The DOI matches its Crossref record; 43 of the 43 references it cites were found in a registry or an index; 12 of the 33 sources its claims cite were checked for support.
An algorithm for removing sensitive information: Application to race-independent recidivism prediction
Open accessAbstract
Predictive modeling is increasingly being employed to assist human decision-makers. One purported advantage of replacing or augmenting human judgment with computer models in high stakes settings—such as sentencing, hiring, policing, college admissions, and parole decisions—is the perceived “neutrality” of computers. It is argued that because computer models do not hold personal prejudice, the predictions they produce will be equally free from prejudice. There is growing recognition that employing algorithms does not remove the potential for bias, and can even amplify it if the training data were generated by a process that is itself biased. In this paper, we provide a probabilistic notion of algorithmic bias. We propose a method to eliminate bias from predictive models by removing all information regarding protected variables from the data to which the models will ultimately be trained. Unlike previous work in this area, our procedure accommodates data on any measurement scale. Motivated by models currently in use in the criminal justice system that inform decisions on pre-trial release and parole, we apply our proposed method to a dataset on the criminal histories of individuals at the time of sentencing to produce “race-neutral” predictions of re-arrest. In the process, we demonstrate that a common approach to creating “race-neutral” models—omitting race as a covariate—still results in racially disparate predictions. We then demonstrate that the application of our proposed method to these data removes racial disparities from predictions with minimal impact on predictive accuracy.
Overview
AI snapshot
Written from the paper’s full text.
TL;DR
The paper proposes a likelihood-based chained transformation that removes information about protected variables from training data, accommodating mixed-scale covariates and targeting mutual independence from race. Applied to Broward County recidivism data, it produces race-independent predictions with only slight loss in predictive accuracy.
Key findings
- Omitting race as a covariate does not yield race-neutral predictions: models trained on unadjusted data still show large racial disparities in predicted re-arrest distributions.
- The proposed adjustment removes most racial dependence: pairwise Cramér’s V statistics are reduced to near zero in adjusted data, and random forest predictions by race become nearly identical.
- Predictive accuracy changes little: random forest AUC was 0.72 unadjusted and 0.71 adjusted, comparable to Northpointe’s reported 0.70 for this dataset.
- Adjustment altered high-risk classification rates: African-American high-risk proportion decreased by 0.07 (0.47 to 0.40), Caucasian increased by 0.14 (0.25 to 0.39), and Hispanic increased by 0.15 (0.27 to 0.42); false-positive-rate variation across race decreased (mad 0.05 to 0.02), while positive predictive value variation increased (mad 0.02 to 0.06).
Methods
- The method transforms covariates \(X\) to new variables \(W\) independent of protected variables \(Z\) by chaining conditional CDF transformations, matching conditional quantiles to marginal quantiles; it handles continuous, count, binary, and other mixed-scale variables and aims for mutual rather than only pairwise independence.
- The application uses ProPublica’s Broward County, Florida recidivism dataset, with race as the protected variable and re-arrest within two years as the outcome; covariates include age, sex, and juvenile/adult prior charge counts.
- The authors estimate conditional distributions with parametric regressions, generate \(K=50\) adjusted datasets, and compare random forest and logistic regression predictions on unadjusted, pairwise-adjusted, and adjusted data using ROC/AUC, calibration-like distribution plots, Cramér’s V, and race-specific accuracy metrics.
Limitations
- The paper states that the conditional distributions \(F_{X_j|V^{(j)}}\) are estimated and may be misspecified in finite samples or asymptotically, so model fit must be checked carefully and a “black box” automated approach is not recommended.
- The paper notes that dependence among variables may change over time, requiring routine re-testing and possible retraining; it also states that re-arrest is a biased measure of re-offense, which motivates the fairness definition but limits outcome validation.
- Evident from the text: the method’s guarantee depends on accurate conditional modeling, and the empirical application is one dataset with a contested demographic-parity fairness criterion; the paper itself notes disagreement among researchers about appropriate fairness notions.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
12 of 33 cited sources checked · 5 supported · 1 flagged · accuracy not yet measured
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“Finally, the procedure requires that all of the covariates X j be continuous variables, further limiting its scope.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
What the sources say
- Disputedcites Feldman et al. (2015) sourceExperimental — accuracy not yet measured
The Experiments section (pp. 16–17) says ordered categorical features were converted to integers and included in the repair; the procedure requires an ordering, not that every covariate be continuous. The Limitations section (p. 23) describes the algorithm as operating on numerical attributes, which does not establish a continuity requirement. (LLM, conf=0.90)
Checked by a reasoning model — may be wrong.
