Oncel Aldanmaz, Spatial Spillovers in Emergency Lending: Evidence from Bayesian Quantile Models of Reconstruction Finance Corporation Lending During the Great Depression (2026)

Summary

Claremont McKenna College senior thesis by Oncel Aldanmaz, submitted to Professor Angela Vossmeyer on April 18, 2026, on spatial spillovers in Reconstruction Finance Corporation lending during the Great Depression. Using a monthly panel of 2,979 counties built from digitized RFC card index records for 1932 to 1936, it estimates Bayesian quantile regression models for binary longitudinal data. The thesis finds that lending in neighbouring counties is positively associated with a county's own lending, most strongly where counties were least likely to receive lending, and that the state-bank channel dominates the national-bank channel. It concludes that judging emergency lending only by direct local exposure understates its impact, with implications for TARP and the Paycheck Protection Program.

Summary drafted by a model from the document's text below and checked by script against that text before publication. It is a navigation aid, not a reading of what the document proves. Where AI is used

Cite this document

“Oncel Aldanmaz, Spatial Spillovers in Emergency Lending: Evidence from Bayesian Quantile Models of Reconstruction Finance Corporation Lending During the Great Depression (2026).” Original at web.archive.org. Archived at Pandemic Darlings, https://pandemicdarlings.com/source-documents/oncel-aldanmaz-spatial-spillovers-in-emergency-lending-evidence-from-b-2b477ff57e/, SHA-256 bf7bd7126a08720e299de6b682be4cb2058cb36b7d7073ff9b01e73fd93482ed.

Full text

Claremont Colleges
Scholarship @ Claremont

CMC Senior Theses CMC Student Scholarship

2026

Spatial Spillovers in Emergency Lending: Evidence from Bayesian
Quantile Models of Reconstruction Finance Corporation Lending
During the Great Depression
Oncel Aldanmaz
Claremont McKenna College

Follow this and additional works at: https://scholarship.claremont.edu/cmc_theses

Recommended Citation
Aldanmaz, Oncel, "Spatial Spillovers in Emergency Lending: Evidence from Bayesian Quantile Models of
Reconstruction Finance Corporation Lending During the Great Depression" (2026). CMC Senior Theses.
4239.
https://scholarship.claremont.edu/cmc_theses/4239

This Open Access Senior Thesis is brought to you by Scholarship@Claremont. It has been accepted for inclusion in
this collection by an authorized administrator. For more information, please contact scholarship@claremont.edu.
Claremont McKenna College

Spatial Spillovers in Emergency Lending: Evidence from Bayesian

Quantile Models of Reconstruction Finance Corporation Lending

During the Great Depression

submitted to

Professor Angela Vossmeyer

Oncel Aldanmaz

for

Senior Thesis

Spring 2026

April 18, 2026
Abstract

This thesis investigates whether emergency lending by the Reconstruction Finance
Corporation (RFC) generated spatial spillovers across U.S. counties between 1932 and
1936. Using a monthly panel of 2,979 counties constructed from digitized RFC card
index records, we estimate Bayesian quantile regression models for binary longitudinal
data (QBLD) that allow spillover effects to vary across the conditional distribution of
lending propensity. Neighboring lending exposure is positively associated with own-
county lending outcomes across a wide range of quantiles and subsamples, with the
strongest effects at the lower tail, where counties are least likely to receive lending.
This pattern is consistent with a mechanism in which nearby lending increases the vis-
ibility of the program, enhances its credibility, and improves its perceived accessibility,
particularly in places with limited prior engagement with federal credit. Decompos-
ing adjacent lending by charter type reveals that the state-bank channel consistently
dominates the national-bank channel, suggesting that spatial diffusion operated primar-
ily through state-chartered banking networks. These findings suggest that evaluating
emergency lending programs solely on the basis of direct local exposure understates
their total impact, with implications extending from the Depression-era RFC to modern
interventions such as TARP and the Paycheck Protection Program.

1
Contents

1 Introduction 4

2 Data 9
2.1 Data Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Data Aggregation and Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Panel Dataset and Variable Construction . . . . . . . . . . . . . . . . . . . . 11
2.4 Descriptive Statistics and Motivating Evidence . . . . . . . . . . . . . . . . . 12

3 Methodology 18
3.1 Overview of Methodological Framework . . . . . . . . . . . . . . . . . . . . . 18
3.2 Model Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.3 Estimation: MCMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4 Results 20
4.1 Main Model: Aggregate Spillover Results . . . . . . . . . . . . . . . . . . . . 20
4.2 Population Subsample Results . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Federal Reserve District Results . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4 Decomposed Model: State-Bank versus National-Bank Spillovers . . . . . . . 25
4.5 Decomposed Model: Federal Reserve District Results . . . . . . . . . . . . . 27
4.6 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5 Alternative Approaches, Robustness, and Limitations 30
5.1 Alternative Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1.1 Fixed-Effects Panel Regression . . . . . . . . . . . . . . . . . . . . . . 31
5.1.2 Panel Quantile Regression . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.3 Tobit with Mundlak Device . . . . . . . . . . . . . . . . . . . . . . . 32
5.1.4 Panel Tobit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.1.5 Summary of Alternative Specifications . . . . . . . . . . . . . . . . . 33

2
5.2 Robustness and Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5.3 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

6 Conclusion 36

3
1 Introduction

Financial crises pose a recurring challenge to economic stability. When banking systems
come under severe stress, the resulting disruptions to credit, confidence, and economic ac-
tivity rarely remain confined to their point of origin. Instead, financial distress tends to
propagate across institutions and regions, transmitted through interbank relationships, de-
positor behavior, trade networks, and shared expectations. Governments have repeatedly
responded to such episodes with emergency lending programs designed to stabilize affected
institutions and prevent localized failures from escalating into broader systemic collapse. A
central question for evaluating these interventions is whether their effects remain limited
to the institutions and localities that directly receive support, or whether they extend into
surrounding areas through various channels.
The Great Depression offers a powerful historical setting for studying this question. Bank
failures, collapsing asset values, shrinking credit, and widespread economic uncertainty cre-
ated conditions in which local financial distress could easily become regional or national
instability. In response, the federal government increasingly relied on emergency interven-
tion to stabilize banking institutions and restore confidence. Among the most important
of these efforts was the Reconstruction Finance Corporation (RFC), which became a major
source of support for banks and other institutions during the early 1930s. Existing literature
has shown that RFC assistance mattered for bank survival and for the functioning of finan-
cial markets, making it a natural setting for studying how crisis-era public lending affected
local economies.
The relevance of this question extends well beyond the 1930s. During the 2007–2009
Global Financial Crisis (GFC), the Troubled Asset Relief Program (TARP) and Federal
Reserve emergency lending facilities channeled hundreds of billions of dollars to banks and
financial institutions. More recently, the COVID-19 pandemic prompted the Paycheck Pro-
tection Program (PPP) and other emergency credit interventions that were distributed
through banking intermediaries to businesses and localities nationwide. In each episode,

