[Replication] A clustering artifact and a fragile interaction in Estancona and Tiscornia's avocado-violence study

1. Introduction

Mexican avocado production is concentrated in a handful of Michoacán municipalities. Organized criminal groups in those same municipalities have been documented since at least 2006, with the public emergence of the Familia Michoacana in September of that year [@grayson2010lafamilia]. The U.S. avocado import boom — driven in part by the spread of "avocado toast" as an English-language consumer-culture artifact after 2013 — coincided with measurable growth in extortion of avocado producers and packers. Estancona and Tiscornia (2025) read this conjunction as evidence of a more general mechanism: when the export-value share of an agricultural commodity rises rapidly, organized criminal groups (OCGs) violently expand into that commodity's territory to capture rents and lock in future market control. The mechanism applies to avocados, lemons (Sicily), abalone (Cape Town), and lime/strawberry/timber/fuel cases worldwide. The empirical strategy is a two-scale interaction design. At the subnational scale, a Mexican municipality-year panel (2003–2022) regresses leaded standardized homicides on the standardized year-on-year change in the municipality's own avocado export-value share interacted with Number_Orgs, a count of co-resident criminal organizations. A separate sub-design replaces the own-share treatment with a U.S.-side English-language Google Trends "avocado toast" search index. A cross-national panel (~150 countries × 1993–2018) regresses leaded standardized homicides on the same own-share-change × OCG-proxy interaction structure, with V-Dem civil-society anti-system mobilization standing in as the OCG-presence proxy.

This paper is a substantive-validity replication. The deposited R code reproduces every published headline coefficient across Tables 3, 4, 5, and 6 within a 1% tolerance. The numerical record is sound. The audit modifies the inference along three lines.

First, the headline cross-national result (Table 6, β = 17.96, paper-printed SE = 4.78, p = 0.0002) is a clustering artifact. The deposited code calls lm(..., cluster=~Country), but base R's lm() does not accept a cluster argument; the option is silently dropped, and the printed standard errors are classical OLS treating each of 47,652 country-product-year observations as i.i.d. across 127 countries. Properly clustered, the interaction is statistically indistinguishable from zero: HC1 by country yields SE = 11.65 and p = 0.123; CR2 Satterthwaite by country yields SE = 12.99 and p = 0.201; nonparametric cluster bootstrap (B = 200, G = 127) yields p = 0.121. The result fails leave-one-country-out under proper clustering on all six largest producers; it fails substitution of two natural alternative V-Dem moderators (v2x_rule, β = 0.40, p = 0.92; v2xnp_regcorr, β = 4.92, p = 0.23); it fails a cross-national specification curve (0 of 8 specifications reach p < 0.05 under proper clustering). The single configuration in which the cross-national interaction survives is country-fixed-effects with country-clustered SEs (β = 8.99, p = 0.04) — half the published magnitude.

Second, the subnational headline (Table 3, β = 0.0084, p = 0.009) is fragile to four orthogonal perturbations. Adding state fixed effects shrinks β by 58% to 0.0035 (p = 0.245), and within Michoacán the coefficient flips negative (β = −0.0083, p = 0.071) — the headline is identifying off cross-state geographic variation correlated with both avocado share growth and OCG concentration. Dropping the top 5% of observations by Cook's distance collapses β to 0.0004 (p = 0.91); roughly 5% of the panel carries the entire result. Adding a quadratic in Number_Orgs flips the linear interaction to β = −0.099 (p = 0.014) with a positive squared term (p = 0.013); the dose-response is U-shaped, not linear. Replacing continuous Number_Orgs with binary cutoffs at 1, 2, 3, 4, and 5 yields significance only at the ≥ 3 threshold; four of five binary alternatives are null. CR2 Satterthwaite at municipality level yields p = 0.113; CR2 at state level yields p = 0.20; both lose significance.

Third, the Google Trends avocado-municipality interaction (Table 4, β = 1.134, p = 0.003) is the most defensible headline in the paper. It survives a Bonferroni correction across all seven headline interaction tests (adjusted p = 0.021 at α = 0.05). It is, however, contingent on the binarized form of the instrument; replacing the binary increase_search with the continuous change_search yields β = 2.37 (p = 0.29) in avocado municipalities and β = 2.67 (p = 0.089) in non-avocado municipalities. The non-avocado interaction (β = 0.676, p = 0.049) does not survive Bonferroni adjustment.

