3  Data and provenance

Every number in this paper traces to one source: the Community Development Financial Institutions (CDFI) Fund’s New Markets Tax Credit Public Data Release, FY2003–FY2022, released June 2024. The data is in the U.S. public domain (17 USC §105) and we redistribute it unmodified alongside the cleaning code.

3.1 The source

Community Development Financial Institutions (CDFI) Fund, U.S. Department of the Treasury. New Markets Tax Credit Public Data Release, FY2003–FY2022. Released June 2024. https://www.cdfifund.gov/documents/data-releases

We did not request or receive any non-public data. Treasury is statutorily required to publish program activity once allocations have run their course; the published transaction-level dataset is the backbone of the CDFI Fund’s program-evaluation literature (Abravanel et al. 2013; Theodos, Stacy, Teles, Davis, and Hariharan 2021; Theodos, Stacy, Teles, Davis, Rajasekaran, et al. 2021).

3.2 What we downloaded

file size purpose
NMTC_Public_Data_Release_FY2003-FY2022.xlsx 1.95 MB the actual transaction- and project-level records
NMTC_Public_Data_Release_Summary_FY2003-FY2022.pdf 0.49 MB the CDFI Fund’s own codebook

Both are in data/raw/ of the project repository, unmodified from the bytes pulled off the CDFI Fund’s web server. The integrity hashes (recorded in PROVENANCE.md):

NMTC_Public_Data_Release_FY2003-FY2022.xlsx
  SHA-256: fa709714e93d67356b90a1c0f98dbed71ec1998d0b686e969ad3bacafc112683

NMTC_Public_Data_Release_Summary_FY2003-FY2022.pdf
  SHA-256: 6e205240f31670fe66095d506f6db08fadd65575b2f198ab52d95882364a1433

3.3 Structure of the source xlsx

The workbook contains two relevant sheets:

Sheet 1 — Financial Notes 1 - Data Set PU (19,907 rows). One row per QLICI transaction, each row a single financial flow from a CDE to a QALICB on a specific date.

Sheet 2 — Projects 2 - Data Set PUBLISH.P (8,024 rows). One row per project. The relationship is one-to-many: a project can receive multiple QLICIs from multiple CDEs in multiple tranches. Leverage is a project-level concept (project cost / total QLICI to that project), so all leverage analysis runs on Sheet 2.

3.4 The cleaning pipeline

Every step is in scripts/describe_nmtc.py (~240 lines, runs in ~25 seconds). In order:

  1. Strip whitespace from column headers. CDFI’s headers have stray whitespace.
  2. Rename columns to short snake_case names. The full mapping is in DATA_DICTIONARY.md.
  3. Normalize the metro flag. CDFI’s source column "Metro/Non-Metro, 2020 Census" contains three literal strings — "Metro", "Non-metro", and "Non-Metro" (inconsistent capitalization across years). We fold them into a clean lowercase metro column with three values: metro, non_metro, or unknown.
  4. Coerce dollar columns to numeric with pd.to_numeric(errors="coerce"). Coercion failures become NaN, which means specifically “the source had a non-numeric value” — a distinction that matters for sample-construction transparency.
  5. Compute project-level leverage:
pr["leverage_ratio"] = pr["project_cost"] / pr["project_qlici"]
pr["leverage_win"]   = pr["leverage_ratio"].clip(lower=1.0, upper=20.0)

Both columns are kept. leverage_win is what we use for summary statistics; leverage_ratio is the raw version for robustness checks.

  1. Justify the [1, 20] winsorization. About 1% of projects report project_cost < project_qlici (leverage < 1), implausible because the QLICI is part of the project cost by construction. About 0.3% report ratios > 20, typically a small QLICI into a very large RE stack. Capping at [1, 20] preserves the shape of the distribution but stops the right tail from dominating the mean.

  2. Group-by rollups produce the seven summary CSVs (by metro, by QALICB type, by year, top 20 CDEs, leverage distribution, multi-CDE crosstab).

  3. Emit headline.json with the program-level numbers.

3.5 Geocoding

The CDFI release gives a 2020 census tract FIPS code for each project but no latitude/longitude. We merge with the U.S. Census 2020 Tract National Gazetteer (16 MB tab-separated file, 85,395 tracts) which contains the internal-point lat/lon for every census tract:

pr = pr.merge(gaz, on="tract_fips", how="left")

8,019 of the 8,024 projects merge cleanly (99.94% match rate). The five unmatched projects have tract codes that were renumbered between the 2010 and 2020 censuses; they remain in the analytical tables but do not appear on the interactive map.

3.6 Reproducibility

From a blank machine, three commands rebuild every figure and every regression:

git clone https://github.com/ihelfrich/us-nmtc-viewer
cd us-nmtc-viewer
pip install -r requirements.txt
python3 scripts/describe_nmtc.py
python3 scripts/make_figures.py
python3 scripts/run_regressions.py

Total runtime: about two minutes on a laptop. Every number, figure, and dot on the map is regeneratable from the SHA-256-anchored raw xlsx in less than two minutes.

3.7 What’s not in the data

For transparency:

  • No QEI flow (investor → CDE side). We see only the downstream deployment.
  • No allocation-round data. CDFI’s allocation awards specify how much each CDE may deploy, but we observe the deployment, not the award.
  • No tract demographics beyond metro status. Poverty rate, MFI, population, and racial composition need to be merged from the American Community Survey for the LIC-eligibility regression- discontinuity extension.
  • Nominal dollars throughout. No CPI deflation in the headline figures.
  • CDE institutional form is unobserved. The data lists CDE names but does not tag them as bank-subsidiary / nonprofit / for-profit / government.