4
the geographic reach and indirect effects of government-supported lending have been central
concerns for policymakers and researchers alike. By studying the spatial dynamics of RFC
lending during the Depression, this thesis speaks directly to the broader question of whether
emergency lending programs generate effects that extend beyond their immediate recipients.
We ask whether RFC-era government lending to banks generated spatial spillovers across
counties between 1932 and 1936. More specifically, we investigate whether the effects of
RFC lending were limited to the counties whose banks directly received support, or whether
those effects extended into neighboring counties as well. This question matters because local
banking systems were not sealed off from one another. Banks and depositors observed nearby
lending decisions, local balance-sheet conditions, and signs of federal support, so information
about government assistance could travel across county lines even when the loans themselves
did not. In this sense, lending in neighboring counties may have increased the visibility of
the program, enhanced its credibility, and improved its perceived accessibility, particularly
in areas with limited prior engagement with federal credit. If RFC lending to banks in one
county stabilized nearby banks through indirect channels such as transparency, information
spillovers, or improved confidence, then analyses that focus only on directly exposed counties
may understate the total impact of federal intervention.
The idea of spatial spillovers is especially relevant in a crisis setting. During a severe
banking contraction, conditions in one local banking market can influence nearby banking
markets through confidence effects, interbank ties, depositor responses, and information
revealed by government support decisions. This makes it plausible that RFC assistance,
even if formally directed at specific banks in specific counties, may have improved conditions
for nearby banks as well.
Empirically, we study county-level monthly panel data from 1932 to 1936. The core
banking data come from digitized card index records of the Reconstruction Finance Corpo-
ration, converted into usable data frames following the methodology of Amujala, Vossmeyer,
and Das (2023), recording county-level bank counts, total assets, and closures. County adja-

5
cency is defined through a Queen-contiguity spatial weights matrix derived from U.S. Census
TIGER shapefiles.
We mainly estimate two Bayesian quantile binary panel models of RFC lending outcomes,
following the QBLD framework of Rahman and Vossmeyer (2019). The first (main) model
regresses a binary indicator of whether banks in a county receive any positive RFC lending
on lagged own-county RFC lending and a lagged binary indicator of whether any adjacent
county’s banks received positive RFC lending. The second (decomposed) model replaces
the composite adjacent indicator with two separate binary indicators: lagged adjacent state-
bank RFC lending and lagged adjacent national-bank RFC lending, allowing us to distinguish
spillovers by bank type given that national banks were more directly connected to Federal
Reserve liquidity channels such as the discount window. This distinction may matter because
differences in outside liquidity access could have affected banks’ need for RFC assistance and
the extent to which RFC lending to banks in neighboring counties translated into subsequent
local lending outcomes. To capture various forms of heterogeneity, both models include
county-specific random intercepts and a full set of sample-month (year×month) time fixed
effects. Both models are estimated for the full sample, bottom-25% and rest-75% population
subsamples at quantiles ranging from 0.1 to 0.9, and for each of the twelve Federal Reserve
districts at quantiles 0.1, 0.5, and 0.9.
Three main findings emerge. First, neighboring lending exposure is positively associated
with own-county lending outcomes across a wide range of specifications, quantiles, and sub-
samples. The evidence is strongest at the lower tail of the conditional distribution, where
counties are least likely to receive lending—precisely where spillover support from neighbors
would matter most. Second, when the composite adjacent indicator is decomposed, the state-
bank channel consistently dominates the national-bank channel in magnitude, suggesting
that the spatial diffusion of lending during the Depression operated primarily through state-
chartered banking networks rather than nationally chartered ones. Third, the relationship
is heterogeneous across county size and Federal Reserve districts: the rest-75% subsample

6
shows stable, significant effects throughout the distribution, while the bottom-25% subsam-
ple and individual district runs reveal more variable patterns with effects concentrated at
the tails.
Historical work has established that the RFC played a central role in financial stabilization
during the Depression. Mason (2003) showed that RFC allocations were not simply driven by
standard rent-seeking measures, which makes geographic patterns of lending and their con-
sequences worth studying more closely. Vossmeyer (2016) provided a multivariate selection-
and-treatment framework tailored to the RFC’s three-stage “apply/approve/perform” pro-
cess, highlighting how program participation itself had positive effects on bank lending and
was endogenous. Calomiris, Mason, Weidenmier, and Bobroff (2013) provided perhaps the
most direct mechanism for this thesis: their evidence that RFC assistance improved the
survival of Michigan banks suggests that emergency lending could stabilize institutions in
ways that might plausibly benefit neighboring counties as well.
Modern studies of the Paycheck Protection Program show that bank intermediation
shaped the geographic allocation and effects of emergency credit during a modern crisis.
Granja, Makridis, Yannelis, and Zwick (2022) and Autor et al. (2022) document substan-
tial heterogeneity in how PPP funds were distributed and transmitted across places and
firms. Anbil, Carlson, and Styczynski (2023) show that the Federal Reserve’s PPPLF fa-
cility causally increased PPP lending, especially among smaller community banks, and that
the facility also operated as a funding backstop by encouraging lending even among banks
that did not borrow from it directly. Their results also suggest a broader pattern relevant
for this thesis: banks that were already familiar with interacting with government liquidity
programs were more likely to engage with a new one when it became available. More re-
cently, Berger, Karolyi, and Roman (2025) study the Bank Term Funding Program during
the 2023 Regional Banking Crisis and find that subsidized BTFP borrowing allowed banks
to substitute cheaper facility funds for deposits while mitigating what would otherwise have
been a sharper contraction in credit provision, broadly consistent with the Bagehot Principle.

7
These studies reinforce the broader idea that government-supported liquidity can generate
uneven and indirect regional effects through financial intermediaries and local conditions.
Related evidence from outside banking points in the same direction: Bloom, Schankerman,
and Van Reenen (2013) document large inter-firm technology spillovers, and Moretti, Stein-
wender, and Van Reenen (2025) show that defense-funded R&D raised productivity beyond
directly funded locations. Together, this literature supports the general proposition that
government-driven shocks can produce meaningful spillover effects beyond directly treated
units.
Cameron and Trivedi (2005) provide the general panel-data and fixed-effects frame-
work needed for county-by-time analysis. Anselin (1988) is foundational for thinking about
adjacency-based spillovers and about how outcomes in one location may depend on condi-
tions in neighboring locations. Lee and Yu (2010) extended spatial autoregressive models
into panel data with fixed effects, which is conceptually close to the central question of this
thesis. The key insight from this literature is that geographically proximate units often
influence one another, and ignoring that structure can understate or distort policy effects.
Reich, Fuentes, and Dunson (2011) developed a Bayesian spatial quantile framework, and
Dai, Yan, Tian, and Tang (2020) developed quantile methods for spatial panel data models
with fixed effects. These papers are relevant because the effect of neighboring lending may
not be constant across the distribution of county outcomes. Rahman and Vossmeyer (2019)
developed the Bayesian quantile regression model for binary longitudinal data (QBLD) that
this thesis employs directly. Their framework places an asymmetric Laplace distribution on
the error term, enabling quantile-specific inference for binary outcomes while accommodating
random effects for panel structure.
This work contributes to the existing literature in three important ways that speak di-
rectly to why spatial spillovers in emergency lending matter for both historical interpre-
tation and modern policy design. First, it brings a spatial perspective to the analysis of
RFC lending, motivated by the idea that financial stabilization during the Great Depression

