The Tax Reform Act of 1986 (TRA 86) substantially changed

The Tax Reform Act of 1986 and the Composition of Consumer Debt The Tax Reform Act of 1986 and the Composition of Consumer Debt Abstract - The Tax Reform Act of 1986 (TRA 86) phased out the deductibility of most nonmortgage interest and also introduced new marginal tax rates that reduced the tax advantage of all types of debt. I estimate that by 1991 aggregate mortgage debt was over 1 percent higher, credit card debt approximately 14 percent lower, and auto loan debt approximately 9 percent lower than they would have been without these changes. The results suggest that consumers attempted to minimize the effects of the deductibility phaseout by reallocating debt toward deductible instruments. Anecdotal evidence suggests that consumers used home equity lines to reallocate their debt. Victor Stango Department of Economics, University of Tennessee, Knoxville, TN 37996 National Tax Journal Vol. LII, No. 4 INTRODUCTION The Tax Reform Act of 1986 (TRA 86) substantially changed the tax treatment of consumer debt. The TRA 86 phased out the deductibility of nonmortgage interest payments, while leaving mortgage interest payments fully deductible. This substantially increased after-tax interest rates on itemized credit card and auto loan debt. The TRA 86 also introduced a new marginal tax rate schedule, resulting in lower marginal tax rates for most taxpayers. This increased after-tax interest rates on all types of debt for consumers who itemize and deduct their interest payments and may also have led to higher disposable income for some households. 1 The policy issue raised by the debt-related provisions of the TRA 86 is that the deductibility of interest payments is a significant subsidy for many households. For example, in 1986, a household holding $900 in credit card debt and $5,800 in auto loan debt at prevailing interest rates could claim more than $850 in deductions. 2 At the most cursory level of analysis, this suggests that the deductibility phaseout would reduce by $170 the after-tax income of a typical household paying a 20 percent tax rate. 1 The TRA 86 was intended to be revenue neutral, and it is not clear (even ex post) whether it significantly affected disposable income for the typical household. Hausman and Poterba (1987) argue that the TRA 86 did not reduce marginal tax rates for very many housholds. In the empirical work below, I discuss the sensitivity of the results to different assumptions regarding the effect of the TRA 86 on disposable income. 2 In all likelihood, these numbers substantially understate the true savings of households who actually itemize. The debt figures themselves are median 717

NATIONAL TAX JOURNAL Households could minimize the losses associated with the deductibility phaseout if they adjusted their holdings of debt following the rise in after-tax interest rates caused by the phaseout. They might accomplish this by unilaterally reducing their credit card and auto loan debt or by shuffling existing credit card and auto loan debt into deductible (mortgagebacked) debt. 3 If consumers did not shuffle, mortgage debt levels would be unaffected by the phaseout; if they did shuffle, mortgage debt levels would rise following the phaseout. A complication is that the TRA 86 also reduced marginal tax rates for many taxpayers. A reduction in marginal tax rates leads to higher aftertax interest rates on deductible debt and reduces the demand for all types of debt. The ultimate effect of the deductibility phaseout and marginal tax rate changes on deductible debt is an empirical one. Maki (1996) uses several pieces of evidence to argue that consumers reallocated debt toward deductible types after implementation of the TRA 86. He presents summary statistical evidence showing that during 1987 91, mortgage debt grew faster and consumer debt grew more slowly than they did from 1981 to 1986. Other work by Maki (1995a, 1995b) uses data from the Consumer Expenditure Survey to show that high-income home-owners reduced their consumer interest payments by 36 percent and increased their mortgage interest payments by 16 percent following implementation of the TRA 86. These shifts appear to be due to substitution between types of debt rather than uncorrelated shifts. Further evidence from the 1983 and 1989 Surveys of Consumer Finances corroborates these claims. Maki highlights the policy relevance of his results by noting that two goals of the deductibility phaseout were raising tax revenue and increasing savings and that both of these goals were frustrated due to consumer shuffling of portfolios following the TRA 86. The estimates suggest that the phaseout caused essentially no increase in savings and that shuffling reduced the expected increase in tax revenue by over 40 percent. In related work, Long (1989) uses data in the 1983 IRS Individual Model Tax File to construct ex ante predictions of changes in deductions due to the new tax schedule. He estimates tax and income elasticities of deductions and uses the elasticities to predict changes in deductions following the implementation of the new marginal tax rate schedule. He does not consider the effects of the deductibility phaseout his work appears to predate the introduction of the deductibility phaseout to the TRA 86. His results suggest that the new tax schedule would reduce deductions on both mortgage and personal debt, which includes auto loan, credit card, and other installment debt, because it increased after-tax interest rates. Overall, he predicts 8.5 and 1.5 percent declines in mortgage and personal interest deductions in 1987. 4 Long also briefly mentions a pre TRA 86 debate on the effect of the TRA 86 on home equity lines as a debt instrument and notes that there was no clear consensus regarding how their use would change after 1986. Other work debt levels from the 1989 Survey of Consumer Finances. However, the average debt level of itemizing households is probably much higher than this because high debt levels make households more likely to itemize. The calculation of deductions uses an 18 percent interest rate on credit card debt and a 12 percent interest rate on auto loan debt (these are close to the prevailing levels in 1986 87). The amount of deductions is rτ, where r is the interest rate and τ is the marginal tax rate of the household. I choose an average marginal tax rate of 20 percent in the calculations because it is close to the rates used later in the paper. Appendix A contains these rates and their derivation. 3 A consumer could shuffle debt both by directly substituting one type of debt for another (via home equity lines of credit, for example) and by changing repayment patterns on existing debt. 4 These estimates consider only the tax rate changes of the TRA 86. When the analysis widens to include the effects of the increase in the standard deduction, the numbers become 16 and 8.7 percent, respectively. 718

