Comparing Homeowner and Lender Estimates of Housing Wealth and Mortgage Terms

Transcription

1 Comparing Homeowner and Lender Estimates of Housing Wealth and Mortgage Terms Brian Bucks and Karen Pence Federal Reserve Board of Governors May 2005 The views expressed in this paper are ours alone and not necessarily those of the Board of Governors or its staff. We thank Brad Case, Karen Dynan, Gerhard Fries, Markus Grabka, Diana Hancock, Andreas Lehnert, Arthur Kennickell, Kevin Moore, Wayne Passmore, Howard Savage, Shane Sherlund, and participants at the Luxembourg Wealth Study Conference on Construction and Usage of Comparable Microdata on Wealth for helpful discussions and suggestions. Contact information: 1

2 Introduction Many studies of housing wealth and mortgage debt are based on household-reported data. The results from these studies could be misleading if household-level data are reported with error. To examine the extent to which this may be a concern, we compare the distributions of house values and mortgage terms as reported by homeowners in the Survey of Consumer Finances (SCF) to the distributions of the same variables as reported by lenders in the Office of Federal Housing Enterprise Oversight (OFHEO) data, the Loan Performance Corporation data, and the Residential Finance Survey (RFS). We also examine the SCF edit or imputation flags to gauge indirectly how many respondents have difficulty answering particular survey questions. We find that an index of house value appreciation based on SCF data matches the aggregate OFHEO index fairly closely. In addition, the SCF edit flags suggest that most homeowners are able to report their house values with some degree of confidence. The SCF measures of mortgage maturity and type also match lender estimates well. However, the distributions of adjustable-rate mortgage (ARM) terms differ substantially across surveys, and a sizeable number of adjustable-rate borrowers report that they are unaware of their mortgage terms. These findings suggest that household-reported surveys, and the Survey of Consumer Finances in particular, capture broad measures of housing wealth and mortgage terms fairly accurately. The comparisons do not necessarily indicate whether individual households report these values correctly, since the errors of individual households could offset each other in such a way that the distribution remains accurate. However, summary statistics from the household-reported data appear valid, and researchers investigating questions such as the effect of housing wealth on consumption may be on safe ground using household-reported data. Household- and lender-reported data paint different pictures, though, of the terms of a typical adjustable-rate mortgage. The discrepancies across surveys may stem partly from differences 1

3 in sample composition, survey design, and methodology. In particular, the complexity of ARMs implies that differences in the structure of surveys or the wording of questions may yield different answers. However, the SCF edit flags indicate that borrower recall problems likely account for some of the discrepancy. For researchers interested in household perceptions of their mortgage terms, household-reported data are still the best choice. For other studies of adjustable-rate mortgages, lender-reported data may be a better option. Previous studies Previous validation studies of household-reported housing data have used one of three approaches. The first approach, which is followed in this paper, compares homeowner estimates in the aggregate to external indexes. The second approach compares individual homeowner reports of house values to either the transaction sales prices or lender appraisals of their homes. The third approach uses panel data to examine whether homeowners describe their homes and mortgages consistently over time. In general, the existing house value studies are optimistic about respondent accuracy, while the sole study of mortgage terms is pessimistic. Turning first to house values, an example of the first approach is DiPasquale and Somerville (1995), who use the homeowner-reported house values and transaction sales prices on the AHS to construct measures of the changes in house prices over time. Although the homeowner estimates are somewhat higher than transaction prices, the two house price series track each other and a National Association of Realtors house price series fairly closely. Two recent examples of the second approach are Goodman and Ittner (1992) and Kiel and Zabel (1999), who compare homeowner estimates of house values on the American Housing Survey (AHS) to transaction sales prices also reported on the survey. They find that homeowners overestimate the value of their homes by 5 to 6 percent. In addition, Kiel and Zabel find that recent 2