8
was unlikely to remain confined within county boundaries. If lending to one bank increases
the probability that nearby banks receive funds, then the RFC’s impact extends beyond
directly treated institutions, implying that studies focusing only on direct exposure may
substantially underestimate the true reach and effectiveness of the program. More broadly,
documenting such patterns provides evidence that government interventions can generate
information spillovers—through mechanisms such as increased transparency, learning, or
signaling—which shape how financial resources are allocated across space. Second, the pa-
per employs a Bayesian quantile regression framework for binary longitudinal data, allowing
spillover effects to vary across the conditional distribution rather than imposing a constant
mean effect; this is particularly important in crisis settings where responses may be highly
heterogeneous across locations. Third, it separates overall adjacent lending exposure into
state-bank and national-bank components, making it possible to identify whether different
institutional channels transmitted these spillovers differently, and thereby offering more pre-
cise insight into how the structure of the banking system influences the diffusion of policy
effects.
The rest of the thesis proceeds as follows. Section 2 describes the data sources, variable
construction, and summary statistics. Section 3 presents the econometric methodology.
Section 4 reports the main estimation results. Section 5 discusses alternative approaches
and robustness. Section 6 concludes.

2 Data

2.1 Data Sources

The empirical analysis uses a county-level monthly panel covering 1932 to 1936, coming from
digitized card index records of the RFC, converted into structured data frames using the ma-
chine learning and natural-language processing pipeline developed by Amujala, Vossmeyer,
and Das (2023). Their methodology leverages AWS Textract for optical character recog-

9
nition and AWS Comprehend for entity classification, enabling the systematic extraction
of bank names, locations, loan dates, loan types, decisions, and amounts from over 22,000
archival card images held at the National Archives. The resulting dataset records county-
level bank counts, total assets, and closures. The panel universe is defined as all counties
with at least one operating bank in the January 1932 snapshot, and these counties are then
followed through all subsequent months to 1936. RFC lending records, including separate
counts and dollar amounts for state-chartered banks, nationally chartered banks, and non-
bank borrowers, are merged from a separate monthly lending panel. County population is
incorporated from historical census data. Because population is not observed at monthly
frequency, it is treated as a time-invariant county characteristic over the 1932–1936 sample
period; each county is assigned its first observed population value, typically corresponding
to the 1930 Census, and that value is carried across all months in the panel.

2.2 Data Aggregation and Cleaning

County FIPS codes are assigned by matching county and state names to the U.S. Census
TIGER county shapefile, with manual corrections applied for historical name changes, inde-
pendent cities, and consolidated city-counties. In particular, special adjustments are required
for cases such as Virginia’s independent cities and New York City boroughs. After geographic
identifiers are standardized, the county panel is used to construct a Queen-contiguity spatial
weights matrix, in which two counties are treated as neighbors if they share either a border
or a vertex. Using this matrix, adjacent lending exposure variables are computed by aggre-
gating lending indicators across all neighboring counties. The resulting adjacency structure
is binary and unweighted, so each neighboring county contributes equally to the spillover
measure.

10
2.3 Panel Dataset and Variable Construction

The final panel contains 142,992 county-month observations: 2,979 counties observed over
48 months from January 1932 through December 1935. The panel is balanced—every county
appears in all 48 months. Because all regressors enter in one-period lagged form, the first
month for each county is consumed by the lag and the QBLD estimation samples contain
at most 47 usable periods per county. For district-specific and population-split subsamples
in the decomposed specification, an iterative rebalancing procedure ensures that every re-
tained county appears in every retained period, which may reduce the estimation sample
further. The core lending variable, FinalAmount Bankit , records the total dollar amount
of RFC lending to banks in county i during month t, with analogous variables record-
ing lending separately to state-chartered banks (FinalAmount StateBankit ) and nationally
chartered banks (FinalAmount NationalBankit ). The panel also includes adjacent-county
aggregates constructed from the Queen-contiguity matrix: Adjacent FinalAmount Bankit ,
Adjacent FinalAmount StateBankit , and Adjacent FinalAmount NationalBankit , which sum
lending across all counties sharing a border or vertex with county i.
From these continuous measures, we construct the binary and log-transformed variables
used in estimation. The main dependent variable is FAB binaryit , equal to one if county
i receives any positive RFC lending to banks in month t. Only 12.7% of county-month
observations record positive RFC lending to banks, reflecting the heavily zero-inflated nature
of the data and motivating the binary outcome approach. The adjacent lending indicator
AFAB binaryit equals one if any neighboring county received positive bank lending; 43.9%
of observations have at least one actively lending neighbor. Log-transformed continuous
versions (log(FAB), log(AFAB), and their state-bank and national-bank decompositions)
are also constructed for use in the alternative specifications described in Section 5. All
regressors enter the models in one-period lagged form (prefixed L1 ) to reduce simultaneity
concerns and better capture whether prior lending activity in a county or its neighbors is
associated with subsequent lending outcomes.

11
County-level controls drawn from census and agricultural sources include total population
(Total Pop). Population is available for 2,968 counties and is treated as time-invariant over
the sample period. Each county is also assigned to one of twelve Federal Reserve districts
using a county-to-district crosswalk, with one manual correction for FIPS 29210 (assigned to
District 8). For the population-based sample split, each county’s first non-missing population
value is used to compute the 25th percentile threshold (approximately 10,644 residents).
The 743 counties at or below this threshold form the bottom-25-percent subsample, while
the remaining 2,225 counties form the rest-75-percent subsample. Both subsets retain the
balanced panel structure. Banking characteristics from the January 1932 baseline snapshot—
including initial bank counts (1932 banks) and total assets (1932 TotalAssets)—are available
as cross-sectional controls capturing pre-sample financial conditions.

2.4 Descriptive Statistics and Motivating Evidence

The descriptive evidence establishes that RFC-related lending during 1932–1936 was geo-
graphically clustered rather than evenly dispersed. Figure 1 shows strong spatial concentra-
tion in total lending: upper-percentile counties (P83–P100) cluster visibly in the Northeast,
Great Lakes, and parts of the Pacific coast, while much of the rural South, Great Plains,
and Mountain West falls in the lowest percentile bins. Figure 1 also shows that the total
lending per capita map partially reshuffles this picture—some Western and Southern coun-
ties move into higher per-capita bins despite modest absolute totals—but the overall pattern
of geographic unevenness persists.

12
(a) Total lending by county (b) Total lending per capita by county

Figure 1: Geographic distribution of RFC total lending, 1932–1936. Counties are binned by
percentile of cumulative lending.

