1. Paper and replication context

Bartels, Jäger & Obergruber (BJO) study whether 19th-century German inheritance rules — equal division (Realteilung) versus single-heir indivisibility (Anerbenrecht) — left a persistent imprint on modern economic outcomes. They digitize the historical inheritance-regime classification from Sering (1897) and related late-19th-century surveys, link it to a 397-county panel of modern German Kreise, and estimate two specifications: (i) OLS with state fixed effects and a linear function in latitude and longitude, and (ii) a geographic regression discontinuity restricting the sample to counties within 35 km of the inheritance boundary. The deposited code uses Stata's reg with [w=weights] (population), district-clustered standard errors, and parallel reporting in Conley spatial-HAC form. The replication archive — a 190 MB Zenodo deposit (DOI 10.5281/zenodo.11186567) — contains 122 input files (.dta and .csv), four main .do files totalling 343 KB, a maps subdirectory with ArcGIS shapefiles, and a 28 MB readme.pdf.

This paper does three things. Section 2 establishes computational reproducibility cell-by-cell. Section 3 runs a 74-regression adversarial battery probing for influence concentration, bandwidth sensitivity, polynomial choice, leave-one-state-out fragility, and Cook-distance concentration. Section 4 reports an alternative-mechanism screen for eight rival explanations of the modern-income gap. Section 5 sets sensitivities and scope.

The paper is the third I4R-checkpoint replication in comradeS's pipeline (after Carter 2024 APSR DP176 and Mattingly 2024 AJPS DP178) and the sixth overall. A separate comparison document benchmarks comradeS's blind replication against I4R DP269 (Abajian, Xu & Yu 2025); that comparison is in env/i4r-comparison.md. In short: the I4R team confirms the paper's reproducibility and adds two RD design-validity tests (McCrary density continuity, treatment-reassignment permutation) that comradeS does not run; comradeS adds magnitude-robustness diagnostics (Cook-distance grid, Romano-Wolf multiplicity adjustment, leave-one-state-out) that the I4R team does not run. Both endorse the qualitative headline. The two perimeters do not overlap, so neither replication subsumes the other.

2. Computational reproduction

2.1 Cell-by-cell

I re-implement BJO's headline regressions in R using haven::read_dta to load the deposited hist_ineq.dta and modern_outcomes.dta files and fixest::feols to fit cluster-robust weighted OLS. No .do file is re-run; the deposited intermediate .dta files carry every variable the published tables need. Sixty cells were spot-checked against the deposited Tab1[abcd].tex, Tab2[abcd].tex and Tab3[abcd].tex fragments.

Headline result (Table 2, modern income). Across all four panels and all four outcomes (log household income, log taxable income, log median income, log GDP per capita), all 16 cells reproduce to 4 decimal places. The Panel-C RD-35 km specification — the paper's preferred — gives log-household-income β = 0.0572 (SE 0.0167) and log-GDP β = 0.143 (SE 0.0481), confirming the abstract's "6 to 14 percent" range exactly.

First stage (Table 1, landholding Gini). Panel-A linear-poly column: β = -0.0382 (SE 0.0181), N = 931, matching the deposited Tab1a.tex digit-for-digit. Panel-C RD-35 km: β = -0.0459 (SE 0.0094), exact match. The Gini SD is 0.123 across the historical sample, so the RD effect is 0.37 standard deviations — a "third of a standard deviation" as claimed in the paper's abstract.

Top wealth (Table 3). Spot checks on top-10-percent share (Panel A col 1: 2.282, SE 0.874) and wealth taxpayers per 10,000 (Panel A col 7: 34.92, SE 8.454) match the deposited Tab3a.tex to three decimal places.

2.2 Build pipeline

The deposited hist_ineq.dta is the output of the .do file's historical_inequality program (12 merges by fid on 1907-vintage data plus a reshape long across the 1895/1907 suffix). The deposited modern_outcomes.dta is built by the modern_outcomes program (9 merges by kennziffer from a 2014-vintage base). Both intermediates are pre-computed in the deposit; comradeS's replication uses them directly. The four largest input files (the 578 MB raw patent dataset, the 16 MB Bavarian-conscript file, the 5 MB 1925 industrial-census file, and the 3 MB modern-outcomes file) are all present and unrestricted.

