Pith. sign in

REVIEW 4 major objections 7 minor 34 references

NY Real Estate Racial Equity Analysis via Applied Machine Learning

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Owners are whiter than residents in 81% of NY tracts

desk verdict The qualitative finding is almost certainly right, but the magnitudes rest on an unvalidated domain transfer and a stress test that is not a valid correction. read the letter →

arxiv 2505.16946 v3 pith:T2A6L46D submitted 2025-05-22 cs.CY cs.LG

Sanjana Chalavadi , Andrei Pastor , Terry Leitch This is my paper · ORCID
classification cs.CYcs.LG
keywords racialequitypropertyownershipraceimputationmachinelearningcensustractsNewYorkStatehomeownershipdisparitycorporate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that property ownership in New York is racially skewed even after accounting for who lives in each neighborhood. Statewide, the average census tract has 5–10 percentage points more White owners than White residents, and in over 81% of tracts the predicted White owner share exceeds the White population share. The gap is steepest in minority-majority neighborhoods: in an average majority-Black tract, White individuals are estimated to own about 39% of properties although they are only 6.7% of residents. The paper reaches these numbers by applying a name- and geolocation-based machine-learning model to millions of parcel records, and it argues that corporate ownership widens the gap further by removing housing stock from minority owner-occupancy. If correct, the results map a racial wealth gap in property control at neighborhood scale and point to concrete housing-policy targets.

What carries the argument

The machine carrying the argument is a deep-learning race/ethnicity imputation system: a character-level LSTM neural network that reads owner names and, in the Full Model, also takes census-tract geolocation as input and passes outputs through an XGBoost filter, trained on labeled voter-registration data and previously validated at 89.2% accuracy. A Name-Only LSTM version is used for New York City, where owner addresses are missing. The model outputs per-owner probabilities over five race/ethnicity categories; the paper assigns the most likely category, aggregates ownership counts and assessed value by tract, and compares them with census population shares. A New York ground-truth sample of 35,134 PPP loan applicants who self-reported race gives the Full Model 87.25% accuracy and per-class false-positive rates, which the paper then applies to stress-test its extreme-disparity examples.

What would settle it

Obtain true race/ethnicity for a random sample of individual property owners in majority-Black New York tracts, from self-identification in a survey, matched administrative records, or field verification, and compare observed owner shares with the model's predictions. If White ownership in those tracts is close to the 6.7% White population share rather than the predicted 39%, the central disparity estimate is largely a measurement artifact.

Watch

Extended reading notes

Core claim

The paper's central discovery is that New York property ownership is systematically whiter than the resident population at the census-tract level. Statewide, the mean tract is about 65% White in population but 72% White in ownership, and in more than 81% of tracts the predicted White owner share exceeds the White population share. The pattern is most dramatic in minority-majority tracts: in an average majority-Black tract, Black residents are 69.2% of the population but Black owners are only 44.6% of owners, while White residents are 6.7% of the population but White owners are 39.3%; majority-Hispanic tracts show Hispanic residents at 64.5% versus 38.3% Hispanic ownership and White residents at 10.2% versus 27.1% White ownership. White owners also hold about 87% of individually-owned assessed property value against roughly 75% of the state population. Corporate ownership widens the effective gap: in majority-Black, majority-Hispanic, and mixed tracts, White individual plus corporate ownership reaches 55.3%, 55.3%, and 56.7% respectively, meaning most housing stock in these communities is not held by the resident majority.

Load-bearing premise

The load-bearing premise is that the race-imputation model, trained on voter rolls from Florida and North Carolina, transfers to New York property owners, and that error rates measured on PPP loan applicants apply directly to tract-level ownership predictions; if those error structures differ, the reported 5–10 point White overrepresentation could be partly a measurement artifact.

