Chapter 6.4 — Reader’s Guide: Fair Lending in the Algorithmic Era (Bartlett et al. 2022; Bhutta, Hizmo & Ringo)

Bartlett, R., Morse, A., Stanton, R., & Wallace, N. (2022). Consumer-Lending Discrimination in the FinTech Era. Journal of Financial Economics, 143(1), 30–56.

Bhutta, N., Hizmo, A., & Ringo, D. (2022). How Much Does Racial Bias Affect Mortgage Lending? Evidence from Human and Algorithmic Credit Decisions. Federal Reserve Finance and Economics Discussion Series (FEDS) Working Paper No. 2022-067. Subsequently published in the Journal of Finance (2025).

[Editor’s choice — not a [CHECK]: both venues above are verified. Cite the 2025 Journal of Finance version of record for a published reference list; cite the 2022 FEDS working paper if you want the open-access PDF link (federalreserve.gov, file 2022067pap.pdf) that students can read for free. Pick one consistently across the book.]

Maya has spent two weeks now circling the same question. In Mentor Session 4 she learned the hard way that a disparity is not a finding: that the gap between “minority borrowers are charged more” and “lenders discriminate against minority borrowers” is the entire intellectual content of fair lending, and that no list of controls can close it by itself. In Chapter 6.x she met Sun & Gao’s same-sex-borrower work, where HMDA’s lack of a credit score forced a design — not a regression — to carry the weight. Now she wants to know something that sounds like good news and may not be: if a computer makes the lending decision instead of a loan officer, does the discrimination go away? A machine has no face to read, no neighborhood prejudice, no discretion to abuse. Surely the algorithm is the cure.

This chapter reads two papers that take that hope seriously and refuse to flatter it. Bartlett, Morse, Stanton & Wallace ask whether FinTech lenders — who price algorithmically, with no human in the loop — charge minority borrowers more than they should. Bhutta, Hizmo & Ringo ask the companion question on the acceptance margin: of the raw racial gap in mortgage denials, how much is legitimate risk and how much is a residual that looks like bias? Read together they map the whole battlefield: pricing versus acceptance, machine versus human, raw gap versus residual. And read together they teach the lesson that the algorithmic era did not retire — it sharpened: the trick is never in the headline disparity, it is in what you were and were not able to hold fixed.

1. The research question

State each paper’s question as a question, naming the outcome and the key right-hand-side variable, exactly as the Week-5 template demands.

Bartlett et al.: Holding creditworthiness fixed, do Black and Hispanic borrowers pay higher mortgage interest rates than otherwise-identical white borrowers — and does it matter whether the loan was priced by an algorithm or by a human? The outcome is the interest rate (and, separately, the probability of rejection); the key regressor is borrower race/ethnicity; the moderator that makes the paper modern is whether the lender is a FinTech (algorithmic, no face-to-face contact) or a traditional lender. The hoped-for answer — algorithms are blind, so they do not discriminate — is exactly the hypothesis the paper sets out to test rather than assume.

Bhutta, Hizmo & Ringo (BHR): Of the large raw racial disparity in mortgage approval, what share is explained by legitimate, observable risk factors — and how much residual “differential treatment” survives once you can finally see the creditworthiness that earlier studies could not? The outcome is loan denial/approval (and, in places, pricing); the key regressor is race; the decisive new ingredient is a near-complete set of underwriting inputs, including the credit score that HMDA famously lacks, plus the verdict of a race-blind automated underwriting system (AUS) that tells you what a purely algorithmic decision would have been.

Notice how the two questions partition the problem. Bartlett et al. live on the pricing margin (conditional on getting a loan, what do you pay?); BHR live primarily on the acceptance margin (do you get the loan at all?). A complete fair-lending verdict needs both, because a lender can launder discrimination from one margin to the other — approve everyone but price the disfavored group higher, or price uniformly but deny more. And both papers put the algorithm itself on trial, which is what makes this a Week-6 chapter and not a re-run of Week 4.

2. The identification strategy

Both papers face the perennial threat Mentor 4 drilled into Maya: a raw disparity confounds the same rule applied to groups that differ in legitimate factors with a different rule applied to the same factors, and only the second is discrimination. Their identification strategies are two different bets on how to push past that confound.

Bartlett et al. — price comparisons under a strong conditioning set, plus the FinTech contrast. The workhorse is a regression of the loan’s rate on a minority indicator, conditioning on the risk variables a fair lender is allowed to use — crucially the GSE pricing inputs that mechanically determine what a loan “should” cost: credit score, loan-to-value (LTV), and so on. The identifying claim is sharp precisely because of the institutional setting: for conforming loans sold to Fannie Mae and Freddie Mac, the default risk is priced by a published formula in the loan-level price adjustments, so two loans with the same score and LTV carry the same credit risk to the investor by construction. Any systematic rate gap that survives is therefore hard to attribute to risk the lender legitimately priced — it points to something on the discrimination pathway, whether a human’s discretion or an algorithm’s proxy variable. The second piece of identification is the FinTech-versus-traditional contrast: if face-to-face discretion is the mechanism, removing the human (FinTech) should shrink the gap; if the gap persists even under algorithmic pricing, the mechanism is something the algorithm itself does — competing less hard in some markets, or pricing on geography that proxies for race.

