Gender, age, and race in subprime America

Abstract

For almost 20 years, evidence from journalists' reports, Congressional testimony, and consumer protection litigation suggested that predatory practices in the subprime market were especially harmful for elderly African American women, many of them widows. Much of this evidence has been dismissed as anecdotal, however, and lending research has generally ignored feminist theory – obscuring the relations among race/ethnicity, gender, and age. In this paper, we draw on two complementary datasets to test the hypothesis that subprime inequalities were intensified for African American women. Analysis of Home Mortgage Disclosure Act (HMDA) data confirms that gender inequalities exacerbate racial/ethnic inequalities in the segmentation of high-cost subprime credit, while the National Mortgage Data Repository provides limited circumstantial evidence of disproportionate representation of elderly African American women. Loan terms among subprime borrowers in the NMDR display only modest variations by gender and race/ethnicity, however, although there is some evidence of bait-and-switch tactics and persistently higher total fees among African American women. The veneer of equal treatment within an exploitative subprime market conceals the wider context of structural inequalities of race/ethnicity, gender, and age in housing and credit.

Keywords:

• elderly
• minorities
• mortgages
• gender

Acknowledgements

We are grateful to Elizabeth Renuart, Patricia McCoy, and Stephen Ross for permitting access to the loan-level records of the National Mortgage Data Repository. The NMDR is supported by the Ford Foundation and the National Consumer Law Center. We are also indebted to Alan White and the anonymous referees for valuable comments and criticisms on earlier versions, and to Jen Douglas at the NCLC for helpful clarifications and advice on the structure of the NMDR files. The usual disclaimer – that the usual disclaimer applies – applies.

Notes

¹Along the continuum between disparate treatment and disparate impact, fairly explicit gender biases do occur. HUD recently secured a settlement with a mortgage lender that rescinded a loan approval when it learned that the customer was on maternity leave – despite the fact that her leave provided full salary. HUD continues to investigate other, similar complaints (Bernard 2011).

²Hyman (forthcoming, 228) wryly observes that with the inclusion of extensive arrays of seemingly objective measures, “the correlated variable would acquire the predictive power of the protected category,” with discrimination erasing its own evidence.

³This is not to say, of course, that the regulatory goal of complete coverage is achieved. Penalties for non-disclosure under HMDA are negligible compared to the potential liability for civil rights violations that HMDA records could signal, and thus it is likely that HMDA reports are most likely to be missing for some of the worst offenders.

⁴Even if it were possible to use Ross and Yinger's (2002) matching procedures to link individual NMDR and HMDA records, the small NMDR sample would be spread too thin across space and time to permit sufficient inferential tests for geographical representation.

⁵The absence of credit history information is a direct result of industry lobbyists' pressures to minimize the expansion of Regulation C in 2002. Immergluck (2004, 219) accurately predicted the public discourse that greeted the release of the new loan-pricing data in early 2005: “…without some credit history data, lenders will dismiss disparities as due primarily to differentials in credit history, without having to offer any evidence in this regard. … banks argued against including such data in HMDA, but later they will almost certainly argue that, without such data, the pricing information cannot be interpreted.” For a description of lobbyists' efforts to spin the new data in early 2005, see Wyly et al. (2007, 2139–41).

⁶This figure is correspondingly reduced when records are excluded on the basis of missing information on key variables.

⁷The TILA APR is a problematic measure of the cost of credit (see Immergluck 2004, 218–19, and Sovern 2010) but it is the only summary expense figure reported for all borrowers for all types of loans. In multivariate analysis, including controls for adjustable-rate and pay-option mortgages helps to mitigate the systemic distortions of the meaning of the APR for these types of instruments. The benchmark for our TILA spreads are the market yields for constant-maturity US Treasury securities of comparable duration (10-year, 15-year, 30-year) from the Federal Reserve's H-15 series; simple linear interpolation was used for the few loan notes with unusual durations. The US Treasury's published extrapolation factors were used to estimate 30-year yields for the 2002–2006 period when the benchmark Treasury long bond was not sold.

⁸Applicants were classified on the basis of the race/ethnicity/sex information of the primary applicant; in the case of loan files with multiple applications, the first document with full demographic information was used.

