Research | ISSUES | The Ohio State University

What 25 Years of Jury Selection Data Reveal About the Questions Raised by the Karmelo Anthony Trial

Fri, 10 Jul 2026 00:00:00 +0000

In Batson v. Kentucky (1986), the Supreme Court held that prosecutors may not exclude potential jurors solely because of their race. Nearly four decades later, questions about racial disparities in jury selection continue to arise. They resurfaced during the 2026 murder trial of Karmelo Anthony, where the defense raised a Batson challenge after all eligible Black jurors were removed from the pool and an all-White jury was ultimately seated.

The Anthony case cannot be evaluated using the Mississippi data analyzed here. It occurred in a different state, a different county, and under different circumstances, and nothing below should be read as a claim about it. However, the case does bring up an important question that the criminal legal system should confront: if race-based juror exclusion is unconstitutional, do measurable racial disparities in jury selection still persist? To examine that question, I reanalyzed a publicly available dataset assembled by APM Reports containing jury selection records from 305 criminal trials in Mississippi’s Fifth Circuit Court District spanning roughly 1991–2017.

Five Key Findings

Racial disparity in prosecution strikes persisted after Batson. Black jurors remained roughly 37–38 percentage points more likely than non-Black jurors to be struck by prosecutors, even as overall strike rates declined.
All-White juries occurred far more often than a race-neutral process would predict. Across 218 trials with complete jury-composition data, 14 juries contained no Black members. Under race-neutral selection, the model predicts roughly 1.3.
The disparity was robust across statistical approaches. The odds of a Black eligible juror being struck are nearly eight times those of non-Black jurors, and the gap persisted after adjusting for county, defendant race, and trial-level clustering.
The disparity persisted even when defendants were White. The pattern cannot be explained solely as an effort to match juror demographics to defendant demographics.
Every county examined showed a statistically significant disparity. The magnitude varied; no county showed the gap disappearing.

These are descriptions of patterns in a particular, non-random sample of trials. They are not, and should not be read as, determinations of intent in any individual decision, nor as population estimates for trials in general.

Overview

In 2018, reporters working on the podcast In the Dark published a detailed statistical investigation of jury selection in Mississippi’s Fifth Circuit Court District. Their analysis documented that, across hundreds of criminal trials in seven counties, prosecutors used their discretionary “peremptory” challenges to remove Black prospective jurors at rates far exceeding those for white jurors. That investigation is available at: features.apmreports.org/in-the-dark/mississippi-district-attorney-striking-black-jurors/

This article independently replicates the findings reported by APM Reports using publicly available jury selection data and extends the analysis through multilevel modeling, county-level decomposition, time-trend analyses, and probability-based estimates of how often the observed outcomes would be expected under race-neutral jury selection. The article concludes by placing the 2026 murder trial of Karmelo Anthony in a broader historical context. While the Mississippi data cannot determine what occurred in that specific case, the questions it raised about Batson challenges, race-neutral explanations for juror strikes, and the seating of an all-White jury are consistent with these historical patterns.

Throughout, the analysis is descriptive and statistical. Racial disparities in aggregate data are documented as patterns; they are not, and should not, be read as determinations of intent in any individual decision.

About the Dataset

The APM Reports jury dataset and analysis are available at github.com/APM-Reports/jury-data. It covers 305 criminal trials across seven Mississippi counties that are within the Fifth Circuit Court District. The trial coverage is roughly 1991 to 2017, although not all data are complete. All cases were ultimately appealed, creating selection bias that makes the sample non-random and may have analytic implications. Nevertheless, anyone can easily replicate the analysis below using the code shown in each code block.

For each trial, the dataset records the defendant’s race, charges, judge, and prosecuting attorneys. For each of the 14,874 individual prospective jurors, it records race, gender, what happened during selection (struck for cause, struck by state, struck by defense, or seated), and a strike_eligibility field indicating which side retained peremptory power at that juror’s position in the process.