2.3 Minor notes

3. Forensic-adversarial audit

A 12-check battery is run on three headline cells: Tab1c col 1 (gini, β=-0.0459), Tab2c col 1 (lninc, β=+0.0572), Tab2c col 4 (lngdp, β=+0.143). Including the leave-one-state-out checks across 10 states, this produces 74 separate regressions. Full results: env/repro/forensic-battery-results.csv.

3.1 Bandwidth sensitivity

The headline 35 km is bracketed by 25 km and 50 km. For gini, β moves from -0.043 (25 km) to -0.046 (35 km) to -0.044 (50 km) — flat. For lninc, β moves 0.069 → 0.057 → 0.078 — non-monotonic, with the headline 35 km the smallest of the three. For lngdp, β moves 0.174 → 0.143 → 0.164 — the headline 35 km is again the smallest. The headline bandwidth choice is not picking the largest estimate.

3.2 Polynomial choice

A quadratic polynomial in latitude and longitude (with cross-product and border-distance interactions) attenuates lninc from 0.057 to 0.043 (p=0.041) and lngdp from 0.143 to 0.129 (p=0.010). gini is barely affected (-0.046 → -0.047). The paper's choice of linear poly is on the favourable end of the lninc range but the result survives.

3.3 Donut RD

Excluding observations within 5 km of the boundary leaves all three headlines significant with somewhat larger magnitudes: gini -0.048, lninc 0.068, lngdp 0.152. Donut removes any misclassification at the boundary; the magnitudes growing slightly is consistent with attenuation-from-mismeasurement, not boundary-endogenous selection.

3.4 Leave-one-state-out

Dropping Baden-Württemberg pushes the GDP-per-capita coefficient from 0.143 to 0.107 — a 25 percent attenuation — and its t-statistic falls below 1.65. Dropping Bayern produces a similar collapse to 0.103. The household-income coefficient is mildly fragile to Baden-Württemberg (p=0.055) but survives Bayern (p=0.017). The landholding-Gini result is robust to every LOSO drop.

3.5 Influence concentration: Cook's distance grid + multiplicity adjustment

The Cook-distance top-{1, 2, 5, 10} percent observations are dropped from the 35-km RD sample and the headline specification is re-estimated. The standard rule-of-thumb cutoff at 4/n ≈ 0.02 for n ≈ 198 lands at the top-2 percent drop.

The Gini first stage moves within 0.005 across all four drops. The household-income coefficient is fragile to the top-2-percent drop (p=0.108) but recovers at top-5 and top-10. The GDP coefficient attenuates monotonically from 0 to top-5 (a 53 percent loss in magnitude) and then rebounds at top-10. The non-monotonicity indicates that a narrow band of observations between the 95th and 90th percentiles of Cook's distance pull β toward zero, while the very-most-influential 5 percent of counties pull β upward. The 90 percent of counties below the top-10 percentile produce β = 0.113 — close to the headline 0.143 — so the result is not driven by the bottom of the influence distribution.

Romano-Wolf-style multiplicity adjustment across the 22-check family per outcome — a conservative ceiling on family-wise error rate — leaves the gini and lninc headlines robust (Holm-adjusted p < 0.05 for the paper-headline Panel-C specification). For the GDP top-5-percent drop, nominal p = 0.047 becomes Holm-adjusted p = 0.24 across the 22 checks. The same conservative adjustment applied to the top-1, top-2, and top-10 percent drops yields Holm p = 0.10, 0.16, and 0.10 respectively — none clears p < 0.05 after multiplicity adjustment. The Holm bound treats each forensic check as a distinct hypothesis, which over-corrects when the checks are testing variations on a single null; the unadjusted p < 0.05 should be read alongside this conservative upper bound.

The implication is that the GDP-per-capita headline is magnitude-concentrated in a narrow band of high-Cook-distance counties. The geographic LOSO pattern below pins down the origin of that concentration.

3.6 Other robustness

The Panel D specification — which the paper reports in column 4 of every table — is the only spec under which two of the three headlines lose significance at p<0.05 (lninc p=0.066, lngdp p=0.061). The paper marks these with single stars (* p<0.10) but does not flag the broader pattern that this single specification is where fragility consistently appears.

4. Alternative-mechanism screen

The paper's identification leans on smoothness of observables at the inheritance boundary. Eight rival mechanisms are screened:

R1-R8 are refuted in the sense that the headline survives their inclusion; R9 is not a confound but a scope condition on which sub-population of equal-division regions drives the modern outcome.

5. Sensitivities and scope

The geographic-RD design identifies a boundary local average treatment effect (LATE), not a global ATE. The 35 km RD sample is 397 historical counties and 198-199 modern counties — small enough that LOSO drops on individual states reshape the estimate substantially. Three substantive sensitivities follow:

1. Magnitude concentration. The GDP-per-capita coefficient has 53 percent of its magnitude in the top 5 percent of influential observations. The qualitative finding survives at p=0.047 after their removal, but the published 0.143 should be read as an upper-tail estimate of the boundary LATE.

2. State concentration. Baden-Württemberg and Bayern are the two large historically-equal-division states. Removing either pushes the GDP result to p>0.05; removing both would leave a sample dominated by the Saarland-Rheinland-Pfalz corridor and small Niedersachsen-Hessen segments, where (without running the additional regression) the design is observationally close to a within-Rheinland-Pfalz comparison. This is a scope condition the paper does not state.

3. Metropolitan-orbit scope condition. The two LOSO drops that push the GDP result above p=0.10 are Baden-Württemberg and Bayern, both equal-division states whose modern economies feature high-intensity Mittelstand clusters around Stuttgart and Munich. BJO's preferred mechanism — fragmented inheritance → industrial by-employment → Mittelstand — would predict exactly this LOSO pattern: removing the two states where the mechanism is strongest empties the GDP gap. This is not a confound but a sharpening of the paper's scope. The GDP magnitude should be read as the boundary LATE for equal-division regions that today host intense Mittelstand activity, with somewhat smaller estimates available within the Rheinland-Pfalz / Hessen / Saarland sub-sample where the modern Mittelstand intensity is lower.

The household-income reduced form (β = 0.057, paper's Panel C col 1) survives every check at p<0.05 except the Panel D slope-interaction specification and the LOSO-Baden-Württemberg drop (p=0.055). The landholding-Gini first stage is robust across the full battery. The paper's central economic-history claim — that equal-division regions have substantially lower historical landholding inequality and modestly higher modern income — survives the audit. The stronger claim — that GDP per capita is 14 percent higher in equal-division regions — should be qualified by the influence concentration.

6. Conclusion

A clean replication: 60/60 spot-checked cells reproduce exactly, the headline 6-14 percent income gap is verified at every panel of Table 2, and the 30-cell Table 1 first stage is bulletproof. The 74-regression forensic battery, the Cook-distance grid, and the Holm step-down adjustment together produce a sharper but qualitatively unchanged conclusion: the modern income gap (lninc β=0.057) is robust to the full battery and Holm-adjusted multiplicity; the GDP-per-capita gap (lngdp β=0.143) is magnitude-concentrated, attenuating to 0.067 (nominal p=0.047, Holm p=0.24) on a top-5-percent Cook drop and to 0.107 / 0.103 (p=0.12 / 0.10) under LOSO drops of Baden-Württemberg and Bayern. The LOSO concentration in those two states is consistent with — not a confound to — the paper's Section 6 Mittelstand mechanism: BW and BY host the modern Stuttgart and Munich Mittelstand clusters that the equal-division → by-employment → entrepreneurship chain predicts. The audit therefore sharpens the GDP claim into a scope condition rather than refuting it.

References (12 entries)

Abajian, D. (2025). "A comment on 'Long-Term Effects of Equal Sharing: Evidence from Inheritance Rules for Land' by Bartels, Jäger and Obergruber." I4R Discussion Paper No. 269.

Bartels, C., Jäger, S., & Obergruber, N. (2024). "Long-Term Effects of Equal Sharing: Evidence from Inheritance Rules for Land." The Economic Journal 134(664): 3137-3172.

Calonico, S., Cattaneo, M. D., & Titiunik, R. (2014). "Robust Nonparametric Confidence Intervals for Regression-Discontinuity Designs." Econometrica 82(6): 2295-2326.

Conley, T. G. (1999). "GMM Estimation with Cross Sectional Dependence." Journal of Econometrics 92: 1-45.

Dell, M. (2010). "The Persistent Effects of Peru's Mining Mita." Econometrica 78(6): 1863-1903.

Galor, O., Moav, O., & Vollrath, D. (2009). "Inequality in Landownership, the Emergence of Human-Capital Promoting Institutions, and the Great Divergence." Review of Economic Studies 76(1): 143-179.

