility (Seid, 2005). While the previous literature investigates the effect of means-tested college financial aid on household asset accumulation (Edlin, 1993; Feldstein, 1995; Dick and Edlin, 1997; Long, 2004; Monks, 2004; Ma, 2005; Reyes, 2008), little attention has been given to the portfolio choices of households that expect to have a child attending college in the near future. Households may mitigate the implicit taxation of financial aid by moving savings into retirement accounts or increasing home equity. Using data from the 2001 Survey of Consumer Finances, this study investigates the effect of financial aid rules on household portfolio choices. Since investing in retirement accounts and having mortgage debt have federal income tax benefits, our empirical analysis also controls for marginal income tax rates. This approach enables us to disentangle the relationships between households portfolio choices, the implicit financial aid tax, and the personal income tax. Our findings verify concerns that the methodology for computing financial aid eligibility distorts households composition of assets. We find that retirement assets compared to other financial assets and home equity compared to taxable assets increase significantly with the marginal financial aid tax rate. At the same time, our findings show that the amounts of taxable financial assets or total taxable assets are not significantly influenced by the marginal financial aid tax rate. We conclude that financial aid computation rules provide households with incentives to adjust their asset portfolios, and therefore deserve more attention from policymakers. In order for the distribution of need-based financial aid to satisfy horizontal equity among households that have the same amount of wealth, the assessment of needs should not be affected by households abilities to deliberately or inadvertently conceal their taxable assets. The remainder of the paper is structured as follows. Section II reviews the previous literature on the impact of the implicit financial aid tax on household assets. Section III provides a detailed explanation of marginal financial aid tax rate computations and summary statistics of the data. Section IV presents estimates of the impact of financial aid rules on portfolio decisions and household savings. Finally, section V summarizes our conclusions and discusses the implications of our findings. II. THE IMPACT OF COLLEGE FINANCIAL AID RULES ON HOUSEHOLD PORTFOLIO DECISIONS AND SAVINGS The federal government is primarily responsible for assessing eligibility for need-based college financial aid from the government. Financial need for each college student is assessed as the difference between the cost of college attendance and the Expected Family Contribution (EFC). The EFC is the amount that the student and her mily are expected to contribute toward educational expenses during each 1 See Maximizing Your Aid Eligibility, The SmartStudent Guide to Financial Aid, at fsa/maximize.phtml. 636

3 The Impact of College Financial Aid Rules on Household Portfolio Choice year the student attends college. The U.S. Department of Education calculates a student s EFC using a methodology based on the information the student provides on the Free Application for Federal Student Aid (FAFSA). The larger the EFC, the smaller the assessment of student s financial need and the amount of aid the student is eligible to obtain. The federal algorithm used to compute financial aid can have adverse effects on household asset accumulation. The marginal financial aid tax rate on assets is the expected change in the amount of financial aid received by the student that results from a one dollar increase in parents current assets. Early studies estimate a large effect of the implicit financial aid tax on household savings (Edlin, 1993; Feldstein, 1995; Dick and Edlin, 1997). 2 Edlin (1993) is the first to illustrate that although the maximum financial aid tax on parental assets is only 5.64 percent in any given year, milies with only one child in college at a time for eight consecutive years may ce a cumulative tax on assets of over 57 percent. In a subsequent study, Dick and Edlin (1997) show that a typical mily with a child at an average-priced college may lose over $2,000 in financial aid as a result of increasing their assets by $10,000. Feldstein (1995) provides additional evidence that the implicit financial aid taxes on assets have a significant adverse effect on the accumulation of financial wealth. Using data from the 1986 Survey of Consumer Finances (SCF), Feldstein (1995) concludes that the assets of an average mily can be taxed at a rate approaching 50 percent over the entire period of only one child s college education. Long (2004) and Monks (2004) question the magnitude of the effect of the marginal financial aid tax rate on parental assets estimated by the previous literature. Long (2004) shows that the effect of financial aid tax on parents assets is sensitive to underlying assumptions used in the computation of the financial aid tax. The magnitude of the effect of the financial aid tax on parents assets varies substantially across models with distinct sets of assumptions regarding the probability of college attendance, college costs, future mily income, contribution from students, and eligibility for simplified needs evaluation. 3 Monks (2004) replicates the empirical analysis in Edlin (1993) and Feldstein (1995), using more recent data collected by the 1997 National Longitudinal Survey of Youth. Monks focuses on the homogenous sample of milies with pre-college aged children and shows that the magnitude of the effect of the implicit financial aid tax on asset accumulation is sensitive to the assumptions utilized in the previous two studies. The effect of the implicit financial aid tax on assets may be more complex than a simple decrease in total savings, because retirement assets and home equity are exempt from the calculations of financial aid and thus nontaxable under the federal methodology. The federal methodology thus creates incentives for milies to adjust their portfolios in order to make their children eligible to receive financial support or to increase the value of their financial aid packages. The previous literature provides evidence that taxes affect the portfolio choices of households. Differences in tax 2 In addition to the financial aid tax rate, income, parents marital status and the number of siblings can affect the amount of parent s savings and financial support for each child s college expenses. For example, Yilmazer (2008) introduces life-cycle savings into a quality and quantity model of fertility and shows that parents support for each of their children s college expenses decreases with the number of children. 3 Federal methodology allows some milies to be evaluated by the Simplified Needs Test. Under this procedure, households with adjusted gross incomes of less than $50,000 who have filed, or are eligible to file, an IRS Form 1040A or 1040EZ or are not required to file an income tax return and are exempt from any contributions from assets. For these households, the marginal financial aid tax rate on assets is zero. 637

4 NATIONAL TAX JOURNAL rates on different assets alter after-tax returns on investments. Theoretical studies of the relationship between taxation and portfolio choices generally conclude that optimal portfolio decisions imply holding relatively highly-taxed assets in tax-protected accounts or shifting highlytaxed assets to tax-privileged assets (Tepper, 1981; Auerbach and King, 1983; Shoven and Sialm, 2004; Dammon, Spatt, and Zhang, 2004). Taxes affect not only the amounts invested in different assets but also decisions regarding which assets to hold (Leape, 1987). The complexity of the impact of taxation on household portfolios has generated a substantial amount of empirical research. For example, using the U.S. President s Commission on Pension Policy data, Hubbard (1985) estimates the significant impact that income tax rates and mandatory participation in the Social Security pension system have on the allocation of assets in households non-pension wealth. Using data from the Survey of Consumer Finances, Poterba and Samwick (2003) investigate whether milies that ce higher income tax rates invest more in tax-deferred accounts or tax advantaged assets and conclude that household income tax rates display a substantial positive correlation with ownership of tax-exempt bonds and with the share of the portfolio that is held in tax-deferred accounts and invested in corporate stock. Scholz (1994) and Maki (2001) use the exogenous incentives introduced by the Tax Reform Act of 1986 to document that some milies restructure household debt into tax-vored mortgages in response to tax policy. A tradeoff between mortgage prepayments and tax-deferred retirement savings is discussed by Amromin, Huang and Sialm (2007), who show that tax arbitrage by reducing mortgage payments and increasing contributions to tax-deferred retirement account is possible. However, their empirical investigation ils to document that a significant percentage 638 of households takes advantage of this tax arbitrage opportunity. III. DATA SUMMARY The empirical analysis uses data from the 2001 Survey of Consumer Finances (SCF). The SCF is a comprehensive triennial survey of U.S. household finances conducted by the Federal Reserve Board. It contains detailed information on household assets, liabilities, income, and demographic characteristics. The 2001 SCF is the most recent cross section of the survey that identifies the ages of children living in the household. The SCF uses multiple techniques of imputation to compensate for missing data, which results in five replicates of data for all households. In this study, we use all five replicates and report aggregated coefficient estimates and standard errors. Our methodology for aggregating the coefficient estimates and standard errors is described in Section III.A. A. Estimation of the Marginal Financial Aid Tax Rate The implicit financial aid tax on assets results from the reduction of federal financial aid for households that have relatively large holdings of taxable assets. The Marginal Financial Aid Tax Rate ( ) is a hypothetical measure of the overall tax imposed on an additional dollar of parents assets. It cumulates the effects of the annual tax rates that households expect to ce for each year of their children s college attendance. depends on the mily s future income during the years each child is attending college, the number of children enrolled in college each year, the amount of the reduction in financial aid given a marginal increase in EFC, and the costs of the colleges in which the children are enrolled. To gather the effects of tax rates from all the years in which children attend college,

5 The Impact of College Financial Aid Rules on Household Portfolio Choice we compute the marginal financial aid tax rate ( ) according to the following formula: (1) MTR ( MTR i ) i= 2001 where MTR i represents the annual evaluation of the implicit college financial aid tax rate in year i. If the mily does not expect to have a child in college in any particular year, their evaluation of the tax rate in this particular year will be zero and the multiplicative term (1 MTR i ) will be 1, and thus will not alter the overall. For example, a mily with two children in 2001, 16 and 10 years old, both of whom will attend college, will ce an overall implicit financial aid tax rate of = 1 (1 MTR 2003 ) * (1 MTR 2004 ) * (1 MTR 2005 ) * (1 MTR 2006 ) * (1 MTR 2009 ) * (1 MTR 2010 ) * (1 MTR 2011 ) * (1 MTR 2012 ). Under the federal financial aid formula, 12 percent of parents discretionary net worth is deemed the expected contribution from assets. Parents discretionary net worth is defined as net worth minus the sum of educational savings and an asset protection allowance. 4 Assets in retirement accounts and home equity are not included in parents net worth. In addition, to contributions from parents assets, contributions from income are also required, and are computed as total income minus tax liabilities, an allowance for Social Security taxes, an income protection allowance, and an employment expense allowance. 5, 6 Contributions from parents assets and income are added to yield the Adjusted Available Income (AAI), which is then used to compute the EFC. The percentage of AAI that is included in EFC increases with the amount of AAI. Families in the lowest bracket of AAI may have a zero EFC, which means that they do not ce any financial aid tax. Families in the upper bracket of AAI are expected to contribute 47 cents of each additional dollar of discretionary net worth. Since exactly 12 percent of discretionary net worth is included in AAI, the maximum marginal tax rate that a household can ce in any given year of a child s college attendance is 0.12 times 0.47, which results in an annual financial aid tax rate of 5.64 percent. 7 This implicit tax is reapplied every year while the student is enrolled in college. Thus, the cumulative multiple-year effect of the tax can be much larger than the maximum 5.64 percent annual rate. 4 Net worth includes cash, savings, checking accounts, net worth of investments and adjusted net worth of businesses and rms. The educational savings and asset protection allowance depends on the age of the older parent. 5 Dividends and other capital gains that are included in total income may also increase the effect of the implicit financial aid tax on portfolio allocations. Dependent students are required to report, among other things, their parents Adjusted Gross Income from their federal tax returns. Thus, dividends and other capital gains are included in the calculation of need assessment. Although this additional income is included in our main proxy for the implicit financial aid tax rate, we do not attempt to quantify separately the importance of capital gains on the magnitude of the financial aid tax. For milies that have large capital gains, the implicit tax on assets can be significantly higher than it is for otherwise identical milies with small capital gains. 6 The income protection allowance depends on the number of mily members and how many of them are college students. For example, for a mily of four with two children in college, the income protection allowance was $17,440 in For milies with two working parents, the employment expense allowance was the smaller of $2,900 or 35 percent of the earned income of the lower-earning spouse. Two-parent milies in which only one parent works do not receive an employment expense allowance. 7 We are interested in the impact of the marginal financial aid tax rates on parents assets. Assets held in the student s name reduce financial need and aid eligibility to a much greater extent than assets held under the parents name. Up to 5.64 percent of parents assets and up to 35 percent of the student s assets are included in the calculations of a dependent student s EFC (Information for Financial Aid Professionals, 2000). Following the bulk of the previous literature, we ignore the student s assets due to the lack of available data. 639

6 NATIONAL TAX JOURNAL We do not observe children s actual college attendance in the SCF. Instead, we predict the probability that each child in our sample who is under age 21 will attend college when s/he reaches college age as follows. First, we model the probability of college attendance as a function of the student s gender, ther s race and predicted mily income when the student reaches age 18, obtaining coefficient estimates βˆ using data from the October Current Population Survey for the years Second, we utilize the estimated coefficients to predict whether each child in our sample would attend college when s/he reaches age Specifically, we predict the probability that a child will attend college as Φ ( xβˆ + ε), where ε is an error term drawn at random from a normal distribution that has a mean value of zero and standard deviation of one. We create a (0/1) indicator variable by assuming that the child attends (does not attend) college for the next consecutive four years if the estimated probability of attendance exceeds (is less than) 50 percent. We add an error term to our prediction of college attendance because, without an error term, the calculations of the indicator variable that measures college attendance would be too closely related to the variables used in the prediction. As a result, for example, most of the children from high income households would be classified as college students, and most of the children from low income households would be classified as non-students. We acknowledge that using this method to create an indicator variable for college attendance and using this variable in the calculation of adds error