jurors_clean <- jurors |>
 mutate(
 race_clean = str_to_lower(str_trim(race)),
 struck_by_clean = str_to_lower(str_trim(struck_by)),
 race_group = case_when(
 race_clean == "black" ~ "Black",
 race_clean %in% c("white", "asian", "latino") ~ "Non-Black",
 TRUE ~ NA_character_
 ),
 state_struck = if_else(struck_by_clean == "struck by the state", 1L, 0L)
 )

jury_df <- jurors_clean |>
 left_join(
 trials |> select(id, county, defendant_race, judge,
 prosecutor_1, prosecutor_2, prosecutor_3,
 verdict, cause_number, batson_claim_by_defense),
 by = c("trial_id" = "id")
 )

article_universe <- jurors_clean |>
 filter(
 race_clean %in% c("black", "white"),
 !struck_by %in% c("Juror not struck", "Struck without notation",
 "Juror excused/absent", "Unknown")
 )

eligible_df <- jury_df |>
 filter(
 !is.na(race_group),
 strike_eligibility %in% c("Both State and Defense", "State")
 ) |>
 mutate(
 race_group = factor(race_group, levels = c("Non-Black", "Black")),
 defendant_race = as.factor(defendant_race),
 county = as.factor(county)
 )

Below, “Non-Black” combines white, Asian, and Latino jurors, who together are overwhelmingly white in these counties; I use “non-Black” when the comparison is the analytic contrast, and “white” only when the underlying group is effectively all White (for example, when describing all-White juries).

Replication of Headline Statistics

APM Reports described a pool of 6,763 prospective jurors. I identify the equivalent group as all jurors recorded as Black or White who received a definitive outcome, excluding ambiguous categories (“Juror not struck,” “Struck without notation,” and similar).

sfc <- article_universe |> filter(struck_by == "Struck for cause")
after_cause <- article_universe |> filter(struck_by != "Struck for cause")
prose_str <- article_universe |> filter(struck_by == "Struck by the state")
def_str <- article_universe |> filter(struck_by == "Struck by the defense")
served <- article_universe |>
 filter(struck_by %in% c("Juror chosen to serve on jury", "Juror chosen as alternate"))
after_pros <- article_universe |>
 filter(struck_by %in% c("Struck by the defense",
 "Juror chosen to serve on jury",
 "Juror chosen as alternate"))

tibble(
 Stage = c("Total prospective pool", "Struck for cause",
 "Pool entering peremptory", "Prosecution strikes",
 "Pool after prosecution", "Defense strikes",
 "Seated (jurors + alternates)"),
 `My N` = c(nrow(article_universe), nrow(sfc), nrow(after_cause),
 nrow(prose_str), nrow(after_pros), nrow(def_str), nrow(served)),
 `Article N` = c("6,763","1,342","~5,421","1,275","-","1,458","2,688"),
 `My % Black` = c(
 percent(mean(article_universe$race_clean == "black"), 0.1),
 percent(mean(sfc$race_clean == "black"), 0.1),
 percent(mean(after_cause$race_clean == "black"), 0.1),
 percent(mean(prose_str$race_clean == "black"), 0.1),
 percent(mean(after_pros$race_clean == "black"), 0.1),
 percent(mean(def_str$race_clean == "black"), 0.1),
 percent(mean(served$race_clean == "black"), 0.1)
 ),
 `Article % Black` = c("39%","57%","35%","71%","21%","9%","32%")
)

# A tibble: 7 × 5
 Stage `My N` `Article N` `My % Black` `Article % Black`
 <chr> <int> <chr> <chr> <chr>
1 Total prospective pool 6795 6,763 39.1% 39%
2 Struck for cause 1342 1,342 56.6% 57%
3 Pool entering peremptory 5453 ~5,421 34.8% 35%
4 Prosecution strikes 1285 1,275 70.9% 71%
5 Pool after prosecution 4168 - 23.7% 21%
6 Defense strikes 1466 1,458 8.7% 9%
7 Seated (jurors + alternates) 2702 2,688 31.9% 32%

All percentage figures replicate to within rounding error. The point of reproducing the previous analysis is to verify before building on it.

The Flow of Jurors

The prosecution strike flow shows that 70.9% of state strikes removed Black jurors. The defense flow runs the other way, 91.3% white, but only because, by that stage, the pool was already heavily white. Together, they show how Black jurors entered the process at roughly their expected share (39.1%) yet were “funneled” away from selection at specific stages of voir dire.

Statistical Analysis

Overall Strike Rates

Among jurors eligible for peremptory challenges, 49.8% of Black jurors were struck by the prosecution, compared with 11.2% of non-Black jurors. The unadjusted odds ratio is 7.84 (95% CI: 6.81–9.04). Adjusting for defendant race and county yields an OR of 8.25, indicating that the disparity persists after accounting for these factors.