The Tax Reform Act of 1986 and the Composition of Consumer Debt using cross-sectional data by Skinner and Feenberg (1990) finds evidence of shuffling toward mortgage debt and away from personal debt for some taxpayers, and Scholz (1994) employs summary data from the Surveys of Consumer Finances to argue that the TRA 86 seems to have caused a reallocation of debt. Thus, while there seems to be general consensus that the TRA 86 caused shifts in consumer debt, to date, there have been no ex post estimates of the magnitude of these shifts, 5 nor has there been an explicit attempt to separate the effects of the deductibility phaseout from the effects of the change in marginal tax rates. One difficulty in discerning the effects of the deductibility and tax incentives of the TRA 86 on debt is that debt levels move due to a host of other factors income and pretax interest rates, as well as relative returns on other components of the household asset/liability portfolio. This paper provides estimates of the overall effects of the TRA 86 on debt allocation and attempts to control for the other primary influences on household debt. While Long (1989) and Maki (1995a, 1995b) primarily use cross-sectional data to estimate the effects of the TRA 86, I use aggregate time-series data from 1980 to 1991. This allows me to estimate the effects of the deductibility phaseout over the entire period during which it occurred. This paper also distinguishes the effects of the deductibility phaseout from the effects of the new tax schedule and assesses their relative impacts on borrowing. Based on the results in the paper, it appears that after 1986 consumers adjusted debt holdings away from auto loan and credit card debt and toward mortgage-backed debt. By 1991, mortgage debt was over one percent higher, credit card debt seven ten percent lower, and auto loan debt nine ten percent lower than they would have been if the deductibility phaseout had not 719 occurred. The tax changes appear to have had a much smaller effect on debt. Only credit card debt appears to have been changed by more than one percent by 1991, it was three four percent lower than it would have been without the tax changes. I present some summary survey and statistical evidence suggesting that consumers used home equity lines to weight their debt portfolios toward deductible debt. Three main points emerge from these results. First, it appears that consumers may have sidestepped the subsidy losses implied by the deductibility phaseout without reducing their overall debt holdings. They accomplished this by either (1) substituting mortgage-backed debt for credit card and auto loan debt or (2) repaying mortgage debt relatively more slowly than other types following the TRA 86. Second, work assessing the effects of tax reform on the subsidy to owneroccupied housing needs to consider the role of the deductibility phaseout despite the fact that it only indirectly affected mortgage borrowing. The results presented here show that in fact the change in mortgage debt due to the new tax schedule was negligible compared to change due to the deductibility phaseout. A third point is that previous work suggesting that the new tax schedule made mortgage debt less attractive might have ignored its cross-price effects, which may have mitigated the own-price disincentive to borrow created by higher after-tax interest rates. The next section of this paper summarizes the relevant provisions of the TRA 1986 and its debt-related incentive effects. I then propose a general model of debt holding and discuss the econometric issues involved in estimating such a model, and present estimation results. The penultimate section elaborates on some points related to the main results, and the 5 The work by Long (1989), Maki, and Skinner and Feenberg estimates the effects of the TRA 86 on interest deductions, but not on debt levels.

NATIONAL TAX JOURNAL final section concludes with some remarks on the implications of the results. THE TRA 1986 AND DEBT HOLDING INCENTIVES The Direct Incentives of Taxes and Deductibility The two most notable provisions of the TRA 1986 regarding consumer debt were (1) the phaseout of personal interest payment deductibility from 1987 to 1991 and (2) the reduction in marginal tax rates. 6 Both of these provisions increased aftertax interest rates on deductible debt; if r is the nominal interest rate on debt, τ is the marginal tax rate faced by the household, and δ is the percentage of interest payments that are deductible, the after-tax interest rate on itemized debt is r at = (1 τδ)r. Thus, a fall in either τ or δ increases the after-tax interest rate. A fall in τ also increases disposable income. While the effects of the TRA 86 clearly varied across households, we can assess its aggregate effect on borrowing incentives using the average marginal tax rate. Barro and Sahasakul (1983) propose an average marginal rate that weights different marginal rates by the proportion of adjustable gross income (AGI) taxed at each rate. They measure the rate as [1] τ am = Σ ( y i / Y ) τ i where ( y i /Y ) is the share of AGI in the tax bracket with marginal rate τ i. 7 Using this measurement, the average marginal tax rate fell substantially after the implementation of the TRA 86. During the pre TRA 86 years, it ranges from a low of 27.2 percent in 1984 to a high of 32.5 percent in 1981. In 1987, the average marginal rate fell to 25.1 percent, and by 1991, it was 22.7 percent (see Appendix A for a complete list of the rates from 1980 to 1991 and a comparison to some alternatives). In order to assess the impact of the TRA 86 on debt holding incentives, I use the values for τ am to calculate the average tax advantage to itemization per dollar of debt for mortgage, auto loan, and credit card debt from 1980 to 1991. 