Editorial extensions

If this is right

  • White overrepresentation is not a suburban or rural artifact; it persists inside Black- and Hispanic-majority tracts, so ownership and resident demographics diverge in the places where the resident majority is non-White.
  • Corporate ownership materially changes the disparity picture: in majority-Black, majority-Hispanic, and mixed tracts, combining White individual owners with corporate owners puts more than half of the housing stock under White-plus-institutional control.
  • The choice between Full and Name-Only models matters for New York City; since the Name-Only model is acknowledged to risk overestimating White ownership, the city's true disparities could be larger than the reported numbers.
  • Policy levers that follow from the analysis include tenant opportunity-to-purchase programs, restrictions on speculative LLC purchases, and support for minority homebuyers in tracts where ownership is most out of line with population.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit: the stress test corrects only four illustrative tracts; running the same error adjustments across all tracts would reveal whether the 81% overrepresentation figure and the 5–10 point average gap survive in aggregate.
  • Because corporate owners are not assigned a race, the 'White + corporate' combined shares rest on the untested assumption that corporate owners are effectively White-controlled; beneficial-ownership records would turn that assumption into a measurement.
  • The analysis does not separate White owner-occupants from absentee White landlords in minority tracts, so it cannot distinguish gentrification-driven displacement from long-run rent extraction; adding owner-address data would split those mechanisms.
  • Applied longitudinally to several years of assessment rolls, the same pipeline could track whether ownership gaps widen or narrow under new housing policies, a follow-up the paper itself identifies.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. This paper applies an LSTM-based race/ethnicity imputation model (Full Model with geolocation and XGBoost; Name-Only Model) to New York property parcel data, inferring the race of individual property owners and comparing the predicted owner composition with census tract population composition. The authors report that White individuals are overrepresented among property owners relative to their population share in over 81% of tracts, with a statewide mean gap of 5-10 percentage points, and that in majority-Black and majority-Hispanic tracts the majority group owns far less than its population share while White ownership is disproportionately high. They further examine corporate ownership and urban-suburban-rural differences, and perform a stress test intended to show that extreme disparities persist under error corrections. The paper also presents ground-truth validation of the imputation models on Paycheck Protection Program loan applicant data.

Significance. If the quantitative claims were established, the paper would make a useful contribution by providing tract-level estimates of racial ownership disparities in New York, supplementing household-level homeownership statistics with property-level evidence and connecting to policy debates on absentee ownership and financialization. The paper is commendable for comparing two model variants, reporting confusion-matrix-based error rates, and discussing limitations candidly, including the difference between owner-occupied and investor-owned properties. However, the central quantitative estimates rest on an unvalidated domain transfer and on a stress-test procedure that is not a valid confusion-matrix correction. These issues do not necessarily invalidate the qualitative direction of the findings, which is consistent with prior homeownership statistics, but they preclude accepting the reported magnitudes as established. The corporate-ownership analysis also relies on an interpretive assumption that corporate entities can be treated as White-controlled.