4 homeowners over-value their homes by about 3 percentage points relative to more established homeowners. However, since these over-estimates do not appear to be correlated with any borrower or property characteristics besides length of tenure, the authors conclude that the bias does not pose a significant problem for research. Both papers also include comprehensive literature reviews that provide further evidence that homeowners do not overstate the value of their homes substantially. Kennickell and Starr-McCluer (1997), an example of the third approach, is the only previous study to examine how accurately respondents report housing data in the SCF. They exploit the unique structure of the SCF panel, which contains cross-sectional information on household portfolios in 1983 and 1989 as well as retrospective questions (asked in 1989) about changes in household portfolios over the period. They found that only 5 percent of households in the panel provided retrospective data about home sales and purchases that was inconsistent with the cross-sectional data. They attribute this high consistency rate to the fact that home purchases and sales are well-defined, highly salient events. Alessie and Zandvliet (1993) draw analogous conclusions from a study based on Dutch data. The literature on the accuracy of respondent-reported mortgage data is much sparser and more pessimistic. The only study we have found is Leary, Newhouse, and Mihaly (2004), who find that 46 percent of mortgages in the Survey of Income and Program Participation data had at least one mortgage term reported inconsistently over time. Thus whereas our study of the accuracy of house values estimates builds upon a long-established literature, our study of the accuracy of mortgage terms explores somewhat new territory. 3

5 Data Our empirical work is based on four datasets: The Federal Reserve s Survey of Consumer Finances is the most comprehensive and highest quality household wealth dataset in the United States. It is designed to be representative of the U.S. household population. The survey has been conducted every three years since 1983, with a consistent survey design since In 2001, the survey included 4,442 households, of which 3,162 were homeowners, 1,557 had fixed-rate mortgages, and 235 had adjustable-rate mortgages. 1 All missing data are imputed using multiple imputation techniques. 2 Aizcorbe, Kennickell, and Moore (2003) provide an overview of the 2001 data. The Residential Finance Survey (RFS) is conducted every ten years by the U.S. Census Bureau. The survey is designed to be representative of all residential properties in the United States. It included data on 17,000 owner-occupied properties in Households who are selected for the RFS sample are required by law to participate, unlike the SCF, where participation is voluntary. Despite this difference in legal status, though, overall response rates are comparable in the two surveys (56 percent in the RFS and x percent in the SCF). The survey collects general information on the property and the mortgage from the homeowner and detailed information on the mortgage from the lender. The lender-reported data are missing for roughly half of all mortgages. These data could be missing because the borrower did not provide information about the mortgage lender or because the mortgage had been sold and the RFS staff could not find the current servicer. In other cases, the RFS was able to find the current servicer, but the servicer did not have access to the original loan documents and thus could not report all variables. The RFS does not impute missing data for most lender-reported variables, although it 1 The data also include 78 balloon mortgages. 2 See Kennickell (1998) for more information on multiple imputation in the SCF. 4

6 does impute information for some household-reported variables with missing data. Our tabulations of RFS variables exclude observations with missing data. 3 The Office of Federal Housing Enterprise Oversight (OFHEO) house price index is a repeat transactions house price index based on mortgages backed by single-family properties that have been purchased or securitized since 1975 by Fannie Mae and Freddie Mac. The index is based only on conforming mortgages, which are those small enough to qualify for purchase by Fannie Mae or Freddie Mac (under $275,000 in 2001); it also excludes government-backed Federal Housing Administration or Veterans Administration mortgages. Tabulations on the Residential Finance Survey indicate that about half of owner-occupied properties are captured by the OFHEO index. Appraisals from home refinancings tend to distort the index over short time periods but not over the longer periods examined in this paper. The Loan Performance data are collected from the administrative records of large mortgage servicers. The mortgages are originated by a wide variety of institutions, including banks and mortgage brokers. All mortgages guaranteed by Fannie Mae or Freddie Mac are represented in the data. In total, the December 2001 data covered about 80 percent of U.S. home mortgages. Homeowner estimates of housing values In the aggregate, the changes in self-reported house values in the Survey of Consumer Finances match well the changes in house values as measured by the national OFHEO house price index (table 1A and figures 1 and 2). To show this finding, we calculate the cumulative percentage change in house price for each SCF household from its self-reported current house value and original purchase price. We then sort these percentage changes by the year of purchase and calculate the 3 The percent of observations missing for a given variable ranges from 41 percent for the mortgage amortization period variable to 62 percent for the cap on interest-rate changes per period variable. 5

7 median and geometric mean house price change for each year of purchase. The median controls better for outliers, whereas the geometric mean replicates the OFHEO methodology more closely and may be more informative than the arithmetic mean for a skewed measure such as house prices. 4 We perform the same exercise for the households in the Residential Finance Survey. The RFS does not provide a clean test of the accuracy of self-reported house price changes, since the house purchase price variable on the RFS is drawn from estimates provided by both the borrower and the lender. 5 However, the RFS results may be precise, in a statistical sense, than the SCF results because of the larger RFS sample size. Finally, in a last series of tests, we limit both the SCF and RFS samples to the conforming, non-government-backed, mortgages backed by single-family properties on which the OFHEO index is based. This sample restriction may be important if the prices of these properties change differently over time than prices in the market as a whole. The price indices based on the medians of the household price changes match the OFHEO index reasonably well. For homes purchased in 2000, the median one-year price change was 7.7 percent in the SCF data and 8.3 percent in the RFS data; the OFHEO index suggests a one-year change of 7.5 percent. For homes purchased in 1991, the SCF suggests a 54 percent cumulative house price change; the RFS suggests a 41 percent change; and the OFHEO index suggests a 50 percent change. For homes purchased more than fifteen years ago, the SCF estimates lie below the OFHEO estimates. This result is consistent with the Kiel and Zabel (1999) finding that new homeowners provide higher house value estimates than do established homeowners. Nonetheless, 1/ n 4 The geometric mean of a sequence { a } i= 1 n i is ai n. The OFHEO index is based on a regression of log house i= 1 prices. Thus, it models the arithmetic mean of log house prices, or equivalently the geometric mean of house prices. 5 It is not possible to infer from the public RFS data whether the reported purchase price is based on data from the borrower or lender. The estimated current value, however, is reported only by the borrower. 6

8 researchers would not be seriously misled by relying on the medians of the household-reported house price changes. In contrast, the geometric means of reported house price changes in the SCF are larger, sometimes substantially so, than the OFHEO index. This discrepancy could occur because more expensive homes, which are represented in the SCF data but not in the OFHEO index, have different price dynamics than other homes; because some homeowners are over-stating the increase in their house value; or because of sampling error. The discrepancy is less pronounced in the RFS, perhaps because the sample size is larger or because the data are edited more heavily and are topcoded. 6 In fact, in the farther-out years the RFS geometric means match the OFHEO index quite closely: over the period, for example, both the RFS measure and the OFHEO index suggest that house values increased 50 percent. To examine whether sample selection skews the results, we repeat this exercise but eliminate government-backed and nonconforming loans (table 1B). This sample restriction reduces the SCF sample of homes purchased since 1985 by roughly 70 percent and reduces the RFS sample by over 50 percent. 7 The estimates in table 1B are thus more statistically variable than the earlier estimates. However, median house-price changes in the SCF still track the OFHEO index reasonably well, particularly for those who purchased their home within the last ten years. The tendency of the geometric mean of the SCF values to lie above the OFHEO estimates is also apparent in this more limited sample. As before, the estimates based on the RFS medians fall below the OFHEO values, 6 30% of RFS households and 16% of SCF households had current house value, purchase value, or year of purchase edited or imputed. Many of the RFS edits stem from reconciling lender- and household-reported data. Most of the SCF edits stem from replacing a respondent-reported range with a single value. Excluding these edits, only 5% of SCF households had any of these values imputed. To protect respondent confidentiality, the highest 3% of house values in the RFS are replaced with the average calculated over these households. 7 These restrictions have a larger effect on the SCF sample because high-wealth households are over-represented in the survey. These households are more likely to have very large mortgages or no mortgage at all and are less likely to have government-backed mortgages. 7