This is where the over-controlling trap from Mentor 4 bites hardest, and Bartlett et al. are walking a knife’s edge they are largely aware of. Controlling for the GSE risk inputs is exactly right — those are legitimate, formula-priced risk. But if you also controlled for a variable that is itself a channel of discrimination — say, the specific product the borrower was steered into — you would regress away the very effect you are hunting (Mentor 4, warm-up 3). The credibility of the residual rate gap rests on the conditioning set being “all the legitimate risk and nothing on the discrimination pathway.” Name the threat in one sentence (CONVENTIONS §4): the rate gap equals differential treatment only as long as the GSE pricing inputs capture all risk the lender legitimately prices and no controls sit on the steering/discrimination channel.

BHR — close the omitted-variable gap by observation, then benchmark against the algorithm. Their bet is the opposite and just as instructive. Where Sun & Gao and the classic HMDA literature had to design around the missing credit score, BHR simply get the score (and much more) by linking to richer loan-level data. Their identifying logic: if the residual racial disparity collapses once you add the creditworthiness variables underwriters actually see, then much of what earlier studies called “unexplained” was omitted-variable bias all along — the OVB of Week 2, finally measured rather than feared. The masterstroke is the race-blind AUS benchmark: the government’s automated underwriting engine produces an approve/refer recommendation from risk inputs without seeing race, so it is a literal mechanical counterfactual for “what a rule that cannot discriminate would decide.” Comparing actual denials to AUS recommendations separates what the risk model says from what the human did on top of it.

Tie this back to the Week-2/Week-3/Week-4 spine. The whole exercise is the selection-on-observables logic of Week 3: both papers are betting they have conditioned on enough of creditworthiness that the residual race coefficient is causal. Week 2 named the danger (OVB from the unobserved score); BHR’s contribution is to shrink the unobserved set until the bias has nowhere to hide; Bartlett et al.’s contribution is to find an institutional setting (formula-priced GSE loans) where the legitimate-risk benchmark is unusually clean. Neither runs a DiD or an RD — and that is itself a reading lesson: in this literature, the identifying work is done by the conditioning set and the institutional benchmark, so the referee’s first question is always “what is in the controls, and is the score in there?”

3. The data

This is the section where the two papers most obviously complement each other, because they see different things.

Bartlett et al. combine HMDA — the Home Mortgage Disclosure Act near-census of U.S. mortgage applications (loan amount, applicant race/ethnicity and income, action taken, and, post-2018, pricing fields) — with GSE / loan-level data that supply the risk inputs and rate detail HMDA alone does not. The unit of observation is the loan (or application); the sample is large enough that small rate differences are estimable with precision. What this combination can see is the loan’s price and the formula inputs that should determine it; the FinTech flag lets them split the sample by lender technology. [CHECK: exact sample period and loan counts; whether the lead pricing analysis is GSE-conforming refinance/purchase loans.]

BHR also start from HMDA but link to GSE and credit-bureau / loan-level records, and the decisive difference from every HMDA-only study is that they observe the credit score and a near-complete underwriting file, plus the race-blind AUS recommendation. The unit is again the loan/application. What this dataset can see that HMDA cannot is precisely the variable whose absence forced Sun & Gao toward design: the applicant’s credit score and the full risk profile the underwriter weighed. [CHECK: exact data vintages, the AUS systems used (e.g., the GSEs’ automated engines), and sample years.]

The contrast is the whole pedagogical point and worth saying plainly to Maya. HMDA alone is blind to the credit score. That single omission is why Mentor 4 argued a design beats more controls in same-sex-borrower work, and why Sun & Gao reached for matched comparison and policy variation rather than a kitchen-sink regression. BHR’s data let them do the thing the HMDA-only researcher cannot: directly test how much of the residual disparity was the missing score in disguise. When you read either paper, locate immediately whether the credit score is in the data — because that one fact determines whether the residual race coefficient is a serious estimate of differential treatment or a placeholder for everything the data could not measure.

4. Table-by-table reading order

Read both papers the professional way — tables first, in the order the argument actually flows: headline disparity, then the controls/decomposition that shrink it, then the mechanism table that assigns blame to human or machine.

For Bartlett et al.:

The headline rate-gap table — read this first. Find the regression of interest rate on the minority indicator with the GSE risk controls in. Read the sign, the magnitude (a rate difference in basis points), and the t-stat off the table before any prose. This is the number that, if it broke, sinks the paper: minority borrowers pay a statistically and economically meaningful premium conditional on formula-priced risk. Do not memorize a precise figure — what is robust is that the conditional premium is positive and nontrivial, on the order of a handful of basis points. [CHECK: the exact headline rate premium in basis points for purchase/refi.]
The acceptance/rejection table — the second margin. Confirm the story is not only about price by reading the parallel rejection-probability result. Both margins matter for a complete verdict.
The FinTech-versus-traditional table — the mechanism. This is the table that earns the paper its title. Read whether the gap is smaller for algorithmic (FinTech) lenders than for face-to-face lenders. The robust finding: algorithmic lending reduces the disparity but does not eliminate it — discretion-based discrimination shrinks, yet a residual survives even when no human is in the loop, pointing to algorithmic pricing on variables that proxy for race or to weaker competition in some markets.