# A tibble: 2 × 5
 race_group n struck_n strike_rate pct
 <fct> <int> <int> <dbl> <chr>
1 Non-Black 3320 373 0.112 11.2%
2 Black 1811 902 0.498 49.8%

County-level Variation

The disparity appears across all counties, with odds ratios ranging from 4.6 (Choctaw) to 17.3 (Carroll). In Carroll County, an eligible Black juror was struck at roughly 17 times the odds of a non-Black juror. All county-level estimates are statistically significant (p < 0.001).

# A tibble: 7 × 5
# Groups: county [7]
 county OR lower upper p.value
 <fct> <dbl> <dbl> <dbl> <dbl>
1 Carroll 17.3 8.09 40.8 3.19e-12
2 Montgomery 12.6 8.62 18.7 9.82e-38
3 Webster 9.97 3.94 26.9 2.21e- 6
4 Grenada 9.21 7.10 12.0 5.93e-61
5 Winston 8.14 5.35 12.6 7.18e-22
6 Attala 6.14 4.64 8.17 2.94e-36
7 Choctaw 4.60 2.92 7.26 4.40e-11

Does Defendant’s Race Matter?

When the defendant is Black, 52.6% of Black jurors are struck versus 9.7% of non-Black jurors. In White-defendant cases, the absolute rates differ (Black: 40.0%, non-Black: 16.2%), but the gap persists. The interaction term (OR = 0.34, p < 0.001) indicates that the disparity is statistically smaller in White-defendant cases but not eliminated.

# A tibble: 4 × 5
 term estimate conf.low conf.high p.value
 <chr> <dbl> <dbl> <dbl> <dbl>
1 (Intercept) 0.108 0.0946 0.123 4.59e-244
2 race_groupBlack 10.3 8.69 12.2 4.63e-162
3 defendant_raceWhite 1.79 1.40 2.27 2.10e- 6
4 race_groupBlack:defendant_raceWhite 0.336 0.242 0.468 9.48e- 11

Time Trend, 1992–2012

Over the span with reasonably complete data, both strike rates drifted downward together, and the gap between them barely moved. As shown by the smoothed (LOESS) trends, the Black-juror strike rate declined modestly while the non-Black rate stayed roughly flat, leaving a gap that began near 38 percentage points and ended near 32, averaging about 37 points across the period. The stability of the gap is the primary takeaway.

County-level Poverty

Six of the seven counties have poverty rates between roughly 22% and 26%, near the Mississippi median (~23%). Carroll County has the lowest poverty rate (~14%) yet the highest odds ratio (17.3). With only seven counties, this cannot support an inferential claim, but there is no visible tendency for poorer counties to show larger disparities, so local economic conditions are not an obvious explanation for the pattern.

Multilevel Model: Accounting for Clustering

Standard logistic regression treats each juror as independent. In reality, jurors in the same trial share a common local context, such as the same prosecutor, judge, defendant, and case facts, creating within-trial correlation. Ignoring this correlation produces artificially narrow confidence intervals. A mixed-effects (multilevel) logistic model accounts for this by estimating a random intercept for each trial.

The multilevel model yields an OR of 8.55 (95% CI: 7.34–9.95), compared with 7.84 from the simple logistic model. The ICC of 4.5% indicates that approximately 4.5% of the variation in strike decisions reflects between-trial differences, although the majority of the variation is within trials.

# A tibble: 3 × 4
 Model OR `95% CI` Notes
 <chr> <dbl> <chr> <chr>
1 Simple logistic (no clustering) 7.84 6.81–9.04 Assumes jurors …
2 Multilevel logistic (jurors nested in trials) 8.55 7.34–9.95 ICC = 4.5%
3 Multilevel logistic (+ defendant race) 8.54 7.33–9.95 Adjusts for def…

The Mathematics of Exclusion

How Likely are All-White Juries?

After strikes for cause, 34.8% of the eligible pool in this district was Black. The question is how often an all-White twelve-person jury should appear under race-neutral selection, and how it changes after prosecution strikes are applied.

Under a reasonable approximation, the chance of seating a twelve-person jury with no Black members from a pool that is 34.8% Black is about 0.6%, roughly 1 in 170. After applying the observed prosecution strike rates, the effective share of Black jurors available at seating falls to about 23.7%, and the probability of an all-White jury rises to about 3.9%, roughly 1 in 26.

All else equal, the observed strike behavior is associated with an all-White jury becoming about 6.6 times more likely than the race-neutral benchmark.

Across Many Trials

Small per-trial differences compound across many trials. The table below shows the probability of observing at least one all-White jury under the race-neutral benchmark and under the observed strike rates.

Number of Trials	Race-Neutral Selection	Observed Strike Rates
10	5.7%	32.7%
25	13.7%	62.8%
50	25.4%	86.1%
100	44.4%	98.1%
500	94.7%	100.0%
1,000	99.7%	100.0%

Observed Versus Expected

Across 218 trials with complete jury-composition data, 14 juries contained no Black members. Under the race-neutral benchmark, the expected count is about 1.3. Therefore, the observed number is roughly 11 times the benchmark. An exact binomial test rejects a race-neutral process at p < .001.

What These Data Can—and Cannot—Tell Us

Every trial here was appealed, and appeal is not random, so the rates (the 49.8% strike rate, the “1 in 26” probability, the per-trial figures) describe this non-random set of trials, not jury selection in the district at large or elsewhere. But selection bias cannot easily explain the disparity’s direction, consistency, and magnitude: an eightfold odds ratio that survives adjustment for county and defendant race, appears in every county at p < 0.001, and holds even when the defendant is White is not what a selection artifact looks like.

The data still cannot establish intent by any prosecutor or why any particular juror was struck; those turn on voir dire responses and case-specific reasons absent here. What it shows is an aggregate disparity: prosecutors struck Black eligible jurors at far higher rates than non-Black jurors, and 14 of 218 trials produced all-White juries, where a binomial model based on each county’s Black share predicts only about 1.3. That does not prove discrimination, but it sharpens the question the Anthony trial revived: if race-based exclusion is unconstitutional, why do disparities this large persist decades after Batson?

Transparency, Data, and the Future of Jury Selection Analysis

Debates about jury selection tend to run on speculation and competing narratives precisely because the underlying records are rarely available for independent evaluation. When data do exist, analyzing them is not straightforward, and the results must be read within an appropriate legal and statistical frame, including an unbiased account of how the data were chosen, cleaned, and analyzed.

Expanding access to court records, voir dire transcripts, and structured jury-selection datasets would let researchers, journalists, attorneys, and the public assess claims with evidence rather than assumption, and would let the kind of selection problem flagged above be measured rather than merely conceded.

How to Analyze Caselaw Using CourtListener's API in R

Fri, 01 Aug 2025 00:00:00 +0000

I previously showed how to pull Indian Child Welfare Act (ICWA) cases using the Caselaw Access Project (Harvard Law). That API has since changed. This post updates the workflow using the Free Law Project’s CourtListener, which now partners with Harvard to make this data freely available.

Finding the Cases

Start at the CourtListener website and search for “Indian Child Welfare Act” (or "icwa"). You’ll see a page of opinion results.

Click Select Jurisdiction → State. Clear all, then select Ohio Supreme Court, Ohio Court of Appeals, and Ohio Court of Claims. This gives you a quick check on the expected volume to validate the code’s output below.

The site reported 36 cases from Ohio. Note that the exact number may change over time as new opinions are added or metadata is corrected.

Getting the Data into R

The Search API returns clean JSON results by querying for opinions that contain the exact phrase "indian child welfare act", restricting to published opinions, ordering by relevance, and (optionally) limiting results to three Ohio courts.

library(httr2)
library(purrr)
library(dplyr)
library(tibble)
library(tidyr)
library(stringr)

SEARCH_BASE <- "https://www.courtlistener.com/api/rest/v4/search/"
UA <- "R/httr2 (your-email-address) – ICWA metadata collection"

query_params <- list(
 q = '"indian child welfare act"',
 type = "o",
 order_by = "score desc",
 stat_Published = "on",
 court = "ohio ohioctapp ohioctcl" # optional: specify court
)

A few helper functions keep the parsing tidy:

pull_chr <- function(x, ...) {
 v <- purrr::pluck(x, ..., .default = NA)
 if (is.null(v) || length(v) == 0) return(NA_character_)
 as.character(v)
}

pull_num <- function(x, ...) {
 v <- purrr::pluck(x, ..., .default = NA_real_)
 suppressWarnings(as.numeric(v))
}

get_json <- function(url) {
 resp <- request(url) |> req_user_agent(UA) |> req_url_query() |> req_perform()
 resp_check_status(resp)
 resp_body_json(resp, simplifyVector = FALSE)
}

This function handles pagination automatically, following the next cursor until all pages are exhausted:

fetch_search_pages <- function() {
 req <- request(SEARCH_BASE) |>
 req_user_agent(UA) |>
 req_url_query(!!!query_params)

 dat <- req |> req_perform() |>
 (\(r){ resp_check_status(r); resp_body_json(r, simplifyVector = FALSE) })()
 pages <- list(dat)

 while (!is.null(dat$`next`)) {
 dat <- request(dat$`next`) |> req_user_agent(UA) |> req_perform() |>
 (\(r){ resp_check_status(r); resp_body_json(r, simplifyVector = FALSE) })()
 pages <- append(pages, list(dat))
 }
 pages
}

pages <- fetch_search_pages()
results <- pages |> map("results") |> flatten()

results is a list of per-case records from the API. Each element includes the case name, filing date, court identifiers, citation strings, a public URL where the caselaw text is available, and other metadata.

For Ohio, this produced ~35 unique case names versus 36 displayed on the site. Minor discrepancies happen because search hits can include multiple opinions or sibling opinions per matter, and because the index is updated over time.

Downloading the Full Text

Each opinion has an API endpoint with plain-text and HTML fields. A small loop can request the text for each case URL, with a short pause and retry between requests, and save one .txt file per opinion:

download_opinion_text <- function(url, out_dir = "opinions") {
 dir.create(out_dir, showWarnings = FALSE)
 slug <- basename(url)
 path <- file.path(out_dir, paste0(slug, ".txt"))

 resp <- request(url) |> req_user_agent(UA) |> req_perform()
 resp_check_status(resp)
 body <- resp_body_json(resp, simplifyVector = FALSE)

 text <- body$plain_text %||% body$html_with_citations %||% ""
 writeLines(text, path)
 Sys.sleep(0.5) # be considerate of rate limits
 invisible(path)
}

walk(map_chr(results, "absolute_url"), download_opinion_text)

Here are the Ohio ICWA cases downloaded to a local folder:

Visualizing the Results

You can use ggplot2 to visualize the results. Below are two examples: filings by year and volume by court.

What You Can Do Next

After downloading all data — including caselaw text and metadata — you can:

Natural Language Processing — examine sentiment or common words across opinions
Topic Modeling — explore latent themes in ICWA cases using LDA or STM
Map filings by jurisdiction — visualize geographic patterns of ICWA litigation
Integrate with other sources — link to child welfare administrative data or Census indicators

Remember to be considerate of rate limits and to include a User Agent string with your contact information in every request.

The complete workflow is available on my GitHub.

Happy Coding!

A PostgreSQL Tutorial for Analyzing Racial and Geographic Disparities in HMDA Applications

Wed, 16 Jul 2025 00:00:00 +0000

My lab recently published a paper examining the association between redlining, reinvestment, racial segregation, and firearm-related injury in Chicago, Illinois. You can read the paper here: Redlining, reinvestment, and racial segregation: a Bayesian spatial analysis of mortgage lending trajectories and firearm-related violence.

To develop reverse redlining measures, we relied on data from the Home Mortgage Disclosure Act (HMDA).

If you’re interested in patterns of lending discrimination, the HMDA dataset is an invaluable resource. It contains millions of individual loan application records submitted annually by financial institutions, making it one of the most extensive and detailed public data sources on mortgage lending in the United States. The 2024 nationwide Modified Loan/Application Register (LAR) alone contains over 12 million records and 85 fields that detail applicant and co-applicant demographics, loan terms, and property characteristics — including race, income, loan amount, interest rate, property value, and loan outcomes.

Each record includes the census tract of the loan origination, which allows researchers to situate applicants within their geographic context. This spatial identifier can be linked to other data sources (e.g., Census or ACS data), enabling robust analyses of racial disparities in lending, neighborhood-level disinvestment, and broader patterns of residential inequality.