9 whereas the estimates based on the geometric mean track the OFHEO index fairly well in the farther-out years. To consider further the accuracy of self-reported housing values, we turn to an examination of response rates to specific questions on the Survey of Consumer Finances (table 2). These response rates are based on the edit or imputation flags provided for nearly all SCF variables. These flags allow users to infer, to a large extent, whether a response has been edited or imputed, as well as the reason why a value was originally missing. These codes also indicate when a respondent gave a range for a dollar value, such as the home s original purchase price. 8 The number in parentheses after the variable name indicates the percent of the sample asked the question. Only the 68 percent of SCF respondents who are homeowners, for example, were asked the value of their house. Responses are divided among five classifications: Original value: the original value provided by the respondent was deemed acceptable and included in the survey, Range: the respondent provided a range rather than an actual value or auxiliary information used in the editing process was sufficient to bound the value, 9 Edited value: the respondent provided enough information in auxiliary or related fields that the correct value could be inferred by the SCF staff with a very high degree of confidence, Don t know/refused: the respondent did not answer the question, 8 See Kennickell (1997) for further description and analysis of the use of range responses in the SCF. 9 The next iteration of this paper will indicate how many range values fall in each category. 8

10 Missing due to editing: either a highly implausible value was set to a missing value or insufficient information was available for a given variable when changes were made to connect an erroneously asked question sequence. The response rates to the SCF house value questions are quite favorable. Ninety-seven percent or more of respondents were able to provide an acceptable response or range when asked the current value of their house, its purchase price, or the year of purchase. A remarkable 99 percent reported acceptable values for the year of home purchase; the house value questions elicited a larger number of range responses. 10 Overall, these tabulations suggest that households have a fair amount of confidence in their knowledge of their approximate house value. Homeowner estimates of mortgage data To examine the accuracy of household-reported mortgage data, we compare the SCF data to lender-reported data. As noted earlier, the Loan Performance data are reported entirely by lenders and other participants in the mortgage industry, whereas the Residential Finance Survey data are reported by both borrowers and lenders, with detailed mortgage data reported almost exclusively by lenders. In all data sets, we examine only first mortgages. For broad measures of mortgage terms, the borrower- and lender-reported distributions match well. For example, all three datasets agree that about 85 percent of first mortgages were fixed in 2001, about 11 percent were adjustable, and the rest were balloon. The distribution of amortization periods is also consistent: about a quarter of all mortgages had an amortization period of 15 years or 10 In the case of house values, where the true value is not known with certainty until the house is sold, a range may be as or more economically meaningful than a value. 9

11 fewer, about 65 percent had a 30-year amortization pattern, and the rest were scattered across different categories. The close match is somewhat remarkable; given the differences in data sources and methodology, we would not necessarily expect the estimates to correspond so well. The three datasets are not benchmarked against each other, so the close correspondence is not a function of the way the sample weights were constructed. For most adjustable-rate mortgage terms, however, the distributions do not match as closely. A notable exception is the measure of how often the interest rate can change: the RFS and SCF data agree that the interest rate can change monthly on about 15 percent of ARMs and annually on about 60 percent of ARMs. However, the SCF estimates of the cap on per-period interest rate changes are somewhat lower than the RFS estimates. For example, 42 percent of SCF respondents and 6 percent of RFS respondents report a cap around one percentage point; 24 percent of SCF respondents and 47 percent of RFS respondents report a cap of two percentage points. Seven and 18 percent of SCF and RFS respondents, respectively, report no caps on changes per period. The estimates of the lifetime cap on interest rate changes are also somewhat lower in the SCF. A lifetime cap less than five percentage points is reported by 58 percent of SCF respondents and 6 percent of RFS respondents. In contrast, 21 percent of SCF respondents and 51 percent of RFS respondents indicate that the lifetime cap is five or six percentage points. Only 2 percent of SCF respondents, but 25 percent of RFS respondents, report a lifetime cap greater than 12 percentage points or no cap at all. Further discrepancies are apparent when respondents are asked what index their ARM is tied to and whether their ARM is convertible to a fixed-rate mortgage. The Loan Performance and RFS data agree that approximately two-thirds of adjustable-rate mortgages are linked to the rates on U.S. Treasury bills, with the rest linked to a bank cost-of-funds index or the London InterBank Offered 10