⁹Significance tests for coefficients in our HMDA models are not reported, for two reasons. First, HMDA is a full enumeration, not a sample. Second, the large number of observations means that nearly every coefficient estimate achieves statistical significance, even if the difference from zero is practically insignificant. In the models for Table 3, only a single estimate (home improvement loans in Model 1) fails to achieve a significance level of 0.05; all other estimates achieve P < 0.001. Multicollinearity diagnostics for Table 3 indicate no problems: all variables have tolerances well above the 0.20 threshold recommended by Menard (2002). Most tolerances are in the 0.75–0.90 range, with the lowest values above 0.26 for Model 1, 0.38 for Model 2, and 0.35 for Model 3.

¹⁰Full county-level regressions are not presented here, but are available on request.

¹¹Multicollinearity diagnostics again indicate no bias: the lowest tolerances for the three models in Table 4 are 0.27, 0.40, and 0.34. All coefficients in Table 4 attain statistical significance at P < 0.05, and all except two at P < 0.001.

¹²Most of these “other” sales go to the special-purpose vehicles (SPVs) established as trusts for large mortgage-backed securities offerings.

¹³Exact correspondence is rendered impossible by the wide temporal range of the NMDR (1994–2008), interacting with loopholes in the Regulation C provisions pertaining to the collection of data on race/ethnicity/gender (see Huck 2001). Some of these loopholes were closed in the 2002 Regulation C revisions. Our estimates compare the NMDR race-gender identifications of the first borrower with the lead-applicant tabulations for 2006 as reported in Table 2.

¹⁴Age-missing borrowers are not representative, and thus do not satisfy the condition of missing completely at random (MCAR) (Allison 2002). But for the rate-spread models, the non-random bias is not tied to race/ethnicity or gender. Age-missing stepwise models (pseudo-R-squared of 0.62) indicate a significantly greater prevalence among higher-income, non-occupant, non-face-to-face applicants for purchase mortgages, who apply for fixed-rate loans and then agree to adjustable-rate notes while avoiding prepayment penalties. This profile is broadly consistent with the high-risk lending industry's gradual reorientation from refi/home improvement to purchase credit, and from urban minority equity stripping to diverse middle-class leveraged home-buying and speculative accumulation (Williams et al. 2005; Immergluck 2009).

¹⁵Overall model fit is encouraging, with re-scaled Nagelkerke (1991) pseudo-R ² values between 0.39 and 0.70.

¹⁶Age information is not reported for almost a third of African American women; this figure is three times the rate for White women, and twice the proportion for White men. The figure is not as high, however, as that for all other applicant types (51.3 percent – see Table 1); this reflects interdependency between non-reporting of age, race/ethnicity, and gender (Huck 2001). Multivariate tests, however, indicate a robust result for our over-65 variables in Table 6. As noted earlier, the age-missing data do not satisfy MCAR conditions (Allison 2002). As Allison (2002, 6–7, emphasis in original) notes, however, “if the probability of missing data on any of the independent variables” in a regression “does not depend on the values of the dependent variable, then regression coefficients using listwise deletion will be unbiased (if all the usual assumptions of the regression model are satisfied.” A stepwise logistic regression of the age-missing indicator (an independent variable in Table 6) does not select the variable for African American women (the dependent variable). The stepwise age-missing model, in fact, does not select any of the gender or race variables; the algorithm instead selects a total of nine variables measuring various aspects of applicant finances and loan terms – but does not give any results that trigger Allison's (2002) warnings. Listwise deletion was therefore applied to a model similar to that in Table 6, Model 1 (quasi-complete separation required omission of one variable, the switch from fixed-rate application to pay-option loan note). The standardized coefficient for over-65 borrowers increases from 0.475 (P = 0.043) to 0.833 (P = 0.038) when age-missing borrowers are excluded.

¹⁷Of the 122 coefficient estimates in Table 6, fourteen dip below the 0.20 threshold that is generally regarded as cause for concern. Most of these involve the age variables in Models 1 and 4. Various alternative specifications of Model 1 reduce the magnitude of the over-65 odds ratio as well as its significance level (to a range between 0.10 and 0.25). In all estimations, however, the effect remains positive for older borrowers.

¹⁸One coefficient in Model 3 yields a tolerance below the 0.20 threshold; eliminating this variable (loan amount) yields an odds ratio for the over-65 variable of 4.87 (P = 0.103).

¹⁹These models achieve pseudo-R ² values of 0.47 when comparing older Black women to Black men, and 0.39 compared to White women. Due to sample size constraints, older African American women are defined as those over age 50 (n = 27). All tolerances are above 0.20, with the sole exception of loan amount in the comparison with Black men (0.19).