I highly recommend The Color of Law by Richard Rothstein for anyone interested in the deep-rooted history of discriminatory housing practices in the United States. The book compellingly explains how federal, state, and local policies intentionally created and reinforced de jure segregation. Chicago, one of the most highly segregated urban areas in the country, is prominently featured.

Due to its sheer volume, the HMDA data is not well-suited for analysis using spreadsheets like Excel — particularly when analyzing multiple years or conducting joins. Instead, its relational structure lends itself to PostgreSQL, an open-source relational database management system (RDBMS) that enables efficient querying, filtering, and aggregation of large datasets. PostgreSQL also supports the PostGIS geospatial extension, enabling direct spatial querying and linking to geographic units such as census tracts or ZIP codes. Download PostgreSQL if you don’t already have it.

Step 1: Download the Dataset

The dataset is freely available to the public.

Navigate to https://ffiec.cfpb.gov/data-publication/modified-lar/2024
Choose Year: 2024 → Nationwide: Combined file
Check Include File Header, then click Download Combined Modified LAR with Header

Save or extract the file to: C:\Users\YOUR_USER_NAME\Downloads\2024_combined_mlar_header\

Step 2: Confirm the File Structure in PowerShell

The file is a pipe-delimited .txt file — it can be tricky to read in spreadsheet tools. Use Windows PowerShell to inspect the structure first:

cd "C:\Users\YOUR_USER_NAME\Downloads\2024_combined_mlar_header"

# View the header (column names)
Get-Content .\2024_combined_mlar_header.txt -TotalCount 1

# Count the number of columns (should be 85)
((Get-Content .\2024_combined_mlar_header.txt -TotalCount 1) -split '\|').Count

Step 3: Create the PostgreSQL Table

Launch your SQL interface (pgAdmin Query Tool or psql):

psql -U your_username -d your_database_name

Then define the 85-column table structure:

DROP TABLE IF EXISTS lar2024;
CREATE TABLE lar2024 (
 activity_year INTEGER,
 lei TEXT,
 loan_type INTEGER,
 loan_purpose INTEGER,
 preapproval INTEGER,
 construction_method INTEGER,
 occupancy_type INTEGER,
 loan_amount BIGINT,
 action_taken INTEGER,
 state_code TEXT,
 county_code TEXT,
 census_tract TEXT,
 applicant_ethnicity_1 INTEGER,
 applicant_ethnicity_2 TEXT,
 applicant_ethnicity_3 TEXT,
 applicant_ethnicity_4 TEXT,
 applicant_ethnicity_5 TEXT,
 co_applicant_ethnicity_1 INTEGER,
 co_applicant_ethnicity_2 TEXT,
 co_applicant_ethnicity_3 TEXT,
 co_applicant_ethnicity_4 TEXT,
 co_applicant_ethnicity_5 TEXT,
 applicant_ethnicity_observed INTEGER,
 co_applicant_ethnicity_observed INTEGER,
 applicant_race_1 INTEGER,
 applicant_race_2 TEXT,
 applicant_race_3 TEXT,
 applicant_race_4 TEXT,
 applicant_race_5 TEXT,
 co_applicant_race_1 INTEGER,
 co_applicant_race_2 TEXT,
 co_applicant_race_3 TEXT,
 co_applicant_race_4 TEXT,
 co_applicant_race_5 TEXT,
 applicant_race_observed INTEGER,
 co_applicant_race_observed INTEGER,
 applicant_sex INTEGER,
 co_applicant_sex INTEGER,
 applicant_sex_observed INTEGER,
 co_applicant_sex_observed INTEGER,
 applicant_age TEXT,
 applicant_age_above_62 TEXT,
 co_applicant_age TEXT,
 co_applicant_age_above_62 TEXT,
 income TEXT,
 purchaser_type INTEGER,
 rate_spread TEXT,
 hoepa_status INTEGER,
 lien_status INTEGER,
 applicant_credit_scoring_model INTEGER,
 co_applicant_credit_scoring_model INTEGER,
 denial_reason_1 TEXT,
 denial_reason_2 TEXT,
 denial_reason_3 TEXT,
 denial_reason_4 TEXT,
 total_loan_costs TEXT,
 total_points_and_fees TEXT,
 origination_charges TEXT,
 discount_points TEXT,
 lender_credits TEXT,
 interest_rate TEXT,
 prepayment_penalty_term TEXT,
 debt_to_income_ratio TEXT,
 combined_loan_to_value_ratio TEXT,
 loan_term TEXT,
 intro_rate_period TEXT,
 balloon_payment INTEGER,
 interest_only_payment INTEGER,
 negative_amortization INTEGER,
 other_non_amortizing_features INTEGER,
 property_value TEXT,
 manufactured_home_secured_property_type INTEGER,
 manufactured_home_land_property_interest INTEGER,
 total_units TEXT,
 multifamily_affordable_units TEXT,
 submission_of_application INTEGER,
 initially_payable_to_institution INTEGER,
 aus_1 INTEGER,
 aus_2 TEXT,
 aus_3 TEXT,
 aus_4 TEXT,
 aus_5 TEXT,
 reverse_mortgage INTEGER,
 open_end_line_of_credit INTEGER,
 business_or_commercial_purpose INTEGER
);