12 Rate (LIBOR). However, only 35 percent of SCF adjustable-rate borrowers report that their mortgage is linked to any of these rates; the modal answer is the prime rate, with 46 percent of respondents, and 20 percent report implausible indexes or colloquial terms (the Consumer Price Index, the going rate, or the Federal Funds rate). Likewise, only 10 percent of adjustable-rate mortgages are reported as convertible to a fixed-rate mortgage on the RFS, yet nearly half of SCF adjustable-rate borrowers indicate that their mortgage is convertible. Differences in question design and in how lenders and borrowers understand certain terms may drive some of these differences. For example, convertible mortgage has a specific, technical meaning to a mortgage lender. Borrowers, however, may consider their mortgages convertible if they can refinance their mortgage without a prepayment penalty. Likewise, the meaning of interest rate cap is probably clear to lenders. Some households, though, have ARMs with negative amortization features and payment caps; with these mortgages, payments may not increase as fast as interest rates, and borrowers may mistake payment caps for interest rate caps. Sample selection issues may also explain some of the differences. Prime, conforming, and fixed-rate mortgages are over-represented in the Loan Performance data. Mortgages that the original lender kept in portfolio, rather than selling to other lenders or servicers, are overrepresented in the RFS data. If the data sources capture different slices of the mortgage universe, the estimates will differ. However, sample selection issues seem unlikely to explain the discrepancies fully because the data sets match closely for many variables. Instead, the SCF edit and imputation flags (table 2) suggest that at least part of the difference in borrower and lender reports of ARM terms stems from borrower uncertainty. Borrowers appear fairly confident in their knowledge of broad mortgage features: almost all borrowers provide a value or range for the year the mortgage was obtained, the amount borrowed, the amount still owed, the 11

13 amortization period, the amount of the regular payment, and whether the mortgage is fixed or adjustable rate. Of the major mortgage terms, response rates are relatively low only for the current interest rate on the mortgage; ten percent of respondents either did not know this information or chose not to provide it. Response rates on the details of adjustable-rate mortgages, however, are considerably lower. Nearly 15 percent of these borrowers do not know how often their interest rate can change, and approximately 40 percent do not know the caps on these changes either at a given time or over the life of the mortgage. Over 20 percent do not know the original interest rate on the mortgage and around 15 percent do not know the index to which their mortgage is linked or whether it can be converted to a fixed-rate mortgage. These response rates suggest that borrower-reported data on ARMs should be interpreted with caution. Conclusion Aggregate estimates of housing wealth and broad mortgage terms from household-reported data match estimates from lender-reported data remarkably well. In fact, given the differences in design and methodology across data sets, it is not obvious that these measures should align so closely. The evidence presented in this paper, at least, suggests that it is reasonable to use these householdreported data for research purposes. Lender- and household-reported data do not match well, however, for most adjustable-rate mortgage terms. This discrepancy may stem from differences in survey methodology or because borrowers have difficulty remembering their loan terms. The household-reported data are still valuable for researchers interested in household perceptions of their mortgage terms. Researchers interested in other questions related to adjustable-rate mortgages, though, may find lender-reported data to be a more useful guide. 12

14 REFERENCES Aizcorbe, Ana M, Arthur B. Kennickell, and Kevin B. Moore, Recent Changes in U.S. Family Finances: Evidence from the 1998 and 2001 Survey of Consumer Finances, Federal Reserve Bulletin, January Alessie, Rob, and Christine Zandvliet, Measurement of Household Saving Obtained from First- Differencing Wealth Estimates, VSB-CentER Savings Project Progress Report 13, DiPasquale, Denise, and C. Tsuriel Somerville, Do House Price Indices Based on Transacting Units Represent the Entire Stock? Evidence from the American Housing Survey, Journal of Housing Economics, 4: , Goodman, John L. Jr., and John B. Ittner, The Accuracy of Home Owners Estimates of House Value, Journal of Housing Economics, 2(4):339-57, December Kennickell, Arthur B., Using Range Techniques with CAPI in the 1995 Survey of Consumer Finances, January Available at Kennickell, Arthur B., Multiple Imputation in the Survey of Consumer Finances, September Available at Kennickell, Arthur B., and Martha Starr-McCluer, Retrospective Reporting of Household Wealth: Evidence from the Survey of Consumer Finances, Journal of Business and Economic Statistics, 15(4):452-63, October Kiel, Katherine A., and Jeffrey E. Zabel, The Accuracy of Owner-Provided House Values: The American Housing Survey, Real Estate Economics, 27(2): , Leary, Jesse B., David Newhouse, and Kata Mihaly, The Dynamics and Wealth Effects of High- Rate Loans, manuscript, Federal Trade Commission,