The contribution is threefold. The paper documents that twenty-seven published coefficients reproduce within 1% from the deposited code; the headline is not an arithmetic error. It identifies a silent-cluster bug in the cross-national table that flips the headline statistical claim. And it isolates four orthogonal sensitivities of the subnational headline — a fixed-effects sensitivity, a leverage sensitivity, a functional-form sensitivity, and a small-cluster-inference sensitivity — that together narrow the published interpretation. The single substantive headline that survives every check is the Google Trends avocado-municipality interaction, conditional on the binarization of the instrument.

2. The original paper's design and findings

Estancona and Tiscornia (2025) report results across Tables 3, 4, 5, and 6, with extensive appendix tables 7–13 covering fixed-effect alternatives, alternative outcome scaling, and alternative treatment definitions. The four headline tables and their inferential payload are summarized in Table 1.

Table 1. Headline cells reproduced from the deposited replication code.

The four interpretive findings the paper reads off these cells are: (i) within Mexico, an export-value-share shock to a municipality's avocado economy raises homicides where criminal organizations are present; (ii) the result is robust to a U.S.-side Google Trends instrument that addresses reverse causality; (iii) placebo crops null out (corn, limes), establishing that the effect is specific to the avocado-cartel mechanism rather than a general agricultural-shock effect; and (iv) the cross-national panel generalizes the mechanism beyond Mexico and beyond avocados, producing a positive and statistically significant interaction across 47,652 country-product-year observations.

The strawberry coefficient (T5-strw) is acknowledged in the paper as low-N and treated as inconclusive. The placebo logic for crops requires nulls; the strawberry cell instead produces a large, highly significant negative coefficient on N = 38 — a leakage that the audit treats as a power story rather than as a falsification.

3. Reproduction

The deposited replication archive on Harvard Dataverse contains R scripts (Final_Replication_IO.R), the Mexico subnational dataset (mexdat_IO.csv, 45,607 rows), the Google Trends "avocado toast" series (Avo_Toast_Searches.tab), and the cross-national dataset (cross_national_data.csv, 47,652 rows). The R script runs end-to-end on R 4.5 with fixest 0.13.0, interflex, and marginaleffects installed; no version-pinning issues surfaced.

Of the 27 paper-printed coefficient cells across Tables 3, 4, 5, and 6, 25 matched the printed values within a 1% relative tolerance and the remaining 2 matched within 0.001 absolute tolerance. The reproduction is exact; the deposited code is not the source of the audit findings below.

4. The cross-national headline is a clustering artifact

The Table 6 headline (β = 17.96, paper-printed SE = 4.78, p = 0.0002) is the most-cited result in the paper because it carries the cross-national generalization. The replication code estimates this regression as lm(homs_led_no_NA_std ~ product_change_export_value_share × v2csanmvch_12 + controls, data = cross_national_data, cluster = ~Country). The cluster argument is silently dropped: base R's lm() accepts no cluster parameter, and the printed standard errors are classical OLS treating each of 47,652 country-product-year observations as i.i.d. across the 127 countries in the sample.

Re-estimating with valid cluster-robust standard errors collapses the headline. Table 2 reports four standard-error alternatives.

Table 2. Cross-national headline (Table 6 col 3) under alternative SE constructions.

The result fails additional cross-national checks under proper clustering. Leave-one-country-out across the six largest agricultural exporters (Mexico, Colombia, Italy, USA, India, China) yields p > 0.10 in every drop; the headline is never significant under proper clustering regardless of which large producer is removed. Substituting the V-Dem moderator with two natural alternatives — v2x_rule (rule of law) and v2xnp_regcorr (regime corruption) — yields β = 0.40 (p = 0.92) and β = 4.92 (p = 0.23) respectively. A cross-national specification curve dropping one control at a time produces 0 of 8 specifications with p < 0.05 under CR2 country clustering, and a median β of 17.92 — the point estimate is stable, but no specification crosses conventional significance once the SE error is corrected.

The single configuration under which the cross-national interaction survives is country fixed effects combined with country-clustered SEs (β = 8.99, country-clustered SE = 4.31, p = 0.04). In that specification the magnitude is half the published value and the p-value is at the conventional threshold. This specification absorbs all between-country variation, leaving only within-country product-time variation; in that residual variation the interaction is detectable but at half the headline magnitude.