For BHR:

The raw disparity table — read this first. The large unconditional racial gap in denial (and the raw pricing gap). This is deliberately the “before” picture; it is meant to look alarming so the decomposition can do its work.
The decomposition / add-the-controls table — the heart. Watch the race coefficient march across columns as the credit score and the full risk set enter, exactly the left-to-right robustness read the Week-5 pack taught with Maya’s debt-to-income mock table. (The formal version of this “how much of the gap do the legitimate factors explain?” split is the Blinder–Oaxaca decomposition — Blinder, A. S. (1973), “Wage Discrimination: Reduced Form and Structural Estimates,” Journal of Human Resources, 8(4), 436–455; and Oaxaca, R. (1973), “Male–Female Wage Differentials in Urban Labor Markets,” International Economic Review, 14(3), 693–709 — which you build by hand in PS 6.4 and rebuild on HMDA in nb6.4.) The robust qualitative result: a meaningful share of the raw gap reflects legitimate risk factors (the coefficient shrinks substantially when the score goes in), but a contested residual remains. Read the bottom-row N across columns to be sure the sample is not silently changing as controls enter.
The AUS-benchmark / mechanism table — what the algorithm would have done. Compare actual outcomes to the race-blind automated recommendation. This is where “how much is bias” gets its sharpest answer: the residual, after benchmarking to a rule that cannot see race, is the paper’s estimate of differential treatment — smaller than the raw gap, and smaller than the unconditional HMDA literature implied.

Read across both papers and the synthesis writes itself: raw gaps are large; legitimate risk explains a lot; algorithms shrink the human discretion channel; a residual disparity survives on both margins and remains the contested core of the debate.

5. What’s clever

Each paper has one genuinely good idea, and naming it is how Maya learns to be clever herself.

Bartlett et al.’s cleverness is using a formula-priced market to manufacture a clean legitimate-risk benchmark, and then using technology as a natural split. Because GSE loans are priced off a published risk formula, the authors do not have to argue about what the “fair” price is — the institution tells them. That converts the impossible “did they price legitimate risk?” into the answerable “did they price more than the formula’s risk?” Then the FinTech-versus-human split turns the abstract worry “is the algorithm the cure?” into a measurable contrast: hold the question fixed and vary only whether a human was in the loop. It is the same instinct as a clean comparison in Week 4 — find a setting where the thing you want to hold fixed is fixed by the world.

BHR’s cleverness is twofold: getting the data that closes the OVB, and inventing a mechanical counterfactual out of the AUS. First, simply observing the credit score is a contribution, because it lets them measure the omitted-variable bias the whole HMDA literature could only fear — turning Mentor 4’s “you can never see the score” into “here is what happens to the coefficient when you finally can.” Second, the race-blind AUS is a found experiment: a real, deployed decision rule that by design cannot use race, serving as the “same-quality applicant, vary only the protected trait” counterfactual that Chapter 3.1 says you can never directly observe. They did not have to build the counterfactual in a lab — the mortgage system built one for them.

The deepest shared cleverness is separating pricing discrimination from acceptance discrimination using big loan-level data. Audit studies (Mentor 4’s Bertrand–Mullainathan résumés) randomize the trait but observe only a callback. These papers cannot randomize, but they observe millions of real loans on both margins, so they can ask where in the process the disparity lives. That decomposition — which margin, which lender technology, how much is risk — is something no small audit can deliver.

6. What’s vulnerable

Push back the way a referee will, and the soft spots are the same ones Mentor 4 promised would be attacked first.

Omitted creditworthiness — the perennial threat. For Bartlett et al., the residual rate gap is discrimination only if the GSE inputs capture all legitimate risk. If lenders price on a soft, unobserved risk signal correlated with race, part of the “premium” is real risk, and the coefficient is biased upward (Week 2 OVB, in the direction that flatters a discrimination finding). BHR partly answer this threat — that is their whole point — but cannot fully escape it: even a near-complete file omits the soft information a human underwriter gleans (the explanatory letter, the local knowledge of Mentor 4’s pre-read), so their residual is still a bound, not a clean causal effect.

The over-controlling trap. The mirror-image danger. If either paper conditions on a variable that lies on the discrimination pathway — a steered product type, a lender’s internal risk grade that itself encodes bias — the race coefficient shrinks for the wrong reason and understates discrimination (Mentor 4, warm-up 3). The honest reader checks every control and asks: confounder I should hold fixed, or channel I should not? With formula inputs, Bartlett et al. are on safe ground; the danger grows with every “richer” control added in the name of precision.

Definitional issues in “discrimination.” BHR’s race-blind-AUS benchmark measures disparate treatment — a different rule applied to the same inputs. But fair-lending law also recognizes disparate impact: a facially neutral rule (the AUS itself) that produces worse outcomes for a protected group can be unlawful if not justified by business necessity. A small residual after conditioning on the score is reassuring about disparate treatment and says little about disparate impact — and a credit score shaped by historical inequities may be a channel, not a clean control. The two papers can both be right and still disagree about whether the market is “fair,” because they are answering different legal questions. State which definition a result speaks to before you quote it.

External validity. Bartlett et al.’s FinTech contrast is tied to a particular set of lenders, products, and years; the GSE-conforming, formula-priced segment is the cleanest part of the market and may be the least discriminatory — jumbo, non-conforming, and subprime segments, where pricing is more discretionary, could look worse. BHR’s residual is specific to their data vintage and the AUS in force; both papers’ magnitudes should be read as of their era, not as constants. The qualitative lessons travel; the precise basis points and percentage-point gaps do not.

7. Three replication exercises

These are scaffolded toward nb6.4, the mortgage-disparity decomposition on HMDA. You will reproduce the spirit of both papers on data a student can actually touch — HMDA, which is public — keeping front-of-mind the one thing HMDA cannot show you: the credit score.

Reproduce a raw-then-conditional disparity decomposition (the BHR move, HMDA-limited). On a HMDA extract, regress the action-taken outcome (or, for recent years, the rate-spread field) on a minority indicator with no controls, then add the legitimate factors HMDA does carry — loan amount, applicant income, loan-to-value where available, and tract/lender fixed effects — watching the race coefficient march across columns exactly as in §4. Report how much of the raw gap the observables explain and how much residual survives. Then write the sentence that is the whole point: what is the one variable I could not add, and which way would adding it move the coefficient? (Answer: the credit score; almost certainly shrinks it.)
Stress the design the way Mentor 4 taught — over-control on purpose, and match instead of regress. Re-run Exercise 1 once adding a control that plausibly sits on the discrimination pathway (e.g., a loan-product or high-cost-loan flag) and show the residual shrink for the wrong reason — a hands-on demonstration of the over-controlling trap. Then build a matched comparison: pair each minority application to a non-minority application identical on income band, loan amount, LTV, and tract, and compare outcomes within pairs. Report what matching buys you over the linear regression, and name the one confound (the unobserved score, selection into applying) it still cannot fix.
Extend to the pricing margin and the technology split (the Bartlett move). Using HMDA’s post-2018 pricing fields, estimate the conditional rate-spread gap rather than the denial gap, then split the sample by lender type — flagging predominantly online/FinTech lenders versus traditional banks — and test whether the conditional gap is smaller where there is no human in the loop. You will be rebuilding Bartlett et al.’s headline contrast on public data: confirm, or fail to confirm, that algorithmic lending reduces but does not eliminate the disparity, and write down why your HMDA-only version is weaker than theirs (no GSE formula inputs, no credit score).

Where this goes next

Open nb6.4 (notebooks/week-06/nb6.4-mortgage-disparity-decomposition.ipynb), the mortgage-disparity decomposition, where you build the raw-then-conditional regression on HMDA, run the matched comparison, and read the residual disparity off your own output — confronting, line by line, the missing-credit-score problem that is the spine of this whole literature. The three exercises above are its backbone; the “Your Turn” extension pushes you toward the Capstone 1 design, a clean DiD/decomposition on HMDA built to a standard that could survive a hostile referee — and, as Sun & Gao’s same-sex-borrower work (Sun, H., & Gao, L. (2019), “Lending practices to same-sex borrowers,” PNAS, 116(19), 9293–9302) reminds you, a regulator, a journalist, and a court. That paper is the cohort’s anchor: it shows what it looks like to take exactly this question — are two otherwise-identical applicants treated differently? — from a HMDA extract all the way to Congressional testimony, precisely because the credit score was missing and a design had to carry the argument. Bartlett et al. and BHR are the algorithmic-era sequel: same question, new defendant.

Referee questions to carry into the lab and the discussion:

BHR’s residual disparity shrinks once the credit score enters, suggesting much of the raw gap is risk. But the credit score may itself encode historical discrimination. Is conditioning on it controlling for a legitimate risk factor, or controlling away a channel of disparate impact — and how would your answer change the policy verdict?
Bartlett et al. find algorithmic lending reduces but does not eliminate the rate gap. If the residual under FinTech lending comes from the algorithm pricing on geography that proxies for race, is that “discrimination,” and would removing the proxy fix it or merely hide it? What evidence would distinguish a biased input from a biased outcome?
Both papers identify off a conditioning set, not a natural experiment. If a state changed fair-lending enforcement at a known date, could you turn either design into the DiD of Chapter 4.1 — and what would parallel trends have to mean for a pricing gap rather than an approval gap?

# Chapter 6.4 — Reader's Guide: Fair Lending in the Algorithmic Era (Bartlett et al. 2022; Bhutta, Hizmo & Ringo) > **Bartlett, R., Morse, A., Stanton, R., & Wallace, N. (2022). Consumer-Lending Discrimination in the FinTech Era. *Journal of Financial Economics*, 143(1), 30–56.** > > **Bhutta, N., Hizmo, A., & Ringo, D. (2022). How Much Does Racial Bias Affect Mortgage Lending? Evidence from Human and Algorithmic Credit Decisions. *Federal Reserve Finance and Economics Discussion Series (FEDS) Working Paper No. 2022-067.* Subsequently published in the *Journal of Finance* (2025).** > > *[Editor's choice — not a `[CHECK]`: both venues above are verified. Cite the 2025 *Journal of Finance* version of record for a published reference list; cite the 2022 FEDS working paper if you want the open-access PDF link (`federalreserve.gov`, file `2022067pap.pdf`) that students can read for free. Pick one consistently across the book.]* Maya has spent two weeks now circling the same question. In Mentor Session 4 she learned the hard way that a disparity is not a finding: that the gap between "minority borrowers are charged more" and "lenders discriminate against minority borrowers" is the entire intellectual content of fair lending, and that no list of controls can close it by itself. In Chapter 6.x she met Sun & Gao's same-sex-borrower work, where HMDA's lack of a credit score forced a design — not a regression — to carry the weight. Now she wants to know something that sounds like good news and may not be: **if a computer makes the lending decision instead of a loan officer, does the discrimination go away?** A machine has no face to read, no neighborhood prejudice, no discretion to abuse. Surely the algorithm is the cure. This chapter reads two papers that take that hope seriously and refuse to flatter it. Bartlett, Morse, Stanton & Wallace ask whether FinTech lenders — who price algorithmically, with no human in the loop — charge minority borrowers more than they should. Bhutta, Hizmo & Ringo ask the companion question on the *acceptance* margin: of the raw racial gap in mortgage denials, how much is legitimate risk and how much is a residual that looks like bias? Read together they map the whole battlefield: pricing versus acceptance, machine versus human, raw gap versus residual. And read together they teach the lesson that the algorithmic era did not retire — it sharpened: *the trick is never in the headline disparity, it is in what you were and were not able to hold fixed.* --- ## 1. The research question State each paper's question as a question, naming the outcome and the key right-hand-side variable, exactly as the Week-5 template demands. **Bartlett et al.:** *Holding creditworthiness fixed, do Black and Hispanic borrowers pay higher mortgage interest rates than otherwise-identical white borrowers — and does it matter whether the loan was priced by an algorithm or by a human?* The outcome is the **interest rate** (and, separately, the probability of *rejection*); the key regressor is borrower **race/ethnicity**; the moderator that makes the paper modern is whether the lender is a **FinTech** (algorithmic, no face-to-face contact) or a traditional lender. The hoped-for answer — algorithms are blind, so they do not discriminate — is exactly the hypothesis the paper sets out to test rather than assume. **Bhutta, Hizmo & Ringo (BHR):** *Of the large raw racial disparity in mortgage approval, what share is explained by legitimate, observable risk factors — and how much residual "differential treatment" survives once you can finally see the creditworthiness that earlier studies could not?* The outcome is **loan denial/approval** (and, in places, pricing); the key regressor is **race**; the decisive new ingredient is a near-complete set of underwriting inputs, including the **credit score** that HMDA famously lacks, plus the verdict of a **race-blind automated underwriting system (AUS)** that tells you what a purely algorithmic decision *would* have been. Notice how the two questions partition the problem. Bartlett et al. live on the **pricing** margin (conditional on getting a loan, what do you pay?); BHR live primarily on the **acceptance** margin (do you get the loan at all?). A complete fair-lending verdict needs both, because a lender can launder discrimination from one margin to the other — approve everyone but price the disfavored group higher, or price uniformly but deny more. And both papers put the *algorithm itself* on trial, which is what makes this a Week-6 chapter and not a re-run of Week 4. --- ## 2. The identification strategy Both papers face the perennial threat Mentor 4 drilled into Maya: a raw disparity confounds *the same rule applied to groups that differ in legitimate factors* with *a different rule applied to the same factors*, and only the second is discrimination. Their identification strategies are two different bets on how to push past that confound. **Bartlett et al. — price comparisons under a strong conditioning set, plus the FinTech contrast.** The workhorse is a regression of the loan's rate on a minority indicator, conditioning on the risk variables a fair lender is *allowed* to use — crucially the GSE pricing inputs that mechanically determine what a loan "should" cost: credit score, loan-to-value (LTV), and so on. The identifying claim is sharp precisely because of the institutional setting: for conforming loans sold to Fannie Mae and Freddie Mac, the **default risk is priced by a published formula** in the loan-level price adjustments, so two loans with the same score and LTV carry the same credit risk to the investor *by construction*. Any systematic rate gap that survives is therefore hard to attribute to risk the lender legitimately priced — it points to something on the discrimination pathway, whether a human's discretion or an algorithm's proxy variable. The second piece of identification is the **FinTech-versus-traditional contrast**: if face-to-face discretion is the mechanism, removing the human (FinTech) should shrink the gap; if the gap persists