major comments (4)
  1. [Section 3.4, Table 3] The stress test is not a valid confusion-matrix inversion. The White FPR is P(predicted=White | true≠White); multiplying the predicted White share by (1-FPR) assumes that all predicted-White labels that are wrong are true non-White, and ignores the base rate of White owners, so it does not produce an upper bound on the true White share. Similarly, dividing minority shares by (1-FNR) does not correctly recover true minority counts unless the false negative count is corrected using the appropriate base rates. Consequently, the 'Stress' columns in Table 3 do not demonstrate that the disparities persist under conservative assumptions; a valid correction would require applying the full confusion matrix (or its inverse) to the predicted counts, or estimating positive predictive values in the target population.
  2. [Appendix A.2, Appendix B, Section 3.1] The imputation model is trained on Florida and North Carolina voter registration data and validated on New York PPP loan applicants, but is then applied to New York property owners, a population that differs in age, income, geography, and name distribution. No validation on property owners with known race is provided, and the headline statewide gap is only 5-10 percentage points, so even a modest differential misclassification rate in the target population could account for the entire gap. The paper should either provide target-population validation, derive an upper bound on the disparity that is robust to plausible misclassification, or explicitly downgrade the quantitative claims to illustrative.
  3. [Appendix A.2, Section 3.1] The Full Model uses census tract demographics as an input feature, and the comparison benchmark is the same census tract demographics. This creates a circularity: predicted owner composition is partially regressed toward the resident population distribution, so the model-based gap is not an independent measurement. The paper also applies the Full Model statewide but the Name-Only Model in NYC, confounding regional comparisons with model differences. A sensitivity analysis that re-runs the full-state analysis with tract demographics removed, or that compares both models on the same tracts, is needed to assess the size of these effects.
  4. [Section 3.5, Table 5] The combined 'White + Corp' ownership metric is presented as if corporate ownership is White-controlled, but the paper itself acknowledges that 'we cannot attribute race directly' to corporations. Adding corporate ownership to White individual ownership therefore overstates the White-controlled share and should be labeled as a scenario or bound, not a measured quantity. In addition, the text reporting Albany tract 11 (Section 3.5) says White individuals own 41.9% of properties, but Table 5's columns (White+Corp 94.1%, Corp Only 89.8%) imply White individual ownership of only 4.3%; this internal inconsistency must be resolved.
minor comments (7)
  1. [Section 3.4] The sentence 'We apply error rate adjustments from our PPP ground truth validation (Section 3.2, Table 9)' misreferences the appendix; Table 9 appears in Appendix B, not Section 3.2.
  2. [Table 2] The header 'T ract (Muni - T ract ID)' contains a typo; it should be 'Tract (Muni - Tract ID)'.
  3. [Figures] Figures 1-3 are referenced in the text but do not appear in the manuscript, making the visual claims unverifiable; the figures or a note on their availability should be included.
  4. [Section 3.6] The Urban-Core classification uses an ad-hoc threshold of '>1000 properties per tract' without sensitivity analysis; the choice should be justified or tested against alternative thresholds.
  5. [Section 4] The median income figures for majority-White and majority-Hispanic tracts ($90k vs $62k) are not reported in any table or tied to a specific ACS year; the source and year should be documented.
  6. [Section 5] The conclusion contains vague statements about a forthcoming technique for estimating error distributions and about the authors having 'conceived a technique'; these passages are unsupported and should be removed or replaced with specific references.
  7. [Appendix B] The text references 'Figure 3' for accuracy by income decile, but no figure is provided in the appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the disparity estimates come from a model validated on external NY PPP data, and the target disparity is not a training input; the tract-demographic feature creates a dependency but not a definitional reduction.

full rationale

The central claim—that White ownership exceeds White population by 5–10 percentage points statewide—is a comparison of model-predicted owner race shares to census population shares. The model is not trained on the target disparity: it is an LSTM+Geo+XGBoost classifier trained on FL/NC voter registration labels and re-validated in this paper on NY PPP loan applicants with self-reported race (Appendix B, Tables 6–9). The target quantity is therefore not a fitted input renamed as a prediction. The Full Model does incorporate census tract demographics as an input feature (Appendix A.2), and the same demographics are the comparison benchmark, so the predicted owner composition is not fully independent of the resident population. However, the paper never defines the owner share as equal to the population share; the model output is a learned function of owner names, geolocation priors, and XGBoost filtering, so the 5–10 point gap is not forced by construction. The Section 3.4 stress test applies PPP-derived FPR/FNR adjustments to extreme-disparity tracts; the adjustment formula is statistically questionable (FPR is not a precision correction and base rates are ignored), but this is a robustness-analysis flaw, not a circular derivation. Self-citations to Chalavadi et al. (2025) supply the model architecture and prior accuracy figures, but the current paper independently estimates accuracy and class-specific FPRs on NY PPP data, so the self-citation is not the sole load-bearing support. The paper also explicitly flags its own limitations, including the statement in Section 3.4 that 'The authors are developing a technique to analyze the impact of errors' and the Conclusion's acknowledgement that 'A weakness in applying models is that the predicted population may differ from the model's origin.' These admissions treat the disparity as an empirical estimate with unresolved error, not as a definitional tautology. Accordingly, no step in the derivation chain reduces to its own inputs, and the circularity score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The quantitative claims depend on the imputation model's generalization and on several ad hoc aggregation choices. No new physical entities are postulated. The most consequential free parameters are the undisclosed model weights and the error rates used in the stress test.