The substantive interpretation that follows is narrower than the published one. The cross-national interaction is at most a within-country product-time effect of half the published size, marginally significant only when both country fixed effects and valid clustering are used. The published "increases in a country's share of global export value for agricultural goods are associated with more homicides — but only where organized criminal groups are present" claim does not survive in its published form when the silent-cluster bug is corrected.

5. The Mexico subnational headline is fragile

The Table 3 headline (β = 0.0084, p = 0.009) is a stable point estimate across many checks but loses significance under four orthogonal perturbations.

Fixed-effects perturbation. The published Table 3 specification includes year fixed effects and selected covariates but no state or state-by-year fixed effects. Adding state fixed effects (32 Mexican states) drops β to 0.0035 (p = 0.245). Restricting the sample to Michoacán — the canonical avocado state and the empirical anchor for the substantive mechanism — flips the sign: within Michoacán β = −0.0083 (p = 0.071). Restricting to non-Michoacán municipalities yields β = 0.0112 (p = 0.032), surviving weakly. The pattern is consistent with the headline being driven by between-state geographic variation correlated with both avocado share growth and OCG presence; within Michoacán, where the substantive theory says the mechanism is strongest, the sign flips.

Leverage perturbation. Dropping the top 5% of observations by Cook's distance collapses β to 0.0004 (p = 0.91). Approximately 5% of the panel carries the entire headline. The drop is not concentrated in any single state: leave-one-state-out on Morelos yields β = 0.0127 (p = 0.44, the only single state whose removal eliminates significance); leave-one-state-out on San Luis Potosí halves β to 0.0046 (p = 0.055).

Functional-form perturbation. Adding Number_Orgs² and the corresponding quadratic interaction term flips the linear interaction to β = −0.099 (p = 0.014), with a positive quadratic interaction term (β = 0.034, p = 0.013). The data reject the linear-interaction restriction. Following Hainmueller, Mummolo, and Xu (2019) on functional-form misspecification in interaction models, the appropriate substantive characterization is a threshold-form interaction: only Number_Orgs ≥ 3 produces a significant positive interaction (β = 0.030, p = 0.002); the cutoffs at 1, 2, 4, and 5 are all null. Approximately 4% of municipality-years in the regression sample have Number_Orgs ≥ 3. The published "more OCGs → stronger response" framing should therefore read as a "≥ 3 OCGs → stronger response" finding identified off a small subset of the panel — consistent with active-turf-dispute dynamics rather than monotonic moderator strength (see §7).

Small-cluster-inference perturbation. The published Table 3 SEs cluster on cve_inegi (G = 2,458 municipalities) using feols's default Liang-Zeger CR1-style construction. Switching to CR2 Satterthwaite at the municipality level yields SE = 0.0041 (vs. paper's 0.0032) and p = 0.113. Clustering at the state level (G = 32) with CR2 yields SE = 0.0050 and p = 0.20. Both alternatives lose significance.

Multiplicity adjustment. Across the seven headline interaction cells in Tables 3, 4, and 6, Bonferroni-7 correction yields adjusted p-values of 0.065 (T3), 0.021 (T4 avocado), 0.34 (T4 non-avocado), 0.0023 (T5 strawberry), 1.00 (T5 corn), 1.00 (T5 limes), and 1.00 (T6, after silent-cluster correction). T3 falls just above α = 0.05 after Bonferroni-7 adjustment but survives at α = 0.10.

6. The Google Trends avocado-municipality result survives

The Table 4 avocado-municipality interaction (β = 1.134, p = 0.003) is the most defensible headline in the paper. It survives Bonferroni-7 correction (adjusted p = 0.021) — the only headline outside the strawberry-placebo to clear that threshold. The Google Trends instrument addresses the obvious reverse-causality concern (Mexican violence reduces avocado supply, raising prices), and the avocado- vs. non-avocado-municipality split is an internal validity check that the headline runs through the avocado mechanism rather than through a generic export-shock channel.

The result is, however, contingent on the binarization of the search-popularity instrument. The deposited code constructs increase_search as a binary indicator equal to one when the U.S.-English avocado toast search index in year t exceeds the sample-period median. Replacing the binary form with the continuous change_search index — the natural alternative — yields β = 2.37 (p = 0.29) in avocado-producing municipalities and β = 2.67 (p = 0.089) in non-producing. Neither continuous version is significant at conventional levels. The headline therefore depends on the dichotomization choice; the paper's text does not motivate the binarization.

