In this article I will argue against the view racial bias in the modern economy negatively impacts black people’s incomes, unemployment rates, educational opportunity, and ability to get a loan. Differences in Income Some think racism is necessary to explain group differences in income. But blacks actually make as much or more money than whites when education, cognitive skill, marital status and other confounding variables are controlled for. Such has been shown multiple times (
Article or tweet topic suggestion: the claim that NBA referees are more likely to call fouls against black players. Commonly cited paper: Price, Wolfers 2010
The law studies are trash too. Using odds ratios to measure association are misleading when event rates are not rare. The report states the odds ratios for Black vs White applicants are over 1000 to 1. However, the admission rates for Black applicants were 23-27%, which are not rare events. Odds ratios tend to overstate associations when event rates are common. As the report notes: "The meaning of logistic regression equations and their associated odds ratios may be difficult to grasp because the equations are complex and hard to explain without resorting to mathematical formulations." The analysis does not appear to check model assumptions or evaluate goodness-of-fit. The report does not mention checking assumptions like linearity, multicollinearity, outliers etc. It also does not report any goodness-of-fit statistics. Without evaluating model fit, it is unclear if the logistic regression model accurately represents the data. As the report states: "Statistical calculations always include what is called a p-value. When results are deemed to be statistically significant, this means that the calculated p-value is less than some pre-determined cutoff level of significance." But p-values alone do not indicate good model fit. The independent variables in the model are inadequate. The logistic regression only included LSAT scores, GPA, gender, residency status and race/ethnicity. Many other relevant variables like family income, high school grades, extracurricular activities, personal statements etc. were omitted. With omitted variable bias, coefficients for included variables like race/ethnicity are overestimated.
So the Bhutta and Hizmo study is not that great of evidence. It assumes that race is ignorable or as-if randomly assigned after conditioning on credit score, income, and lender. However, this assumption of no unmeasured confounding is highly unlikely to hold in this context. Race involves many complex factors beyond just skin color that can influence outcomes.
Characteristics like dialect, cultural knowledge, beliefs, and experiences that correlate with race are not fully captured by controls like income. It's been demonstrated through omitted variable bias formulas, any unobserved variable that differs across race and affects the outcome will invalidate estimates. In the mortgage lending context, factors like financial literacy, ability to navigate bureaucracy, neighborhood and family wealth levels, social connections, and perceptions of applicants are likely associated with race and also influence lending decisions. But the Bhutta and Hizmo study cannot claim these factors are balanced across racial groups or fully captured by credit scores. By assuming away unmeasured confounding, the paper essentially claims race is randomly assigned within income level and lender. But race involves far more than skin color, so this assumption fails. Differences between minority and white borrowers on unmeasured factors related to race still exist. Ignorability does not hold in this setting. Without ignorability, the paper's regression estimates will be biased by omitted variable bias. Claims about no racial discrimination after controlling for income and lender are unfounded, since race still correlates with many other factors influencing mortgage lending. The models fail to adequately adjust for critical unobservables connected to race. It also only analyzes approved mortgage borrowers, conditioning on loan applicants successfully receiving a mortgage. This induces collider bias and leads to an unrepresentative sample. They even suggest race affects the initial loan application decision: "One important margin where racial bias could affect outcomes for prospective minority borrowers is in service quality...in responding to mortgage inquiries" and cites evidence of discrimination in response rates to mortgage inquiries." By conditioning on approved loans, the paper essentially restricts analysis to that. This creates a non-representative sample, as minorities approved differ from those rejected. Comparing approved minorities and whites does not accuratelv reflect population differences. For example, suppose lenders reject risky minority applicants but accept risky white applicants. By studying only approvals, risky whites are compared to only the least risky minorities. This makes minorities appear less risky, masking discrimination. The paper's estimates on this selective sample tell us nothing about population-level differences, only differences in the distorted approvals subsample. Collider bias renders approved minorities and whites incomparable. By assuming loan approval is unaffected by race, the paper induces significant bias. Claims about no pricing discrimination on approvals alone are hence invalid. Analyses must account for sample selection processes and discrimination in initial application decisions. Bhutta and Hizmo also assume race has a constant effect across individuals and contexts.
But, race is multidimensional, so its impact likely varies significantly across settings. The paper estimates average differences in interest rates and points paid between minorities and whites.