free parameters (5)
  • LSTM/XGBoost imputation model weights = not disclosed
    Trained on Florida and North Carolina voter registration data; all owner-race predictions depend on these fitted model parameters.
  • White FPR stress correction, Full Model = 0.069
    Used in Section 3.4 to reduce predicted White ownership; measured on NY PPP ground truth in Appendix B, Table 9.
  • White FPR stress correction, Name-Only Model = 0.138
    Used for NYC extreme-disparity tracts; applied directly to predicted White shares without positive predictive value correction.
  • Minority FNR stress corrections = Full: Black 0.0993, Hispanic 0.1525; Name-Only: Black 0.2062, Hispanic 0.2771
    Used to inflate minority ownership in stressed scenarios; derived from confusion matrices in Table 9.
  • Urban-Core property count threshold = top quartile, >1000 properties per tract
    Hand-chosen cutoff in Section 3.6 for labeling Urban-Core tracts; affects urban-suburban-rural comparisons.
assumptions (6)
  • domain assumption Voter registration race labels from Florida and North Carolina generalize to New York property owners.
    Invoked in Appendix A.2 to justify training the imputation model on out-of-state voter files and applying it to NY property records.
  • domain assumption PPP loan applicants' self-reported race is a valid ground truth for property owner race.
    Appendix B uses the NY PPP dataset to evaluate both models; PPP applicants are business owners and may not represent the property-owner population.
  • ad hoc to paper Model error rates from PPP validation can be applied as direct adjustments to tract-level predicted ownership shares.
    Section 3.4 subtracts White FPR from White predicted share and divides minority shares by (1-FNR) without computing predictive values or renormalizing.
  • domain assumption Including tract demographics as model input does not invalidate comparing predicted owner race to resident population.
    The Full Model uses census tract demographics as a feature (Appendix A.2), so predictions and the comparison benchmark share a common input.
  • ad hoc to paper Corporate-owned properties can be counted as White or external control in combined ownership metrics.
    Section 3.5 adds White individual ownership to corporate ownership; the paper concedes corporations have no attributed race.
  • domain assumption Comparing total properties, including rentals, to resident population is a meaningful disparity measure.
    The paper does not separate owner-occupied from investor-owned properties (Section 4), so rental stock and absentee landlords can drive part of the gap.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NY Real Estate Racial Equity Analysis via Applied Machine Learning." pith.science (2026). https://pith.science/paper/T2A6L46D

@misc{pith2026250516946,
  author       = {Pith},
  title        = {Pith review of: NY Real Estate Racial Equity Analysis via Applied Machine Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T2A6L46D}},
  note         = {Machine review of arXiv:2505.16946}
}
read the original abstract

This study analyzes tract-level real estate ownership patterns in New York State (NYS) and New York City (NYC) to uncover racial disparities. We use an advanced race/ethnicity imputation model (LSTM+Geo with XGBoost filtering, validated at 89.2% accuracy) to compare the predicted racial composition of property owners to the resident population from census data. We examine both a Full Model (statewide) and a Name-Only LSTM Model (NYC) to assess how incorporating geospatial context affects our predictions and disparity estimates. The results reveal significant inequities: White individuals hold a disproportionate share of properties and property value relative to their population, while Black, Hispanic, and Asian communities are underrepresented as property owners. These disparities are most pronounced in minority-majority neighborhoods, where ownership is predominantly White despite a predominantly non-White population. Corporate ownership (LLCs, trusts, etc.) exacerbates these gaps by reducing owner-occupied opportunities in urban minority communities. We provide a breakdown of ownership vs. population by race for majority-White, -Black, -Hispanic, and -Asian tracts, identify those with extreme ownership disparities, and compare patterns in urban, suburban, and rural contexts. The findings underscore persistent racial inequity in property ownership, reflecting broader historical and socio-economic forces, and highlight the importance of data-driven approaches to address these issues.