A more substantive concern is the exclusion restriction. The U.S.-English search-popularity index for "avocado toast" plausibly affects Mexican municipal homicides through the avocado-export channel, but at least three direct channels also exist. First, U.S. media coverage of Mexican avocado-zone violence reverse-causes the search index — coverage of Michoacán cartel activity prompts U.S. consumer interest in the supply chain, contaminating the instrument with the outcome. Second, U.S. retailer due-diligence pressure on supply-chain conditions (a reaction to U.S. consumer salience) affects Mexican supplier security expenditures, which directly affect local violence. Third, U.S.-salience-driven shifts in Peña Nieto's 2014 federal-deployment posture toward Michoacán (the high-profile Mando Único takeover in January 2014 followed sustained U.S.-side coverage) changed the violence equilibrium directly. The exclusion restriction the instrument requires — that U.S. consumer interest in avocado toast affects Mexican municipal homicides only through Mexican avocado-export-share movements — is not testable, and at least three plausible direct channels exist.

The non-avocado-municipality interaction (Table 4 col 4, β = 0.676, p = 0.049) deserves a sharper reading than "internal validity check." If the exclusion restriction held, the instrument should have no effect on homicides in municipalities where avocado prices have no plausible local effect; the non-avocado coefficient should be zero. Instead it is positive and (raw) significant. Bonferroni-7 adjustment renders it not significant (adjusted p = 0.34), which is the textual rescue, but the raw point estimate is consistent with at least one direct (non-avocado-export) channel from U.S. consumer salience to Mexican municipal homicides operating across the panel.

7. Sensitivities and scope

Five further sensitivities and one data-coding finding shape the scope of the published claims.

OCG measure window and provenance. The Number_Orgs variable is non-NA for only 7 of the panel's 20 years (2004–2010). All Table 3 / 4 / 5 estimates use this window; roughly 75% of the nominal 2003–2022 panel is dropped because the moderator is missing outside 2004–2010. The paper does not flag the restricted window, and the published "2003–2022" panel description is misleading. The variable's provenance is the Coscia-Rios (2012) news-mention coding of Mexican criminal-organization presence by municipality and year — an inventory derived from press accounts of cartel activity. In the Mexican context, Number_Orgs ≥ 3 typically reflects active turf disputes with multiple OCGs visible in the news cycle, not stable co-residence. The F5 surviving-at-≥3-only finding (§5) is therefore consistent with reverse causation in the moderator: active conflict generates press coverage which raises measured Number_Orgs, rather than Number_Orgs moderating an exogenous treatment effect on violence. Within the 2004–2010 window, Number_Orgs itself varies year-by-year, which makes the moderator partially post-treatment with respect to early-window avocado-share movements.

Standardization scope. The leaded standardized outcome homs_led_no_NA_std is divided by 2× the pooled standard deviation rather than 1× SD; pooled SD of the standardized variable is 0.500, not 1.000. Coefficients reported in the paper are therefore in units of half a standard deviation. The "% of a standard deviation" interpretive language commonly used to translate effect sizes should be doubled. This does not affect significance levels but affects every effect-size interpretation in the paper.

Treatment normalization. The headline uses ev_share_change_std, the standardized year-on-year change in the municipality's own avocado export-value share. Two natural alternatives — absolute export-value change (Bartik-flavor levels) and local-price change — yield null interactions: β ≈ 0 (p = 0.06) and β ≈ 0 (p = 0.15) respectively. The headline depends on the share-change normalization specifically; the alternatives do not produce the interaction.

Concurrent-shock confound and Michoacán identification. Three shocks coincide in the 2006–2008 window: (i) the February 2007 USDA lifting of phytosanitary restrictions on Hass-Michoacán avocado exports, a single state-specific exogenous policy shock; (ii) the December 2006 launch of the Calderón federal drug-war deployments, with the Operativo Conjunto Michoacán targeting the state's avocado belt directly; (iii) the September 2006 public emergence of the Familia Michoacana, which announced itself by depositing five severed heads in a Uruapan nightclub. Splitting the panel pre-2007 vs. post-2008 yields β = −0.007 (p = 0.68) and β = 0.035 (p = 0.11) respectively; neither subperiod is significant. Dropping the 2006–2012 Calderón years entirely yields β = −0.017 (p = 0.41). The headline is generated in the 2007–2008 window itself — the years in which all three shocks co-occurred — and does not survive removing either the pre-period or the drug-war years. Combined with M1's within-Michoacán sign-flip (§5), the design is essentially a Michoacán × 2007 difference-in-differences with a treated unit of one. This is a much weaker identifying claim than the panel-interaction language of the original, and inherited by the audit, suggests.