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passes- Disputed
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Much of the previous work has primarily focused on settings where the outcome or protected variable is binary.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
What the sources say
- Not checked — no full text of the sourcecites Kamiran and Calders (2009) source
- Partly supportedcites Calders and Verwer (2010) sourceExperimental — accuracy not yet measured
The Introduction states that this paper assumes a binary class attribute and a binary sensitive attribute, so it supports the claim as an example of earlier work using binary variables. It does not establish that much previous work primarily focused on such settings. (LLM, conf=0.83)
- Supportedcites Feldman et al. (2015) source
Section 6 (p. 16) describes datasets used in earlier fairness studies with binary outcomes and protected-attribute categories, including GOOD/BAD credit decisions and a YOUNG/OLD age split. Section 3 (p. 4) also frames the paper’s disparate-impact analysis using a binary outcome and two protected-class values. Together, these passages support the broad characterization of the settings emphasized in the work discussed. (LLM, conf=0.84)
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesThe cited sources support only part of this claim — read each source’s verdict below.
In the paper
“In addition to not guaranteeing demographic parity when p > 1, the algorithm has several limitations.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
What the sources say
- Partly supportedcites Feldman et al. (2015) sourceExperimental — accuracy not yet measured
Section 5 (pp. 11–13) establishes fairness for a single numerical attribute and says multiple attributes are repaired individually. This supports a limitation of the stated guarantee, but the paper does not explicitly say that demographic parity is not guaranteed when p > 1. (LLM, conf=0.78)
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passes- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“These two notions of fairness are indeed distinct.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
- Supportedcites Chouldechova (2017) source
Section 2 (p. 2) defines test fairness as a score reflecting the same recidivism likelihood regardless of group. Section 2.1 (p. 2) explains that when recidivism prevalence differs between groups, a test-fair score cannot also have equal false-positive and false-negative rates across those groups. This supports the distinction between the two fairness notions. (LLM, conf=0.93)
- Supportedcites Kleinberg, Mullainathan and Raghavan (2017) source
Section 1.3 (p. 8) distinguishes calibration from symmetry in error rates across groups and says the paper’s results show they can be incompatible. Theorem 1.1 (p. 5) identifies the limited cases in which its fairness conditions can all hold, supporting the claim that these notions are distinct. (LLM, conf=0.91)
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passes- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“For example, Chouldechova (2017) shows the same positive predictive value by protected class and equal false negative and false positive rates cannot both be achieved when the outcome prevalence depends on protected characteristics.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
- Supportedcites Chouldechova (2017) source
Section 2.1 (p. 2) states that when recidivism prevalence differs between groups, a test-fair score cannot have equal false-positive and false-negative rates across those groups. The same section defines test fairness here as equal positive predictive value across groups, matching the cited claim. (LLM, conf=0.94)
Strength:boosted· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passes- Supported
In the paper
“There is strong evidence to reject the null hypothesis of independence for all of the pairs, even when adjusting for multiplicity using the method of Benjamini and Hochberg (1995).”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5. Application: Removing racial dependence in recidivism risk assessment.
What the sources say
- Not checked — no full text of the sourcecites Benjamini and Hochberg (1995) source
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesThe paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Since functions of independent random variables are independent, X ⊥ ⊥ Z implies Y ⊥ ⊥ Z.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passes- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“and W is mutually independent of Z. Consequently, we refer to (3.2) as a transformation to mutual independence.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.2. Multivariate adjustments via chaining.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passes- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“models fit to the unadjusted data omitting race produce drastically different predictive distributions of the probability of re-offense by race-thus empirically demonstrating the insufficiency of omitting race from the analysis when the goal is statistical parity”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesA result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In the example we present below, we have found traditional, parametric regression models to be successful at estimating F X|Z if the analyst employs appropriate model selection and fit diagnostic techniques.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.4. Unknown conditional distributions.
Strength:hedged· not yet calibratedCausal:conditional· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passes- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
- Probably not a substantial claim. 29
- It may state two separate claims as one. 33
- The quote only partly states this claim. 1
The cited sources that could be checked couldn’t confirm or dispute this claim.
In the paper
“This scenario in which there is significant racial bias in re-arrest is specifically that which we seek to address in this paper, and thus the objections of Chouldechova (2017) to demographic parity are not directly relevant to our application.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
What the sources say
- Could not verifycites Chouldechova (2017) sourceExperimental — accuracy not yet measured
The paper discusses trade-offs between test fairness and equal error rates (Section 2.1, p. 2; Discussion, p. 4), but the passages found do not discuss demographic parity or caution about racial bias in re-arrest. I searched for both terms and for bias in recorded recidivism outcomes. Whether Chouldechova’s analysis is relevant to the citing authors’ application is their own judgment, not a claim the source can verify. (LLM, conf=0.74)
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
- The authors state this as an assumption: The average level of criminality is independent of race.