Step 4: Import the Data

In psql, use the following single-line import command:

\copy lar2024 FROM 'C:/Users/YOUR-USER-NAME/Downloads/2024_combined_mlar_header/2024_combined_mlar_header.txt' WITH (FORMAT csv, DELIMITER '|', HEADER true, NULL '');

This tells PostgreSQL to:

Read a pipe-delimited file
Use the header row to match columns
Treat empty strings as NULL

Note: This must be entered as a single line. You will get a confirmation message when the data import succeeds.

Step 5: Verify the Import

The official CFPB download page indicates that the 2024 nationwide file contains approximately 12 million rows. After importing, verify with:

SELECT COUNT(*) FROM lar2024;
SELECT * FROM lar2024 LIMIT 5;

If the row count is substantially less than 12 million, the import did not complete successfully.

Create the Chicago, Illinois Subset

With the full dataset loaded, you can create a focused subset with a CREATE TABLE AS SELECT statement. The query below selects conventional, owner-occupied, site-built single-family loans in Cook County (FIPS 17031), excluding missing or exempt values:

DROP TABLE IF EXISTS public.il2024f1;
CREATE TABLE public.il2024f1 AS
SELECT
 l.activity_year,
 l.state_code,
 l.county_code,
 l.census_tract,
 l.debt_to_income_ratio,
 l.combined_loan_to_value_ratio,
 l.income,
 l.property_value,
 l.loan_amount,
 l.interest_rate,
 l.applicant_ethnicity_1,
 l.applicant_race_1,
 l.denial_reason_1,
 l.rate_spread,
 l.hoepa_status,
 l.lien_status
FROM public.lar2024 l
WHERE
 l.loan_purpose = '1' -- Home purchase
 AND l.action_taken = '1' -- Loan originated
 AND l.loan_type = '1' -- Conventional
 AND l.occupancy_type = '1' -- Owner-occupied
 AND l.state_code = 'IL'
 AND l.county_code = '17031' -- Cook County
 AND l.property_value ~ '^[0-9]+$'
 AND l.income ~ '^[0-9]+$'
 AND l.loan_amount IS NOT NULL
 AND l.loan_amount <> 99999
 AND l.construction_method = '1' -- Site-built
 AND l.total_units IN ('1','2','3','4','');

Prepare for Additional Analysis

Once exported to CSV, I divided the number of Black applicants in each census tract by the total number of applicants in that tract to get the share of Black applicants. I also downloaded Median Household Income data using the tidycensus package in R (ACS 5-year estimates 2018–2023).

The maps below compare neighborhood income levels with the geographic concentration of Black mortgage applicants receiving loans in 2024. Do you observe any patterns?

The spatial overlap — or rather, the striking lack of it — between high-income areas and Black loan recipients is consistent with decades of research documenting structural barriers to homeownership for Black families in Chicago.

Conclusion

The HMDA dataset offers an unparalleled opportunity to examine lending patterns, racial disparities, and neighborhood-level investment across the United States. By importing the 2024 Modified LAR into PostgreSQL, researchers can efficiently manage, filter, and analyze millions of records while linking them to geographic and demographic data.

Next steps may include:

Join with Census shapefiles to visualize patterns by tract
Summarize by loan purpose, race, or geography
Export to PostGIS for spatial queries
Link to neighborhood typologies or redlining indices

Happy Coding!