Intermediaries in blended finance: CDEs, MGS, and pass-through structures

The headline result you will recognise by the end of this chapter

A blended-finance program runs from 2003 to 2019. Federal allocation, billions of dollars, place-based tax credit. The headline question every policy office asks: does the program reach rural areas?

The raw correlation says no. Run an OLS of log private-capital leverage on a rural dummy, controlling for year and industry, and the rural coefficient lands at -0.262 with three stars. Rural projects pull in roughly 26 percent less private capital per dollar of subsidised allocation than urban projects do.

That number has been quoted in policy briefs since the GAO report cycle of the mid-2010s. It frames the entire rural-NMTC debate: rural projects are harder to leverage, so we need bigger subsidies, looser eligibility, more flexible exit windows.

Now add one variable. The identity of the Community Development Entity that intermediates the deal. Add CDE fixed effects to the same regression. The rural coefficient collapses to -0.047 with a p-value of 0.64. The “rural penalty” was never about rurality. It was about which CDEs got the rural allocations.

That is the Helfrich (2026) result, and it is the spine of this chapter. The point is not that the NMTC has a peculiar institutional quirk. The point is that almost every blended-finance program in the world is a pass-through, and pass-through programs cannot be evaluated without modeling the pass-through.

Why intermediaries deserve their own chapter

Most blended-finance capital does not move directly from the public source to the final beneficiary. It moves through a chain. The US Treasury does not write a check to a hardware store in a low-income census tract. It allocates tax-credit authority to a Community Development Entity, which raises matched private capital, which it then deploys to the hardware store via a Qualified Equity Investment.

The European structures are analogous. Portuguese and Spanish small-business credit programs flow through Mutual Guarantee Societies (Sociedades de Garantia Mútua, SGM in Portuguese, SGR in Spanish). Irish SME finance flows through the Strategic Banking Corporation of Ireland (SBCI), which on-lends to commercial banks. German agricultural credit flows through the Landwirtschaftliche Rentenbank into cooperative banks like the Raiffeisen network. Italian SME guarantees flow through Cassa Depositi e Prestiti via Mediocredito Centrale. French regional development capital flows through Bpifrance into local banking networks.

In every case there is at least one intermediary, often two, and the intermediary is not a neutral conduit. The intermediary makes selection decisions: which projects to take into the pipeline, which sectors to specialise in, how aggressively to leverage, how to price risk. Those selection decisions are the program. Studying the program effect without modeling the intermediary is studying the average of two very different forces: the allocation across intermediaries, and the deployment within each intermediary. Average those two and you can hide almost anything.

The canonical pass-through structure

Three layers. Always three layers, sometimes four.

Layer A is the source of subsidised capital. A federal program (NMTC, LIHTC, Opportunity Zones). An EU envelope (InvestEU, Cohesion Funds, CAP Pillar II). A national promotional bank (Bpifrance, SBCI, KfW, BPF in Portugal). A sovereign development institution (DFC, EBRD, EIB). Layer A has a budget, a policy mandate, and an allocation rule.

Layer B is the intermediary. A CDE, an SGM, a cooperative bank, a microfinance institution, an on-lending commercial bank, a regional development agency. Layer B receives an allocation or a guarantee envelope from Layer A and is responsible for deploying it to final projects.

Layer C is the final beneficiary. A rural SME, a low-income census tract real-estate project, a farmer, a community health clinic, a manufacturing firm with under fifty employees.

The flow is A \to B \to C. The interesting econometric structure is that selection happens at both joints. At A \to B, some intermediaries get the allocation and others do not. At B \to C, each intermediary deploys its allocation to a self-chosen subset of eligible projects. A naive program evaluation that regresses outcomes at Layer C on program eligibility dummies confounds both selections into one number.

The decomposition this chapter is built around separates the two. Intermediary fixed effects absorb the A \to B selection. The remaining coefficient on project covariates identifies the B \to C deployment, holding the intermediary constant.