- Could not verify
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Although Adler et al. (2016) makes some improvements to the handling of categorical variables, the procedure is not appropriate for continuous Z or Z with many distinct values.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
What the sources say
- Partly supportedcites Adler et al. (2016) sourceExperimental — accuracy not yet measured
In “Computing Influence” (p. 3), the paper says the numerical obscuring procedure requires the feature being removed to be categorical. “Removing numerical features” (p. 3) then describes binning numerical features so the procedure can handle them; the paper does not state that continuous features are categorically inappropriate or identify many distinct values as a limitation. (LLM, conf=0.78)
- Could not verifycites Feldman et al. (2015) sourceExperimental — accuracy not yet measured
In §6 (pp. 16–17), the paper discretizes age into two protected groups and determines repair quantiles using the smallest protected group. Neither that discussion nor §7’s limitations (p. 23) says the procedure is inappropriate when the protected variable is continuous or has many distinct values; the §7 restriction concerns the attributes being repaired. The quoted citing sentence also names Adler et al., rather than Feldman et al. (LLM, conf=0.72)
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- Partly supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Kleinberg, Mullainathan and Raghavan (2017) shows theoretically that these notions of fairness are usually incompatible.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
- Supportedcites Kleinberg, Mullainathan and Raghavan (2017) source
The Introduction (p. 2) describes calibration within groups and balance conditions that generalize equal false-positive and false-negative rates. Theorem 1.1 (Section 1.2, p. 5) shows that all three conditions can hold together only with perfect prediction or equal base rates, supporting the claimed usual incompatibility. (LLM, conf=0.88)
Strength:boosted· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Under this definition, a model is typically considered fair if differences in the distribution of the model's predictions conditional on the protected variable do not exceed some pre-determined threshold, as measured by some appropriate notion of distance between probability distributions.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
- Partly supportedcites Feldman et al. (2015) sourceExperimental — accuracy not yet measured
The Introduction (p. 2) defines disparate impact using the probability of a positive outcome conditional on protected status, and Section 4.1 (p. 7) applies a threshold to that measure for a classifier. The paper uses a specific ratio of positive-outcome rates, not a general distance between prediction distributions, so that part of the paraphrase is broader than the source supports. (LLM, conf=0.85)
- Not matchedcites Barocas and Selbst (2016)
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
- Partly supported
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“This may or may not be unique, but all such random variables are, informally, as close as possible to X while also being independent of Z.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3. Transformations to independence.
Load-bearingStrength:hedged· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“The procedure we propose is guaranteed to achieve (3.1) when X is univariate.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesThe quote only partly states this claim.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In contrast, the joint adjustment results in nearly indistinguishable predictive distributions, achieving the goal of the procedure.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 4. Simulation example.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“The ordering X 1 , . . ., X p of the X variables is arbitrary, though some orderings may be practically convenient for a given application.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.2. Multivariate adjustments via chaining.
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✗ Based: failsProbably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The pairwise adjustment does reduce some of the discrepancy in the predictive distributions of ŷ | z, but the distribution of ŷ | z = 1 has a much larger right mode than the distribution of ŷ | z = 0.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 4. Simulation example.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Despite achieving W 2 ⊥ ⊥ Z, dependence between W 2 and X 2 is retained. For example, if we select G 2 = Bernoulli(1/2), then cor(W 2 , X 2 ) = 2/3.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.1.2. Atomic example.
Strength:plain· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIt may state two separate claims as one.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“Under this transformation, W 1 | Z ∼ G 1 and, because the distribution is the same for every value of Z, W 1 ⊥ ⊥ Z, so W 1 ∼ G marginally.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.1.1. Continuous example.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“This procedure could be applied independently to each covariate X j to achieve pairwise independence with Z, though it is not guaranteed that the resulting set of independently transformed covariates will achieve mutual independence with Z.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.2. Multivariate adjustments via chaining.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Though the multivariate procedure does not have the optimality guarantee of the univariate procedure, empirically we find it is very successful at achieving independence with minimal distortion of the distribution of X.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The criminal record variables-juv_misd_count, juv_fel_count, juv_other_count, and priors_count-are highly dispersed counts, and there is evidence of substantial dependence between most pairs of variables.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Strength:plain· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In particular, the positive predictive value is reduced by adjustment for Caucasian and Hispanic people and increased for African-Americans.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Strength:plain· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The mad is unchanged for accuracy, and decreases from 0.05 to 0.02 for fpr.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“On the other hand, adjustment appears to decrease variation by race in false positive rates and only slightly increase variation in accuracy.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The results of the logistic regression analysis are qualitatively very similar to those of RF and are deferred to the Appendix.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For the unadjusted data, the area under the curve (AUC) was 0.72, and for the adjusted data it was 0.71.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“It is worth noting that the Hispanic group is by far the smallest of the three, so while this group experiences the largest percent increase in high risk classification, the number of individuals whose status changes as a result of adjustment is relatively small.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Strength:plain· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Thus, similar accuracy is achieved by offsetting improvements and degradation of the two metrics, and the direction of the shift differs for African-Americans compared to all others.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For African-Americans, the change in tpp of 0.29 -0.33 = -0.04 is largely offset by the change in tnp of 0.38 -0.35 = 0.03, for a net change of only -0.01 in total accuracy.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Conversely, the negative predictive value is reduced by adjustment for African-Americans and increased by adjustment for Caucasians and Hispanics.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Strength:plain· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes? Load-bearing: unclear✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Predictions made by training RF on data adjusted using our procedure eliminate almost all racial disparities, as evidenced by the nearly identical distributions by race in the two panels on the right.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The fpr is decreased for African-Americans and increased for Hispanics and Caucasians.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“When c(x, w) = d q (x, w) for q ≥ 1, the transportation distance is related to the Wasserstein-q distance by W q q (μ, μ) = ρ c (μ, μ), so the optimal coupling-when it exists-is also the coupling achieving the W q distance.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX B: OPTIMAL COUPLING AND TRANSPORT MAPS
Strength:plain· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it. It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Unlike previous work in this area, our procedure accommodates data on any measurement scale.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✗ Based: failsProbably not a substantial claim. The paper may give no reason why it holds.
- The authors state this as an assumption: The average level of criminality is independent of race.
- Rests on an assumption: In the overwhelming majority of applications, X and Z are dependent.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In this setting, a model may very successfully predict that an individual is likely to be re-arrested, but the model has no way of disentangling a higher risk of re-arrest due to elevated latent criminality from higher risk due to increased police attention.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
Load-bearingStrength:hedged· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The key to this approach is reliably estimating the ζ j , which in turn requires good choices of G ← and good estimators of F X j |V (j ) .”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.4. Unknown conditional distributions.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim. It may state two separate claims as one.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“This greatly expands the range of datasets that can be transformed and thus expands the universe of problems to which the procedure may be applied.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- The authors state this as an assumption: The average level of criminality is independent of race.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“If F X j |V (j ) were the exact conditional distribution F X j |V (j ) , then W would satisfy W ⊥ ⊥ Z.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Load-bearingStrength:hedged· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✗ Based: failsProbably not a substantial claim. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“This framework also allows us to make transformed data W that are mutually independent of Z, rather than only pairwise independent, thereby guaranteeing that the predictions of any algorithm trained on the transformed data will be independent of Z.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes? Based: unclearIt may state two separate claims as one.
- The authors state this as an assumption: The average level of criminality is independent of race.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“It is typical in applied statistics and regression modeling to transform predictors prior to model fitting for computational reasons or to obtain better predictive accuracy, which would neutralize any choice we make for G.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.3. Defining G.
Strength:hedged· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✗ Checkable: fails✗ Load-bearing: fails✗ Based: failsProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our approach allows for any number of mixed-scale variables to be adjusted and naturally accommodates one or more protected variables that may be categorical or continuous.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
- The authors state this as an assumption: The average level of criminality is independent of race.
- Rests on an assumption: In the overwhelming majority of applications, X and Z are dependent.
A claim is checked only against the cited paper’s full text, never its abstract, and none is available for its sources.
In the paper
“However, for a variety of reasons, information on true baseline rates of crime against which to”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
What the sources say
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
In the paper
“First, it works only for categorical Z, and reasonable performance requires a large number of observations taking each possible level of Z.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > Abstract
What the sources say
- Not checked yetcites Feldman et al. (2015) source
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Estimation of F X j |V (j ) is an exercise in regression modeling.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.4. Unknown conditional distributions.
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In practice, how one chooses or estimates G is less critical than using the conditional CDF to transform X.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 3.3. Defining G.
Strength:plain· not yet calibratedTests✓ Asserted: passes✗ Checkable: fails✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.
- The authors state this as an assumption: The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In constructing the chain of conditional models, we always include discretized versions w * 1 and w * 2 of w 1 or w 2 , respectively, whenever the w 1 or w 2 are included in the model.This captures nonlinearity in the condi-tional mean of the other variables.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Strength:plain· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✗ Based: failsProbably not a substantial claim. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Of course, F X j |V (j ) is an estimate, and thus it will differ from F X j |V (j ) in finite samples, and even asymptotically when F X j |V (j ) is misspecified.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
None of the sources cited for this claim have been checked yet.
In the paper
“Methodology for achieving statistical parity has centered on removing information about the protected variables from the training data by transforming the training data.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
In the paper
“In general, we have found that notions of fairness can be divided into three camps.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
- Not matchedcites Berk (2016)
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
- Not matched
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“If deployed in the real world, the approach we have proposed does not obviate the need for good practices with respect to validation and evaluation.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 6. Discussion.
Strength:plain· not yet calibratedTests✓ Asserted: passes✗ Checkable: fails✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“By doing the adjustment, a dataset could be released to multiple organizations to build prediction models, and regardless of the details of their model, we would be guaranteed that the predictions would be fair under the definition we support in this case.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 6. Discussion.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes? Based: unclearIt may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Furthermore, this approach limits the risk of malicious actors intentionally biasing the model against protected groups, as model builders would not be able to infer protected characteristics.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 6. Discussion.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✗ Based: failsProbably not a substantial claim. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In this case, assessment of how well our model predicts Y using any notion of model performance is not especially well motivated, as Y is a biased measure of the phenomenon it is meant to measure.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“The extension to allow for the adjustment of discrete variables is itself an advancement, as previous proposals for adjusting of the training covariates were only designed for adjusting continuous variables.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 6. Discussion.
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The nature of dependence among variables is likely to change over time (perhaps, even as a result of the application of such a risk assessment model), so it would be important that these tests are routinely applied to ensure that dependence were adequately removed on an ongoing basis.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 6. Discussion.
Strength:hedged· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✗ Checkable: fails✗ Load-bearing: fails✗ Based: failsProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds. It may state two separate claims as one.
In the paper
“This procedure is limited to binary outcome data and the adjusted data is not re-usable in the sense that one could not then use the covariates to estimate relationships with other outcome variables and still be ensured of nondiscriminatory outcomes.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
- Not checked — no full text of the sourcecites Kamiran and Calders (2009) source
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
In the paper
“In this case also, the methods are restricted to binary protected classes and binary outcome variables.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
What the sources say
- Not checked yetcites Calders and Verwer (2010) source
Strength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“This needs to be done separately for every instantiation of a risk assessment model, as the best mathematical characterization of fairness will likely vary with prediction type and local community understanding of fairness and the goals of implementing the model.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 6. Discussion.
Strength:hedged· not yet calibratedCausal:correlational· not yet calibratedTests✓ Asserted: passes✗ Checkable: fails✗ Load-bearing: fails✗ Based: failsProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“because African Americans are more likely to be re-arrested for re-offending, neither definition seems particularly appropriate for this setting.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5. Application: Removing racial dependence in recidivism risk assessment.
Tests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✗ Based: failsProbably not a substantial claim. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Indeed, differences in the marginal rate of arrest could manifest via processes that are entirely unbiased-most obviously, genuine differences in rates of re-offense, though others are possible.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5. Application: Removing racial dependence in recidivism risk assessment.
Strength:hedged· not yet calibratedCausal:conditional· not yet calibratedTests✓ Asserted: passes✗ Checkable: fails✗ Load-bearing: fails✗ Based: failsProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We expect our procedure would also be of value in data privacy and anonymization.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 6. Discussion.
Strength:hedged· not yet calibratedTests✓ Asserted: passes✗ Checkable: fails✗ Load-bearing: fails✗ Based: failsProbably not a substantial claim. It may be too vague to check against evidence. The paper’s conclusions may not depend on it. The paper may give no reason why it holds.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In applying our proposed procedure and the pairwise adjustment, some relevant information is lost.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5.3. Transformations to independence.
Strength:plain· not yet calibratedCausal:direct· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✗ Load-bearing: fails✓ Based: passesProbably not a substantial claim. The paper’s conclusions may not depend on it.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“because African Americans are more likely to be re-arrested for re-offending, neither definition seems particularly appropriate for this setting.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > 5. Application: Removing racial dependence in recidivism risk assessment.
Load-bearingStrength:hedged· not yet calibratedTests✓ Asserted: passes? Checkable: unclear✓ Load-bearing: passes✓ Based: passesIt may state two separate claims as one.
The paper states this without a source or data behind it here, so there is nothing to check it against. The quote shows it is in the paper.
In the paper
“In most cases, the allowable difference is zero, which is equivalent to the requirement that the predictive distribution is independent of the protected variable.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim. It may state two separate claims as one.
A conclusion the paper draws from its own results — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our paper operates within this definition of fairness and builds upon the extant methodology for mitigating disparate impact.”
§ AN ALGORITHM FOR REMOVING SENSITIVE INFORMATION: APPLICATION TO RACE-INDEPENDENT RECIDIVISM PREDICTION 1 > APPENDIX A: OVERVIEW OF NOTIONS OF FAIRNESS
Load-bearingStrength:plain· not yet calibratedTests✓ Asserted: passes✓ Checkable: passes✓ Load-bearing: passes✓ Based: passesProbably not a substantial claim. It may state two separate claims as one.
The authors state this as an assumption.
The average level of criminality is independent of race.
“the most reasonable approach is to assume that the average level of criminality is independent of race.”
The authors state this as an assumption.
The joint and conditional distributions of (X, Z) are known, and the target distribution G is fixed and user-specified.
“In the following motivating sections, we assume the joint and conditional distributions of (X, Z) are known, and the desired target distribution, G, is fixed and user-specified.”
The paper relies on this without support.
In the overwhelming majority of applications, X and Z are dependent.
“In the overwhelming majority of applications, X and Z are dependent, and thus we must take additional measures to ensure Y is fair.”
The authors state this as an assumption.
The distribution of risk is independent of race.
“Thus, as we argue above, a reasonable way to proceed is to assume that the distribution of risk is independent of race.”
The paper relies on this without support.
The outcome variable (re-arrest) is observed with bias with respect to the protected variable (race).
“Given that the outcome variable is observed with bias with respect to the protected variable”
The authors state this as an assumption.
A suitable conditional model can be found to adequately describe the full conditional distribution of the permitted variables given the protected variables.
“so long as a suitable conditional model can be found to adequately describe the full conditional distribution of the permitted variables given the protected variables”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“Applications in perturbationbased data repair [98,125,136,152,154,172] have shown that accuracy is not signiicantly afected.”
From Fairness in machine learning: A survey (2024) · cites this paper as [154]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Application reports seven transformed variables but specifies six
MinorThe overview says the recidivism models use seven transformed variables, while the application identifies six covariates to transform: sex, age, and four criminal-history counts.
“random forest or logistic regression fit to only seven transformed variables-mostly pertaining to an individual's criminal history”
“Of the six x j , one (sex) is binary, one [log(age)henceforth simply "age"] is continuous, and the other four, which relate to prior criminal record, are counts.”
Citations
Cites
Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing
Benjamini, Y., & Hochberg, Y. (1995) · Journal of the Royal Statistical Society Series B: Statistical Methodology · doi:10.1111/j.2517-6161.1995.tb02031.x · Cited by 109,482
Multiple Imputation for Nonresponse in Surveys
Rubin, D. B. (1987) · Wiley Series in Probability and Statistics · doi:10.1002/9780470316696 · Cited by 20,701
mice: Multivariate Imputation by Chained Equations in R
Buuren, S. V., & Groothuis-Oudshoorn, K. (2011) · Journal of Statistical Software · doi:10.18637/jss.v045.i03 · Cited by 13,870
Multiple imputation using chained equations: Issues and guidance for practice
White, I. R., Royston, P., & Wood, A. M. (2011) · Statistics in Medicine · doi:10.1002/sim.4067 · Cited by 9,792
Dwork, C., Hardt, M., Pitassi, T., Reingold, O., & Zemel, R. (2012) · Proceedings of the 3rd Innovations in Theoretical Computer Science Conference · doi:10.1145/2090236.2090255 · Cited by 3,501
Barocas, S., & Selbst, A. D. (2014) · doi:10.2139/ssrn.2477899 · Cited by 3,048
Equality of Opportunity in Supervised Learning
Hardt, M., Price, E., & Srebro, N. (2016) · arXiv · doi:10.48550/arxiv.1610.02413 · Cited by 2,536
Fair prediction with disparate impact: a study of bias in recidivism prediction instruments
Chouldechova, A. (2017) · Big Data · doi:10.1089/big.2016.0047 · Cited by 2,159
Certifying and Removing Disparate Impact
Feldman, M., Friedler, S. A., Moeller, J., Scheidegger, C., & Venkatasubramanian, S. (2015) · Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining · doi:10.1145/2783258.2783311 · Cited by 1,827