Figure 2 shows that Districts 2 (New York) and 7 (Chicago) each received over $1 billion
in total lending, with District 4 (Cleveland) close behind at approximately $800 million.
By contrast, Districts 9 (Minneapolis), 10 (Kansas City), and 11 (Dallas) each received
under $150 million. Figure 2 also highlights a notable bank-share gradient across districts:
Districts 4 and 5 show relatively low bank shares (around 40–50%), while Districts 9 and 12
show high bank shares (above 80%), suggesting that the institutional composition of lending
and the economic stress of the Great Depression varied substantially across regions.

13
Figure 2: Total RFC lending by Federal Reserve district, 1932–1936. Bar shading indicates
the bank share of total district lending.

Figure 3 reveals an important asymmetry between state-bank and national-bank lending.
The state-bank map uses five percentile bins starting at P0–P33, indicating that roughly a
third of counties fall in the lowest category but with meaningful variation above. The
national-bank map, by contrast, collapses its bottom bin into P0–P50—half of all coun-
ties received so little national-bank lending that the percentile breaks could not differenti-
ate them—with upper-percentile national-bank counties concentrated in the Northeast and
scattered urban centers. This descriptive pattern foreshadows the model result that the
state-bank channel is the more geographically diffuse and statistically dominant source of
neighboring lending exposure.

14
(a) State-bank lending (b) National-bank lending

Figure 3: Geographic distribution of RFC lending by bank charter type, 1932–1936.

Figure 4 reinforces this point using per-capita measures. State-bank lending per capita
is more evenly spread across the country, with visible upper-percentile counties throughout
the Midwest, parts of the South, and the Pacific coast. National-bank lending per capita
remains heavily concentrated in the Northeast, with large swaths of the South and interior
West in the P0–P50 bin.

(a) State-bank lending per capita (b) National-bank lending per capita

Figure 4: Per capita RFC lending by bank charter type, 1932–1936.

Figure 5 further highlights structural differences across the two population subsamples.
The top-75% population map shows dense, well-differentiated lending activity across the

15
eastern half of the country, with large gray (NA) areas in the West where small counties are
excluded. The bottom-25% population map is its mirror image: small counties scattered
across the West and rural South, many in the lowest percentile bins, with vast gray areas
in the East where larger counties dominate. This visual contrast motivates the population-
based sample splits in the model estimation.

(a) Top 75% population counties (b) Bottom 25% population counties

Figure 5: Total lending by county population subsample, 1932–1936. Gray (NA) areas
indicate counties excluded from the respective subsample.

Finally, Figures 6, 7, and 8 provide direct visual motivation for the formal models. Fig-
ure 6 shows that the “good spillover” panels for total lending include counties like FIPS 34013
(Essex, NJ), 36061 (Manhattan, NY), and 22071 (Orleans Parish, LA) highlighted within
neighborhoods where nearly every adjacent county falls in the P83–P100 bin—high-lending
counties embedded in high-lending clusters. Figure 6 also shows the opposite configuration
in the “bad spillover” panels: counties like FIPS 45019 (Charleston, SC) and 16001 (Ada,
ID) are highlighted at upper percentiles but surrounded by neighbors predominantly in the
P0–P17 or P0–P33 range—relatively active lending counties isolated from a low-lending
neighborhood. Figures 7 and 8 show analogous patterns within the state-bank and national-
bank channels. These maps do not prove causal spillovers, but they illustrate precisely

16
the kind of spatial configuration—local lending embedded in or isolated from a neighboring
lending environment—that the QBLD models are designed to detect.

(a) Good spillover examples (b) Bad spillover examples

Figure 6: Total lending spillover examples. Good spillovers: high-lending counties embedded
in high-lending neighborhoods. Bad spillovers: high-lending counties isolated in low-lending
neighborhoods.

(a) Good spillover examples (b) Bad spillover examples

Figure 7: State-bank lending spillover examples.

17
(a) Good spillover examples (b) Bad spillover examples

Figure 8: National-bank lending spillover examples.

3 Methodology

3.1 Overview of Methodological Framework

The empirical strategy requires a framework that can handle three features of the data si-
multaneously: a binary dependent variable (whether a county receives any positive bank
lending), a panel structure with repeated county-month observations, and a desire to assess
heterogeneity across the conditional distribution of lending propensity rather than focusing
only on the conditional mean. Standard linear or logit panel models address the first two
features but restrict attention to mean effects. Quantile regression methods allow distribu-
tional heterogeneity but are not directly suited to binary outcomes with panel structure.
The Bayesian quantile regression model for binary longitudinal data (QBLD), developed by
Rahman and Vossmeyer (2019) and implemented through the qbld package in R, addresses
all three requirements.

18
3.2 Model Specification

The QBLD model is based on an asymmetric Laplace likelihood, which enables quantile-
specific inference for a binary outcome. Because the dependent variable is binary, the esti-
mated coefficients are interpreted as effects on the latent propensity underlying the observed
lending indicator at different points of the conditional distribution. Each county enters the
model with a random intercept to absorb unobserved time-invariant heterogeneity, while a
full set of time fixed effects—one dummy for each sample-month (year×month) period—
captures common macroeconomic shocks, seasonal forces, and policy changes that affect all
counties simultaneously.

Main Specification: Let yit ∈ {0, 1} denote whether county i receives any positive bank
lending in month t. Define the latent variable:

yit∗ = β1 (τ ) L1 AFAB binaryit + β2 (τ ) L1 FAB binaryit + γ(τ )′ dt + αi + εit (τ ), (1)

where L1 AFAB binaryit is a one-period lag indicating whether any adjacent county received
positive bank lending, L1 FAB binaryit is the lagged own-county lending indicator, dt is a
vector of time fixed effects (one indicator per sample-month period), αi is a county-specific
random intercept, and εit (τ ) follows an asymmetric Laplace distribution indexed by quantile
τ . The observed outcome is:
yit = 1{yit∗ > 0}. (2)

Decomposed Specification: The decomposed model replaces the composite adjacent
indicator with separate state-bank and national-bank channels:

yit∗ = β1 (τ ) L1 AFASB binaryit + β2 (τ ) L1 AFANB binaryit
(3)
′
+ β3 (τ ) L1 FAB binaryit + γ(τ ) dt + αi + εit (τ ),

19
where L1 AFASB binaryit and L1 AFANB binaryit are lagged binary indicators for adjacent
state-bank and national-bank lending, respectively.

3.3 Estimation: MCMC

Posterior inference is based on the unblocked Gibbs sampler implemented in the qbld pack-
age. The MCMC algorithm draws from the full conditional distributions of the fixed effects,
latent variables, random effects, and the random-effect variance parameter. Following the
qbld implementation of Rahman and Vossmeyer (2019), we place a multivariate normal prior
on the fixed-effects coefficients, β ∼ N (b0 , B0 ), with b0 = 0 and B0 = I. We also use the
package-default hyperparameters for the prior on the random-effect variance, φ2 , with c1 = 9
and d1 = 10. We use 10,000 MCMC draws for most specifications and 20,000 draws for the
lower quantiles (0.1 and 0.2), where convergence is slower. A burn-in of 20 percent is applied
in all cases. Trace plots for the key coefficients and φ2 across all quantiles show stable mixing
around stationary values with no visible drift or trend, supporting adequate convergence of
the sampler. Statistical significance is assessed using 95 percent posterior credible intervals,
with coefficients treated as significant when the interval excludes zero.

4 Results

4.1 Main Model: Aggregate Spillover Results

The main model estimates the association between a county’s probability of receiving RFC
lending to banks and the composite binary indicator of whether any adjacent county’s banks
received positive RFC lending in the previous month, controlling for the county’s own lagged
RFC lending status. Interpreted substantively, a positive adjacent-county coefficient is con-
sistent with spillovers across nearby banking markets: once banks in neighboring counties
receive government support, nearby banks may become more likely to receive support as

20
well because information about federal intervention, bank condition, or program accessibil-
ity diffuses across space.
In the full sample, both AFAB and FAB are positive and statistically significant across
all nine quantiles from 0.1 to 0.9. The effects are strongest at the lowest quantile: at τ = 0.1,
AFAB is approximately 2.0 and FAB approximately 1.6, both with 95% credible intervals
well above zero. From τ = 0.2 onward, both coefficients compress to a stable band—AFAB
ranges from roughly 0.45 to 0.60, FAB from roughly 0.55 to 0.75—before FAB rises again
at τ = 0.9 to approximately 1.45. Importantly, the spillover effect is not trivial relative to
a county’s own serial dependence in lending. Across much of the distribution, the AFAB
coefficient is of roughly the same order of magnitude as the FAB coefficient: it is larger at
the lowest quantile, somewhat smaller across most middle quantiles, and still clearly positive
throughout. In other words, neighboring-county lending exposure is not merely a weak
secondary correlate once own-county persistence is taken into account; at several quantiles
it is comparable to, and at τ = 0.1 even stronger than, the effect of the county’s own lagged
lending status. This makes the substantive result stronger: the probability of receiving
RFC lending appears to reflect not only persistence within a county, but also meaningful
transmission across nearby counties.
Thus, adjacent-county lending appears to exert the stronger effect at the lower tail,
where counties are least likely to receive lending, while at the remaining quantiles own-
county lending persistence generally becomes somewhat more important than neighboring
exposure. One interpretation is that neighboring lending mattered most in places with the
lowest baseline propensity to receive RFC support because those counties may have been
less familiar with the program or less connected to the channels through which information
about federal credit circulated. In such settings, observing nearby banks obtain RFC as-
sistance may have reduced uncertainty about how the program worked, whether banks like
theirs could access it, or whether applying was worthwhile at all. This is consistent with an
information-spillover mechanism: neighboring exposure is most valuable where familiarity

21
is weakest. By contrast, at the middle and upper quantiles, counties may already have had
stronger banking networks, better access to information, or greater prior integration into
RFC lending channels, so own-county persistence becomes the more stable predictor of con-
tinued lending. Even there, however, the AFAB estimates remain economically meaningful
because their magnitudes stay fairly close to the FAB estimates through much of the con-
ditional distribution rather than collapsing toward zero. Figure 9 shows this pattern clearly
and suggests that neighboring lending exposure matters most in the lower tail of the con-
ditional distribution, whereas own-county persistence is the more stable force through the
middle and upper parts of the distribution.

Figure 9: Baseline specification: AFAB and FAB coefficient estimates across quantiles, full
sample. Triangles denote AFAB (adjacent lending); circles denote FAB (own-county lend-
ing). Black markers indicate statistically significant estimates; gray markers indicate statis-
tically insignificant estimates.

4.2 Population Subsample Results

Figure 10 shows that the rest-75% population subsample closely mirrors the full sample:
both coefficients are positive and significant at every quantile, with the same large lower-tail
spike and stable middle-range values. As in the full sample, AFAB exceeds FAB at τ = 0.1,
but beyond the lowest quantile the FAB coefficient is generally as large as or larger than
AFAB. Still, the two are broadly similar in magnitude across much of the distribution, which

22
indicates that adjacent-county lending is not negligible relative to own-county persistence.
Rather, neighboring exposure appears to be an important predictor in its own right, even
when it does not dominate the county’s own lagged lending history. This again points to a
mechanism in which neighboring counties matter most when baseline familiarity or access
is relatively limited. In counties that were not already strongly positioned to receive RFC
lending, nearby exposure may have served as a signal that the program was accessible,
legitimate, and practically usable, helping diffuse knowledge across local banking markets.
Once counties moved away from that lower-tail setting, however, the role of neighboring
exposure appears to diminish somewhat relative to the county’s own lending history, which
likely captures persistence in local banking conditions, ongoing relationships, or established
access to RFC channels.
The bottom-25% subsample tells a different story. AFAB is very large at τ = 0.1 (ap-
proximately 3.5, significant), but FAB is weaker. Through the middle quantiles (0.4–0.6),
both coefficients shrink toward zero and lose statistical significance. Significance re-emerges
at the upper tail (0.7–0.9), but magnitudes are smaller than at the lower tail. This suggests
that in the smallest counties, the spillover relationship was less stable and more concentrated
at the extremes of the distribution.

(a) Rest 75% Population (b) Bottom 25% Population

Figure 10: Baseline specification: AFAB and FAB coefficient estimates across quantiles by
population subsample. Triangles denote AFAB (adjacent lending); circles denote FAB (own-
county lending). Black markers indicate statistically significant estimates; gray markers
indicate statistically insignificant estimates.

23
4.3 Federal Reserve District Results

Figure 11 shows that the twelve Federal Reserve district results reinforce the finding of strong
lower-tail spillovers but with meaningful regional heterogeneity. In all twelve districts, both
AFAB and FAB are large and significant at τ = 0.1, typically ranging from 2 to 5 for AFAB
and 1 to 5 for FAB. At τ = 0.5, the pattern is more mixed: AFAB becomes statistically
insignificant in several districts (notably 02, 05, and 10), while FAB also loses significance in
District 06. At τ = 0.9, most districts recover positive significant effects for both coefficients,
though magnitudes are smaller than the lower tail. The strongest combined effects appear
in Districts 03, 04, 09, and 12, while District 05 shows the weakest spillovers at the median
and upper tail.

24
(a) District 1 (Boston) (b) District 2 (New York) (c) District 3 (Philadelphia)

(d) District 4 (Cleveland) (e) District 5 (Richmond) (f) District 6 (Atlanta)

(g) District 7 (Chicago) (h) District 8 (St. Louis) (i) District 9 (Minneapolis)

(j) District 10 (Kansas City) (k) District 11 (Dallas) (l) District 12 (San Francisco)

Figure 11: Baseline specification: AFAB and FAB coefficient estimates by Federal Reserve
district at τ ∈ {0.1, 0.5, 0.9}. Triangles denote AFAB (adjacent lending); circles denote
FAB (own-county lending). Black markers indicate statistically significant estimates; gray
markers indicate statistically insignificant estimates.

4.4 Decomposed Model: State-Bank versus National-Bank Spillovers

The decomposed model separates the composite AFAB indicator into adjacent state-bank
lending (AFASB) and adjacent national-bank lending (AFANB). This decomposition reveals

25
an important asymmetry: the state-bank channel is consistently the greater driver of the
aggregate spillover effect. This distinction is also substantively meaningful because national
banks had direct access to Federal Reserve discount window liquidity, whereas state banks
generally did not, so the two bank types may have responded differently to RFC assistance
and local financial stress.
In the full sample, AFASB is approximately 1.95 at τ = 0.1 and then settles into a
range of roughly 0.45–0.65 through the remaining quantiles, remaining statistically significant
throughout. AFANB is smaller at every quantile: approximately 0.8 at τ = 0.1, then roughly
0.25–0.40 from τ = 0.2 onward. Both channels are significant across all quantiles in the full
sample, but AFASB is consistently about twice the magnitude of AFANB. Figure 12 shows
this pattern clearly. The stronger and more diffuse state-bank effect is consistent with the
idea that banks without comparable Federal Reserve liquidity access may have been more
dependent on RFC support, making neighboring state-bank lending a more important source
of spillover variation.

Figure 12: Decomposed specification: AFASB and AFANB coefficient estimates across quan-
tiles, full sample. Circles denote AFASB (state-bank channel); triangles denote AFANB
(national-bank channel). Black markers indicate statistically significant estimates; gray
markers indicate statistically insignificant estimates.

Figure 13 shows that the rest-75% subsample replicates this pattern closely: AFASB
around 2.0 at the lower tail and 0.45–0.65 elsewhere, AFANB around 0.85 at the lower tail

26
and 0.25–0.45 elsewhere, both significant throughout. In the bottom-25% subsample, the
asymmetry becomes even more pronounced. AFASB is very large at τ = 0.1 (approximately
3.85, significant) and remains significant through the 0.2–0.4 range before weakening in the
middle quantiles. AFANB is significant at τ = 0.1 (approximately 1.4) but loses significance
through most of the middle quantiles (0.4–0.7), only weakly re-emerging at the upper tail.
For small counties, the state-bank channel carried the spillover while the national-bank
channel was largely irrelevant outside the lower tail. This pattern is again consistent with the
institutional difference between the two bank types: national banks’ access to the discount
window may have reduced their dependence on RFC-linked neighboring dynamics relative
to state banks.

(a) Rest 75% Population (b) Bottom 25% Population

Figure 13: Decomposed specification: AFASB and AFANB coefficient estimates across quan-
tiles by population subsample. Circles denote AFASB (state-bank channel); triangles denote
AFANB (national-bank channel). Black markers indicate statistically significant estimates;
gray markers indicate statistically insignificant estimates.

4.5 Decomposed Model: Federal Reserve District Results

Figure 14 sharpens this picture further at the district level. At τ = 0.1, AFASB is large
and significant in nearly every district, often exceeding 3.0 (for example, approximately 4.3
in District 06, 4.7 in District 09, 5.0 in District 12), while AFANB at the same quantile
is often smaller and in several districts insignificant (Districts 01, 02, 05, 07, and 10 show
AFANB with credible intervals crossing zero at τ = 0.1). At τ = 0.5, both channels typically

27
compress toward zero with wider uncertainty; AFANB is insignificant in most districts and
AFASB retains significance in only a few (Districts 03, 04, 07, and 11). At τ = 0.9, moderate
positive effects re-emerge for both channels in some districts (notably 03, 04, 07, 08, 09, and
11), but the state-bank channel again tends to be the larger and more reliably significant of
the two.

28
(a) District 1 (Boston) (b) District 2 (New York) (c) District 3 (Philadelphia)

(d) District 4 (Cleveland) (e) District 5 (Richmond) (f) District 6 (Atlanta)

(g) District 7 (Chicago) (h) District 8 (St. Louis) (i) District 9 (Minneapolis)

(j) District 10 (Kansas City) (k) District 11 (Dallas) (l) District 12 (San Francisco)

Figure 14: Decomposed specification: AFASB and AFANB coefficient estimates by Federal
Reserve district at τ ∈ {0.1, 0.5, 0.9}. Circles denote AFASB (state-bank channel); triangles
denote AFANB (national-bank channel). Black markers indicate statistically significant
estimates; gray markers indicate statistically insignificant estimates.

4.6 Synthesis

Three main findings emerge from the results. First, neighboring lending exposure is pos-
itively associated with own-county lending outcomes across a wide range of specifications,

29
quantiles, and subsamples. The evidence is strongest at the lower tail of the conditional
distribution, where counties are least likely to receive lending—precisely where spillover sup-
port from neighbors would matter most. This pattern is consistent with a mechanism in
which nearby lending increases the visibility of the program, enhances its credibility, and
improves its perceived accessibility, particularly in counties with limited prior exposure to
federal credit. Second, when the composite adjacent indicator is decomposed, the state-bank
channel consistently dominates the national-bank channel in both magnitude and statistical
significance. This suggests that the spatial diffusion of lending during the Depression oper-
ated primarily through state-chartered banking networks rather than nationally chartered
ones. Third, the relationship is heterogeneous across county size and Federal Reserve dis-
tricts: the rest-75% subsample shows stable, significant effects throughout the distribution,
while the bottom-25% subsample loses significance across the middle quantiles and signifi-
cance re-emerges only at the upper tail, suggesting that spillovers in the smallest counties
were less stable and more concentrated at the extremes. Individual Federal Reserve dis-
trict runs reinforce this heterogeneity, with effect sizes and significance varying considerably
across regions.

5 Alternative Approaches, Robustness, and Limitations

5.1 Alternative Models

The main QBLD results are estimated on a binary dependent variable using a Bayesian
quantile framework. To assess whether the core finding—a positive association between
neighboring lending exposure and own-county lending outcomes—is robust to alternative
modeling choices, we estimate a series of frequentist specifications using the continuous (log)
measure of lending, log(FAB)it , as the dependent variable and lagged log adjacent lending,
L1 log AFABit , as the main regressor. These models vary in how they handle censoring,
fixed effects, and distributional heterogeneity, providing complementary perspectives on the

30
spillover relationship.

5.1.1 Fixed-Effects Panel Regression

We begin with a standard linear panel regression with two-way fixed effects (county and
month), estimated via reghdfe:

log(FAB)it = β L1 log AFABit + αi + δt + εit . (4)

The estimated coefficient on lagged adjacent lending is β̂ = 0.034 (t = 17.90, p < 0.001),
with a within R2 of 0.002. The effect is statistically significant but small in magnitude,
suggesting that after absorbing county and time heterogeneity, a one-unit increase in log
adjacent lending is associated with a modest increase in own-county log lending. The low
within R2 is expected given the high proportion of zero-lending observations in the panel;
the linear model does not account for the censored nature of the dependent variable.

5.1.2 Panel Quantile Regression

To allow the spillover relationship to vary across the conditional distribution of log lending,
we estimate a panel quantile regression with county fixed effects and time fixed effects at
τ ∈ {0.10, 0.50, 0.90} using qregpd. The results are:

τ = 0.10 τ = 0.50 τ = 0.90

L1 log AFAB 0.044** 0.831*** 0.256**
(0.014) (0.005) (0.078)

The coefficient is positive and significant at all three quantiles, but its magnitude varies
substantially. At the median (τ = 0.50), the effect is largest at 0.831, indicating a strong
positive association in the middle of the conditional lending distribution. The effect is
considerably smaller at the lower tail (τ = 0.10, coefficient 0.044) and upper tail (τ = 0.90,
coefficient 0.256). This contrasts with the QBLD results on the binary outcome, where

31
spillover effects were strongest at the lower tail. The difference is consistent with the fact
that the continuous specification focuses on the intensive amount of lending, whereas the
binary QBLD framework focuses on whether counties receive any lending at all.

5.1.3 Tobit with Mundlak Device

Because log(FAB)it is left-censored at zero for counties receiving no bank lending, a Tobit
specification is more appropriate than OLS for the continuous outcome. To approximate
fixed effects within the Tobit framework, we apply the Mundlak (1978) device, decomposing
the regressor into within-county and between-county components:

^
log(FAB)it = βW L1 log AFABit + βB L1 log AFABi + γ ′ dt + εit , (5)

^
where L1 log AFABit is the within-county deviation and L1 log AFABi is the county mean.
The within-county coefficient is β̂W = 0.220 (t = 14.17, p < 0.001) and the between-
county coefficient is β̂B = 1.928 (t = 29.74, p < 0.001), both with standard errors clustered
at the county level. The large between-county effect indicates that counties embedded in
persistently higher-lending neighborhoods tend to have substantially higher own lending,
while the within-county effect confirms that time-varying changes in adjacent lending are
also positively associated with own-county outcomes after accounting for censoring.

5.1.4 Panel Tobit

We estimate panel Tobit models in both random-effects and mixed-effects (approximat-
ing fixed effects) form. The random-effects specification (xttobit) yields a coefficient of
β̂ = 0.745 (z = 54.82, p < 0.001), with an estimated intraclass correlation of ρ = 0.223, con-
firming meaningful county-level heterogeneity. The mixed-effects Tobit (metobit) produces
a larger coefficient of β̂ = 0.946 (z = 68.58, p < 0.001). Both specifications confirm a strong
positive association between adjacent and own-county lending in a censored-data framework,

32
and the larger fixed-effects estimate suggests that the random-effects model may attenuate
the spillover coefficient by conflating within- and between-county variation.

5.1.5 Summary of Alternative Specifications

Across all four alternative approaches, the core finding is consistent: neighboring lending
exposure is positively associated with own-county lending. The linear FE panel, panel quan-
tile, Mundlak Tobit, and panel Tobit models all return positive and significant coefficients
on lagged adjacent lending. The magnitudes and patterns differ across specifications in ways
that reflect their respective assumptions about censoring, distributional heterogeneity, and
the relevant margin of adjustment, but the directional finding is robust.
Importantly, the alternative specifications complement and reinforce the main QBLD re-
sults in several ways. At the same time, they are not estimating exactly the same relationship
as the main models, since they do not include lagged own-county lending (L1 FAB) as a re-
gressor. As a result, the alternative specifications should be interpreted as showing whether
neighboring lending exposure is positively associated with own-county outcomes in a broader
reduced-form sense, rather than as direct counterparts to the QBLD coefficients that are esti-
mated conditional on own-county persistence. The QBLD models found that spillover effects
were strongest at the lower tail of the conditional distribution—precisely where counties are
on the margin of receiving any lending at all. The panel quantile regression’s large median
coefficient (0.831) and smaller tail coefficients on the continuous outcome highlight that the
QBLD’s binary framework captures a qualitatively different and policy-relevant margin—the
probability of receiving any lending—that continuous specifications may partially obscure
by focusing on lending intensity. Meanwhile, the Mundlak Tobit’s large between-county
component (β̂B = 1.928) echoes the QBLD finding that counties persistently embedded in
high-lending neighborhoods show stronger lending propensities, consistent with the spatial
clustering documented in the descriptive evidence of Section 2.4. Taken together, these alter-
native models confirm that the QBLD’s focus on the binary outcome and quantile-specific

33
inference is not merely a methodological choice but captures an economically meaningful
dimension of how neighboring lending exposure appears to operate.

5.2 Robustness and Extensions

One strength of this analysis is that it does not rely on a single specification. The main
QBLD results are supported by four alternative frequentist approaches—linear FE panel,
panel quantile, Mundlak Tobit, and a panel Tobit—all of which return a positive association
between neighboring lending and own-county outcomes. Beyond these alternative estimators,
the QBLD framework itself is tested across nine quantiles, twelve district-specific runs, and
two population-based splits, providing a multi-dimensional view of the spillover relationship.
The consistency of a positive neighboring pattern across these many settings and modeling
choices strengthens the argument that the results are not purely driven by a single sample,
specification, or estimator.
Future extensions could include explicitly spatial panel specifications following Anselin
(1988) or Lee and Yu (2010), alternative spatial weight structures, or direct connections
between county-level lending outcomes and downstream economic variables such as bank
survival, employment, or local recovery.

5.3 Limitations

The first limitation is interpretive. The models identify associations consistent with spatial
spillovers, but they do not establish a definitive causal mechanism. Neighboring lending
exposure may proxy for broader regional financial conditions rather than a direct spillover
from one county to another. Vossmeyer (2016)’s work on RFC selection and treatment
highlights that reduced-form panel approaches cannot fully resolve endogeneity concerns.
Second, adjacency is only one way to define interdependence. Economic spillovers may
also operate through trade linkages, banking networks, urban hierarchies, or state-level in-
stitutional structures that are not fully captured by a simple geographic neighbor definition.

34
As a result, the adjacency-based measure may omit important channels of transmission.
Third, the empirical framework relies on a binary measure of lending, which abstracts
from variation in lending intensity. This choice is partly driven by the data structure and
methodological constraints: the lending outcome is heavily zero-inflated and censored at
zero, and there is no straightforward implementation of a censored panel quantile regression
with fixed effects that accommodates this setting. As a result, the QBLD approach focuses
on whether a county receives any lending rather than the magnitude of lending. While this
allows us to model the most policy-relevant margin in a sparse environment, it limits direct
interpretation of effect sizes in terms of lending amounts.
Fourth, the district-level results are based on a limited set of quantiles and often exhibit
wider credible intervals. These estimates provide suggestive evidence of regional hetero-
geneity, but they should not be interpreted with the same precision as the full-sample and
population-split results, which form the core empirical evidence.
Finally, the alternative specifications serve as useful robustness checks but come with
important limitations. Because they rely on the continuous log lending outcome, they do
not accommodate the large point mass at zero as naturally as the main binary framework.
They also do not estimate the same object as the main QBLD models: the alternative
regressions omit lagged own-county lending (L1 FAB), so they do not isolate neighboring
exposure conditional on own-county persistence. In addition, the panel quantile regressions
are based on the continuous outcome and therefore describe heterogeneity in lending intensity
rather than in the probability of receiving any lending, making them not directly comparable
to the binary quantile estimates. The linear fixed-effects model is a very basic benchmark
and does not account for censoring at zero, while the Tobit-based models impose stronger
functional-form assumptions and only approximate fixed effects through the Mundlak device
rather than fully absorbing county-specific heterogeneity. For these reasons, the alternative
models should be interpreted as simplified complementary checks rather than substitutes
for the main analysis. We therefore place the greatest weight on the QBLD results, since

35
that framework is the one that best matches the central features of the data: a binary
outcome, substantial zero inflation, panel structure, county heterogeneity, and interest in
quantile-specific effects.