The decomposition logic in two equations

Write the outcome equation for project i funded by intermediary c(i) in year t:

Y_{it} = \alpha_{c(i)} + \gamma_t + X_i\beta + \varepsilon_{it}

Y_{it} is whatever you care about: log leverage ratio, log loan size, default rate, jobs created per dollar. \alpha_{c(i)} is the intermediary fixed effect. \gamma_t is the year effect. X_i is the vector of project characteristics, including the policy-relevant one (a rural dummy, an SME-size dummy, a sector indicator). \beta is the within-intermediary coefficient on X.

Compare to the no-FE specification:

Y_{it} = \gamma_t + X_i\beta^{\text{OLS}} + u_{it}

The gap \beta^{\text{OLS}} - \beta^{\text{FE}} tells you how much of the apparent X-effect is actually intermediary selection. By the omitted-variable-bias logic, if intermediaries that specialise in rural projects also leverage less aggressively on average, then \beta^{\text{OLS}} picks up part of the negative correlation between rural intensity and CDE leverage style, and \beta^{\text{FE}} does not.

The Helfrich (2026) numbers make this concrete. \beta^{\text{OLS}} = -0.262 (***). \beta^{\text{FE}} = -0.047 (p = 0.64). Decomposition:

\frac{\beta^{\text{OLS}} - \beta^{\text{FE}}}{\beta^{\text{OLS}}} = \frac{-0.262 - (-0.047)}{-0.262} \approx 0.82

About 82 percent of the raw rural penalty was between-CDE selection. The within-CDE rural penalty is roughly zero and statistically indistinguishable from zero. Rural projects underperformed in the aggregate because rural-specialised CDEs leveraged at lower ratios on average, not because the same CDE deployed less leverage when it deployed in rural areas.

You can call this a decomposition. You can also call it a diagnostic test on what the policy debate has actually been arguing about.

What this means for policy

Two policy fixes look identical from outside the model. They lead to very different actions.

Market-structure fix. Believe the raw -0.262. Conclude that rural projects are inherently less leverageable, the market is failing in rural areas, and the remedy is more subsidy per rural project, looser underwriting requirements for rural deals, or a rural-specific tax credit at a higher rate. This is expensive. It also leaves the structural problem in place, because next round’s allocations go back to the same low-leverage rural-specialised CDEs.

Intermediary-allocation fix. Believe the FE-controlled -0.047. Conclude that the within-CDE rural effect is statistically nil and the entire aggregate gap comes from which CDEs deploy rural capital. The remedy is one of two things. Route NMTC allocations to high-leverage urban-experienced CDEs with rural deployment requirements attached. Or capitalise and recruit talent at the rural-specialised CDEs so they leverage at urban rates. Either fix costs almost nothing relative to the market-structure fix, and either fix attacks the binding constraint.

If you are a policy economist reading this, the distinction is not academic. The CDFI Fund has a finite allocation. Spending it on richer per-project subsidies because of a misdiagnosed market failure is a real opportunity cost.

Identification: what the within-intermediary estimator does and does not buy you

The FE estimator identifies the within-CDE effect of X under one assumption: conditional on the CDE choosing to deploy at all, the projects the CDE selects across X are exogenous within the CDE. That is, the same CDE deploying in a rural vs urban project does not face systematically different unobserved project quality.

This assumption is not innocent. It breaks if CDEs strategically cherry-pick projects within X: a CDE may take only the highest-quality rural projects (because rural deals are harder to underwrite) and a broader cross-section of urban projects. If this happens, the within-CDE rural coefficient is biased toward zero by a self-selection mechanism. The “no rural penalty within CDE” finding could be partly a “CDEs only pick the unicorn rural deals” finding.

Diagnostics that help:

• Project-level controls. Add observable project characteristics that proxy for underlying quality (firm age, prior revenue, collateral, sector). If the within-CDE rural coefficient is stable across specifications that add these controls, the cherry-picking story is harder to sustain.
• Eligibility-based instruments. Use exogenous eligibility variation (the income threshold of the census tract, a discontinuity in the SME size cutoff) to instrument for actual deployment. This isolates an exogenous deployment shock within the CDE.
• Bounds analysis. Use the Oster (2019) \delta bound to ask how strong unobservable selection would have to be relative to observable selection in order to overturn the within-CDE result. If the answer is \delta > 2, you are reasonably safe.
• Placebo intermediaries. If you can find a set of “non-treated” intermediary deals (a CDE deploying outside the program, an SGM guaranteeing a non-subsidised loan), check whether the rural coefficient is similar there. It should not be, if the program is doing real work.