Kusner, M. J., Loftus, J. R., Russell, C., & Silva, R. (2017) · arXiv · doi:10.48550/arxiv.1703.06856 · Cited by 995
Flagged first, then by citations
Cited by
Sorted by citations
Items 101–144 of 144
On the Power of Randomization in Fair Classification and Representation
Agarwal, S., & Deshpande, A. (2022) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3531146.3533209 · Cited by 4
Muhlenbach, F., Phuoc, L. N., & Sayn, I. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2007.04824 · Cited by 4
Deep Reinforcement Learning for Unknown Anomaly Detection.Preprint
Pang, G., van den Hengel, A., Shen, C., & Cao, L. (2020) · arXiv (Cornell University) · Cited by 4
Propose and Review: Practical Bias Mitigation for Machine ClassificationPreprint
Li, T., Tang, Z., Lu, T., & Zhang, X. (2022) · SSRN Electronic Journal · doi:10.2139/ssrn.4139244 · Cited by 5
Touitou, M. (2025) · Acta Oeconomica · doi:10.1556/032.2025.00134 · Cited by 4
De-biasing the bias: methods for improving disparity assessments with noisy group measurements
Wastvedt, S., Snoke, J., Agniel, D. M., Lai, J., Elliott, M. N., & Martino, S. C. (2024) · Biometrics · doi:10.1093/biomtc/ujae155 · Cited by 3
Optimal Data-Driven Hiring With Equity for Underrepresented Groups
Zhu, Y., & Ryzhov, I. O. (2024) · Production and Operations Management · doi:10.1177/10591478231224942 · Cited by 5
Fairness Perception from a Network-Centric Perspective
Masrour, F., Tan, P., & Esfahanian, A. (2020) · Proceedings · doi:10.1109/icdm50108.2020.00145 · Cited by 5
Optimal Pre-Processing to Achieve Fairness and Its Relationship with Total Variation BarycenterPreprint
Farokhi, F. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2101.06811 · Cited by 3
Predictive Justice in Criminal Proceedings
Mohanta, A., Garikapati, K., Sneha, Mohanty, S., & Shreya (2024) · Advances in electronic government, digital divide, and regional development book series · doi:10.4018/979-8-3693-7205-0.ch011 · Cited by 2
Wang, X. (2025) · Journal of Politics and Law · doi:10.5539/jpl.v18n4p68 · Cited by 2
Stability selection via variable decorrelation
Nouraie, M., Smith, C., & Muller, S. (2026) · Statistics and Computing · doi:10.1007/s11222-026-10916-7 · Cited by 2
Algorithmic Administrative Justice
Appel, S. M., & Coglianese, C. (2021) · Oxford University Press eBooks · doi:10.1093/oxfordhb/9780190903084.013.24 · Cited by 2
Metric-Independent Mitigation of Unpredefined Bias in Machine Classification
Tang, Z., Lu, T., & Li, T. (2024) · Intelligent Computing · doi:10.34133/icomputing.0083 · Cited by 2
Enforcing Conditional Independence for Fair Representation Learning and Causal Image Generation
Hwa, J., Zhao, Q., Lahiri, A., Masood, A., Salimi, B., & Adeli, E. (2024) · IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw63382.2024.00015 · Cited by 2
Law and Empathy in the Automated State
Coglianese, C. (2023) · Cambridge University Press eBooks · doi:10.1017/9781009334297.014 · Cited by 2
Learning Fair and Interpretable Representations via Linear OrthogonalizationPreprint
He, Y., Burghardt, K., & Lerman, K. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1910.12854 · Cited by 4
Everything is Relative: Understanding Fairness with Optimal TransportPreprint
Kwegyir-Aggrey, K., Santorella, R., & Brown, S. M. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2102.10349 · Cited by 3
Nyawambi, T. M., & Muchiri, H. (2024) · International Conference on Smart Sensors and Application (ICSSA) · doi:10.1109/icssa62312.2024.10788666 · Cited by 1
Responsible AI for Air Traffic Management: Application to Runway Configuration Assistance Tool
Memarzadeh, M., Wang, Z., Shalmani, F. M., Razzaghi, P., & Kalyanam, K. (2025) · Aerospace · doi:10.3390/aerospace12100872 · Cited by 1
Krantsevich, C., Hahn, P. R., Zheng, Y., & Katz, C. M. (2023) · The Annals of Applied Statistics · doi:10.1214/22-aoas1657 · Cited by 3
Non-Invasive Fairness in Learning Through the Lens of Data Drift
Yang, K., & Meliou, A. (2024) · IEEE International Conference on Data Engineering (ICDE) · doi:10.1109/icde60146.2024.00172 · Cited by 2
Strong Group Fair Classification via Optimal Transport and Mixed-Integer Linear Programming
Tao, Y., Chen, C., Jiang, S., & Wang, Q. (2025) · Journal of Global Optimization · doi:10.1007/s10898-025-01525-1 · Cited by 1