8 The first three columns of Table 1 show the results of this calculation. Before the TRA, itemizing credit card debt yielded the greatest savings, because high interest rates lead to a high tax advantage to itemization; in 1986, itemized credit card debt cost $5.11 less (annually) per hundred dollars of debt than nonitemized debt, while the savings to itemizing mortgage debt was only $3.06 per hundred dollars of debt. Of course, in gross terms, the savings to mortgage debt itemization are far greater because mortgage debt is the largest component of debt. Following the TRA 86 the tax advantage of itemizing interest on all types of debt fell because of the new tax schedule. The deductibility phaseout accentuated this 6 The deductibility phaseout allowed 65 percent of interest payments to be deducted in 1987, followed by 40, 20, 10, and 0 percent in the succeeding years. Thus, in the calculations that follow, δ = 0.65 in1987, 0.40 in 1988, 0.20 in 1989, 0.10 in 1990, and 0 in 1991. Technically, the marginal tax rate structure was simplified the number of tax brackets was reduced from ten to three. In practice, marginal tax rates fell for many taxpayers, particularly those in high income brackets. The TRA 86 also increased the standard deduction from $3,540 to $5,000 for married joint filers and from $2,390 to $3,000 for single filers. This change primarily affected lower income households, giving some who had previously itemized an incentive to take the standard deduction. 7 See Appendix A for a detailed discussion of the methodology and a comparison of this measure to some others and Appendix C for a complete list of pretax and after-tax interest rates on mortgage, auto loan, and credit card debt. 8 The tax advantage to itemization is measured as rτδ, the difference between the interest rate on nonitemized debt (i.e., the nominal rate) and the interest rate on itemized debt. For example, in 1986, the nominal interest rate on credit cards averaged 18.26 percent and the average marginal tax rate was 28 percent. Credit card interest payments were fully deductible, so the interest rate on itemized credit card debt was 13.15 percent. The difference between these two rates is 5.11 percent; thus, a household would save $5.11 per $100 of credit card debt by itemizing. 720

The Tax Reform Act of 1986 and the Composition of Consumer Debt TABLE 1 TAX ADVANTAGE TO INTEREST PAYMENT ITEMIZATION BY TYPE OF DEBT AND SPREADS BETWEEN AFTER-TAX INTEREST RATES ON ITEMIZED CONSUMER DEBT AND MORTGAGE DEBT, 1980 91 Average Tax Advantage to Itemization, Dollars per $100 of Debt Interest Rate Spread Between Itemized Personal Debt and Mortgage Rates Year Credit Card Auto Loan Mortgage Credit Card Auto Loan 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 $5.26 5.78 5.46 5.19 5.11 5.13 5.11 2.93 1.66 0.89 0.44 0.00 $4.35 5.38 4.97 3.84 3.73 3.60 3.17 1.70 1.01 0.60 0.28 0.00 Calculations based on data contained in Appendexes A and B. 3.75 4.62 4.27 3.36 3.24 3.06 2.76 2.25 2.04 2.39 2.32 2.05 3.47% 2.40 2.85 4.78 4.99 5.47 6.07 8.30 9.40 9.80 10.35 11.25 1.37% 1.56 1.66 1.27 1.31 1.46 1.08 2.05 3.11 4.14 4.16 4.16 effect for personal interest payments, with striking effects by 1988, for example, the benefit to itemizing credit card debt had fallen from $5.11 to $1.66 per hundred dollars of debt. Of course, 1991 saw the complete elimination of the tax advantage for personal debt. The table also reveals that the mortgage debt itemization advantage fell as well, due to the new tax schedule. Thus, given these own-price effects, we might expect levels of all types of debt to fall between 1987 and 1991. The last two columns of Table 1 broaden the analysis to consider the cross-price effects of the TRA 86 (and the deductibility phaseout in particular) by looking at the spread between credit card and auto loan rates and the mortgage rate (for households who itemize). We can see that although the tax cut reduced the tax advantage to every type of debt, personal debt became far more costly relative to mortgage debt due to the deductibility phaseout. 9 For credit card debt, this margin increased from 6.07 percent in 1986 to 11.25 percent by 1991. Put another way, $100 of credit card debt cost $11.25 more than $100 of mortgage debt for an itemizing household (paying the average marginal rate) in 1991. Auto loan debt became much more costly as well; in fact, between 1986 and 1991, the margin between auto loan debt and mortgage debt nearly tripled. Based on median debt levels, in 1989, a typical household could save $240 per year by replacing its auto loan debt with mortgage-backed debt and $88 per year by replacing its credit card debt with mortgage-backed debt. 10 We can conclude this section with the following observations. The tax reform of the TRA 86 created a powerful set of incentives to change holdings of debt and to allocate debt differently. The changes in after-tax interest rates had both income effects due to the higher cost of all types of debt and substitution effects due to relative price changes between mortgage debt 9 Movements in after-tax spreads may not measure the effects of taxes if pretax spreads changed over the relevant time period, due to changes in default risk or some other factor. In this case, this does not appear to be a problem: the spread between the credit card rate and the mortgage rate was virtually constant between 1986 and 1991 at between eight and nine percent. The auto loan/mortgage rate spread was 2.38 percent at the end of 1986 and 2.41 percent at the end of 1991 but, on average, seems to have risen slightly (100 basis points) following the TRA 86. A more conservative estimate of the effects of the TRA 86 on the auto/mortgage spread would reduce the numbers in the auto loan spread column of Table 1 by 100 basis points. 10 These figures are based on median household debt levels as recorded in the 1989 Survey of Consumer Finances and reported in Kennickell and Shack-Marquez (1992). 721

and personal debt. There may also have been direct income effects for some consumers for whom disposable income rose due to lower tax rates. Of course, a model of debt holding that includes only these effects would be incomplete; a more complete behavioral model of debt holding should consider any other potentially important effects on household portfolio decisions. The next section discusses these effects and outlines their importance with regard to the TRA 86. Other Behavioral Influences on Debt Holding 722 NATIONAL TAX JOURNAL Perhaps equally important as interest rates in determining debt levels are asset prices. Even holding interest rates constant, changes in auto and housing prices could affect debt levels. High asset prices might increase debt levels if wealth effects induce higher borrowing or if demand for the asset is inelastic. Alternatively, high asset prices might reduce current debt levels if demand for the asset is elastic. Poterba (1990) discusses some other housing-related effects of the TRA 86. He notes that many of them (changes in depreciation provisions and capital gains taxes, for example) should primarily affect rental housing markets, but in general equilibrium, these changes could affect the relative user cost of homeownership. His estimates suggest that these changes would reduce the real price of housing following the TRA 86 and that this fall could be as large as ten percent. This highlights the need to control for asset prices in any empirical work investigating the effects of the TRA 86 on mortgage borrowing. Returns on other household portfolio items might also affect the level of household debt. For example, returns on household assets that are not related to consumer debt could affect household debt holding decisions. Engen and Gale (1996) argue that increased 401(k) plan participation may have been funded through higher mortgage borrowing. Similar issues would arise if rapidly growing stock prices induced households to become more leveraged. These factors might increase levels of all types of debt and might change relative debt shares as well. A Note on Itemization Poterba (1990) and others note that the TRA 86 increased the standard deduction and may have changed itemization behavior. Indeed, following the TRA, the percentage of returns itemizing deductions fell, from 39 percent in 1986 to 28 percent in 1989. This drop might conceivably have affected post TRA 86 debt allocation due to the different after-tax interest rates paid by itemizers and nonitemizers. However, an itemization variable (measured as the percentage of returns that itemize interest payments) was insignificant in the regressions presented below, and its inclusion did not change any relevant coefficient estimates. This suggests that the shifts in itemization following the TRA 86 played little role in inducing debt reallocation. This finding may seem puzzling, given that the marginal effect of itemization on after-tax interest rates dwarfs the effects of the deductibility phaseout. However, it is possible that its aggregate effects are small because few households consider itemization to be a choice variable for most households with significant mortgage debt, itemization is a given. Poterba (1990) notes this in his discussion of higher standard deductions and mortgage borrowing, and the pattern of changes in itemization following the TRA 86 fits this story. Data from the Statistics of Income Bulletin show that the number of itemized returns fell from 40.7 million in 1985 to 32.2 million in 1990, a 21 percent fall. However, the number of returns itemizing interest payments fell by only 8.5 percent. Moreover, most of the reduction in item-

The Tax Reform Act of 1986 and the Composition of Consumer Debt ization came from non-debt-related areas the number of returns itemizing medical/dental expenses, taxes paid, and charitable contributions fell by 52.3, 20, and 19.3 percent, respectively. This makes sense considering the relative levels of each type of deduction: interest payments are the largest per return $10,176 in 1985, compared to $575, $3,240, and $1,206 for medical/dental, taxes paid, and contributions, respectively. It appears that households that chose to forgo itemization following the implementation of the TRA 86 were probably those with significant medical/ dental, taxes paid, or contribution levels, but little or no mortgage interest. The changes in behavior of this group probably had little effect on debt aggregates. The next section lays out a model that more precisely estimates the impact of taxes and deductibility on debt composition. It controls for (1) changes in personal income, (2) changes in pretax interest rates, (3) cross-price effects, and (4) asset prices. It also provides separate estimates of the effects of the new tax schedule and the deductibility phaseout. THE MODEL This section develops a model that yields estimates of the effects of the TRA 86 on post TRA 86 debt levels. It combines two techniques: counterfactual analysis and event studies. Event studies measure the effects of an exogenous event (here, the event will be the TRA 86) on the dependent variable in the regression (debt in this case). The approach involves estimating a time-series model that includes intercept dummies for the time of the event. The coefficients on the intercept dummies measure any abnormal changes in the dependent variable during the time period in question, abnormal meaning unaccounted for by the rest of the model. Consider a simple time-series representation of a debt holding equation: [2] DEBT t = β i. X it + γ i. D it + ε t where the debt levels are quarterly from 1980 to 1991, X it is a vector of explanatory variables, and D i is the set of annual dummies for the years 1986 91 (the period of the event). 11 The model essentially computes the β coefficients based on the 1980 85 observations, because the observations with dummy variables equal to 1 i.e., any year from 1986 to 1991 are effectively dropped from the sample. It then tries to fit the 1986 91 data using these coefficients and the X vector. Including the dummies allows the model to fit these observations nearly perfectly, because any deviations between the actual values of the dependent variable and the values predicted by βˆ. i X it will be imputed into the dummy coefficients. 