Key Findings
Robustness: Size Coefficient Across Specifications Inferential
Point estimates and 95% confidence intervals for the within-bank log-assets coefficient under six alternative specifications. Ordered by evidentiary strength; dashed line marks zero.
The first-difference and lagged-controls specifications provide the strongest reassurance. Full robustness details →
NIM Trend by Size Decile
Mean net interest margin by quarter, grouped by bank size decile (1 = smallest, 10 = largest). 8,032 banks, 64 quarters, 2010–2025.
Each line represents one size decile of banks. The persistent gap between small-bank (top) and large-bank (bottom) lines reflects the cross-sectional size penalty.
The core finding: larger banks have lower NIM both cross-sectionally and within-bank over time. But this is not the full story — large banks offset the NIM penalty with higher fee income and noninterest revenue, producing higher overall ROA. The earnings model looks different, not necessarily worse.
Key caveat: the raw-NIM specification makes the size coefficient insignificant, so the strongest reassurance comes from the first-difference and lagged-control specifications rather than the winsorized baseline alone.
Time variation matters: in rolling 20-quarter fixed-effects windows, the within-bank size penalty weakens to about −0.033 in the mid-2010s, steepens to about −0.144 in 2019–2024, and then eases in the latest window.
Cross-Section: NIM vs. Bank Size Descriptive
Binscatter of average NIM against average log assets (20 equal-sized bins). This is descriptive cross-sectional evidence—not the within-bank estimate. The downward slope visualizes the between-bank size penalty.
Reading Path
1. Strongest Evidence
Start with the within-bank baseline, then check six robustness specifications and the time-varying coefficient path. This is where the core claim is tested.
Baseline & robustness →2. Mechanisms
NIM decomposition, franchise dilution, rate-cycle heterogeneity, and the acquisition event study. These sections ask why the penalty exists and when it strengthens.
Mechanisms & dynamics →3. Data & Appendices
Sample construction, distributions, geography, threshold crossings, extensions, and summary statistics. Supporting detail for reviewers and replicators.
Data & appendices →Reproducibility
Every data source is free and federally published. The pipeline is scripted end to end, but full reruns still depend on external-data availability and explicit setup prerequisites. See the GitHub repository for setup instructions.