to the measurement of our variable of interest. Therefore, we create 100 values for this indicator variable by imputing each child s college attendance 100 times using error terms drawn from the standard normal distribution. Note that the SCF already includes five replicates of the data, so we practically impute the data 500 times. For each imputation, we recalculate EFC and since EFC depends on the number of children attending college each year, and is equal to zero for each year that the household does not have any children attending college. We follow the methodology in Reiter (2003) to compute the coefficient estimates and standard errors in our models. In terms of the summary statistics below, we report the averages across 500 imputations. The standard errors of the coefficient estimates are computed by averaging the variances of the imputations and adding the variance of the estimated coefficients divided by the number of imputations. We report the square root of this number for each coefficient as the standard error. Our method produces probabilities of college attendance ranging from and percent across 500 iterations when the children in our sample are 18-years old. Our predictions of college attendance decisions are consistent with the percentage of 18-year old individuals who attend college reported in the October Current Population Survey. For instance, the percentage of 18-year olds that attend college in the Current Population Surveys is 42.0 percent and the percentage of 18-year olds that are predicted to attend college in our sample is on average 55.5 percent. The difference between these percentages can be explained by the oversampling of highincome households in the SCF. Using the ordinary least squares method to estimate the correlation between 8 We predict the mily income used as a covariate in the prediction of college attendance as follows. We fit a simple age-income curve to the data drawn from the 2001 SCF: log(y i ) = α + β 1 t + β 2 t 2 + β 3 t3, where Y = the household income in 2001, t = age of the ther minus 15 if the ther is not a high school graduate, 22 if the ther is a college graduate, and 18 otherwise. Details of both the probability of college attendance and future income computations are available from the authors upon request. 640

7 The Impact of College Financial Aid Rules on Household Portfolio Choice and portfolio choice is problematic, since the implicit tax variable depends on the level of taxable assets. For instance, a decrease in the taste for asset accumulation decreases and thus introduces a correlation between and the error term. To prevent this potential endogenity problem, we first predict parents net worth based on the characteristics of the household head and spouse. 9 We then use the predicted values of net worth to compute the contribution of assets to the EFCs calculated for each year between when milies in our sample expect to have a child in college. This method creates a good instrument for that is uncorrelated with the error term. Note that this method is slightly better than the traditional instrumental variable estimation procedure since the resulting value of reflects all of the information about future income, the number and the age of children, and the age of the parents to calculate the allowances from assets, college expenses and financial aid. Note that this variable is a rough estimate of the level of net worth that is not influenced by the taste for asset accumulation. To compute the contributions to EFC from mily income, we estimate the future path of income and wages. We use the same fitted age-income curve utilized in the estimation of the probability of college attendance. The prediction of mily income for each year between when milies in our sample expect to have a child in college was explained previously. Although the need for financial assistance under the federal methodology is defined as the difference between the cost of college and the EFC, in reality the full amount of financial aid needed is rarely funded. This gapping between the assessed and the actually financed need results from the scarcity of aid funds, which are distributed on a first-come, firstserved basis, or by distributing the total amount of assistance available among the beneficiaries. We do not possess information on the actual amounts or the form of financial aid received by the households in our sample. We utilize the coefficient estimates of (3) in Long (2004) to estimate the amount of financial aid that each mily in our sample receives when their child attends college. 10 The parameters are estimated for the dependent students in the restricted-use version of the National Postsecondary Student Aid Study (NPSAS). We predict the value of financial aid as a function of EFC, college cost, EFC-squared, college cost squared, and a number of interaction terms between the independent variables. We first predict the cost of college for each child using data from the NPSAS. 11 We next obtain MTR i the annual evaluation of the financial aid tax by calculating the decrease in finan- 9 The regression estimates are available from the authors upon request. The model fit statistic R 2 is equal to and the median value of predicted net worth is $56,759. We predict net worth as the exponent of the fitted value of the semi-logarithmic model. However, the exponent of the fitted value of the semi-logarithmic models tends to underestimate the dependent variables (Wooldridge, 2003). We check the robustness of our estimates by computing the predicted value of the net worth using the consistent estimator of the fitted value described in Wooldridge (2003). Our results are robust to use of this estimator of net worth in the calculation of EFC. 10 We thank Mark Long for providing us the coefficient estimates and details of the equation. 11 We model the cost of college attendance as a function of student s age, student s gender, ther s race, ther s education, mother s education, the size of the household, and mily income. To account for the trend in college cost, we increment the predicted college cost by 5 percent every year, which corresponds to the average annual increase in college expenses in 1990s. The average predicted cost of college per household who has at least one child in college in 2001 amounts to $14,660, with the median value equal to $14,195. Details of the computations are available from the authors upon request. 641

