The Trouble With the Offshore Leaks How a Graph-Based Test of 54 FCPA Defendants Shows the ICIJ Archive Cannot Support the Comparisons It Is Used For

A shell company, a Cyprus law firm, and the trouble with the offshore leaks.

This started as an attempt to find out whether companies caught in foreign-bribery cases restructure their offshore holdings in response. The ICIJ Offshore Leaks appeared to show a pattern. Rigorous testing dissolved it. Comparison against similar companies that were never caught found no difference. The real story turned out to be about the leaks themselves. They are not a random sample of the offshore world. They are whatever got leaked, and the pattern inside them is largely the pattern of what reporters got their hands on. A check against the UK’s near-complete corporate registry shows how skewed ICIJ is: Barbados appears there at 200 times its actual share of UK-connected offshore structures. That matters for anyone building claims about corporate behavior on these leaks.

US and UK sanctions authorities designated the Cyprus law firm Christodoulos G. Vassiliades and Co. LLC for its role servicing Russian oligarch corporate structures. Fifteen years before those sanctions, Vassiliades appeared in ICIJ records as intermediary for offshore subsidiaries of two US-listed FCPA-defendant multinationals: Amersham plc, the British life-sciences firm later acquired by GE Healthcare, and Lucent Technologies, later merged with Alcatel. The finding is suggestive. It is the kind of thing that opens a piece.

It is also wrong. The two ICIJ entities in question are “Amersham Consulting Inc.” (BVI, incorporated 2005) and “Lucent Investment Group Ltd.” (BVI, incorporated 1997), and both are name-collision shells rather than affiliates of the operating multinationals. Amersham Consulting Inc.’s sole director, Matthew Charles Stokes, based in the United Arab Emirates, also serves as sole director of 236 other unrelated ICIJ entities. Lucent Investment Group Ltd. has no officer data recorded at all. Someone incorporated the shells because the names sounded plausible. Vassiliades was likely providing nominee-services on standard corporate-formation contracts, not specialized advisory work for two Fortune 500 companies. What looked like a fifteen-year pre-sanctions relationship between a now-sanctioned Cyprus law firm and two US multinationals was two coincidences at one address.

This report set out to test a question of interest to FCPA specialists: when American corporations settle Foreign Corrupt Practices Act cases with the SEC, do they subsequently reorganize their offshore holdings to reduce scrutiny? The obvious dataset is the ICIJ Offshore Leaks (Panama Papers, Bahamas Leaks, Offshore Leaks, Paradise Papers, Pandora Papers, collectively 814,000 leaked offshore entities). Initial analysis said yes. Post-conduct incorporations elevated. Jurisdictional composition shifted from Bermuda, Barbados, and Bahamas to BVI, Malta, and St Kitts. The temporal test came back at p=0.005. The kind of finding that would justify a paper.

Every part of that story fell apart on inspection.

The temporal elevation degraded from p=0.005 to p=0.059 after two obvious corrections: deduplicating Tyco’s double-counted matches, and rebuilding the null model to reflect that each of the five source leaks covers different jurisdictions across different date ranges. The final p-value not only exceeds the conventional 0.05 significance threshold; it fails the Bonferroni-corrected threshold of 0.0167 (three temporal comparisons: pre-conduct, during-conduct, post-conduct) by a wider margin still. The direction of the pattern in the raw data is preserved. The statistical case for a pattern across defendants generally is not.

The jurisdictional shift turned out to be an artifact of when different leaks became available rather than a behavioral pattern. The 2013 Offshore Leaks are BVI-heavy. The 2016 Panama Papers are Panama-heavy. What looked like corporations pivoting from mature offshore havens to privacy-focused domiciles was the ICIJ archive itself acquiring different jurisdictional coverage as different registries cracked open in sequence. The pivot was in the leaks, not the defendants.

Even the case FCPA-compliance teaching literature holds up as the exemplar of post-settlement offshore restructuring, Weatherford International’s 2014 Bermuda-to-Ireland reorganization one year after its $152.6 million SEC settlement, turns out on close reading to be more ambiguous than the textbook framing suggests. Weatherford had already relocated from Bermuda to Switzerland in 2009, four years before the settlement, during the period the SEC subsequently alleged was part of the underlying misconduct. The 2014 move from Switzerland to Ireland was characterized in the trade press at the time as a tax inversion driven by the Irish corporate tax regime, one of several completed by oilfield-services peers in the same period, not as a compliance response to the SEC action. UK Companies House confirms the resulting Irish parent is the Person with Significant Control of at least seven UK-registered companies. The Ireland structure is documentably real. Whether it is a settlement response is genuinely uncertain.

A matched-control comparison was built to answer the obvious peer-review question: how do we know FCPA defendants are unusual at all? For each of 54 defendants, three non-FCPA industry peers were selected of comparable size and era, then filtered against a 197-name exclusion list covering corporate FCPA defendants from both DOJ criminal and SEC civil actions (equivalent to the Stanford Law School FCPA Clearinghouse). 84 unique clean controls resulted. Baseline ICIJ hit rate: 81.5 percent of defendants versus 72.6 percent of clean controls, Fisher’s exact p=0.31. Not significant. Source-leak distribution still differs sharply (chi-square p<10^-78), but defendants are clustered in the 2013 Offshore Leaks and under-represented in post-2017 Paradise Papers subdivisions, which is what Sections 02 and 06 established about archive coverage rather than defendant behavior. Temporal distribution without the Universal Corporation outlier is statistically identical between groups (p=0.99). The matched comparison, properly controlled, does not distinguish FCPA defendants from non-FCPA industry peers in their ICIJ-leak footprint. An initial control-pool build that filtered only against this report’s own 54-defendant list produced an apparent OR=2.5 (p=0.038) elevation. That was an artifact. Halliburton, Pfizer, Eli Lilly, 3M, Honeywell, HP, Cisco, Microsoft, SAP and other post-2013 FCPA defendants were being counted as “clean controls” because they fell outside the pre-2013 SEC Litigation Release channel this report draws from. The clean comparison above comes from the rebuild. Section 11 documents it in full.

The ICIJ Offshore Leaks are routinely treated as a representative sample of the global offshore market. They are not, and the gap is measurable at scale. The UK Companies House Persons with Significant Control snapshot (15.6 million records covering UK-registered companies subject to PSC disclosure) shows Barbados representing 0.13 percent of the offshore market servicing UK-connected corporate structures. The ICIJ leaks put Barbados at 27 percent of pre-conduct alignments for FCPA defendants. Barbados appears in leaked records roughly 200 times more often than it appears in the real offshore market. Some fraction of that gap reflects a genuine US-multinational preference for Barbados during the era (the 1984 US-Barbados tax treaty made Barbados IBCs unusually attractive for American companies in ways that did not apply to UK-connected companies). No plausible account of corporate behavior can generate a 200-fold deviation. The gap is the leak, not the market.

Three findings that initially looked publishable dissolved under rigorous testing. A matched-control comparison, rebuilt to exclude the other FCPA defendants that had contaminated the original control pool, found no distinguishing signal. The 200x Barbados discrepancy ties the dissolution together and warns anyone using ICIJ data for behavioral inference about specific corporate populations.

Scope of what this report can defend. A negative result on the initial FCPA-post-settlement-behavior question, with the statistical work shown transparently. A methodology finding about the ICIJ Offshore Leaks corpus being unfit for representative claims about corporate offshore behavior, corroborated externally by UK Companies House data at 200x scale for the Barbados case alone, and corroborated internally by a matched-control comparison (Section 11) that finds no distinguishing ICIJ-footprint signal between FCPA defendants and non-FCPA industry peers once the control pool is properly scrubbed of other FCPA defendants. One ambiguous individual case (Weatherford) that appears in the FCPA-teaching literature as post-settlement restructuring but that this analysis cannot confirm as such. Nothing in this report should be read as an allegation of unindicted criminal conduct against any FCPA defendant, or as a claim that any specific service provider was knowingly complicit in FCPA violations.

Who this report is for. Three audiences treating the ICIJ Offshore Leaks as a representative proxy for the global offshore market, each with a different stake in the methodology lesson: legal scholars and FCPA-methodology researchers who use leaked corporate archives as evidence for corporate-behavior claims, including in casebook treatment of specific defendants; investigative journalists who build narratives about corporate offshore conduct from ICIJ-match hit lists; and compliance economists who construct firm-level empirical analyses, FCPA-risk models, or offshore-exposure scores on ICIJ data. For all three, the operational claim is identical: any corporate-behavior inference built on this corpus must be matched-control-validated against a non-FCPA peer population explicitly scrubbed of other FCPA defendants, externally corroborated against a near-complete reference registry at the relevant jurisdiction and date range, or both. This report demonstrates both validations against its own initial positive-signal hypothesis and reports in detail that the signal survives neither.

What we did, why we did it, and what we expected to find.

What we did

Matched SEC EDGAR corporate FCPA defendants against the full ICIJ Offshore Leaks Database. The ICIJ corpus aggregates five separate leaks (Panama Papers, Bahamas Leaks, Offshore Leaks, Paradise Papers, Pandora Papers) covering 814,315 offshore entities across dozens of secrecy jurisdictions. Fellegi-Sunter probabilistic record linkage with independence guards produced 128,591 candidate matches, narrowed through five iterations of matcher refinement to 118 defensible alignments across 26 defendants. For each alignment we recorded the offshore entity’s incorporation date and jurisdiction, then pulled the 1-hop network of officers, intermediaries, and registered addresses around each entity to characterize the surrounding infrastructure.

Why we did it

SEC FCPA filings name the defendant corporation and describe the bribery conduct, but they rarely trace the offshore vehicles through which payments moved or through which subsidiary assets were held. Third-party facilitators are typically not named. Corporate FCPA cases routinely involve offshore subsidiaries in jurisdictions with limited public disclosure, and post-2016 leak coverage made a subset of those subsidiaries visible in public data for the first time. No prior published work joins the SEC FCPA docket systematically to the ICIJ corpus. The temporal and structural properties of that intersection are what this analysis characterizes.

What we expected to find, and what we found

The naive expectation, given the standard narrative that offshore structures facilitate bribery, is that FCPA-defendant-linked offshore entities would concentrate in the years before and during the alleged conduct period. Under the null hypothesis of no association, incorporation dates would distribute uniformly across pre, during, and post conduct windows in proportion to the length of each window and the base rate of offshore incorporation over time.

The observed pattern departs from both expectations, initially in the direction of a real effect. Post-conduct incorporation was elevated on the initial null model (p=0.005 as originally reported). Pre- and during-conduct incorporation were not elevated. That inversion of the naive expectation motivated the second analysis: if the offshore infrastructure was not being built to facilitate the conduct, what was it being built for, and where.

Two sensitivity checks were then run to test the robustness of the initial finding: Tyco deduplication (correcting for case-level duplicates that inflated the entity count) and source-stratified null modeling (correcting for the joint distribution of ICIJ source leak and incorporation date). Under either check the p-value fails Bonferroni correction. Under both applied together, the p-value fails even conventional significance. The direction of the pattern is preserved. The statistical support for interpreting it as a defendant-behavior signal is not. Sections 02, 03, and 07 walk through what the checks showed and what they mean for the finding.

How we did it

Two complementary analyses on the same 118 alignments, plus the two sensitivity checks described above.

Directional pattern preserved. Bonferroni significance not.

Of the 118 defensible alignments between 26 FCPA-defendant corporations and ICIJ offshore entities, 32 (27.1%) involve offshore entities incorporated after the alleged conduct period had ended. On the original free-permutation null model this exceeded the 95% confidence interval and produced a Bonferroni-significant post-conduct effect. Two subsequent sensitivity corrections have been run: Tyco deduplication (correcting the fact that two SEC case records for Tyco each matched the same 11 ICIJ entities, inflating the entity count) and source-stratified null modeling (correcting the fact that the ICIJ corpus aggregates leaks with different date and jurisdiction coverage). Under either correction alone, the finding does not survive Bonferroni. Under both together, it does not reach conventional significance.

Sensitivity: what happens when the corrections are applied

Two things stand out. First, the direction of the elevation is preserved across all four tests. The post-conduct share ranges from 27.1% to 29.0% observed against 25.8% to 27.9% expected. There is a small elevation, and its direction is consistent. Second, the statistical significance depends heavily on which null model is used. Deduplicating Tyco removes 11 case-level duplicates and shrinks the sample from 118 to 107, which lowers statistical power enough that the finding no longer clears Bonferroni. Stratifying the null model by ICIJ source leak raises the null expectation from 25.8 to 27.9 because entities in later-collected leaks (Pandora Papers, later Paradise Papers subsets) skew toward later incorporation dates for reasons unrelated to defendant behavior.

The observed shift is largely a leak-coverage artifact.

Splitting the 118 alignments by both temporal category and jurisdiction produces an eye-catching descriptive pattern. Pre-conduct entities cluster in mature Big Four accounting jurisdictions (46 of 67 = 69% in Barbados, Bahamas, or Bermuda). