US Department of the Treasury · CDFI Fund · FY2001–FY2022
A blended-finance first look over 8,019 projects — the public dollar,
the private dollar, and the rural–urban leverage gap.
↓ scroll
The New Markets Tax Credit is the US federal program whose mechanics most closely match the "blended finance" story you want to tell. The federal government offers investors a 39% tax credit (claimed over 7 years) in exchange for putting their equity into a certified Community Development Entity (CDE). The CDE is a regulated intermediary — usually a bank subsidiary or a nonprofit community-development financial institution — that re-deploys the money as loans or equity into qualifying businesses and projects located in low-income census tracts. Those projects are called QALICBs.
What we observe is the CDE → QALICB flow (the QLICI): who got how much, where, in which year, and what the total project cost was. The investor-side flow (QEI, and the tax credit they claim) is upstream and off the public data release.
The NMTC statute requires CDEs to direct at least 20% of their QLICIs to non-metropolitan census tracts. That creates a testable prediction: CDEs will bunch against the 20% line unless they have a strong pull to go further.
The leverage ratio of a project is total project cost / QLICI. If a $5M
project received a $1M QLICI, it leveraged the federal dollar 5×, meaning $4 of
non-federal capital showed up per $1 of credit. That's the mobilization number.
Two sharp discontinuities the law gives us for free: (1) tracts qualify only if poverty ≥ 20% or median family income ≤ 80% of area median — an RDD cutoff; (2) the 20% rural mandate — a bunching test.
Everything below is computed from the CDFI Fund's public data release
(xlsx,
codebook)
via the cleaning pipeline in
scripts/describe_nmtc.py.
See DATA_DICTIONARY.md for every
column and PROVENANCE.md for SHA-256
hashes, license terms, and a step-by-step recreate-from-scratch recipe.
Each dot is one project, placed at its 2020-census-tract centroid. Metro · Non-metro. Dot size ∝ √(QLICI $). Hover for detail. Scroll-zoom to explore.
The program ramped from almost nothing in FY2001 to ~$5 B/yr through the 2010s, then settled into a $3–4 B/yr steady state. The stacked bar shows metro vs. non-metro dollar share; the panel below it pulls the non-metro share out as a line, against the 20% statutory mandate.
This is the empirical headline of the first-look brief. Non-metro projects sit heavier on the 1.0× floor (nearly 100% NMTC-financed, zero private capital mobilized) while metro projects show a fatter right shoulder (more private debt and equity stacked on top).
Same credit, same 39% federal subsidy, same statute. But the top-20 CDEs — which together account for more than half of all NMTC dollars — deploy wildly differently along the rural margin. This is where the paper probably lives.
| CDE | QLICI $M | tx | non-metro share | deployment |
|---|
Sorted by non-metro share (descending). The top row — Rural Development Partners LLC at ~80% non-metro — and the bottom few (ESIC, Consortium America, Capital Impact) at <5% are the same federal instrument deployed to different worlds. A within-CDE fixed-effects specification absorbs that selection.
Four specifications, all identifiable from the public release plus a routine Census-tract merge.
Project-level leverage is directly observed:
$$ \operatorname{Leverage}_i \;=\; \frac{\text{ProjectCost}_i}{\text{QLICI}_i} \qquad\text{and}\qquad \operatorname{Mobilization}_i \;=\; \operatorname{Leverage}_i - 1 $$A leverage of 1 means the project was 100% NMTC-financed. A leverage of 3 means each federal dollar pulled in $2 of non-federal capital.
Let $R_i \in \{0,1\}$ indicate non-metro status. A naive mean comparison is
$$ \operatorname{Leverage}_i \;=\; \alpha \;+\; \beta \, R_i \;+\; \varepsilon_i $$and the brief reports $\hat\beta \approx -0.26$ (mean) or $-0.12$ (median, via a quantile regression). But that's contaminated by which CDE deploys where and what type of project. The fixed-effects version:
$$ \operatorname{Leverage}_{i} \;=\; \alpha \;+\; \beta \, R_i \;+\; \gamma_{c(i)} \;+\; \delta_{t(i)} \;+\; \eta_{q(i)} \;+\; \varepsilon_i $$with $\gamma_{c(i)}$ a CDE fixed effect, $\delta_{t(i)}$ origination-year, and $\eta_{q(i)}$ a QALICB-type fixed effect. The quantity of interest is whether $\hat\beta$ survives the CDE fixed effect — i.e. does the same CDE mobilize less private capital in non-metro than in metro?
A census tract is NMTC-eligible if either condition holds:
$$ \text{LICeligible}_\ell \;=\; \mathbf{1}\!\left\{\;\text{Poverty}_\ell \ge 0.20 \;\;\vee\;\; \tfrac{\text{MFI}_\ell}{\text{AreaMFI}_\ell} \le 0.80\;\right\} $$so we have a sharp cutoff at poverty = 20%. Run a local-linear RDD separately for metro and non-metro tracts:
$$ Y_\ell \;=\; \alpha \;+\; \tau_{R} \cdot \mathbf{1}\{P_\ell \ge 0.20\} \;+\; f(P_\ell - 0.20) \;+\; R_\ell \cdot g(P_\ell - 0.20) \;+\; u_\ell $$where $Y_\ell$ is, say, tract-level NMTC dollars per capita or mean leverage. $\hat\tau_{R=\text{metro}}$ vs. $\hat\tau_{R=\text{non-metro}}$ is exactly the "does the policy bite differently in rural markets" question.
Let $s_j$ be CDE $j$'s realized non-metro share of QLICIs. The statute requires $s_j \ge 0.20$. Under no-mandate counterfactual, $s_j$ would be smooth around $0.20$. Under a binding mandate, we expect a visible mass at $s_j = 0.20$. Formally, compare the empirical density $\hat f(s)$ to a polynomial fit that excludes a window around the cutoff:
$$ B \;=\; \int_{0.20-h}^{0.20+h} \bigl[\hat f(s) - \tilde f(s)\bigr]\, ds $$$B > 0$ is the "excess mass" due to the mandate — standard Chetty et al. (2011) / Kleven (2016) machinery. CDEs that exceed the mandate voluntarily reveal willingness-to-deploy rurally; CDEs that pin the mandate are constrained. The marginal rural project identifies off the constrained CDEs.
The empirical paper is short: (2) establishes that the rural gap isn't a composition artefact (CDE FE absorbs identity, QALICB FE absorbs type); (3) shows that the RDD treatment effect differs between metro and non-metro (the interaction we care about); (4) validates the identifying assumption by showing CDEs bunch exactly where the law says they would.
See the full brief: briefs/katia_nmtc_v1.md.