8 NATIONAL TAX JOURNAL cial aid when taxable assets are increased by one dollar. The average value of financial aid for milies in our sample who have a child in college in 2001 is $4,011 per household, while the median value is $3,243. It should also be noted that milies might assign different values to the different forms of financial assistance that they receive. Non-returnable grants are clearly valued more than loans. The value of a student loan can be described as the value of the subsidy supplied by the federal government, which can reflect interest rate caps, deferment of payments, and exemption from interest while the student is enrolled in college. We follow the previous literature (Case and McPherson, 1986; Edlin, 1993; Dick and Edlin, 1997; Long, 2004), which assumes that loans have a value of 50 cents per dollar of loan. 12 Our approach takes into account the unmet needs and varying valuation of types of financial aid received, since the composite dependent variable is the sum of 100 percent of aid received in the form of grants and 50 percent of aid received in the form of loans. In the final step, we calculate by compounding MTR i for years in which milies expect to have a child in college, as specified in (1). In addition to not being included in the calculation of EFC, investments in retirement assets and holding a home mortgage have federal income tax advantages. The amounts contributed to retirement accounts are typically not subject to income tax until withdrawn from the account. Also, interest payments on mortgage loans are deductible against federal income tax liability for households that itemize tax deductions. Therefore, federal income tax considerations may also provide households with incentives to restructure their assets similar to those of the implicit financial aid tax. In order to rule out the possibility that proxies for the effect of marginal income tax on household portfolio choices, we include estimated Marginal Income Tax Rates (MTR in ) as explanatory variables in our empirical analysis. 13 B. Summary Statistics The 2001 SCF includes 4,442 households. Several sample restrictions are imposed on the data. Specifically, 1,362 single-headed households, 1,514 households with no children younger than 25 or households with only independent children (living elsewhere), and finally 74 households with a retired spouse or partner are excluded from the sample. These exclusions are made because the marginal financial aid tax rate should not affect the saving behavior of households without (dependent) children, and households that include retired individuals may already have started liquidating their retirement assets. Additionally, 159 households that reported negative income, negative home equity, and/or households with total financial assets below $100 are excluded. 12 Feldstein (1995) assumed a higher value for loans, equivalent to 60 cents per dollar. Since we use the coefficient estimates in Long (2004) to predict financial aid, we followed his assumption that loans have a value of 50 cents per dollar of loan. 13 Household tax liabilities are not reported in the 2001 SCF. We calculate MTR using detailed information on in household income and demographic characteristics. We calculate the income tax rates using only income from wages and salaries and ignore all investment income from other sources such as taxable interest, dividends, capital gains, tax-exempt interest earnings, etc. to prevent the bias resulting from the ct that MTR may in itself be affected by portfolio choices. We subtract allowances for dependents from adjusted gross income and assume that all households take advantage of standard deductions. In 2001, only 34.2 percent of tax return files were accompanied by Schedule A for itemizing tax deductions (U.S. Department of the Treasury, 2002). Almost 28.3 percent of the total number of tax filers in 2001 itemized interest payments, which include interest payments on home mortgage loans. 642