Figures

Figures reproduced from arXiv: 2505.16946 by the authors.

Figure 1
Figure 1. Population vs. Predicted Owner Racial Composition in Tracts by Majority [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Ownership Disparity (Ownership Share - Population Share) vs. Population [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Accuracy by census-tract income decile for the Full Model and Name Only [PITH_FULL_IMAGE:figures/full_fig_p032_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

  1. [1]

    Aalbers, M. (2016). The Financialization of Housing: A Political Economy Approach . Routledge

    work page 2016

  2. [2]

    Argyle, L. P. and Barber, M. (2023). Misclassification and bias in predictions of individual ethnicity from administrative records. American Political Science Review , pages 1--9

    work page 2023

  3. [3]

    censusgeo: Census geocoding api

    Census Geocoder (2020). censusgeo: Census geocoding api. https://geocoding.geo.census.gov/geocoder/. Census api converting address to geography information (county/tract/block)

    work page 2020

  4. [4]

    Chalavadi, S., Pastor, A., and Leitch, T. (2025). Lstm+geo with xgboost filtering: A novel approach for race and ethnicity imputation with reduced bias. arXiv:2504.21259. arXiv preprint arXiv:2504.21259. Available at https://arxiv.org/pdf/2504.21259

    work page Pith review arXiv 2025

  5. [5]

    Dantzler, P., Korver-Glenn, E., and Howell, J. (2022). Introduction: What does racial capitalism have to Do with Cities and Communities ? City & Community , 21(3):163--172

    work page 2022

  6. [6]

    N., Morrison, P

    Elliott, M. N., Morrison, P. A., Fremont, A., McCaffrey, D. F., Pantoja, P., and Lurie, N. (2009). Using the census bureau's surname list to improve estimates of race/ethnicity and associated disparities. Health Services and Outcomes Research Methodology , 9(2):69--83

    work page 2009

  7. [7]

    and Raymond, E

    Fields, D. and Raymond, E. (2021). Racialized geographies of housing financialization. Progress in Human Geography , 45:030913252110092

    work page 2021

  8. [8]

    and Uffer, S

    Fields, D. and Uffer, S. (2016). The financialisation of rental housing: A comparative analysis of new york city and berlin. Urban Studies , 53(7):1486--1502

    work page 2016

Show all 34 references
  1. [9]

    Goodman, L., Zhu, J., and Guild, M. (2023). A profile of institutional investor--owned single-family rental properties. Technical report, Urban Institute

    2023

  2. [10]

    M., Duranti-Mart\' i nez, J., Winston, F., Anderson, S., Udell, J., Kirk, C., and Hendra, R

    Greenberg, D. M., Duranti-Mart\' i nez, J., Winston, F., Anderson, S., Udell, J., Kirk, C., and Hendra, R. D. (2022). Gambling with homes, or investing in communities: How speculation drives evictions and poor housing quality, and how affordable housing protects neighborhoods ...

    2022

  3. [11]

    and Tegeler, P

    Haberle, M. and Tegeler, P. (2020). A call to remedy segregation and advance housing justice: Federal strategies for 2021 and beyond. Poverty & Race , 29(6). Poverty & Race Research Action Council (PRRAC), November/December 2020 Issue

    2020

  4. [12]

    and Korver-Glenn, E

    Howell, J. and Korver-Glenn, E. (2020). The increasing effect of neighborhood racial composition on housing values, 1980–2015. Social Problems , 68(4):1051--1071

    2020

  5. [13]

    and Nguyen, M

    Kwon, J. and Nguyen, M. (2023). Four decades of research on racial equity and justice in urban planning. Journal of Planning Literature , 38(2):1--27