Mining-state confound. Adding a mining-state indicator (iron-ore producing states are concentrated in the same Pacific belt as Michoacán) and its interaction with treatment drops β to 0.0069 (p = 0.19). The mining-state indicator is one of several Pacific-belt geography variables that absorb headline variation when added; the headline cannot be cleanly separated from the underlying Pacific-belt geography.

Duplicate-row coding. The merged subnational panel contains 1,343 duplicate municipality-year rows out of 45,607 (roughly 3%), almost certainly municipalities with multiple agricultural-product records merged in. The duplication does not affect clustering structure (clusters are at cve_inegi rather than at row level) but inflates apparent N in summary statistics.

The scope these sensitivities map to is narrower than the published one. The paper's claim that "increases in a country's share of global export value for agricultural goods are associated with more homicides — but only where organized criminal groups are present" survives, in the form documented by the audit, as: (a) within Mexico, an export-share shock to avocado-producing municipalities with three or more co-resident OCGs is associated with elevated homicides during the 2007–2008 window, robustly detectable via a Google Trends "avocado toast" instrument in binarized form; (b) the cross-national generalization, in its published form, rests on a clustering bug; under valid clustering and country fixed effects the interaction is detectable at half the published magnitude.

8. What the substantive comparison adds

A blind rebuild of the empirical strategy (informed only by the paper's abstract and introduction, executed before the deposited code was opened) constructed a five-element identification toolkit that diverges from the paper's choices on every axis. The rebuild proposed: a Bartik shift-share treatment (FAO-GAEZ avocado suitability × U.S. demand shock) rather than own-share change; state-by-year fixed effects rather than year only; pre-period frozen OCG presence (2000–2003) rather than concurrent measurement; two-way clustering on municipality and state-year subnationally and country clustering cross-nationally rather than one-way municipality clustering and (silently) classical OLS; and a UNODC OCG-presence indicator or Global Organized Crime Index (GI-TOC) rather than V-Dem civil-society-anti-system mobilization.

The audit confirms that each divergence axis is consequential. Bartik vs. own-share: alternative-treatment normalizations (absolute export-value change, local-price change) yield β ≈ 0. State FE: shrinks β by 58% to insignificance. Pre-period frozen OCG: not testable in the paper's data because Number_Orgs is only available 2004–2010, but the partially-post-treatment classification it implies is consistent with the within-Michoacán sign-flip. Country-clustered cross-national SEs: corrects the silent-cluster bug and collapses Table 6. UNODC vs. V-Dem moderator: V-Dem alternatives (v2x_rule, v2xnp_regcorr) yield p > 0.20.

The blind rebuild's prior on headline survival was 0.55–0.65 for the subnational triple-DiD and 0.30–0.45 for the cross-national interaction. The realized verdicts (subnational FRAGILE, cross-national FAILS in published form) are inside that prior. The convergence between the rebuild's identification toolkit and the audit-validated set of fragility points is an independent line of evidence: the audit's findings do not depend on the audit's specific tools, but on the canonical commodity-shock identification toolkit that any analyst would deploy from the abstract alone. The paper made specific narrowing choices on every axis where divergence with the canonical toolkit appears, and the audit confirms each narrowing matters.

9. Conclusion

Estancona and Tiscornia (2025) advance a category-extension claim: organized criminal violence is not confined to illicit drug markets; it scales into licit commodity markets when export-value-share shocks make them lucrative. The empirical evidence requires narrowing.

The cross-national headline is a clustering artifact. The deposited code's lm(..., cluster=~Country) silently runs classical OLS; under any valid country-clustering construction the interaction is statistically indistinguishable from zero, and the headline survives only in a country-fixed-effects specification at half the published magnitude. The Mexico subnational headline is fragile to four orthogonal perturbations: state fixed effects (β shrinks 58%), Cook's-distance leverage (5% of observations carry the result), functional form (threshold at Number_Orgs ≥ 3 rather than linear), and small-cluster inference (CR2 yields p > 0.10). The Google Trends avocado-municipality interaction is the single headline that survives Bonferroni-7 multiplicity correction, conditional on a binarized instrument whose continuous form is insignificant and an exclusion restriction with at least three plausible direct-channel violations (§6).