But these average effects poorly represent heterogeneous effects across borrowers and lenders. For example, one lender discriminates only against low-income minorities, while another discriminates based on neighborhood demographics. Estimated average effects masks variation in discrimination. Similarly, discrimination differs depending on borrower attributes like financial knowledge, dress, speech patterns, etc. Lenders treat minorities with certain qualities differently. Average effects fail to capture such nuances. By modeling race as having a uniform impact, the paper incorrectly assumes away effect heterogeneity. If discrimination manifests only in certain contexts, the model will underestimate its pervasiveness. The paper cannot conclusively claim no pricing discrimination overall when effects likely exhibit significant heterogeneity across lenders, borrowers, neighborhoods, and originator policies. Average effects poorly represent variable treatment effects across contexts. Claims about average impacts oversimplif a complex phenomenon. Bhutta and Hizmo also assume no interference between units, but discrimination involves spillovers where the outcomes for one borrower affect others. Again it says "One important margin Where racial bias could affect outcomes for prospective minority borrowers is in the loan application accept/reject decision" and cites evidence of discrimination in application denials.
This implies the number of minorities approved affects outcomes for other minority applicants through spillovers. For example, if lenders aim to limit total minority approvals, then denying one minority borrower increases the chance of approving the next. Minorities face negative externalities from other minorities' outcomes. By assuming SUTVA holds, the paper rules out these spillovers. But discriminatory application decisions create interference across borrowers. Lenders manage overall minority approval rates, not just individual decisions. Models assuming no interference misrepresent the nature of discrimination. If lenders limit total minority approvals, pricing outcomes for one minority depend on outcomes for other minorities. This interconnectivity and strategic behavior is ignored. The presence of spillovers means the stable unit treatment value assumption fails. The paper's inferences relying on SUTVA are hence invalid. Analyses must account for interference and interdependencies created by lenders managing minority approval rates. Bhutta and Hizmo also assume race is manipulable, but race is an immutable characteristic. This raises philosophical questions about the meaning of causal effects of race. The paper attempts to estimate counterfactual differences in mortgage pricing if a minority borrower were white. But race is fixed, so this counterfactual is not clearly defined. Race shapes lifetime experiences and opportunities, so it is not possible to instantaneously change race. Claiming to estimate the effect of making someone "white" lacks concrete meaning. Race is multifaceted, involving culture, beliefs, skills, and more. Isolating skin color alone as a manipulable treatment ignores this complexity. By modeling race as manipulable, the paper assumes away deep conceptual issues about the nature of race. At best, the paper reveals pricing differences correlated with race. Claims about causal effects of race itself are on shaky philosophical ground given the inability to physically manipulate race. The paper does not wrestle with this inherent difficulty in studying race effects. Bhutta and Hizmo also assume causal sufficiency - that the model fully accounts for the treatment assignment process. But researchers cannot credibly claim full control over the treatment of race. Race is complex, involving an interrelated bundle of factors like skin tone, culture, beliefs, etc. Isolating and manipulating any one dimension independently is impossible. The paper's models assume race can be specified and assigned independently of other factors. This falsely suggests race is simple and fully accounted for in the analysis. In reality, race cannot be randomized or cleanly specified. Models attempting to estimate race effects fail to address this lack of researcher control. By assuming causal sufficiency, the paper claims to fully model race and its assignment. But race is far more intricate than any model can capture. Researchers fundamentally lack control over race's many entangled facets. Models pretending to isolate race alone are hence oversimplified. The paper does not wrestle with this inherent lack of causal sufficiency when studying race. Claims about estimating causal effects of race require strong and unfounded assumptions about researcher control over race. There examples of the naive estimates that suffer from bias in the paper. In the Data section, the paper states: "In the HMDA data, lenders must report the APR - a single measure of loan cost that is a function of the interest rate, points, fees, and insurance premiums –if the APR exceeds a threshold." So estimates of fees/points are only available for a subset of borrowers. The Data section also states: "We obtained administrative data from the Federal Housing Administration (FHA) covering the universe of FHA-insured home purchase loans originated in 2014 and 2015." However, this means the data misses discrimination in non-FHA lending. In the Results, the paper finds a black-white interest rate gap of 0.03% and Hispanic-white gap of 0.015% after controlling for lender and risk factors. However, the naive model likely suffers from omitted variable bias and does not address unobserved differences across race. The Results also state "we find no evidence of systematic discrimination in fees by race or ethnicity." But again, this conclusion is based only on partial fee data for FHA borrowers, ignoring selection issues. The data limitations and narrowly specified models mean the paper's estimates are naive. The conclusions ignore bias from unobservables and selectively analyzed data. More complex causal analyses are needed to make credible claims about discrimination.
"Comment for the algorithm."
Article or tweet topic suggestion: the claim that NBA referees are more likely to call fouls against black players. Commonly cited paper: Price, Wolfers 2010
The law studies are trash too. Using odds ratios to measure association are misleading when event rates are not rare. The report states the odds ratios for Black vs White applicants are over 1000 to 1. However, the admission rates for Black applicants were 23-27%, which are not rare events. Odds ratios tend to overstate associations when event rates are common. As the report notes: "The meaning of logistic regression equations and their associated odds ratios may be difficult to grasp because the equations are complex and hard to explain without resorting to mathematical formulations." The analysis does not appear to check model assumptions or evaluate goodness-of-fit. The report does not mention checking assumptions like linearity, multicollinearity, outliers etc. It also does not report any goodness-of-fit statistics. Without evaluating model fit, it is unclear if the logistic regression model accurately represents the data. As the report states: "Statistical calculations always include what is called a p-value. When results are deemed to be statistically significant, this means that the calculated p-value is less than some pre-determined cutoff level of significance." But p-values alone do not indicate good model fit. The independent variables in the model are inadequate. The logistic regression only included LSAT scores, GPA, gender, residency status and race/ethnicity. Many other relevant variables like family income, high school grades, extracurricular activities, personal statements etc. were omitted. With omitted variable bias, coefficients for included variables like race/ethnicity are overestimated.
So the Bhutta and Hizmo study is not that great of evidence. It assumes that race is ignorable or as-if randomly assigned after conditioning on credit score, income, and lender. However, this assumption of no unmeasured confounding is highly unlikely to hold in this context. Race involves many complex factors beyond just skin color that can influence outcomes.
Characteristics like dialect, cultural knowledge, beliefs, and experiences that correlate with race are not fully captured by controls like income. It's been demonstrated through omitted variable bias formulas, any unobserved variable that differs across race and affects the outcome will invalidate estimates. In the mortgage lending context, factors like financial literacy, ability to navigate bureaucracy, neighborhood and family wealth levels, social connections, and perceptions of applicants are likely associated with race and also influence lending decisions. But the Bhutta and Hizmo study cannot claim these factors are balanced across racial groups or fully captured by credit scores. By assuming away unmeasured confounding, the paper essentially claims race is randomly assigned within income level and lender. But race involves far more than skin color, so this assumption fails. Differences between minority and white borrowers on unmeasured factors related to race still exist. Ignorability does not hold in this setting. Without ignorability, the paper's regression estimates will be biased by omitted variable bias. Claims about no racial discrimination after controlling for income and lender are unfounded, since race still correlates with many other factors influencing mortgage lending. The models fail to adequately adjust for critical unobservables connected to race. It also only analyzes approved mortgage borrowers, conditioning on loan applicants successfully receiving a mortgage. This induces collider bias and leads to an unrepresentative sample. They even suggest race affects the initial loan application decision: "One important margin where racial bias could affect outcomes for prospective minority borrowers is in service quality...in responding to mortgage inquiries" and cites evidence of discrimination in response rates to mortgage inquiries." By conditioning on approved loans, the paper essentially restricts analysis to that. This creates a non-representative sample, as minorities approved differ from those rejected. Comparing approved minorities and whites does not accuratelv reflect population differences. For example, suppose lenders reject risky minority applicants but accept risky white applicants. By studying only approvals, risky whites are compared to only the least risky minorities. This makes minorities appear less risky, masking discrimination. The paper's estimates on this selective sample tell us nothing about population-level differences, only differences in the distorted approvals subsample. Collider bias renders approved minorities and whites incomparable. By assuming loan approval is unaffected by race, the paper induces significant bias. Claims about no pricing discrimination on approvals alone are hence invalid. Analyses must account for sample selection processes and discrimination in initial application decisions. Bhutta and Hizmo also assume race has a constant effect across individuals and contexts.
But, race is multidimensional, so its impact likely varies significantly across settings. The paper estimates average differences in interest rates and points paid between minorities and whites.