    2023

  6. [14]

    Lees, L. (2008). Gentrification and social mixing: towards an inclusive urban renaissance? Urban studies , 45(12):2449--2470

    2008

  7. [15]

    Massey, D. S. and Denton, N. A. (1993). American apartheid: Segregation and the making of the underclass . Harvard University Press

    1993

  8. [16]

    Existing-home sales data

    National Association of Realtors (2024). Existing-home sales data. Reported by Investopedia, accessed 2024-05-18

    2024

  9. [17]

    Property Valuation and Assessment Data -- Tax Classes

    New York City Department of Finance (2025). Property Valuation and Assessment Data -- Tax Classes . https://data.cityofnewyork.us/City-Government/Property-Valuation-and-Assessment-Data-Tax-Classes/8y4t-faws/about_data. accessed: 2025-05-14

    2025

  10. [18]

    Property Assessment Data from Local Assessment Rolls

    New York State (2025). Property Assessment Data from Local Assessment Rolls . https://data.ny.gov/Government-Finance/Property-Assessment-Data-from-Local-Assessment-Rol/7vem-aaz7/about_data. accessed: 2025-05-15

    2025

  11. [19]

    Annual housing market report 2021

    New York State Association of REALTORS (2022). Annual housing market report 2021. Accessed 2024-05-18

    2022

  12. [20]

    Inquiry into redlining practices

    New York State Department of Financial Services (2022). Inquiry into redlining practices. Accessed May 2025

    2022

  13. [21]

    Annual minority mortgage lending report

    New York State Department of Financial Services (2024). Annual minority mortgage lending report. Accessed May 2025

    2024

  14. [22]

    State of the city’s housing stock 2023

    NYU Furman Center (2023). State of the city’s housing stock 2023. Accessed 2024-05-18

    2023

  15. [23]

    Homeownership rates in new york

    Office of the New York State Comptroller (2021). Homeownership rates in new york. Accessed: May 18, 2025

    2021

  16. [24]

    Sampson, R. J. (2012). Great American city: Chicago and the enduring neighborhood effect . University of Chicago Press

    2012

  17. [25]

    Smith, N. (1979). Toward a theory of gentrification: A back to the city movement by capital, not people. Journal of the American Planning Association , 45(4):538--548

    1979

  18. [26]

    Smith, N. (2002). New globalism, new urbanism: Gentrification as global urban strategy. Antipode , 34(3):427--450

    2002

  19. [27]

    Taylor, K.-Y. (2019). Race for Profit: How Banks and the Real Estate Industry Undermined Black Homeownership . University of North Carolina Press, Chapel Hill, NC

    2019

  20. [28]

    Tegeler, P. (2020). Public or community control of rental housing policy brief no. 1: Right of first refusal – tenant opportunity to purchase acts. Accessed 2025-04-13

    2020

  21. [29]

    Census Bureau (2000)

    U.S. Census Bureau (2000). Historical census of housing tables: Home values. Accessed 2024-05-18

    2000

  22. [30]

    Census Bureau (2021)

    U.S. Census Bureau (2021). 2020 census state profile: New york. Accessed: May 18, 2025

    2021

  23. [31]

    Small Business Administration (2021a)

    U.S. Small Business Administration (2021a). Economic injury disaster loan (eidl) data. https://data.sba.gov/dataset/covid-19-eidl. Accessed May 2025

    2021

  24. [32]

    Small Business Administration (2021b)

    U.S. Small Business Administration (2021b). Paycheck protection program (ppp) loan data. https://data.sba.gov/dataset/ppp-foia. Accessed May 2025

    2021

  25. [33]

    Xie, F. (2021). Rethnicity: Predicting ethnicity from names. arXiv preprint arXiv:2109.09228

    2021 arXiv

  26. [34]

    Zillow Research (2023). U.s. housing market surpasses \ 51.9 trillion in value. Accessed 2024-05-18

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.