even under algorithmic pricing, the mechanism is something the algorithm itself does — competing less hard in some markets, or pricing on geography that proxies for race. This is where the **over-controlling trap** from Mentor 4 bites hardest, and Bartlett et al. are walking a knife's edge they are largely aware of. Controlling for the GSE risk inputs is *exactly right* — those are legitimate, formula-priced risk. But if you also controlled for a variable that is *itself* a channel of discrimination — say, the specific product the borrower was steered into — you would regress away the very effect you are hunting (Mentor 4, warm-up 3). The credibility of the residual rate gap rests on the conditioning set being "all the legitimate risk and nothing on the discrimination pathway." Name the threat in one sentence (CONVENTIONS §4): *the rate gap equals differential treatment only as long as the GSE pricing inputs capture all risk the lender legitimately prices and no controls sit on the steering/discrimination channel.* **BHR — close the omitted-variable gap by observation, then benchmark against the algorithm.** Their bet is the opposite and just as instructive. Where Sun & Gao and the classic HMDA literature had to *design around* the missing credit score, BHR simply *get the score* (and much more) by linking to richer loan-level data. Their identifying logic: if the residual racial disparity collapses once you add the creditworthiness variables underwriters actually see, then much of what earlier studies called "unexplained" was **omitted-variable bias** all along — the OVB of Week 2, finally measured rather than feared. The masterstroke is the **race-blind AUS benchmark**: the government's automated underwriting engine produces an approve/refer recommendation from risk inputs *without* seeing race, so it is a literal mechanical counterfactual for "what a rule that cannot discriminate would decide." Comparing actual denials to AUS recommendations separates *what the risk model says* from *what the human did on top of it*. Tie this back to the Week-2/Week-3/Week-4 spine. The whole exercise is the selection-on-observables logic of Week 3: both papers are betting they have conditioned on enough of creditworthiness that the residual race coefficient is causal. Week 2 named the danger (OVB from the unobserved score); BHR's contribution is to *shrink the unobserved set* until the bias has nowhere to hide; Bartlett et al.'s contribution is to find an institutional setting (formula-priced GSE loans) where the legitimate-risk benchmark is unusually clean. Neither runs a DiD or an RD — and that is itself a reading lesson: in this literature, the identifying work is done by the **conditioning set and the institutional benchmark**, so the referee's first question is always "what is in the controls, and is the score in there?" --- ## 3. The data This is the section where the two papers most obviously complement each other, because they see different things. **Bartlett et al.** combine **HMDA** — the Home Mortgage Disclosure Act near-census of U.S. mortgage applications (loan amount, applicant race/ethnicity and income, action taken, and, post-2018, pricing fields) — with **GSE / loan-level data** that supply the risk inputs and rate detail HMDA alone does not. The unit of observation is the **loan (or application)**; the sample is large enough that small rate differences are estimable with precision. What this combination *can* see is the loan's price and the formula inputs that should determine it; the FinTech flag lets them split the sample by lender technology. [CHECK: exact sample period and loan counts; whether the lead pricing analysis is GSE-conforming refinance/purchase loans.] **BHR** also start from **HMDA** but link to **GSE and credit-bureau / loan-level records**, and the decisive difference from every HMDA-only study is that they observe the **credit score** and a near-complete underwriting file, *plus* the **race-blind AUS recommendation**. The unit is again the loan/application. What this dataset can see that HMDA cannot is precisely the variable whose absence forced Sun & Gao toward design: the applicant's credit score and the full risk profile the underwriter weighed. [CHECK: exact data vintages, the AUS systems used (e.g., the GSEs' automated engines), and sample years.] The contrast is the whole pedagogical point and worth saying plainly to Maya. **HMDA alone is blind to the credit score.** That single omission is why Mentor 4 argued a *design* beats more controls in same-sex-borrower work, and why Sun & Gao reached for matched comparison and policy variation rather than a kitchen-sink regression. BHR's data let them do the thing the HMDA-only researcher *cannot*: directly test how much of the residual disparity was the missing score in disguise. When you read either paper, locate immediately *whether the credit score is in the data* — because that one fact determines whether the residual race coefficient is a serious estimate of differential treatment or a placeholder for everything the data could not measure. --- ## 4. Table-by-table reading order Read both papers the professional way — tables first, in the order the argument actually flows: headline disparity, then the controls/decomposition that shrink it, then the mechanism table that assigns blame to human or machine. **For Bartlett et al.:** 1. **The headline rate-gap table — read this first.** Find the regression of interest rate on the minority indicator with the GSE risk controls in. Read the sign, the magnitude (a rate difference in *basis points*), and the t-stat *off the table* before any prose. This is the number that, if it broke, sinks the paper: minority borrowers pay a statistically and economically meaningful premium conditional on formula-priced risk. *Do not memorize a precise figure* — what is robust is that the conditional premium is positive and nontrivial, on the order of a handful of basis points. [CHECK: the exact headline rate premium in basis points for purchase/refi.] 2. **The acceptance/rejection table — the second margin.