But these average effects poorly represent heterogeneous effects across borrowers and lenders. For example, one lender discriminates only against low-income minorities, while another discriminates based on neighborhood demographics. Estimated average effects masks variation in discrimination. Similarly, discrimination differs depending on borrower attributes like financial knowledge, dress, speech patterns, etc. Lenders treat minorities with certain qualities differently. Average effects fail to capture such nuances. By modeling race as having a uniform impact, the paper incorrectly assumes away effect heterogeneity. If discrimination manifests only in certain contexts, the model will underestimate its pervasiveness. The paper cannot conclusively claim no pricing discrimination overall when effects likely exhibit significant heterogeneity across lenders, borrowers, neighborhoods, and originator policies. Average effects poorly represent variable treatment effects across contexts. Claims about average impacts oversimplif a complex phenomenon. Bhutta and Hizmo also assume no interference between units, but discrimination involves spillovers where the outcomes for one borrower affect others. Again it says "One important margin Where racial bias could affect outcomes for prospective minority borrowers is in the loan application accept/reject decision" and cites evidence of discrimination in application denials.
This implies the number of minorities approved affects outcomes for other minority applicants through spillovers. For example, if lenders aim to limit total minority approvals, then denying one minority borrower increases the chance of approving the next. Minorities face negative externalities from other minorities' outcomes. By assuming SUTVA holds, the paper rules out these spillovers. But discriminatory application decisions create interference across borrowers. Lenders manage overall minority approval rates, not just individual decisions. Models assuming no interference misrepresent the nature of discrimination. If lenders limit total minority approvals, pricing outcomes for one minority depend on outcomes for other minorities. This interconnectivity and strategic behavior is ignored. The presence of spillovers means the stable unit treatment value assumption fails. The paper's inferences relying on SUTVA are hence invalid. Analyses must account for interference and interdependencies created by lenders managing minority approval rates. Bhutta and Hizmo also assume race is manipulable, but race is an immutable characteristic. This raises philosophical questions about the meaning of causal effects of race. The paper attempts to estimate counterfactual differences in mortgage pricing if a minority borrower were white. But race is fixed, so this counterfactual is not clearly defined. Race shapes lifetime experiences and opportunities, so it is not possible to instantaneously change race. Claiming to estimate the effect of making someone "white" lacks concrete meaning. Race is multifaceted, involving culture, beliefs, skills, and more. Isolating skin color alone as a manipulable treatment ignores this complexity. By modeling race as manipulable, the paper assumes away deep conceptual issues about the nature of race. At best, the paper reveals pricing differences correlated with race. Claims about causal effects of race itself are on shaky philosophical ground given the inability to physically manipulate race. The paper does not wrestle with this inherent difficulty in studying race effects. Bhutta and Hizmo also assume causal sufficiency - that the model fully accounts for the treatment assignment process. But researchers cannot credibly claim full control over the treatment of race. Race is complex, involving an interrelated bundle of factors like skin tone, culture, beliefs, etc. Isolating and manipulating any one dimension independently is impossible. The paper's models assume race can be specified and assigned independently of other factors. This falsely suggests race is simple and fully accounted for in the analysis. In reality, race cannot be randomized or cleanly specified. Models attempting to estimate race effects fail to address this lack of researcher control. By assuming causal sufficiency, the paper claims to fully model race and its assignment. But race is far more intricate than any model can capture. Researchers fundamentally lack control over race's many entangled facets. Models pretending to isolate race alone are hence oversimplified. The paper does not wrestle with this inherent lack of causal sufficiency when studying race. Claims about estimating causal effects of race require strong and unfounded assumptions about researcher control over race. There examples of the naive estimates that suffer from bias in the paper. In the Data section, the paper states: "In the HMDA data, lenders must report the APR - a single measure of loan cost that is a function of the interest rate, points, fees, and insurance premiums –if the APR exceeds a threshold." So estimates of fees/points are only available for a subset of borrowers. The Data section also states: "We obtained administrative data from the Federal Housing Administration (FHA) covering the universe of FHA-insured home purchase loans originated in 2014 and 2015." However, this means the data misses discrimination in non-FHA lending. In the Results, the paper finds a black-white interest rate gap of 0.03% and Hispanic-white gap of 0.015% after controlling for lender and risk factors. However, the naive model likely suffers from omitted variable bias and does not address unobserved differences across race. The Results also state "we find no evidence of systematic discrimination in fees by race or ethnicity." But again, this conclusion is based only on partial fee data for FHA borrowers, ignoring selection issues. The data limitations and narrowly specified models mean the paper's estimates are naive. The conclusions ignore bias from unobservables and selectively analyzed data. More complex causal analyses are needed to make credible claims about discrimination.
Tl;dr on why credit scores underpredict risk of default for minorities on the low end?