The within-intermediary estimator only speaks to the population of funded projects. It says nothing about rejected applicants or about projects that never applied. If the rural penalty operates at the application margin (rural firms do not apply because they expect rejection), the FE estimator misses it. This is a real limitation. Be honest about it in the writeup.

The Mutually Guaranteed Society (SGM) structure

The Spanish and Portuguese SGM model emerged in the late 1970s and 1980s to address SME credit access. The structure is mutual: SMEs are members of the SGM and pay a small capital contribution to join. The SGM issues guarantees to commercial banks on behalf of its members, allowing the SMEs to borrow at terms they could not access on their own merits. The SGM in turn is counter-guaranteed by a public-private vehicle. In Portugal that vehicle was historically SPGM (Sociedade Portuguesa de Garantia Mútua), now consolidated under Banco Português de Fomento (BPF). In Spain it is CESGAR (Confederación Española de Sociedades de Garantía).

The Portuguese system pre-merger had four operating SGMs, each with a sectoral or regional specialisation:

SGM	Specialisation	Notes
Lisgarante	Greater Lisbon and Tagus Valley	Urban SME focus, services, light industry
Norgarante	Northern Portugal	Manufacturing-heavy, textile and metalworking SME base
Garval	Centre and South Portugal	Diversified, includes Alentejo and Algarve
Agrogarante	Agriculture and agribusiness, national	Sector-specialised, NUTS-3 spread across rural Portugal

For a researcher who has internalised the Helfrich CDE decomposition, the four-SGM structure is almost too good to pass up. The natural research question is whether the apparent rurality penalty in Portuguese SME guarantees is a within-SGM effect or a between-SGM selection effect. If Agrogarante leverages at systematically lower ratios than Lisgarante, and Agrogarante handles disproportionately rural deals, the raw rural coefficient will be biased downward for exactly the reasons it was biased downward in NMTC.

The mechanics differ from a CDE in one important way. An SGM issues a guarantee, not an equity investment. The leverage variable is therefore the ratio of guaranteed bank loan principal to the SGM’s counter-guarantee, not the ratio of private equity to tax credit. The decomposition logic is identical. The economic interpretation of the coefficient changes.

The Irish parallel: SBCI as a pure intermediary platform

The Strategic Banking Corporation of Ireland was established in 2014 as an on-lending vehicle. SBCI does not lend directly to SMEs. SBCI raises wholesale funding (from the EIB, from KfW, from the Council of Europe Development Bank, from Irish institutional investors) at sovereign-backed rates and on-lends to a roster of commercial banks at a small spread, with conditions on the rates and terms those banks must offer to qualifying SMEs.

The on-lending banks vary. AIB, Bank of Ireland, Permanent TSB, Ulster Bank (until exit), plus specialist lenders like Finance Ireland, Bibby Financial Services, Linked Finance, and a handful of asset-finance houses. Each on-lending bank applies its own credit underwriting on top of the SBCI program rules. The SBCI sets the maximum rate, the minimum term, the SME definition, and the eligible-purpose categories. The on-lending bank decides who actually clears underwriting.

For decomposition purposes, the Irish setup is the cleanest version of the structure in this chapter. Each SBCI loan has a unique on-lending bank identifier. A panel regression of borrower outcomes on borrower covariates with bank fixed effects mechanically isolates the within-bank deployment from the between-bank allocation.

What the data will probably show, if anyone has run this carefully: between-bank dispersion in deployment patterns is large. AIB and Bank of Ireland have very different rural branch networks, very different agricultural sector exposure, and very different SME underwriting cultures. Pooling them in a non-FE regression averages those differences into the residual, and the resulting “program effect” is mostly an artefact of the cross-bank composition of the sample.

A research agenda for Portugal and Ireland

Five steps, ordered by the level of effort.

Step one. Gather SGM-level disbursement data. In Portugal this means combining IFAP records (for agriculture-related guarantees), legacy SPGM records (for the pre-2020 period), and current BPF operating data (post-merger). The unit of observation should be the individual guarantee at the project level, with fields for amount, guaranteed bank loan size, SME sector (NACE rev. 2), SME size (turnover, headcount, total assets), NUTS-3 location, SGM identifier, year of approval, and outcome variables where available (loan performance, firm survival, employment growth). In Ireland the equivalent is the SBCI annual loan tape with on-lending bank identifier.

Step two. Run the Helfrich-style decomposition. The outcome is log leverage (guaranteed bank principal divided by SGM counter-guarantee for Portugal, or loan size divided by SBCI on-lent funds for Ireland). The treatment variable is a rural NUTS-3 indicator, or better, a continuous rurality index based on population density and OECD rural-urban typology. Run OLS without intermediary FE, then with intermediary FE, then with intermediary \times year FE. Report all three side by side. Compute the share of the raw rural effect that is absorbed by adding intermediary FE.

Step three. Name the intermediaries. If the decomposition shows that 60 or 80 percent of the apparent rural penalty is between-SGM allocation, the policy argument requires you to say which SGMs deploy at which leverage ratios. Lisgarante mean leverage vs Agrogarante mean leverage, plotted across years, with confidence intervals. This is the figure that travels into the policy debate. Anonymising the intermediaries kills the policy contribution.

Step four. Replicate in Ireland with SBCI on-lender data. The data structure is less rich than the CDE-level NMTC data because SBCI on-lenders are commercial banks that also do non-SBCI business, so the bank-level deployment style mixes program and non-program activity. Use bank-by-program cohort FE if cohort identifiers are available.

Step five. Cross-country comparison. Do similar EU directives and common SME definitions (Commission Recommendation 2003/361/EC) produce similar intermediary-selection effects across Portugal, Spain, Ireland, and Italy? Or is the intermediary-selection effect institution-specific, driven by the historical legacy of each national guarantee system? My prior is that the pattern is institutional, not directive-driven, but I would happily be talked out of it by the data.

Common traps in intermediary-based studies

A short list of the failure modes that recur in this design.

Confusing intermediary FE with treatment effect. \alpha_{c} is not the “effect” of being intermediary c. It is the conditional mean of Y for projects funded by c, holding everything else fixed. You cannot interpret \alpha_{\text{Agrogarante}} - \alpha_{\text{Lisgarante}} as the causal effect of routing a project through one SGM rather than the other unless you have a quasi-experimental source of variation in routing.

Inadequate intermediary identification. Many programs do not publish recipient-intermediary linkages in their public-use datasets. Researchers proxy by matching on project location, sector, or approval window. If you do this, the intermediary identifier is measured with error, and the FE estimator is biased toward the no-FE result (attenuation bias on the intermediary effect, which leaks into the covariate of interest). Be transparent about how the intermediary identifier was constructed and report sensitivity to alternative matching rules.

Reverse causality at the intermediary level. High-leverage intermediaries may attract better projects (selection on outcomes), which inflates their apparent leverage further. Within-intermediary controls for project quality help, but the cleanest defense is to use plausibly exogenous variation in the intermediary \to project assignment (a rotation rule, a geographic catchment, a credit-rating discontinuity).

Generalisation beyond the funded sample. The FE estimator speaks to the conditional distribution of Y given that the project was funded. It says nothing about the marginal applicant who was rejected, and it says nothing about the firm that never applied. If the policy question is about extensive-margin participation (“are rural firms applying at all?”), the FE estimator is the wrong tool. Pair it with an application-margin analysis using eligible-population denominators.

Pooling structurally different intermediaries. A CDE makes equity investments. An SGM issues guarantees. A cooperative bank takes deposits and lends. A microfinance institution lends without collateral. Do not pool these in the same FE regression and call the result a “blended finance intermediary effect.” The selection mechanisms and incentive structures are different, and pooling washes out everything interesting. Report intermediary-type subsamples separately.

Confounding intermediary FE with geographic FE. If Agrogarante is the only SGM operating in a given NUTS-3 region, the SGM FE absorbs the regional FE and you cannot separate “Agrogarante is a low-leverage SGM” from “Alentejo is a low-leverage region.” Test for the rank deficiency directly: do a rank check on the intermediary \times region design matrix before interpreting the coefficients.

A worked example: synthetic Portuguese SGM data

What follows is a synthetic dataset that illustrates the decomposition without requiring proprietary BPF data. The structure is calibrated to look like a plausible Portuguese SGM panel.

Four SGMs (Lisgarante, Norgarante, Garval, Agrogarante). Eight years (2015 through 2022). About 150 projects per SGM per year, for a total of roughly 4800 observations. Each project has a NUTS-3 identifier, a rural dummy (1 if the NUTS-3 is classified rural by OECD typology), a sector indicator, a log SME asset size, and an outcome of log leverage ratio. The data-generating process is rigged so that:

• The within-SGM rural effect is small (about -0.05 in true units of log leverage).
• The between-SGM differences in average leverage are large (Lisgarante deploys at log-leverage 1.6, Agrogarante at 1.1, others in between).
• The rural share of each SGM’s portfolio is highly skewed: 70 percent of Agrogarante deals are rural, 10 percent of Lisgarante deals are rural.

This setup mechanically generates a raw rural penalty of around -0.25 in a no-FE regression, which collapses to about -0.05 once SGM FE is included. The decomposition recovers the truth.

R code

library(fixest)
library(modelsummary)
library(dplyr)
library(tidyr)

set.seed(2026)

n_sgms      <- 4
n_years     <- 8
n_per_sgm   <- 1200  # 150/year for 8 years

sgm_names   <- c("Lisgarante", "Norgarante", "Garval", "Agrogarante")
sgm_mean    <- c(1.6, 1.4, 1.3, 1.1)        # SGM-specific mean log leverage
sgm_rural_p <- c(0.10, 0.25, 0.40, 0.70)    # rural share by SGM

df <- expand.grid(
  sgm_id    = sgm_names,
  obs       = 1:n_per_sgm,
  KEEP.OUT.ATTRS = FALSE
) %>%
  mutate(
    year      = 2015 + (obs %% n_years),
    sector    = sample(c("manuf","services","agri","retail"),
                       n(), replace = TRUE),
    log_assets = rnorm(n(), 13, 1.2)
  ) %>%
  group_by(sgm_id) %>%
  mutate(
    rural = rbinom(n(), 1, sgm_rural_p[match(first(sgm_id), sgm_names)])
  ) %>%
  ungroup() %>%
  mutate(
    sgm_fe       = sgm_mean[match(sgm_id, sgm_names)],
    year_fe      = (year - 2015) * 0.01,
    rural_within = -0.05 * rural,
    asset_eff    = 0.08 * (log_assets - 13),
    noise        = rnorm(n(), 0, 0.20),
    log_leverage = sgm_fe + year_fe + rural_within + asset_eff + noise
  )

# Model 1: OLS, no intermediary FE
m_ols <- feols(log_leverage ~ rural + log_assets + i(sector) | year,
               data = df, cluster = ~sgm_id)

# Model 2: with SGM FE
m_fe  <- feols(log_leverage ~ rural + log_assets + i(sector) | sgm_id + year,
               data = df, cluster = ~sgm_id)

# Model 3: with SGM x year FE
m_fe2 <- feols(log_leverage ~ rural + log_assets + i(sector) | sgm_id^year,
               data = df, cluster = ~sgm_id)

modelsummary(
  list("OLS (no SGM FE)"   = m_ols,
       "+ SGM FE"          = m_fe,
       "+ SGM x year FE"   = m_fe2),
  stars = TRUE,
  coef_map = c("rural" = "Rural NUTS-3",
               "log_assets" = "log(assets)"),
  gof_omit = "AIC|BIC|Log|Pseudo|R2 Within|RMSE"
)

# Decomposition: share of raw rural effect absorbed by SGM FE
b_ols <- coef(m_ols)["rural"]
b_fe  <- coef(m_fe)["rural"]
absorbed_share <- (b_ols - b_fe) / b_ols
cat(sprintf("Raw rural coef:        %.3f\n", b_ols))
cat(sprintf("Within-SGM rural coef: %.3f\n", b_fe))
cat(sprintf("Share absorbed by SGM: %.1f%%\n", 100 * absorbed_share))

The output should show a raw rural coefficient near -0.24 and a within-SGM rural coefficient near -0.05, with about 80 percent of the raw effect absorbed by adding SGM FE. The decomposition mirrors the NMTC result on synthetic Portuguese data because the data-generating process was rigged to mirror it. That is the point: when the world looks like this, the FE diagnostic recovers what is going on.