12 Thus, the γˆ i coefficients will measure the deviation of post TRA 86 debt from that predicted by βˆ. i X it. The second part of the technique counterfactual analysis determines the X vector of explanatory variables. Such a vector should include tax and deductibility variables, insofar as they affect after-tax interest rates and disposable income. The interpretation of the dummy coefficients will vary depending on the tax deductibility terms. If the vector of explanatory variables includes the (counterfactual) after-tax and interest rate variables that would have existed if the TRA 86 had not taken place, then the dummy coefficients will represent the deviations of actual post TRA 86 debt from the levels predicted under the 11 I include 1986 in the dummy set to allow for shifts in debt in anticipation of the TRA 86. 12 The model fits the post TRA 86 observations nearly perfectly because the dummies are annual. This restricts the four quarterly deviations of actual debt from predicted debt in each year to be equal (and have the same standard error). In practice, the difference between this approach and using quarterly dummies is likely to be very small. 723

NATIONAL TAX JOURNAL assumption that the TRA 86 had not been implemented. That is, the model will be loan, or credit card) and TOTAL t is the sum of the three types of debt. [3] DEBT t = β i. X it cf + γ i. D it + ε t where the superscript on X cf indicates that the vector of explanatory variables is counterfactual. The coefficients on the dummy variables are estimates of the post TRA 86 shifts in debt. Note that this approach is not very restrictive in that the dummy coefficients may reflect both the influence of the TRA 86 and the influence of other factors on debt holding. For example, the effects of any regulatory or institutional changes that occurred between 1986 and 1991 would be imputed into the dummy coefficients. Choosing the Dependent Variable In order to assess the robustness of the results to alternative specifications, I use two dependent variables to measure debt. The first, DEBT t, is the level of real debt per capita. One criticism of using this measure is that real debt levels may be nonstationary. This is important because nonstationarity of the debt series would positively bias the dummy variable coefficients measuring post TRA 86 changes in borrowing. To test for this, a time dummy was included in the debt equations. The coefficient on this dummy was not statistically significant nor did its inclusion in the model change the dummy coefficients. The second dependent variable gets closer to the issue of debt composition and is also more likely to be stationary. This variable is the share of total debt held in each type of debt, measured as ln SHARE t = ln (DEBT t /TOTAL t ), where DEBT t is the level of debt (mortgage, auto Specifying the Model Following much of the work on mortgage holding, I use a partial adjustment model of debt holding. 13 In this context, the partial adjustment model accounts for liquidity constraints; quite often people cannot change their debt holdings as quickly as they would like. We can think of a standard demand-side debt holding model as one that controls for interest rates, income, other interest rates, taxes, and asset prices. 14 The model thus allows for income effects on debt holding, as well as own- and cross-price effects of interest rate and tax changes. If we take logs, we can write the model as [4] ln DEBT t (or SHARE t ) = β 0 + β 1 ln DEBT t 1 (or SHARE t 1 ) + β 2 ln DISPOSABLE INCOME cf t + β 3 ln MORTGAGE RATE cf t + β 4 ln AUTO RATE cf t + β 5 ln CREDIT CARD RATE cf t + β 6 ln ASSET PRICE t + β j ANNUAL DUMMY jt + ε t where DEBT t is a measure of either mortgage, auto loan, or credit card debt, and the superscripts indicate that the interest rate and disposable income variables are counterfactual. The last control is an asset price variable reflecting the effects 13 The partial adjustment model postulates that consumers desire a value of the dependent variable y*, which is determined by the independent variable X, i.e., y t * = βx t. However, actual balances can only approach desired balances by a factor γ, i.e., y t y t 1 = (1 α) (y t * y t 1 ) + v t. Solving for y t yields y t = β(1 α)x t + αy t 1 + ε t. Mayo (1981) discusses partial adjustment models of mortgage holding. 14 The model also implicitly controls for inflation by using real interest rates, debt levels, and income levels. Using nominal interest rates left the results unchanged. 724

The Tax Reform Act of 1986 and the Composition of Consumer Debt mentioned earlier. The variable is defined as the current real asset price for the relevant type of debt (measured as the purchase price of a house or new car in the mortgage and auto loan equations, respectively). This specification is an appealing way of measuring the effects of the TRA 86 for several reasons. First, the standard errors on the dummy coefficients estimate the statistical significance of the deviations of actual debt from predicted debt. Second, because the model is in logs, the dummy variable coefficients will measure deviations of actual debt from predicted debt in percentages. 15 One potential problem with this specification is that interest rates and asset prices may be endogenous in the debt holding equations, because they are simultaneously determined by lender supply and consumer demand. The model employs instrumental variables for the current interest rates and current asset prices to correct for this potential bias. The instrument list included the exogenous variables present in the holding equation, lagged values of the interest rates and asset prices, and macrovariables such as unemployment and the three-month Treasury bill rate. 16 Defining the Counterfactual Variables In partial equilibrium, we can define the counterfactual after-tax interest rate figures simply as [5] r cf = r pt (1 τ cf ) where r cf is the counterfactual interest rate, r pt is the (actual) pretax interest rate, and τ cf is a counterfactual tax term that assumes no deductibility or tax rate changes. This term is equal to the true term τδ until 1987; then it keeps δ = 1 and fixes τ at its 1986 level for 1987 91 for all interest rate variables. 17 In general equilibrium, it is possible that the TRA 86 may also have affected pretax interest rates. Poterba (1990) notes this in his discussion of the housing-related effects of the TRA 86. To test for this, I conduct a series of event studies with the pretax interest rates as dependent variables. Appendix B contains the results of these event studies. These event studies provide weak evidence that the TRA 86 affected pretax auto loan and mortgage rates and virtually no evidence that it affected credit card interest rates. Nonetheless, the summary evidence on pretax interest rates is suggestive of general equilibrium effects: the quarterly change in the real pretax interest rate was negative in eight of eight quarters in 1986 7 for auto loan rates, six of eight quarters for mortgage loan rates, and six of eight quarters for credit card rates. In order to correct for any potential bias caused by the improper measurement of post TRA 86 interest rates, I construct counterfactual forecasts of rates using the coefficients from the event study regres- 15 Technically, the coefficient α on a dummy variable measures the percentage change in Y of a 1 value for the dummy in log-differences. The standard percentage change is e α 1. In practice, the two are nearly identical; for example, if α = 0.05 (a representative value from the regressions in Table 3), e α 1 = 0.051. It happens that α < e α 1 for any α, so using the coefficients in Table 3 will underestimate the true values of the coefficients. 16 The Treasury bill may also be endogenous if the amount of household debt affects the demand for government debt. Dropping this variable from the instrument list did not change the results. 17 A more precise counterfactual definition for τ would be the average marginal tax rate (as calculated in equation 1) that would have existed if the marginal tax rate structure had not changed. Unfortunately, it is not possible to construct this figure because the Internal Revenue Service Statistics of Income Bulletin reports only the share of income in each tax bracket. Because the number of brackets fell in 1987, this makes disaggregation into shares in each pre TRA 86 bracket impossible. Another point is that the average marginal tax rate may be endogenous, because an increase in borrowing will lead to an increase in itemized deductions and a smaller share of taxpayers in high marginal tax rate brackets. Tests for the endogeneity of the tax rate did not support this hypothesis. 725

NATIONAL TAX JOURNAL sions. Regressions using these rates (which are typically higher in the post TRA 86 period than the actual series) yield virtually the same results as regressions using the actual rates. Appendix B presents results of these regressions. Returning to the specification, we can simplify the model because for each interest rate term: [6] ln[r cf ] = ln [r pt (1 τ cf )] = ln r pt + ln (1 τ cf ). In addition, if disposable income is included in the model, we note that We can combine the four terms on the right-hand side (RHS) of the tax-adjusted regression equation that have a tax term 18 and write the final model as [9] ln DEBT t (or ln SHARE t ) = β 0 + β 1 ln DEBT t 1 (or ln SHARE t 1 ) + β 2 ln PERSONAL INCOME t + β 3 ln MORTGAGE RATE pt t + β 4 ln AUTO RATE pt t + β 5 ln CREDIT CARD RATE pt t [7] ln DISPOSABLE INCOME cf t + β 6 ln (1 t cf ) t + β 7 ln ASSET PRICE t = ln PERSONAL INCOME t + ln (1 τ cf ) which yields [8] ln DEBT t (or SHARE t ) = β 0 + β 1 ln DEBT t 1 (or ln SHARE t 1 ) + β 2 ln PERSONAL INCOME t + β 2 ln (1 τ cf ) t + β 3 ln MORTGAGE RATE pt t + β 3 ln (1 τ cf ) t + β 4 ln AUTO RATE pt t + β 4 ln (1 τ cf ) t + β 5 ln CREDIT CARD RATE pt t + β 5 ln (1 τ cf ) t + β 6 ln ASSET PRICE t + β j ANNUAL DUMMY jt + ε t + β j ANNUAL DUMMY jt + ε t where the interest rate terms are pretax interest rates and (1 τ cf ) t is the counterfactual deductibility and tax term. This is a useful transformation of the model because it separates the tax term from the pretax interest rates. This means that the t-statistic on the tax term tests the hypothesis that debt aggregates are sensitive to tax policy against the null that they are responsive only to pretax interest rates. 19 It also means that the tax rate coefficient captures the net effects of all other terms in the model that are transformed by taxes. This implies that the following restriction should hold: [10] β 6 = β 2 + β 3 + β 4 + β 5. The restriction could not be rejected in any specification; furthermore, the restricted coefficients are very close to the unrestricted coefficients. The restriction was imposed in every equation. 18 The tax terms differ slightly. The tax terms on the interest rate variables are marginal rates, while the term on disposable income is the average tax rate. Estimating the model with separate terms for the two tax rates did not appreciably change the coefficients, although it did introduce collinearity problems that rendered the tax terms statistically insignificant. 19 If the tax term were not separate from the interest rate terms, the appropriate test would be a likelihood ratio test of a model using the pretax interest rates against a model using the after-tax interest rates. 726

The Tax Reform Act of 1986 and the Composition of Consumer Debt Taking logs allows the coefficients to be interpreted as elasticities. The coefficient on lagged balances estimates the speed at which people adjust their debt holdings. The personal income coefficient estimates the effect of income levels on debt levels. The three interest rate terms measure ownand cross-price terms elasticities of debt. Also, note that because the term is measured as (1 τ) rather than τ, if β 6 > 0, a fall in the tax rate will increase borrowing. The Data The model uses quarterly data for the period 1980 91. 