6 Conclusion

Financial crises routinely prompt large-scale government interventions in credit markets. A
recurring question—from the Depression-era RFC through TARP and the Paycheck Protec-
tion Program—is whether such interventions affect only directly targeted institutions and
localities, or whether their effects diffuse into surrounding areas. This thesis has examined
that question in the context of RFC lending during 1932–1936, focusing on whether county-
level lending outcomes were associated not only with a county’s own prior lending, but also
with lending activity in adjacent counties.
Using county-level monthly panel data, descriptive geographic evidence, and Bayesian
quantile regression models for binary longitudinal data following Rahman and Vossmeyer
(2019), the main result is that neighboring lending exposure is positively associated with
own-county lending outcomes. The main (AFAB) model shows that adjacent lending matters
across much of the conditional distribution, with particularly strong effects at the lower quan-
tiles and positive effects persisting into the upper tail. The decomposed (AFANB/AFASB)
model further indicates that both adjacent state-bank and national-bank channels contribute
to these patterns, though the state-bank channel is consistently stronger and more sta-
ble—often approximately twice the magnitude of the national-bank channel. The results
also reveal meaningful heterogeneity across county population groups and Federal Reserve
districts, indicating that spatial spillovers were not uniform.
These findings matter for both historical interpretation and broader policy reasoning.
Historically, they suggest that the effects of Depression-era emergency lending were not
confined to directly exposed counties. The dominance of the state-bank channel highlights

36
that the institutional structure of local banking systems shaped how financial stabilization
diffused across space. More broadly, the results imply that evaluating emergency lending
programs solely on the basis of direct exposure likely understates their total impact, as
interventions can generate indirect regional effects that extend beyond treated units.
The results are also consistent with an information or familiarity channel. One interpre-
tation is that lending to banks in one county increased the visibility, credibility, or perceived
accessibility of the RFC program for banks in nearby counties, lowering informational or
administrative barriers to participation. This interpretation aligns with modern evidence
from Anbil, Carlson, and Styczynski (2023), who show that pre-existing familiarity with
government lending frameworks increased both participation and readiness to participate in
the PPPLF, with effects arising even when the facility was not actively used. While this
thesis does not directly observe these mechanisms, the positive association between nearby
lending and own-county outcomes is consistent with the idea that government programs can
diffuse through information and expectations as well as through direct financial channels.
The contribution is therefore twofold. First, it provides evidence that RFC lending
exhibited a geographically diffuse component, operating in part through neighboring banking
networks and plausibly reinforced by informational spillovers. Second, it demonstrates the
value of combining historical policy analysis with Bayesian quantile methods for binary
outcomes to uncover heterogeneous effects across the distribution. Taken together, these
findings underscore that the effectiveness of emergency lending depends not only on where
funds are allocated, but also on how both capital and information propagate across space—an
insight that is central to understanding crisis interventions in both historical and modern
contexts.

37
References

Amujala, S., Vossmeyer, A., & Das, S. R. (2023). Digitization and data frames for card index
records. Explorations in Economic History, 87, 101469.

Anselin, L. (1988). Spatial Econometrics: Methods and Models. Kluwer Academic Publishers.

Anbil, S., Carlson, M., & Styczynski, M.-F. (2023). The effect of the Federal Reserve’s lending
facility on PPP lending by commercial banks. Journal of Financial Intermediation, 55,
101042.

Autor, D., Cho, D., Crane, L. D., Goldar, M., Lutz, B., et al. (2022). An evaluation of the
Paycheck Protection Program using administrative payroll microdata. Journal of Public
Economics, 211, 104664.

Berger, A. N., Karolyi, S. A., & Roman, R. A. (2025). Par for the course: Did the Bank
Term Funding Program (BTFP) support credit intermediation? Working Paper, Federal
Reserve Bank of Philadelphia.

Bloom, N., Schankerman, M., & Van Reenen, J. (2013). Identifying technology spillovers
and product market rivalry. Econometrica, 81(4), 1347–1393.

Calomiris, C. W., Mason, J. R., Weidenmier, M., & Bobroff, K. (2013). The effects of
Reconstruction Finance Corporation assistance on Michigan’s banks’ survival in the 1930s.
Explorations in Economic History, 50(4), 526–547.

Cameron, A. C., & Trivedi, P. K. (2005). Microeconometrics: Methods and Applications.
Cambridge University Press.

Dai, X., Yan, Z., Tian, M., & Tang, M. (2020). Quantile regression for general spatial panel
data models with fixed effects. Journal of Applied Statistics, 47(1), 45–60.

Granja, J., Makridis, C., Yannelis, C., & Zwick, E. (2022). Did the Paycheck Protection
Program hit the target? Journal of Financial Economics, 145(3), 725–761.

38
Lee, L., & Yu, J. (2010). Estimation of spatial autoregressive panel data models with fixed
effects. Journal of Econometrics, 154(2), 165–185.

Mason, J. R. (2003). The political economy of Reconstruction Finance Corporation assistance
during the Great Depression. Explorations in Economic History, 40(2), 101–121.

Moretti, E., Steinwender, C., & Van Reenen, J. (2025). The intellectual spoils of war?
Defense R&D, productivity, and international spillovers. The Review of Economics and
Statistics, 107(1), 14–27.

Mundlak, Y. (1978). On the pooling of time series and cross section data. Econometrica,
46(1), 69–85.

Rahman, M. A., & Vossmeyer, A. (2019). Estimation and applications of quantile regression
for binary longitudinal data. Advances in Econometrics, 40B, 157–191.

Reich, B. J., Fuentes, M., & Dunson, D. B. (2011). Bayesian spatial quantile regression.
Journal of the American Statistical Association, 106(493), 6–20.

Vossmeyer, A. (2016). Sample selection and treatment effect estimation of lender of last
resort policies. Journal of Business & Economic Statistics, 34(2), 197–212.

File and source

File: aldanmaz-2026-spatial-spillovers-in-emergency-lending-evidence-from-bayesian-quantile-mode.pdf
Size: 6,871,829 bytes
SHA-256: bf7bd7126a08720e299de6b682be4cb2058cb36b7d7073ff9b01e73fd93482ed
Our copy: aldanmaz-2026-spatial-spillovers-in-emergency-lending-evidence-from-bayesian-quantile-mode.pdf
Original: web.archive.org