** Confirm the story is not only about price by reading the parallel rejection-probability result. Both margins matter for a complete verdict. 3. **The FinTech-versus-traditional table — the mechanism.** This is the table that earns the paper its title. Read whether the gap is *smaller* for algorithmic (FinTech) lenders than for face-to-face lenders. The robust finding: **algorithmic lending reduces the disparity but does not eliminate it** — discretion-based discrimination shrinks, yet a residual survives even when no human is in the loop, pointing to algorithmic pricing on variables that proxy for race or to weaker competition in some markets. **For BHR:** 1. **The raw disparity table — read this first.** The large unconditional racial gap in denial (and the raw pricing gap). This is deliberately the "before" picture; it is meant to look alarming so the decomposition can do its work. 2. **The decomposition / add-the-controls table — the heart.** Watch the race coefficient *march across columns* as the credit score and the full risk set enter, exactly the left-to-right robustness read the Week-5 pack taught with Maya's debt-to-income mock table. (The formal version of this "how much of the gap do the legitimate factors explain?" split is the **Blinder–Oaxaca decomposition** — Blinder, A. S. (1973), "Wage Discrimination: Reduced Form and Structural Estimates," *Journal of Human Resources*, 8(4), 436–455; and Oaxaca, R. (1973), "Male–Female Wage Differentials in Urban Labor Markets," *International Economic Review*, 14(3), 693–709 — which you build by hand in PS 6.4 and rebuild on HMDA in nb6.4.) The robust qualitative result: **a meaningful share of the raw gap reflects legitimate risk factors** (the coefficient shrinks substantially when the score goes in), **but a contested residual remains.** Read the bottom-row $N$ across columns to be sure the sample is not silently changing as controls enter. 3. **The AUS-benchmark / mechanism table — what the algorithm would have done.** Compare actual outcomes to the race-blind automated recommendation. This is where "how much is bias" gets its sharpest answer: the residual, after benchmarking to a rule that *cannot* see race, is the paper's estimate of differential treatment — smaller than the raw gap, and smaller than the unconditional HMDA literature implied. Read across both papers and the synthesis writes itself: raw gaps are large; legitimate risk explains a lot; algorithms shrink the human discretion channel; a residual disparity survives on *both* margins and remains the contested core of the debate. --- ## 5. What's clever Each paper has one genuinely good idea, and naming it is how Maya learns to be clever herself. **Bartlett et al.'s cleverness is using a formula-priced market to manufacture a clean legitimate-risk benchmark, and then using technology as a natural split.** Because GSE loans are priced off a *published* risk formula, the authors do not have to argue about what the "fair" price is — the institution tells them. That converts the impossible "did they price legitimate risk?" into the answerable "did they price *more* than the formula's risk?" Then the FinTech-versus-human split turns the abstract worry "is the algorithm the cure?" into a measurable contrast: hold the question fixed and vary only whether a human was in the loop. It is the same instinct as a clean comparison in Week 4 — find a setting where the thing you want to hold fixed *is* fixed by the world. **BHR's cleverness is twofold: getting the data that closes the OVB, and inventing a mechanical counterfactual out of the AUS.** First, simply observing the credit score is a contribution, because it lets them *measure* the omitted-variable bias the whole HMDA literature could only fear — turning Mentor 4's "you can never see the score" into "here is what happens to the coefficient when you finally can." Second, the race-blind AUS is a found experiment: a real, deployed decision rule that by design cannot use race, serving as the "same-quality applicant, vary only the protected trait" counterfactual that Chapter 3.1 says you can never directly observe. They did not have to build the counterfactual in a lab — the mortgage system built one for them. The deepest shared cleverness is **separating pricing discrimination from acceptance discrimination using big loan-level data.** Audit studies (Mentor 4's Bertrand–Mullainathan résumés) randomize the trait but observe only a callback. These papers cannot randomize, but they observe *millions* of real loans on *both* margins, so they can ask where in the process the disparity lives. That decomposition — which margin, which lender technology, how much is risk — is something no small audit can deliver. --- ## 6. What's vulnerable Push back the way a referee will, and the soft spots are the same ones Mentor 4 promised would be attacked first. **Omitted creditworthiness — the perennial threat.** For Bartlett et al., the residual rate gap is discrimination *only if* the GSE inputs capture all legitimate risk. If lenders price on a soft, unobserved risk signal correlated with race, part of the "premium" is real risk, and the coefficient is biased upward (Week 2 OVB, in the direction that flatters a discrimination finding). BHR partly *answer* this threat — that is their whole point — but cannot fully escape it: even a near-complete file omits the soft information a human underwriter gleans (the explanatory letter, the local knowledge of Mentor 4's pre-read), so their residual is still a *bound*, not a clean causal effect. **The over-controlling trap.** The mirror-image danger. If either paper conditions on a variable that lies *on* the discrimination pathway — a steered product type, a lender's internal risk grade that itself encodes bias — the race coefficient shrinks for the wrong reason and *understates* discrimination (Mentor 4, warm-up 3). The honest reader checks every control and asks: confounder I should hold fixed, or channel I should not? With formula inputs, Bartlett et al. are on safe ground; the danger grows with every "richer" control added in the name of precision. **Definitional issues in "discrimination."** BHR's race-blind-AUS benchmark measures *disparate treatment* — a different rule applied to the same inputs. But fair-lending law also recognizes **disparate impact**: a facially neutral rule (the AUS itself) that produces worse outcomes for a protected group can be unlawful if not justified by business necessity. A small residual after conditioning on the score is reassuring about disparate *treatment* and says little about disparate *impact* — and a credit score shaped by historical inequities may be a channel, not a clean control. The two papers can both be right and still disagree about whether the market is "fair," because they are answering different legal questions. State which definition a result speaks to before you quote it. **External validity.** Bartlett et al.'s FinTech contrast is tied to a particular set of lenders, products, and years; the GSE-conforming, formula-priced segment is the cleanest part of the market and may be the *least* discriminatory — jumbo, non-conforming, and subprime segments, where pricing is more discretionary, could look worse. BHR's residual is specific to their data vintage and the AUS in force; both papers' magnitudes should be read as *of their era*, not as constants. The qualitative lessons travel; the precise basis points and percentage-point gaps do not. --- ## 7. Three replication exercises These are scaffolded toward **nb6.4**, the mortgage-disparity decomposition on HMDA. You will reproduce the *spirit* of both papers on data a student can actually touch — HMDA, which is public — keeping front-of-mind the one thing HMDA cannot show you: the credit score. 1. **Reproduce a raw-then-conditional disparity decomposition (the BHR move, HMDA-limited).** On a HMDA extract, regress the action-taken outcome (or, for recent years, the rate-spread field) on a minority indicator with *no controls*, then add the legitimate factors HMDA *does* carry — loan amount, applicant income, loan-to-value where available, and tract/lender fixed effects — watching the race coefficient march across columns exactly as in §4. Report how much of the raw gap the observables explain and how much residual survives. Then write the sentence that is the whole point: *what is the one variable I could not add, and which way would adding it move the coefficient?* (Answer: the credit score; almost certainly shrinks it.) 2. **Stress the design the way Mentor 4 taught — over-control on purpose, and match instead of regress.** Re-run Exercise 1 once *adding* a control that plausibly sits on the discrimination pathway (e.g., a loan-product or high-cost-loan flag) and show the residual shrink for the wrong reason — a hands-on demonstration of the over-controlling trap. Then build a **matched comparison**: pair each minority application to a non-minority application identical on income band, loan amount, LTV, and tract, and compare outcomes within pairs. Report what matching buys you over the linear regression, and name the one confound (the unobserved score, selection into applying) it still cannot fix. 3. **Extend to the pricing margin and the technology split (the Bartlett move).** Using HMDA's post-2018 pricing fields, estimate the *conditional rate-spread* gap rather than the denial gap, then split the sample by lender type — flagging predominantly online/FinTech lenders versus traditional banks — and test whether the conditional gap is smaller where there is no human in the loop. You will be rebuilding Bartlett et al.'s headline contrast on public data: confirm, or fail to confirm, that algorithmic lending *reduces but does not eliminate* the disparity, and write down why your HMDA-only version is weaker than theirs (no GSE formula inputs, no credit score). --- ### Where this goes next Open **`nb6.4`** (`notebooks/week-06/nb6.4-mortgage-disparity-decomposition.ipynb`), the mortgage-disparity decomposition, where you build the raw-then-conditional regression on HMDA, run the matched comparison, and read the residual disparity off your own output — confronting, line by line, the missing-credit-score problem that is the spine of this whole literature. The three exercises above are its backbone; the "Your Turn" extension pushes you toward the **Capstone 1** design, a clean DiD/decomposition on HMDA built to a standard that could survive a hostile referee — and, as Sun & Gao's same-sex-borrower work (Sun, H., & Gao, L. (2019), "Lending practices to same-sex borrowers," *PNAS*, 116(19), 9293–9302) reminds you, a regulator, a journalist, and a court. That paper is the cohort's anchor: it shows what it looks like to take exactly this question — *are two otherwise-identical applicants treated differently?* — from a HMDA extract all the way to Congressional testimony, precisely *because* the credit score was missing and a design had to carry the argument. Bartlett et al. and BHR are the algorithmic-era sequel: same question, new defendant. **Referee questions to carry into the lab and the discussion:** 1. BHR's residual disparity shrinks once the credit score enters, suggesting much of the raw gap is risk. But the credit score may itself encode historical discrimination. Is conditioning on it controlling for a legitimate risk factor, or controlling away a *channel* of disparate impact — and how would your answer change the policy verdict? 2. Bartlett et al. find algorithmic lending reduces but does not eliminate the rate gap. If the residual under FinTech lending comes from the algorithm pricing on geography that proxies for race, is that "discrimination," and would removing the proxy fix it or merely hide it? What evidence would distinguish a biased input from a biased outcome? 3. Both papers identify off a conditioning set, not a natural experiment. If a state changed fair-lending enforcement at a known date, could you turn either design into the DiD of Chapter 4.1 — and what would parallel trends have to mean for a *pricing* gap rather than an approval gap?