20 Interest rates, debt aggregates, and macroeconomic variables such as personal income were obtained from the Bureau of Economic Analysis Survey of Current Business all debt and interest rates are real values. Appendix C presents these series. Average marginal tax rates were calculated using data from the Internal Revenue Service Statistics of Income Bulletin. Population data were obtained from the Bureau of Labor Statistics. The model also included seasonal dummy variables to correct for any consistent seasonal variation in borrowing behavior; for convenience, they are suppressed in the regression tables. RESULTS Table 2 presents the results of the model. For each type of debt, column (I) contains results using the level of real debt per capita as the dependent variable and column (II) contains results from the equations with the share variable as the dependent variable. The results from the two sets of estimates are fairly similar, and the discussion below refers to both sets except where noted. Additionally, tests for autocorrelation indicated that significant autocorrelation of the error term existed in the mortgage equations (but not the other equations). Because autocorrelation biases all of the RHS coefficients when there is a lagged dependent variable in the regression, the Cochrane Orcutt transformation was used to correct for autocorrelation. In every equation, the interest rate variable corresponding with the dependent variable for example, the credit card interest rate in the credit card debt level equation is negative and significant. The personal income coefficient is positive and significant at five percent in the credit card equations as well; this suggests that the level and share of credit card debt increase as income increases. The Net Effect of Taxes The average marginal tax rate is significant only in the mortgage share equation. This coefficient measures the net elasticity of debt with respect to taxes; for example, in the mortgage share equation, the coefficient of 0.075 states that a one percent fall in taxes will reduce mortgage s share of total debt by 0.075 percent (it will reduce the share because the tax rate variable is defined as 1 τ rather than τ, so a decrease in t increases 1 τ). The model finds a positive net effect of taxes on auto and credit card borrowing, but these coefficients are not significant. It is important to note that these results do not indicate that holders of these types of debt find taxes irrelevant. Rather, it indicates that taxes matter, but that their countervailing incentives cancel. For example, the zero net tax coefficients in the auto loan equations result from the fact that the own-price effect of after-tax auto loan rates is offset by the cross-price effect of after-tax mortgage rates. Cross-Price Effects and Asset Prices The model finds some evidence that debt levels respond to cross-prices. In both the auto loan and mortgage equations, the 20 The series starts in 1980 because that is when the series on credit card interest rates was first published. 727

NATIONAL TAX JOURNAL TABLE 2 DEBT EQUATIONS; ANNUAL DUMMY COEFFICIENTS INDICATE PERCENTAGE DEVIATIONS OF ACTUAL DEBT FROM COUNTERFACTUAL DEBT ESTIMATES; TWO-STAGE LEAST-SQUARES ESTIMATES; INDEPENDENT VARIABLES IN LOGS; STANDARD ERRORS IN PARENTHESES; ALL INTEREST RATES ARE REAL Variable Lagged balances Mortgage (I) (II) 0.700** (0.094) Dependent Variable: Log (Debt or Share of Debt) 0.651** ( 0.036) Auto Loan (I) (II) 0.944** ( 0.071) 0.929** ( 0.073) Credit Card (I) (II) 0.865** ( 0.061) 0.765** ( 0.084) Personal income per capita 0.181 (0.103) 0.072** ( 0.014) 0.291 ( 0.259) 0.118 ( 0.219) 0.692** ( 0.282) 0.703** ( 0.283) Mortgage interest rate 0.409** (0.100) 0.068** ( 0.019) 0.401* ( 0.231) 0.426** ( 0.205) 0.142 ( 0.272) 0.058 ( 0.259) Auto loan interest rate 0.300** (0.088) 0.057** ( 0.018) 0.482** ( 0.246) 0.460* ( 0.225) 0.201 ( 0.274) 0.066 ( 0.263) Credit card interest rate 0.087* (0.053) 0.009 ( 0.010) 0.011 ( 0.136) 0.004 ( 0.145) 0.324** ( 0.023) 0.393** ( 0.138) 1 t cf Average marginal tax 0.160 (0.122) 0.075** ( 0.023) 0.222 ( 0.387) 0.088 ( 0.351) 0.308 ( 0.394) 0.302 ( 0.389) Real asset price 0.119 (0.082) 0.042** ( 0.011) 0.121 ( 0.175) 0.124 ( 0.175) 1986 dummy 0.016** (0.008) 0.007** ( 0.002) 0.038* ( 0.022) 0.037* ( 0.021) 0.035 ( 0.028) 0.034 ( 0.026) 1987 dummy 0.010 (0.012) 0.012** ( 0.003) 0.052* ( 0.031) 0.051* ( 0.029) 0.069* ( 0.037) 0.072** ( 0.033) 1988 dummy 0.003 (0.014) 0.009** ( 0.003) 0.019 ( 0.031) 0.016 ( 0.031) 0.054 ( 0.036) 0.066** ( 0.034) 1989 dummy 0.013 (0.015) 0.014** ( 0.003) 0.037 ( 0.031) 0.036 ( 0.031) 0.102** ( 0.035) 0.118** ( 0.033) 1990 dummy 0.045** (0.019) 0.016** ( 0.003) 0.069* ( 0.037) 0.080** ( 0.037) 0.125** ( 0.042) 0.151** ( 0.039) 1991 dummy 0.044* (0.024) 0.013** ( 0.003) Notes: Equations (I) and (II) use real debt per capita and share of total debt as the dependent variable, respectively. Real asset price variables are the average purchase prices of a home and a financed new car in the mortgage and auto loan equations, respectively, deflated by the consumer price index (CPI). *Indicates significance at ten percent or more. **Indicates significance at five percent or more. 728 0.088* ( 0.046) 0.090** ( 0.050) 0.136** ( 0.033) cross-elasticity terms auto loan and mortgage rates, respectively are positive and significant at ten percent or more. The credit card interest rate term is significant in only the mortgage level equations. The results suggest that consumers reallocate mortgage debt and auto loan debt based on relative interest rates. No interest rates other than the credit card rate are significant in the credit card debt equation. Current real asset prices enter the mortgage share equation as significant and negative, but they are not significant in any other specification. Counterfactual Estimates of Shifts in Debt 0.162** ( 0.051) The dummy variable coefficients in columns (I) and (II) confirm that debt

