Mortgage Discrimination ECN741: Urban Economics, Professor Yinger.

Mortgage Discrimination

ECN741: Urban Economics, Professor Yinger


Outline of Class

Formal Definitions of Discrimination

The Boston Fed Study

The Importance of Disparate-Impact Discrimination

Blending Loan Characteristics and Loan Performance Data to Obtain More Precise Estimates of Discrimination


A Hint of Trouble

The Home Mortgage Disclosure Act (HMDA) provides information on virtually all mortgage applications in the country.

These data reveal that for many years, blacks and Hispanic applicants have been far less likely than white applicants to have their applications approved.

This does not prove that discrimination exists, but it is a difficult regularity to explain without discrimination.


1995

1996

1997

1998

1999

2000

2001

2002

2003

2004

2005

2006

2007

2008

2009

2010

2011

0

0.5

1

1.5

2

2.5

3

Minority-White Loan-Denial Ratios, Conventional Loans, 1995-2011

Black Hispanic Asian

Loa

n-D

en

ial

Ra

tio

New Rules for Report-ing Race and Ethnicity


Definition of Discrimination

The Behavior

Lenders receive applications for mortgage loans and decide whether to approve them (and what terms to offer) based on the characteristics of the applicant, the loan, and the property.

This underwriting process may involve automated underwriting tools, including credit scores.

It results in both an approval decisions and, if the loan is approved, a set of terms and conditions.


Discrimination in Mortgage Lending

Disparate-Treatment Discrimination

Using different criteria for evaluating the loan applications (or setting terms) of people in different groups.

Disparate-Impact Discrimination

Using underwriting standards that have a disparate, negative impact on people in protected classes without being justified on the basis of “business necessity.”

Business Necessity

To meet the business necessity test, a standard must be directly linked to the profits a lender could make in a competitive, group-neutral market.


The Loan-Approval Decision

Definitions

L = characteristics of the loan, such as its interest rate and loan-to-value ratio

A = characteristics of the applicant, such as his or her past credit problems

P = characteristics of the property, such as whether it is in a declining neighborhood.

π = loan profitability

Decision rule based on best available information

*

*

accept if : ( , , )

reject if: ( , , ) .

L A P

L A P


The Loan-Approval Decision, 2

Decision rule based on incomplete statistical model

or

*

*

accept if : ( , , )

reject if: ( , , ) .

E

E

L A P

L A P

*

*

accept if : ( , , ) [ ( , , ) ( , , )]

reject if: ( , , ) [ ( , , ) ( , , )] .

E

E

L A P L A P L A P

L A P L A P L A P


Formal Definitions of Discrimination

Disparate-Treatment Discrimination

This type of discrimination exists if membership in a protected class (= M) affects the accept/reject decision after controlling for L, A, and P.

*

*

accept if : ( , , ) ( ) [ ( , , ) ( , , )]

reject if: ( , , ) ( ) [ ( , , ) ( , , )] .

E

E

L A P D M L A P L A P

L A P D M L A P L A P



This type of discrimination is buried in the difference between πE and π.

A difference between πE and π implies that the lender is using criteria that, to some degree, are not related to loan profitability.

For disparate-impact discrimination to exist, the criteria leading to this difference must be particularly unfavorable to applicants in a protected class.


The Boston Fed Study (1990 Data) CONTROL VARIABLES IN THE BOSTON FED STUDY

Risk of Default Loan Characteristics Housing Expense/Income Two- to Four-Family House Total Debt Payments/Income Lender ID Net Wealth Consumer Credit History Mortgage Credit History Public Record History Probability of Unemployment Self-Employed Loan/Appraised Value Is Low Loan/Appraised Value Is Medium Loan/Appraised Value Is High Costs of Default Personal Characteristics Denied Private Mortgage Insurance Black or Hispanic Census Tract Dummies


Criticisms of Boston Fed Study’s Methodology

Potential flaws in the Boston Fed Study include:

omitted variables, data errors in the explanatory variables, mis-classification in the dependent variable, incorrect specification, and endogenous explanatory variables.


Ross and Yinger, The Color of Credit (2002)

R/Y investigate each of these potential flaws.

No legitimate data-cleaning exercise or revised statistical procedure has a significant impact on the estimated minority status coefficient (or its statistical significance)

They confirm the Boston Fed’s conclusion:

Blacks and Hispanics are 82% more likely than whites to be turned down for a loan, controlling for L, A, and P.


R/Y’s Re-Interpretation of Boston Fed Study

The Boston Fed Study’s results could reflect Disparate-Treatment Discrimination

If all lenders use the same underwriting standards, any group-based difference in treatment implies that the standards are applied differentially across groups.

Or



A significant minority-status coefficient could arise if different lenders use different underwriting standards and deviations from average standards cannot be justified on the basis of business necessity.

Legitimate Variation in Underwriting Standards Across Lenders

A significant minority status coefficient could reflect non-discriminatory, cost-based differences in underwriting standards across lenders.


Accounting for Variation in Underwriting Standards

R/Y re-estimate the Boston Fed Study’s

equations to account for potential variation in underwriting standards across lenders.

Some regressions have interactions with

characteristics of lenders’ portfolios

Some regressions allow different coefficients for

different lenders


All methods yield the same two main results:

Underwriting weights differ significantly across lenders.

Accounting for this variation in underwriting weights has no significant impact on the estimated minority-white disparity in loan approval.

Unless legitimate variation in underwriting weights in not linked to a lender’s portfolio, these results indicate that either disparate-treatment or disparate-impact discrimination exists.


Subsequent Literature

The Boston Fed Study has not been replicated

Later studies include

Studies of loan approval with a small sample of lenders.

Studies of discrimination in “overages” for a small sample of lenders.

Explorations of the HMDA data, which cannot be used to study discrimination due to lack of controls.


Data Problems

Scholars have not been able to get ahold of the necessary data.

Problems include:

Lenders will not release the data to scholars.

Regulators will not provide the data to scholars.

Loans are difficult to follow because they are sold an re-sold, so samples with approval and performance data (more on this later) are very hard to put together.


The Need to Study Disparate-Impact Discrimination

People who want to practice disparate-treatment discrimination (but are prevented from doing so) can achieve virtually the same outcomes using disparate-impact discrimination.

A firm can rely on traits that are correlated with race to

“predict” a person’s race and then to construct rules that are unfavorable for people in a certain racial group.

This is particularly true in lending, because so many financial variables are highly correlated with race.


Disparate Impact in Automated Underwriting

An automated or statistically based underwriting system (or “scoring scheme”) uses sample of outstanding loans to explain loan performance, P (=default probability or loan profitability) on the basis of credit characteristics, X (the same as L, A, P from earlier slides).

Then it predicts loan performance for applicants based on their characteristics and this analysis.

No scheme is supposed to use membership in a protected class, M.


A Full-Information Scoring Scheme A full-information scheme estimates

Then calculates S1 = full-information loan score

Whenever γ is significant, this scoring scheme involves disparate-treatment discrimination because it considers M.

1

N

i ii

P X M

1

1

ˆˆ ˆN

i ii

S X M


A Non-Discriminatory Scoring Scheme

A non-discriminatory scoring scheme is:

where is the share of loans held by minority households.

This scheme is based on the same performance

regression as S1, but it does not differentiate between minority and white applicants.

2

1ˆ ) ˆˆ(

N

i ii

MS X

M


A Seemingly “Group-Neutral” Scoring Scheme Suppose a credit-scoring company tries to be

race-neutral by estimating

and using the “group-neutral” scoring scheme:

1

N

i ii

P X

3

1

ˆˆN

i ii

S X


Bias in the “Group-Neutral” Scheme

Using expected values, it is easy to show that this performance equation is biased by the omission of M.

where bi is the correlation between M and Xi.

This as an application of the standard formula for omitted variable bias.

3 2

1) )( ( ( )i i

N

ii

b XE S E S X


The Trouble with Lawyers This “group-neutral” scheme changes the

weight placed on each X, not on the basis of a business necessity, but instead on the basis of correlation with minority status.

Lawyers seem to conclude that this approach is neutral.

But, in fact, this is disparate-impact

discrimination.


Blending Performance and Time-of-Application Data This analysis leads to new, more precise tests for

discrimination, which have not yet been implemented, due to lack of data.

The most straightforward application is to loan approval, but the same logic applies to automated underwriting or credit scoring schemes, which have never been evaluated by scholars, and to loan pricing.

This approach picks up both disparate-treatment and disparate-impact discrimination.


A New Test for Discrimination

Step 1:

Obtain a sample of applications to many lenders; observe the number of approvals (A).

Step 2:

Obtain a sample of approved loans drawn from the same pool; observe their performance over time; estimate a loan-performance model using the best method available.


A New Test for Discrimination, 2

Step 3:

Use the model from step 2 to obtain a loan-performance score for each application in step 1.

Rank the applications in step 1 by their loan-performance score; identify the highest-ranking A applications.


A New Test for Discrimination,3

Step 4:

Compare the minority composition of the approved applications in step 1 with the minority composition of the highest-ranking applications (which might not be approved).

This difference is a measure of discrimination (and can

be evaluated with a difference-of-means test).

It cannot separate disparate-treatment and disparate-impact discrimination, but picks up both.


Accounting for Legitimate Variation Across Lenders

There are two ways to account for legitimate variation in underwriting standards across lenders:

1. Implement the above procedure with data for a single lender.

2. Estimate the above model with data pooled across lenders, with lender dummies, and with variables to describe a lender’s portfolio (so long as they predict within-group performance).


Extension to Loan Terms

This approach can be extended to cover loan terms, which is an important extension in a world with automated underwriting.

Estimate a model of loan performance as before.

Predict performance for a sample of new loans.

Regress interest rate (including points and fees) on predicted performance and group membership.


Extension to Loan Terms, 2

I recently came across a working paper that uses his approach for a large sample of loans from Florida and California in 2005.

“Race, Redlining, and Subprime Loan Pricing,” by A. C. Ghent, R. Hernandez-Murillo, and M. T. Owyang. Available at: http://www.public.asu.edu/~aghent/research/ .

This study finds evidence of discrimination in loan pricing, especially for loans originated by mortgage brokers.

http://www.public.asu.edu/~aghent/research/





Extension to Loan Terms, 3

The impact of discrimination on blacks and Hispanics in this study is up to 29 basis points, which translates into an increase in the monthly payment of about $60 or increase in the mortgage amount of about $20,000.

They also find evidence of redlining against minority neighborhoods, which is also illegal.

Mortgage Discrimination ECN741: Urban Economics, Professor Yinger.

Documents

Transcript of Mortgage Discrimination ECN741: Urban Economics, Professor Yinger.