9 The Impact of College Financial Aid Rules on Household Portfolio Choice The working sample comprises 1,333 households. 14 In the subsequent analysis of the effect of on households composition of assets, we estimate models that include retirement assets and home equity as dependent variables. We also investigate the effect of on mortgage payments. Retirement assets include IRAs, Keogh, and any other retirement accounts provided by a current employer including tax-vored thrift and savings plans, 401(k) and 403(b) plans, Supplemental Retirement Annuities, etc. Home equity is defined as the market value of a primary residence less outstanding mortgage debt. We are also interested in the effect of on retirement assets or home equity relative to taxable financial assets, total taxable assets or total financial assets. Taxable financial assets include cash, checking, saving, money market or call accounts, certificates of deposit, stocks, bonds, mutual funds and annuities. Retirement assets and the cash value of life insurance are not included in taxable financial assets since both of these assets are not included in the calculation of EFC. Total financial assets include taxable financial assets, retirement assets and cash value of life insurance. Total taxable assets include taxable financial assets, the net value of investments in real estate other than the primary residence, and the adjusted value of business interests. Finally, total assets include total financial assets and nonfinancial assets. Summary statistics for the variables used in this study are presented in Table 1. The 2001 SCF identifies households that expect major financial expenditures for children s education within the next 5-10 years and households that save for these expenses. Households that expect expenditures for children s education constitute about half of our sample. Out of 688 households that expect expenditures for children s education, 169 report that they do not save and 519 report that they save for these expenditures. There are significant differences in financial measures across these groups. For example, households that expect expenditures for children s education but do not save for them have lower values of income, retirement assets, home equity, and total assets than households that save for these expenditures. Interestingly, however, financial characteristics are not very different between those who do not expect educational expenditures and those who do not save for these expenditures. Figure 1 presents the distribution of by household income in Households in our sample with incomes greater than $221,000 always ce zero because EFC in any given year of their child s college attendance exceeds college costs. In our sample, 20.6 percent of households have non-zero values for. Our calculation of ranges from 0 to , with a mean value of for households that have non-zero. The greatest concentration of nonzero values is among households with income between $20,000 $80,000. Household income influences the implicit tax in several ways. It constitutes a major contribution to EFC, but it also has an impact on our estimate of through the probability of college attendance and the level of college costs. We compare of different income groups to demonstrate how income affects EFC, college attendance and college cost and, thus. For example, among households who only have one child and income below $25,000, the average equals Only 2 percent of these households have positive. The average esti- 14 The sample size and the number of households excluded from the sample are reported for the first replicate of the data. However, the sample sizes of the remaining four replicates are almost the same as the first replicate. 643

10 NATIONAL TAX JOURNAL Age Age in 20s or less Age in 30s Age in 40s Age in 50s or more Black Hispanic Number of children Homeownership Amount of annual mortgage payments Post-college degree Completed college Some college Completed high school No high school Self-employed Managerial or professional occupation Approve to borrow for educational expenses Approve to borrow for living expenses Long-term planning horizon Above average risk tolerance Income MTR in Employer contributions to husband s plan Employer contributions to wife s plan Retirement assets Home equity Taxable financial assets Total taxable assets Total financial assets Total assets Table 1 Descriptive Statistics Mean Median Mean Median Mean Median Mean Median Mean Median Mean Median Expect Expenses for Children s Education = 0 N= , , ,792 8,500 89,046 40,000 85,132 8, ,444 12, ,595 34, , ,900 Expect Expenses for Children s Education = 1 N= * 43.0*** *** *** *** *** ,865*** * *** ,182** *** ,159*** 18,000*** 116,011** 61,000*** 146,115 13,120*** 316,074 19,460*** 259,789* 56,000*** 713,235** 285,200*** Expect Expenses for Children s Education = 1 Save for Expected Expenses for Children s Education = 0 N= , , ,235 4,900 75,490 36, ,109 8, ,009 5, ,742 39, , ,500 Save for Expected Expenses for Children s Education = 1 N= c , b a a c a a a c 130,846 c a b ,077 a 29,600 a 134,743 a 77,400 a 167,273 18,130 a 390,531 27,100 a 305,113 80,600 a 841,947 b 335,730 a Notes: Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels (for the comparison of Expect Expenses for Child s Education = 0 to Expect Expenses for Child s Education = 1), 1% ( ), 5% ( ), and 10% ( ) levels (for the comparison of Save for Expected Expenses for Child s Education = 0 to Expect Expenses for Child s Education = 1), and 1% (a), 5% (b), and 10% (c) levels (for the comparison of Save for Expected Expenses for Child s Education = 0 to Save for Expected Expenses for Child s Education = 1). Data are taken from the 2001 Survey of Consumer Finances. Descriptive statistics are weighted using the sampling weights. Tests of medians are based on Mann-Whitney non-parametric test procedure. mated probability of college attendance is and the average estimated cost of college is $11,988 if the child attends college. Among one-child households with income between $25,001 $50,000, the average is , and 34 percent of these households have a positive. The probability of college attendance in this group is estimated to be , and the average cost of attendance is estimated to be $12,482 if the child attends college. Households with one child and income between $50,001 $75,000 have an average of Over 72 percent of these households ce a positive. 644

