A Common Fair Lending Misunderstanding
Fair Lending is an endlessly fascinating topic for us at Visible Equity. The main goal of Fair Lending is to determine through data whether or not evidence of disparate treatment exists within a portfolio for a protected class of borrowers. One of the best ways to meet this goal is through Visible Equity’s Pricing Discrimination report in its Fair Lending module.
We’re going to go over a scenario we often see that people struggle to understand. Above is an example of what you might see in a pricing discrimination table in Fair Lending (the values in this table were completely made up). In the table we have the races/groups along the left: black, white, Asian/Pacific Islander, American Indian, Hispanic, not reported, and overall. Next, we have applicable loan count. This column consists of the number of loans for each group that are used in the Fair Lending calculations. Some loans are not included typically due to missing data, such as credit scores or original loan amounts. The next column is the average interest rate for each group followed by the average expected interest rate. Average expected interest rate is a special column. Expected interest rate is a Visible Equity model built to estimate what we would expect a borrower’s interest rate to be for a given loan based mostly on the borrower’s lending attributes: credit score, original loan amount, loan to value, etc. The only attribute used in the model not based on the borrower is a market variable to capture economic effects at the time of origination. It’s crucial to understand that race is not taken into account with the expected interest rate model. Again, it’s essentially built only using borrower attributes related to lending. The final column is the most important column. This is the difference between averages (DBA), which is simply the difference between the previous two columns. We’ll talk more about this column later.
Something else to notice is the coloring of each cell for the races. We won’t go into great detail as to how these colors are determined (for that we suggest you read our Fair Lending white paper). The important takeaway is that groups in green are considered to not show evidence of disparate treatment, groups in red do show evidence of disparate treatment, groups in yellow show evidence of favoritism, and groups in gray do not have enough applicable loans to carry out the calculations that determine if the group should be green, red, or yellow (or the tests used to determine the color do not agree).
And now for the misunderstanding. Based on this table, a common question is, “How could the black group be experiencing favoritism if their average expected interest rate (7.09%) is higher than the white group’s (6.75%) who aren’t experiencing favoritism or disparate treatment?” Or even, “How could the Asian/Pacific Islander group be experiencing disparate treatment if their average expected interest rate (6.63%) is lower than the white group’s?” The answer is that we shouldn’t even be paying attention to the expected interest rate column! These questions are comparing the wrong averages. The column that truly matters is the last column. This DBA column is the key to pricing discrimination analysis.
Instead of comparing average expected interest rates of one race with another, we need to compare DBA. Not only that, we should not be comparing DBAs between races; we should compare each race’s DBA with the overall DBA in the bottom right of the table (2.08%).
Recall that the expected interest rate model is built using mostly borrower lending attributes and that race is never taken into consideration. So, in essence, when we find DBA by subtracting off the expected interest rate estimate from the actual interest rate, we are left with the portion of the interest rate not explained by the borrower attributes listed previously. Now we have these residuals for each race and for the overall population. We want to see how these residuals for each group compare with the overall residuals. We won’t go into detail for the calculation, but know that we determine the green/red/yellow status of a group by comparing the group’s DBA with the overall DBA. Let’s look at a plot to further drive this point home.
Here we have the distributions of differences for each group to hopefully give you a better idea of what’s going on when determining the green/red/yellow status (note that this plot is on a decimal scale as opposed to a percentage). The dashed black lines represent the DBA for each respective group. The solid black line going down the middle is the overall DBA. Again, we want to compare each group’s DBA with the overall DBA. Any group whose DBA falls within the green section is not significantly different from the group DBA (white and American Indian). If a group DBA falls within the yellow section, then we consider this evidence for favoritism (black and Hispanic). If a group DBA falls within the red section, then we consider this evidence for disparate treatment (Asian/Pacific Islander). In the yellow and red cases, we carry out further statistical testing to determine if the group DBA is significantly different from the overall DBA. If we do in fact determine statistical significance, we suggest further investigation on the institution’s end.
So the big takeaway is when looking at a pricing discrimination report for Fair Lending, do not compare average expected interest rates amongst races. Instead, compare each race’s DBA with the overall DBA.