The Tax Reform Act of 1986 and the Composition of Consumer Debt holdings shifted measurably following the TRA 86. Based on both the share and level equations, mortgage debt slowly increased from 1987 to 1991. This pattern accords with TRA 86 s gradual change in borrowing incentives from the deductibility phaseout. The level equation coefficients for 1986, 1990, and 1991 are significant, and the group of dummy coefficients for 1987 91 is jointly significant at one percent. The share equation coefficients all are significant at five percent or more. The share coefficients suggest that by 1991 mortgage debt was more than one percent higher than it otherwise would have been. The auto loan and credit card equations show substantial and significant drops in debt levels and shares following the TRA 86. Nearly all of the coefficients are significant, and as in the mortgage equations, the pattern of larger effects over time accords with the incentives of the TRA 86. The debt level equations estimate, by 1991, more than a 13 percent reduction in credit card debt and nearly a 9 percent reduction in auto loan debt. The share equations corroborate these results and, for many years, estimate even higher percentage reductions in debt. Because the model fixes taxes and deductibility at their 1986 levels, the dummy variables should measure any shifts in debt caused by these provisions of the TRA 86. However, they might also measure the influence of other exogenous factors that caused debt to shift after 1987. The next section provides an alternative measure of the shifts in debt in order to corroborate the estimates discussed above. An Elasticity-Based Measure of Shifts in Debt The counterfactual estimates presented above are useful because they exploit the ex post nature of the analysis by comparing counterfactual predictions of post TRA 86 debt levels to actual post TRA 86 debt levels. However, it is possible to construct a second set of estimates that use different information. First, we can predict debt from 1986 to 1991 using the coefficients from Table 2 and the counterfactual tax variable (1 τ cf ). 21 Then, we can predict debt from 1986 to 1991 using the coefficients in Tables 2 and the actual tax and deductibility variables. 22 The difference between the two is an estimate of the effects of the TRA 86 on debt. This approach is more restrictive than the counterfactual event study approach above, because it excludes the possibility of other exogenous shifts in debt between 1986 and 1991. It instead provides estimates based solely on the elasticity of debt with respect to after-tax interest rates. In fact, Long s (1989) work uses this approach. He uses data from 1983 to construct elasticities of deductions with respect to taxes and income and uses these elasticities and data from 1986 to predict changes in deductions following the implementation of the new tax schedule. The first five rows of Table 3 present the results of this approach in the Elasticity column and compare them to the results of the counterfactual/event study approach of the previous section the Counterfactual column. The results suggest that by 1991 mortgage debt increased by 1.16 percent due to the tax and deductibility changes. It also suggests that by 1991 auto loan and credit card debt had fallen by 10.46 and 9.17 percent, respectively. These figures are quite close to the counterfactual estimates of 1.31, 8.85, and 13.6 percent. This reinforces the results somewhat, because it suggests that the dummy coefficients in Table 2 are picking up the effects of the TRA 86. 21 All predictions were calculated dynamically; that is, they used last period s predicted value of debt as the lagged dependent variable. Static predictions yielded virtually identical results. 22 For this approach, I use the more conservative of the estimates from columns (I) and (II): the share coefficients from the mortgage debt share equation and the level coefficients from the auto loan and credit card equations. 729

NATIONAL TAX JOURNAL TABLE 3 ESTIMATED SHIFTS IN DEBT CAUSED BY TAX REFORM ACT. COMPARISON OF DUMMY VARIABLE ESTIMATES AND OUT-OF-SAMPLE PREDICTION ESTIMATES; SHIFTS IN PERCENT (LOG-DIFFERENCE) PERCENT Mortgage Auto Loan Credit Card Elasticity Counterfactual Elasticity Counterfactual Elasticity Counterfactual 1987 1988 1989 1990 1991 0.44 0.64 1.2 1.22 1.16 Due to Deductibility Phaseout and Tax Rate Changes 1.18 0.94 1.42 1.62 1.31 4.34 6.51 9.92 10.45 10.46 5.17 1.91 3.68 6.93 8.85 3.66 5.42 9.05 9.4 9.17 6.9 5.4 10.2 12.5 13.6 Due to Deductibility Phaseout 1987 1988 1989 1990 1991 0.73 1.11 1.54 1.64 1.69 0.73 0.97 1.36 1.75 1.35 4.08 6.08 9.61 10.06 9.97 5.35 1.92 3.81 6.99 8.8 2.26 3.2 7.43 7.42 6.61 4.99 2.68 8.09 9.79 10.19 Due to Tax Rate Changes 1987 1988 1989 1990 1991 0.29 0.33 0.16 0.53 0.19 0.45 0.03 0.06 0.13 0.04 0.26 0.43 0.31 0.39 0.49 0.18 0.01 0.13 0.06 0.05 1.4 2.22 1.62 1.98 2.56 1.91 2.72 2.11 2.71 3.41 The story told by these results is that following the TRA 86 consumers shifted debt away from credit cards and auto loans. The key impetus for doing so was that these types of debt became more expensive than they had been prior to the TRA 86, due to both the deductibility phaseout and the new tax schedule. The limitation of these estimates is that they do not reveal how much each provision of the TRA 86 the deductibility phaseout and the new tax schedule relatively influenced debt allocation. The next section addresses this issue. Decomposing the Post TRA 86 Shifts in Debt We can further decompose each estimate of post TRA 86 shifts in debt by relaxing the counterfactual assumptions one at a time rather than simultaneously. That is, instead of estimating the model (or predicting out of sample) under the assumption that neither taxes nor deductibility changed, we can estimate the model under the assumption that only taxes or deductibility changed, but not both. 23 This separates the contribution of each policy to the net shift in debt. The bottom ten rows of Table 3 show the results of these calculations. The results show that the effects of the deductibility phaseout dominated the effects of the new tax schedule on all three types of debt. The deductibility phaseout increased mortgage s share of total debt by roughly 1.7 percent by 1991. This increase is primarily due to its cross-price effects on after-tax auto loan rates. The effect of the deductibility phaseout on credit card and auto loan debt is even more striking. The estimates suggest 23 This means (1) estimating a model using a variable τ cf in which τ equals its true value but δ = 1 for the credit card and interest rate variables and then (2) estimating a model using a variable τ cf in which τ is fixed at its 1986 level but δ equals its true value. The first model will predict debt assuming that taxes changed but deductibility did not, and the second will predict debt assuming that taxes did not change but deductibility did. 730