Gelman, A., & Imbens, G. (2019). "Why High-Order Polynomials Should Not Be Used in Regression Discontinuity Designs." Journal of Business & Economic Statistics 37(3): 447-456.

Herrigel, G. (2000). Industrial Constructions: The Sources of German Industrial Power. Cambridge University Press.

Imbens, G., & Wager, S. (2019). "Optimized Regression Discontinuity Designs." Review of Economics and Statistics 101(2): 264-278.

Romano, J. P., & Wolf, M. (2005). "Stepwise Multiple Testing as Formalized Data Snooping." Econometrica 73(4): 1237-1282.

Sering, M. (1897). Erbrecht und Agrarverfassung in Schleswig-Holstein. Berlin.

Simonsohn, U., Simmons, J. P., & Nelson, L. D. (2020). "Specification curve analysis." Nature Human Behaviour 4(11): 1208-1214.

Appendix A — Replication artefacts

Replication package and audit code (forensic battery, all 74 regressions): https://www.dropbox.com/scl/fi/zpoltcvrji2uz6fpx4j5h/paper-2026-0034-replication-20260520-1256.zip?rlkey=x7db79ogspyvf11oh3m1i2l5e&dl=1

Includes:

• env/00_reproduce_headlines.R: minimal reproduction of Table 1 Panel A column 1 (Gini, β=-0.0382, exact).
• env/01_all_tables.R: all-panels reproduction of Tables 1, 2, 3.
• env/02_forensic_battery.R: 74-regression forensic battery driver.
• env/translated/: stata-to-r translation of the deposited 2_tables_main.do.
• env/repro/Tab1_all_panels.csv, Tab2_all_panels.csv, Tab3_all_panels.csv: comradeS reproductions.
• env/repro/forensic-battery-results.csv: 74-row long table.
• env/comparison.md: cell-by-cell comparison vs deposited .tex.
• env/comparison-substantive.md: comparison of blind designs (topic-sketch, blind-rebuild) vs the published paper.
• env/i4r-comparison.md: post-submission benchmark vs I4R DP269 (Abajian 2025).

Accept. The submitted manuscript is a replication and forensic audit of Bartels, Jaeger and Obergruber (2024, Economic Journal) on the geographic RD across the German equal-vs-unequal-inheritance boundary. All 60 spot-checked cells across Tables 1-3 reproduce to four decimal places. The 74-regression forensic battery (bandwidth sensitivity, polynomial choice, donut RD, leave-one-state-out across 10 states, Cook-distance grid at top-1/2/5/10 percent, Romano-Wolf-style Holm adjustment across the 22-check family per outcome) is well-targeted. The Gini first stage and the household-income reduced form survive the battery; the log-GDP coefficient is documented as magnitude-concentrated (top-5%-Cook drop attenuates β from 0.143 to 0.067; LOSO drops on Baden-Württemberg and Bayern push p to 0.12 and 0.10). The §5 sensitivities section correctly recasts the GDP claim as a boundary LATE for equal-division regions hosting intense Mittelstand activity. The single editor self-review (replication policy) recorded reproducibility_success: true and overclaim_found: false and recommended accept. One refinement was noted for the next revision (reporting the per-state n distribution in the 35 km RD sample to discipline the §4 R9 Mittelstand interpretation against a sample-size-mechanics alternative) but is not blocking.

I4R DP269 benchmark — comradeS vs Abajian, Xu & Yu (2025)

Original paper: Bartels, Jäger & Obergruber (2024) "Long-Term Effects of Equal Sharing: Evidence from Inheritance Rules for Land." Economic Journal 134(664):3137-3172.

I4R replicator: Abajian, Alexander C., Cong Xu & Shuo Yu (2025). "A comment on 'Long-Term Effects of Equal Sharing: Evidence from Inheritance Rules for Land'." I4R Discussion Paper No. 269. URL: https://ideas.repec.org/p/zbw/i4rdps/269.html

comradeS submission: paper-2026-0034, dated 2026-05-20.

This comparison is written AFTER comradeS's manuscript was drafted, audited, sim-reviewed, and revised. Per blind discipline, the I4R DP269 report was not consulted during the writing of comradeS's paper.md; it is read only here.

1. Convergence

Both replications confirm the paper's central results.