15 FIGURE 1: HOUSEHOLD-REPORTED HOUSING APPRECIATION, GEOMETRIC MEAN COMPARISON Cumulative % Increase in Home Value Years Since Home Purchase OFHEO Index SCF Geom. Mean RFS Geom. Mean 14

16 FIGURE 2: HOUSEHOLD-REPORTED HOUSING APPRECIATION, MEDIAN COMPARISON Cumulative % Increase in Home Value Years Since Home Purchase OFHEO Index SCF Median RFS Median 15

17 TABLE 1A. COMPARISON OF HOUSE VALUE APPRECIATION ACROSS DATA SOURCES ALL HOUSES Year OFHEO 2001 Survey of Consumer Finances 2001 Residential Finance Survey Purchased Index (Percent) Count Median Geom. Mean Count Median Geom. Mean NOTE. The Residential Finance Survey does not provide the exact year of purchase for houses purchased before

18 TABLE 1B. COMPARISON OF HOUSE VALUE APPRECIATION ACROSS DATA SOURCES OVER HOUSES ELIGIBLE FOR INCLUSION IN THE OFHEO INDEX Year Purchased OFHEO 2001 Survey of Consumer Finances 2001 Residential Finance Survey Index (Percent) Median Geom. Mean Count Median Geom. Mean NOTES. The Residential Finance Survey does not provide the exact year of purchase for houses purchased before Samples include unattached, mortgaged properties that are not financed by government-insured loans or by loans that exceed the conforming loan limits. These restrictions rely in part on variables that are unavailable on the public SCF data. Counts are suppressed for the SCF estimates for disclosure protection reasons. 17

19 TABLE 2. REPORTING RATES FOR HOUSING AND MORTGAGE CHARACTERISTICS: 2001 SCF Variable (Percent of SCF sample asked question) Original Value Range Edited Value Don t Know/ Refused Missing due to Editing Housing Variables Current value of home (67.7%) Purchase price of home (67.7%) Year home purchased (67.7%) <0.1 Mortgage Characteristics Year mortgage obtained (42.3%) Amount borrowed or refinanced (42.3%) Amount still owed (42.3%) Amortization period (42.3%) Amount of regular payment (41.3%) Annual interest rate (42.3%) Adjustable rate (42.3%) Adjustable-Rate Details Frequency rate can change (4.8%) <0.1 Maximum rate can rise at once (4.8%) Maximum rate can be charged (4.8%) Original interest rate (4.8%) On what index does it depend (3.8%) Convertible mortgage (4.8%) 85.0 <

20 TABLE 3. DISTRIBUTION OF MORTGAGE TERMS IN THREE DATA SETS Loan Performance (December 2001) 2001 Residential Finance Survey 2001 Survey of Consumer Finances Over all first mortgages Mortgage type Fixed Adjustable Balloon Amortization period Over all adjustable rate first mortgages Frequency with which interest rate can be changed Monthly Quarterly or every six months 10 5 Annually or 5 years Other 3 9 Caps on interest rate changes per period Less than 2 ppts ppts Between 2 & 9 ppts ppts No caps 18 7 Caps on lifetime interest rate changes Less than 5 ppts ppts ppts ppts ppts GT 12 ppts No caps 12 ARM Index Treasury bills Cost of funds Index LIBOR or CD 14 7 Prime 46 Consumer Price Index 10 Going rate 5 Federal funds rate 5 Other 8 Convertible to a fixed-rate mortgage? Yes 9 47 No