9 Limitations – The Rural Mobilization Gap

9.1 Selection on observables, not causal

The fixed-effects strategy is a decomposition of the rural-vs-urban gap into observable components, not a causal estimate of “the effect of rural location on leverage.” Even within-CDE, deal allocation is not random. A CDE that decides to take on a rural deal has selected that deal from its opportunity set; the unobserved characteristics of the deal differ from those of its metro deals in ways the regression cannot absorb.

What the result tells us is: among the deals CDEs actually do, there is no within-CDE rural penalty. It does not tell us what would happen if a CDE were forced to do a randomly-selected rural deal. The LATE- style interpretation requires the LIC-eligibility regression discontinuity, which is a planned extension.

9.2 The LIC RDD is not yet in the paper

The fully causal piece of the empirical strategy compares NMTC investment outcomes for tracts at 19.9% poverty (ineligible) vs. 20.1% poverty (eligible), separately for metro and non-metro samples, with the differential local average treatment effect as the rural-versus- urban interaction estimator. This requires merging in tract-level demographics from the American Community Survey and is the next data move. Until that merge is complete the paper is a within-program decomposition, not a causal RDD.

Imbens and Lemieux (2008) is the methodological reference; Freedman (2012) and Harger and Ross (2016) establish the canonical NMTC-RDD design at the LIC eligibility cutoff.

9.3 CDE institutional form is unobserved

The data lists CDE names but does not tag them as bank subsidiary, nonprofit CDFI, for-profit specialized, or government. The fixed effect \gamma_c absorbs whatever organizational characteristics are constant within a CDE — but cannot tell us which organizational characteristics drive the between-CDE selection effect. Hand-classifying the top 50 CDEs by institutional form, building on the typology from Theodos et al. (2021), is a planned extension that would let us interact CDE form with rural status and pin down the mechanism.

9.4 Allocation-round data is unavailable

The CDFI Fund publishes allocation awards (which CDEs got how much in each annual round) separately from the deployment data. We use only the deployment side. Bringing in allocation-round data would let us test the bunching hypothesis at the right margin (commitment, not realization).

9.5 Nominal dollars throughout

No CPI deflation. The aggregate program-level numbers ($66.6 B total deployed, $120.9 B project cost) are nominal. A robustness check with real 2022 dollars is straightforward and does not change the within- program decomposition results — inflation affects rural and metro projects equally and is absorbed by year FE — but the headline program-aggregate numbers should be read as nominal.

9.6 Winsorization choice

Leverage is winsorized at [1, 20]. The lower bound is structurally meaningful (project cost ≥ QLICI by construction). The upper bound is chosen so the top ~0.3% of observations don’t dominate the mean. The median is unaffected by either bound. We report robustness to the unwinsorized version in §5.3 of the formal paper; the headline result holds.

9.7 CDE consolidation

CDEs are listed as-named in the source data; affiliated CDEs (for example, multiple bank-subsidiary vehicles owned by the same parent) are not consolidated. A larger CDE family operating multiple registered CDEs would appear in our data as several separate units. This is unlikely to change the decomposition direction but may affect the precise magnitudes.

9.8 The 20% mandate is not the only rural rule

There are also overrides for “high-migration rural counties” and certain “targeted population” rules under the LIC framework. We treat these as part of the residual variation. A more granular institutional analysis would distinguish them.

9.9 External validity beyond NMTC

Whether the within-CDE-equals-zero finding generalizes to other place- based subsidy programs (LIHTC, Opportunity Zones, IRA energy credits) is a question the present paper does not answer. The methodological framework transfers; whether the empirical answer transfers is open. A comparative-program extension is forthcoming.

External validity beyond the United States is a separate question. European blended finance involves a different intermediary architecture — state-owned banks like Caixa Geral de Depósitos in Portugal, the European Investment Bank, EU Cohesion Fund disbursements through managing authorities — and the rural-vs-urban allocation gap there has its own institutional setting (European Investment Bank 2024). The U.S. finding generates a testable hypothesis for those settings; whether the answer is the same is empirically open and worth working out elsewhere.