11 The Impact of College Financial Aid Rules on Household Portfolio Choice Figure 1 as a Function of Household Income in more Household income in 2001 ($ Thousands) Note: The ends of the whiskers represent maximum and minimum values. The bottom and top of the boxes represent the 25th and 75th percentile, and the horizontal line in the middle of the boxes represents the 50th percentile. For milies in higher income groups, the probability of college attendance and costs grow steadily, while and the percentage of milies with positive tend to decline. One-child milies in the $100, ,000 income group are characterized by an average of , but only 29 percent of these milies ce positive implicit taxes. All one-child households in our sample with income exceeding $200,000 ce a zero due to the ct that their EFC exceeds their estimated costs of college attendance. Figure 1 and Table 2 demonstrate that several ctors drive the variation in within the income groups. These ctors include household composition (the number of children attending college, the age difference between children, the number of household members, etc.) and household savings. For example, one-child households with income between $50,001 $100,000 ce an average of , while similar two-children households have an average of Additional information on the distribution of by household income and taxable financial assets is provided in Table 2. At lower levels of household income and taxable financial assets, is an increasing function of both income and assets. However, when the increased contributions from income or assets raise EFC over the cost of college, becomes zero, which results in a regressive trend in the average for households in the upper tails of the income and taxable financial assets distri- 645

12 NATIONAL TAX JOURNAL Income in 2001 $0 $25,000 $25,001 $50,000 $50,001 $100,000 >$100, Table 2 by Household Income and Taxable Financial Assets in 2001 Taxable Financial Assets in 2001 $0 $10,000 $10,001 $20,000 $20,001 $40,000 $40,001 $80,000 $80,001 $160,000 $160,000 and more All Percent of Sample Percent of Sample All Source: Authors computations based on the 2001 Survey of Consumer Finances. Percent of Sample Percent of Sample Percent of Sample Percent of Sample Percent of Sample

13 The Impact of College Financial Aid Rules on Household Portfolio Choice bution. 15 For example, among the income group between $25,001 $50,000, equals for households with taxable financial assets lower than $10,000, then increases steadily, reaching an average of for households with taxable assets between $20,001 $40,000, and then lls to almost zero for households with taxable financial assets more than $160,000. IV. RESULTS A. Retirement Assets Table 3 reports the parameter estimates of the Tobit regressions for retirement assets. In Model I, we use the ratio of retirement assets to taxable financial assets as the dependent variable. The coefficient of is positive but not significant. In Model II, we estimate an analogous equation with the log of retirement assets as the dependent variable. This specification supports a significant effect of on retirement assets. On average, each additional percentage point increase in increases retirement assets by 9.45 percent, other things being constant. 16 We next focus our attention on households that do not yet have a child in college but that might be preparing for college years. We limit the sample to households with only dependent children younger than 18 (Model III) and regress the ratio of retirement assets to taxable financial assets on the same set of independent variables. The coefficient of is positive but again not significant. Finally, we regress the share of retirement assets in total financial assets using our initial sample and the set of control variables (Model IV). This specification reveals that a one percentage point increase in the implicit financial aid tax rate is associated with an average increase of 58 basis points in the share of retirement assets in total financial assets. At the sample means, a household with equal to 4 percent would hold about 32.6 percent of financial assets in retirement savings, while the identical household with a equal to 3 percent would have a share of retirement assets equal to about 32.0 percent of the financial assets. 17 The magnitude of the increase in the share of retirement assets in total financial assets with an increase in is small. It would be interesting to analyze the relationship between and contributions to retirement assets. Unfortunately, the SCF does not include information on contributions to IRAs or 401(k) plans. The only relevant information on contributions to retirement accounts is on whether the household head and/or wife contribute to an employer-sponsored retirement account and the amount of contributions for those that report that they contribute. The number of households that report that they contribute to an employer-sponsored retirement plan (10.2 percent of husbands and 9.2 percent of wives) is too small for a robust empirical analysis. We did not find any significant relationship between 15 Of course, calling regressive is not totally accurate since households in the upper tails of the income and asset distribution do not usually receive financial aid. 16 To interpret the coefficient estimates of the Tobit regressions, we multiply the estimates by the adjustment ctor given by Φ(x i βˆ/ˆσ), where x i βˆ/ˆσ = (βˆ0 + βˆ1 x βˆk x k )/ˆσ, ˆσ is the estimated standard deviation of the error term, and Φ(.) denotes the standard normal cumulative distribution function. When we evaluate the adjustment ctor at the sample means for Model II in Table 3, Φ(x i βˆ/ˆσ) = The estimated coefficient of is 9.82, and multiplication of the estimate with the adjustment ctor evaluated at the sample means is We compute the difference between Tobit fitted values evaluated at = 0.03 and = 0.04, respectively, with values of other explanatory variables fixed at sample means. The fitted value of the variable explained by Tobit model is given by ŷ i = Φ(x i βˆ/ˆσ)x i βˆ + ˆσφ(x i βˆ/ˆσ), where x i βˆ/ˆσ = (βˆ0 + βˆ1 x βˆk x k )/ˆσ, ˆσ is the estimated standard deviation of the error term, and Φ(.) and φ(.) denote standard normal cumulative distribution function and standard normal probability density function, respectively. 647