Stata code

* Equivalent Stata workflow using reghdfe and estout

clear all
set seed 2026
set obs 4800

* Build the panel
gen sgm_id = ceil(_n / 1200)
label define sgm 1 "Lisgarante" 2 "Norgarante" 3 "Garval" 4 "Agrogarante"
label values sgm_id sgm

gen year = 2015 + mod(_n, 8)

gen log_assets = rnormal(13, 1.2)
gen sector = ceil(runiform() * 4)

* Rural share varies by SGM
gen rural_p = .
replace rural_p = 0.10 if sgm_id == 1
replace rural_p = 0.25 if sgm_id == 2
replace rural_p = 0.40 if sgm_id == 3
replace rural_p = 0.70 if sgm_id == 4
gen rural = runiform() < rural_p

* SGM fixed effect on log leverage
gen sgm_fe = .
replace sgm_fe = 1.6 if sgm_id == 1
replace sgm_fe = 1.4 if sgm_id == 2
replace sgm_fe = 1.3 if sgm_id == 3
replace sgm_fe = 1.1 if sgm_id == 4

gen log_leverage = sgm_fe + 0.01 * (year - 2015)        ///
                + -0.05 * rural + 0.08 * (log_assets - 13) ///
                + rnormal(0, 0.20)

* Model 1: no SGM FE
reghdfe log_leverage rural log_assets i.sector, ///
    absorb(year) vce(cluster sgm_id)
estimates store m_ols

* Model 2: with SGM FE
reghdfe log_leverage rural log_assets i.sector, ///
    absorb(sgm_id year) vce(cluster sgm_id)
estimates store m_fe

* Model 3: with SGM x year FE
reghdfe log_leverage rural log_assets i.sector, ///
    absorb(sgm_id#year) vce(cluster sgm_id)
estimates store m_fe2

estout m_ols m_fe m_fe2,                              ///
    cells(b(star fmt(3)) se(par fmt(3)))              ///
    stats(N r2_a, fmt(0 3) labels("N" "Adj R-sq"))    ///
    legend label collabels(none) varlabels(_cons Constant)

The Stata output should align with the R results: raw rural coefficient around -0.24, within-SGM around -0.05, with the SGM FE absorbing most of the gap.

A reporting checklist for intermediary-FE papers

If you publish a paper using this design, include the following. Reviewers will ask anyway.

1. Intermediary deployment distribution. A table showing how each intermediary’s portfolio breaks down by the covariate of interest. For NMTC, CDE \times rural share. For SGM, SGM \times rural share, SGM \times sector, SGM \times size class.

2. Side-by-side regression table. Raw OLS, with intermediary FE, with intermediary \times year FE, ideally with intermediary \times year \times region FE. Show that the coefficient of interest is stable (or unstable) across specifications.

3. Variance decomposition. Total variance of Y, between-intermediary share, within-intermediary share, within-intermediary-within-year share. ANOVA-style. This communicates the relative magnitudes of the two layers of variation.

4. Drop-largest robustness. Re-estimate dropping the single largest intermediary. If the result depends on one big player, the policy interpretation is weaker. If the result survives, it is much stronger.

5. Weighted vs unweighted. Re-estimate weighting by deployment size and unweighted. Small intermediaries make a lot of noise per dollar; they should not drive the headline number unless that is what you want to show.

6. Direct interpretation paragraph. A single paragraph in plain language that says, in effect: “The apparent X effect is mostly intermediary allocation, not within-intermediary deployment. The policy implication is Y rather than Z.” Reviewers and policy readers should not have to do this work themselves.

7. Caveats on the funded sample. A paragraph that says the FE estimator speaks only to the funded sample and acknowledges the application-margin question as a separate analysis.

Why this transfers beyond NMTC and SGM

The methodological discipline this chapter advocates is not specific to blended finance. Anywhere capital, regulatory authority, or service delivery flows through a pass-through intermediary, the same decomposition applies.

• Microcredit. Banerjee, Karlan, and Zinman (2015) note that microcredit RCTs typically estimate program effects pooled across microfinance institutions. The Helfrich logic suggests reporting between-MFI vs within-MFI decomposition. The treatment effect of microcredit may be small on average and large within some MFIs.
• CAP Pillar II. EU agricultural rural-development payments flow through Member State paying agencies. A common observation is that some Member States have higher absorption rates than others. The decomposition into paying-agency selection vs within-paying-agency deployment is the right way to interpret the absorption gap.
• EU Cohesion Funds. Becker, Egger, and von Ehrlich (2010, 2018) study eligibility-discontinuity effects in regional aid. Their identification strategy is at the regional level, but managing-authority selection within country is the analogous pass-through layer. The within-managing-authority effect of an additional euro of cohesion funding is the policy parameter.
• Banking-channel pass-through. Khwaja and Mian (2008) showed that bank-specific capital shocks pass through to firms via bank-firm relationships. The bank fixed effect is the intermediary effect, the firm-level outcome is the deployment effect, and the within-bank shock identifies the lending channel cleanly. The Helfrich CDE design is the public-finance analogue.
• Insurance regulation. Solvency II and US state-level insurance regulation operate through insurance carriers. Carrier fixed effects in a study of regulatory pass-through are doing the same work as CDE FE in NMTC.
• Hospital-level healthcare delivery. Medicare reimbursement and quality programs flow through hospitals. Hospital FE in a study of patient outcomes isolates the policy variation that is operating within a hospital from the cross-hospital selection.

The general principle is that whenever a treatment is delivered through an intermediary with discretion, the within-intermediary effect is the policy-relevant parameter and the between-intermediary variation is a distinct question (it is the question of how intermediaries are recruited and incentivised).

If you only learn one thing from this chapter, it is this. The intermediary’s deployment decisions are the program. Studying the program without modeling the intermediary is studying noise.

References

Banerjee, A., Karlan, D., and Zinman, J. (2015). “Six randomized evaluations of microcredit: introduction and further steps.” American Economic Journal: Applied Economics, 7(1), 1–21.

Becker, S. O., Egger, P. H., and von Ehrlich, M. (2010). “Going NUTS: the effect of EU Structural Funds on regional performance.” Journal of Public Economics, 94(9–10), 578–590.

Becker, S. O., Egger, P. H., and von Ehrlich, M. (2018). “Effects of EU regional policy: 1989–2013.” Regional Science and Urban Economics, 69, 143–152.

Bertoni, F., Croce, A., and D’Adda, D. (2021). “The role of guarantees in SME credit markets: evidence from the Italian Central Guarantee Fund.” Small Business Economics, 57, 1879–1900.

Freedman, M. (2012). “Teaching new markets old tricks: the effects of subsidized investment on low-income neighborhoods.” Journal of Public Economics, 96(11–12), 1000–1014.

Harger, K., and Ross, A. (2016). “Do capital tax incentives attract new businesses? Evidence across industries from the New Markets Tax Credit.” Journal of Regional Science, 56(5), 733–753.

Helfrich, I. T. (2026). “Rural mobilization gap in US place-based tax credit: a between-CDE vs within-CDE decomposition.” Working paper, SSRN forthcoming.

Khwaja, A. I., and Mian, A. (2008). “Tracing the impact of bank liquidity shocks: evidence from an emerging market.” American Economic Review, 98(4), 1413–1442.

Oster, E. (2019). “Unobservable selection and coefficient stability: theory and evidence.” Journal of Business and Economic Statistics, 37(2), 187–204.

Pombo, P., Molina, H., and Ramírez, J. (2008). “El sistema de garantía iberoamericano: una visión integradora.” CESGAR working paper, Madrid.

SPGM / Banco Português de Fomento (various years). “Relatório de Atividade do Sistema Português de Garantia Mútua.” Lisbon.