Incentives Needed for Low-Cost Fair Lateral Data Reuse
Maio, R., & Chaintreau, A. (2020) · ACM-IMS Foundations of Data Science Conference (FODS) · doi:10.1145/3412815.3416890 · Cited by 1
Duwe, G. (2021) · International Journal of Offender Therapy and Comparative Criminology · doi:10.1177/0306624x20986523 · Cited by 2
Cabreros, I., Snoke, J., Osoba, O., Khan, I., & Elliott, M. N. (2023) · RAND Corporation eBooks · doi:10.7249/rra1542-1 · Cited by 1
The impact of overbooking on a pre-trial risk assessment toolPreprint
Lum, K., Boudin, C., & Price, M. (2020) · arXiv (Cornell University) · doi:10.2139/ssrn.4478818 · Cited by 1
Self-citationScalable Assessment and Mitigation Strategies for Fairness in RankingsPreprint
Nandy, P., Sepehri, A. A., Basu, K., Logan, H., Agarwal, D., & Karoui, N. E. (2020) · arXiv (Cornell University) · Cited by 1
On the Fairness of 'Fake' Data in Legal AIPreprint
Boswell, L., & Prakash, A. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2009.04640 · Cited by 1
Costs and Benefits of Wasserstein Fair Regression.Preprint
Zhao, H. (2021) · arXiv (Cornell University) · Cited by 1
Framework for developing algorithmic fairness
Kristiadi, D. P., Sunarya, P. A., Ismanto, M., Dylan, J., Santoso, I. R., & Warnars, H. L. H. S. (2020) · Bulletin of Electrical Engineering and Informatics · doi:10.11591/eei.v9i4.2028 · Cited by 1
Fair Learning: an optimal transport based approach
Gordaliza, P. (2020) · UVaDOC UVaDOC University of Valladolid Documentary Repository (University of Valladolid) · doi:10.35376/10324/43392
Social Determinants of Recidivism: A Machine Learning Solution.Preprint
Shirvaikar, V., & Lakshminarayan, C. K. (2020) · arXiv (Cornell University)
Enhancing Data-Driven Decision Support with Multi-Perspective Solutions
Wang, Y. (2020) · University of Minnesota Digital Conservancy (University of Minnesota)
Fairness for Cooperative Multi-Agent Learning with Equivariant Policies.Preprint
Grupen, N. A., Selman, B., & Lee, D. D. (2021) · arXiv (Cornell University)
Costs and Benefits of Fair Regression
Zhao, H. (2024) · doi:10.55277/researchhub.mz9wqvk1
Lerch, L., Wich, M., Eder, T., & Groh, G. (2022) · Ethik in mediatisierten Welten · doi:10.1007/978-3-658-36446-5_17
Bednarz, Z., & Zalnieriute, M. (2023) · Cambridge University Press eBooks · doi:10.1017/9781009334297.013
Data Quality: What if Deming Were Born Today?
Lin, D. K. J., & Rios, N. (2025) · Applied Stochastic Models in Business and Industry · doi:10.1002/asmb.70025
From fair graphs to fair data: a DAG-based approach to mitigating bias in AI systems
Jiang, V. W., Batista, G., & Bain, M. (2026) · Knowledge and Information Systems · doi:10.1007/s10115-026-02817-y
Fair Learning : an optimal transport based approach
Gordaliza, P. (2020) · theses.fr (ABES) · doi:10.70675/9b44aae6z5b37z47d7zae40ze61595fa014d
Falsifying Discriminant Validity of Predictive Algorithms
Coston, A. (2026) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3805689.3812310
Zhang, Y. (2026) · Scientific Research Bulletin · doi:10.71052/srb2024/elvi1174
Metadata
- Authors
- Johndrow, J. E., & Lum, K.
- Affiliations
- Stanford University (US)
- Venue
- The Annals of Applied Statistics
- Publisher
- Institute of Mathematical Statistics
- Published
- 2019-03-01
- Volume
- 13
- Issue
- 1
- Pages
- 189-220
- ISSN
- 1932-6157 (ISSN-L)
- 1941-7330
- OpenAlex
- W2600853029
- Cited by
- 141
- References
- 40
- Citations by year
- 2026: 4 · 2025: 7 · 2024: 12 · 2023: 14 · 2022: 16 · 2021: 32 · 2020: 28 · 2019: 19
- FWCI
- 41.52
- Type
- journal article
- Language
- en
- Open access
- bronze
- doi.org
- Locations
- The Annals of Applied Statistics (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- Project Euclid (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Social Sciences › Sociology and Political Science › Criminal Justice and Corrections Analysis
- Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
- Keywords
- algorithmic bias, sensitive information removal, recidivism prediction, protected characteristics, parole decisions, sentencing, racial disparities, criminal justice
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-11-09, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-09 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