14 NATIONAL TAX JOURNAL Dependent Variable: Sample: Sample Size: Intercept Expect expenses for children s education Age in 30s Age in 40s Age in 50s or more Black Hispanic Number of children Homeownership Post-college degree Completed college Some college Self-employed Managerial or professional occupation Approve to borrow for educational expenses Approve to borrow for living expenses Long term planning horizon Above average risk tolerance log(income) Income below 33th percentile * log(income) Income above 66th percentile * log(income) MTR in Employer contributions to husband s plan Employer contribution to wife s plan Sigma Log likelihood Proportion of + obs. Mean value of the dependent variable (positive values) Table 3 Tobit s for Retirement Assets Model I Model II Model III Model IV Retirement Assets / Taxable Financial Assets log(retirement Assets) , Notes: Asterisks denote significance at the 1% (***), 5% (**), and 10% (*) levels. Retirement Assets / Taxable Financial Assets Retirement Assets / Total Financial Assets Number of Families with Children Full Sample Full Sample over 18 or Independent Full Sample N=1,333 N=1,333 N=620 N=1, * * *** ** * *** ** *** , ** ** ** *** *** *** *** *** *** *** ** ** *** * *** *** *** *** , * * *** * * ** ** * *** *** *** ** ** ** *** *** *** * *** * *** *** *** *** 648

15 The Impact of College Financial Aid Rules on Household Portfolio Choice contributions to employer-sponsored retirement plans and. Consistent with the previous literature on income tax and portfolio choice (Poterba and Samwick, 2003), MTR in seems to encourage households to increase savings in tax-deferred retirement accounts. Regressing retirement savings on our set of explanatory variables (Model II) shows that a one-percentage point increase in the income tax rate results in an increase in retirement savings of 10.6 percent. In reality, the marginal income tax rate has a discrete distribution, so a more realistic interpretation is that an average household in the 28 percent income tax bracket accumulates over 3 times more retirement savings than an identical household in the 15 percent income tax bracket. We control for a non-monotonic relationship between retirement assets and household income by including nonlinear income terms. Specially, we interact log of income with two dummy variables that represent the households that have income below the 33 rd percentile and households that have income above the 66 th percentile. The estimated coefficients in Table 3 show that household income has a negative impact on the ratio of retirement assets to total financial assets for the base sample (Model IV) and a larger negative impact for households in the 33 rd percentile of income distribution. Retirement savings, regardless of being scaled by taxable assets or by total financial assets, are higher when an employer makes a contribution to the husband s (Models I-IV) or to the wife s (Models I, II and IV) retirement plan. In addition, Hispanic households have lower levels of retirement assets (Model II). Retirement assets increase with the age and education of the household head and decrease with the number of children (Model II). B. Home Equity The coefficient estimates for Tobit models that investigate the relationship between home equity and are presented in Table 4. has a positive and significant effect on the ratio of home equity to total taxable assets (Model I). A one percentage point increase in results in increase in the ratio of home equity to total taxable assets at the sample means, 18 which implies that a household with $20,000 in total taxable assets and a equal to 4 percent has about $7,361 more in home equity than an identical household with equal to 3 percent. The magnitude of the impact of the on the ratio of home equity to total taxable assets is higher when the sample is limited to milies that have children younger than 18 (Model III). When we regress the log of home equity on our set of independent variables, we observe a positive but insignificant effect of (Model II). The effect of on home equity should primarily be driven by the current or relatively recent financial decisions of households. The value of home equity might reflect the ct that milies have resided in their homes for a long period, which increases the value of home equity because a large part or all of mortgage debt has been repaid. Also, using home equity as the dependent variable makes the estimation more vulnerable to macroeconomic conditions that affect the values of homes for which we are unable to control. To address these problems and to account for the current decisions of households, we regress the log of the 18 The adjustment ctor at the sample means for the first model in Table 4, Φ(x i βˆ/ˆσ), is equal to The estimated coefficient of is 65.12, and multiplication of the estimate with the adjustment ctor evaluated at the sample means is