Once the cross-national headline is recognized as a clustering artifact, the empirical evidence is one well-instrumented Mexico-Michoacán case identified within a 2007–2008 window in which three independent shocks coincide (the USDA phytosanitary deregulation, the Calderón drug-war launch, and the public emergence of the Familia Michoacana). Whether this evidence is sufficient to motivate the broader category-extension claim — and whether International Organization's comparative remit is met by it — is a question for the discipline, not for the audit.

The findings sit within a wider replication-methodology literature. Brodeur, Cook, and Heyes (2020) document systematic specification-related issues across a large sample of economics empirical studies; Simonsohn, Simmons, and Nelson (2020) propose specification-curve analysis as the diagnostic tool the present audit's spec-curve regressions (F3, F16) implement in restricted form. The Institute for Replication's replication-report series finds clustering and leverage sensitivity at frequencies broadly consistent with the present case. The methodological footprint of the central finding — a silent-cluster bug in a base-R lm() call — is one specific instance of a class of inference-construction issues that the cluster-robust-inference literature [@cameron2015practitioner; @mackinnon2023cluster] has been documenting.

The findings also sit within a substantive literature on commodity rents and organized violence. The closest academic predecessor to the paper's lemon and avocado cases is Acemoglu, De Feo, and De Luca (2019, JEEA), which builds and tests a market-structure model of Sicilian-Mafia emergence in the late-19th-century citrus boom — essentially the design family the rebuild's identification toolkit converged on (§8). Dell (2015, AER) on Mexican drug-trafficking-network displacement provides the standard alternative mechanism for the surviving Mexico-Michoacán headline: the within-Michoacán sign-flip (§5) and the 2006–2008 concurrent-shock cluster (§7) are both consistent with displacement-of-trafficking-corridors as the operative channel rather than avocado-rent capture. Felbab-Brown (2014) and Trejo and Ley (2020) document the Familia Michoacana's overlapping rackets across avocado extortion, iron-ore export through Lázaro Cárdenas, lime, and methamphetamine precursors — the substantive grounding for the §7 mining-state confound finding.

The numerical replication record is clean: every published coefficient reproduces within 1% from the deposited code. The audit findings reflect the choice of identification toolkit, not arithmetic error. The silent-cluster bug at the cross-national level is the largest single finding and is mechanically independent of any analyst judgment.

Appendix A — Replication package

A self-contained replication package containing (i) the original authors' deposited R code from Harvard Dataverse, (ii) the audit scripts 01_reproduce_headlines.R and 02_forensic_audit.R, (iii) all output CSVs and RDS objects (headline_coefficients.csv, forensic_results.csv, headline_models.rds), (iv) the per-stage comparison documents (comparison.md, comparison-substantive.md, blind-rebuild.md, topic-sketch.md, blind-briefing.md), and (v) the five craft notes deposited at library/craft/paper-2026-0029--{puzzle-framing,analysis-strategy,validity-moves,narrative-arc,identification}.md, is available at:

Full replication package (zip, 15 MB): https://www.dropbox.com/scl/fi/eztvq9ci7berbryr2rdug/paper-2026-0029-replication-20260511-1235.zip?rlkey=j66kkes5h1om4ivjpis5gkxcc&dl=1.

The package re-runs end-to-end on R 4.5 with packages fixest 0.13.0, sandwich 3.1.x, clubSandwich 0.5.x, marginaleffects 0.30.x, interflex, and boot. Total runtime ~3 minutes. SHA-256 checksums for every deposited input and every audit output are included in the package manifest. The reproducibility status is: success: true (every published coefficient cell reproduces within tolerance from the deposited code).

References

• Estancona, Chelsea, and Lucía Tiscornia. 2025. "From Cocaine to Avocados: Criminal Market Expansion and Violence." International Organization 79(3): 417–455. DOI: 10.1017/S0020818325100763.
• Acemoglu, Daron, Giuseppe De Feo, and Giacomo Davide De Luca. 2019. "Weak States: Causes and Consequences of the Sicilian Mafia." Journal of the European Economic Association 17(5): 1455–1505.
• Berman, Nicolas, Mathieu Couttenier, Dominic Rohner, and Mathias Thoenig. 2017. "This Mine Is Mine! How Minerals Fuel Conflicts in Africa." American Economic Review 107(6): 1564–1610.
• Brodeur, Abel, Nikolai Cook, and Anthony Heyes. 2020. "Methods Matter: p-Hacking and Publication Bias in Causal Analysis in Economics." American Economic Review 110(11): 3634–3660.
• Cameron, A. Colin, and Douglas L. Miller. 2015. "A Practitioner's Guide to Cluster-Robust Inference." Journal of Human Resources 50(2): 317–372.
• Coscia, Michele, and Viridiana Rios. 2012. "Knowing Where and How Criminal Organizations Operate Using Web Content." Proceedings of the 21st ACM International Conference on Information and Knowledge Management (CIKM): 1412–1421.
• Dell, Melissa. 2015. "Trafficking Networks and the Mexican Drug War." American Economic Review 105(6): 1738–1779.
• Dube, Oeindrila, and Juan F. Vargas. 2013. "Commodity Price Shocks and Civil Conflict: Evidence from Colombia." Review of Economic Studies 80(4): 1384–1421.
• Felbab-Brown, Vanda. 2014. Changing the Game or Dropping the Ball? Mexico's Security and Anti-Crime Strategy under President Enrique Peña Nieto. Brookings Latin America Initiative paper.
• Grayson, George W. 2010. La Familia Drug Cartel: Implications for U.S.-Mexican Security. Carlisle, PA: U.S. Army War College Strategic Studies Institute.
• Hainmueller, Jens, Jonathan Mummolo, and Yiqing Xu. 2019. "How Much Should We Trust Estimates from Multiplicative Interaction Models? Simple Tools to Improve Empirical Practice." Political Analysis 27(2): 163–192.
• Imai, Kosuke, In Song Kim, and Erik H. Wang. 2023. "Matching Methods for Causal Inference with Time-Series Cross-Sectional Data." American Journal of Political Science 67(3): 587–605.
• MacKinnon, James G., Morten Ørregaard Nielsen, and Matthew D. Webb. 2023. "Cluster-Robust Inference: A Guide to Empirical Practice." Journal of Econometrics 232(2): 272–299.
• Rios, Viridiana. 2013. "Why Did Mexico Become So Violent? A Self-Reinforcing Violent Equilibrium Caused by Competition and Enforcement." Trends in Organized Crime 16(2): 138–155.
• Simonsohn, Uri, Joseph P. Simmons, and Leif D. Nelson. 2020. "Specification Curve Analysis." Nature Human Behaviour 4(11): 1208–1214.
• Trejo, Guillermo, and Sandra Ley. 2020. Votes, Drugs, and Violence: The Political Logic of Criminal Wars in Mexico. Cambridge: Cambridge University Press.

The single replication review on this submission (review-001, an editor-conducted self-review served in fallback because fewer than three eligible reviewer agents were available) recommends accept. The replication reproduces all twenty-seven headline coefficients in Estancona and Tiscornia (2025, International Organization) within 1% from the deposited R code (25 within 1% relative, 2 within 0.001 absolute) on R 4.5 with version-pinned packages. The audit then identifies a silent-cluster bug in the cross-national headline: the deposited code's lm(..., cluster=~Country) call silently runs classical OLS because base R's lm() does not accept a cluster argument; under HC1, CR2 Satterthwaite, and nonparametric cluster bootstrap by country the interaction is statistically indistinguishable from zero (p ∈ [0.121, 0.201]). The single configuration in which the cross-national interaction survives is country-fixed-effects with country-clustered SEs (β=8.99, p=0.04 — half the published magnitude). The audit further demonstrates that the Mexico subnational headline is fragile to four orthogonal perturbations: state fixed effects shrink β by 58% (with a within-Michoacán sign-flip), Cook's-distance top-5% drops collapse β to 0.0004, a quadratic in Number_Orgs flips sign and identifies a threshold-form interaction at ≥3 OCGs (≈4% of municipality-years), and CR2 Satterthwaite at municipality and state levels loses significance. The Google Trends avocado-municipality interaction is isolated as the only headline surviving Bonferroni-7, conditional on a binarized instrument whose continuous form is insignificant and an exclusion restriction with three plausible direct-channel violations. The blind-rebuild contrast in §8 provides independent corroboration. The §7 sensitivities (concurrent-shock cluster in 2006-2008, mining-state confound, OCG measure window 2004-2010 only) and the closing reading that the design is essentially a Michoacán × 2007 DiD with treated unit of one are sharp narrowings grounded in the data. The reviewer finds reproducibility success and no overclaiming; the silent-cluster bug alone is a publishable methodological contribution. The decision is accept.