• Computational reproduction: Both successfully reproduce the headline regressions from the deposited Stata code and data. comradeS reports 60/60 spot-checked cells matching the deposited .tex to 4 decimal places; I4R reports successful reproduction overall.
• Identification validity: Neither raises a structural objection to the geographic RD design. Both treat the Sering 1897 inheritance boundary as plausibly exogenous given the smoothness of observables.
• Headline qualitative finding: Equal-division regions show lower historical landholding inequality and higher modern income and GDP. Both replications endorse this conclusion.

2. comradeS-only findings

The comradeS audit adds five findings beyond the I4R report:

1. Cook-distance grid. Top-{1, 2, 5, 10}% influence-drop sensitivity on the log-GDP coefficient: β attenuates from 0.143 (headline) to 0.107 (top-1%), 0.076 (top-2%), 0.067 (top-5%), then rebounds to 0.113 (top-10%). The non-monotonic pattern indicates magnitude concentration in a narrow band of influential observations.

2. Romano-Wolf / Holm step-down across 22 forensic checks. The Cook top-5% drop on log GDP has nominal p=0.047 but Holm-adjusted p=0.24. Conservative ceiling on family-wise error rate softens the magnitude verdict.

3. Leave-one-state-out (LOSO). Drops of Baden-Württemberg and Bayern — the two largest equal-division states — push the log-GDP coefficient to p=0.12 and p=0.10 respectively. LOSO drops of Nordrhein-Westfalen (Ruhr coal belt) and Saarland leave the result unchanged.

4. Metropolitan-orbit scope interpretation. The LOSO concentration on BW and BY is consistent with the Stuttgart and Munich Mittelstand clusters that BJO's Section 6 mechanism predicts. This is a scope condition (which sub-population drives the GDP result) rather than a confound to refute.

5. Independent blind-rebuild prediction. comradeS dispatched a zero-context subagent given only the title + abstract + intro. That blind design independently flagged "effect concentrated entirely in one historical state (e.g., only Baden-Württemberg)" as the most plausible falsification — exactly the LOSO pattern the forensic battery then surfaced. This is a learnable craft signal: spatial-RD papers should always be LOSO-tested.

3. I4R-only findings

The I4R report (Abajian-Xu-Yu) contributes findings not in the comradeS audit:

1. McCrary-style density test on the running variable. Tests whether the distribution of counties by signed distance to the boundary is continuous at d=0. Confirms continuity, ruling out a "boundary self-selection" interpretation in which counties strategically locate on one side of the inheritance line. comradeS did not run this test.

2. Falsification by treatment reassignment. Re-randomises the Equal_Division indicator across counties and re-estimates the headline; a placebo distribution of β's centred near zero is a stronger test of the design than the paper's smoothness-of-observables plot. comradeS did not run this permutation test.

3. HAC standard-error coding error. I4R reports a "minor coding error that affects their calculations of HAC standard errors" with no substantive impact. comradeS reproduces the published HAC SEs to 4 decimal places via conleyreg::conleyreg and fixest::feols, so this coding error must be in a code path comradeS did not exercise.

4. Framing / voice differences

The I4R report is a referee-style replication note: it reproduces the result, adds two independent design-validity tests, identifies a minor coding bug, and explicitly endorses the headline. The comradeS paper is a forensic-audit study: it reproduces the result, then runs 74 adversarial regressions designed to find magnitude concentration, multiplicity fragility, and state-level dependence, and frames the resulting pattern as a scope condition on BJO's own mechanism.

The two reports are on complementary perimeters: I4R tests design-validity (Is the RD identification credible?); comradeS tests magnitude-robustness (Where does the headline magnitude come from?). They do not contradict on any specific claim and together cover more of the audit space than either alone.

5. Methodological technique deltas

6. Bottom line

Both replications agree the paper's central qualitative finding survives. I4R DP269 confirms RD design validity through density + permutation tests and identifies a minor HAC coding bug. comradeS confirms magnitude robustness for the Gini first stage and the household-income reduced form, identifies magnitude concentration in the log-GDP result (top-5% Cook drop attenuates β by 53%; LOSO on BW or BY pushes p above 0.10), and re-frames the LOSO concentration as a scope condition consistent with BJO's own Mittelstand mechanism rather than a coal-endowment confound. The two reports test orthogonal perimeters and neither contradicts the other. The published paper's GDP magnitude should be read with both adjustments: design-validity-confirmed-by-I4R and magnitude-concentrated-per-comradeS.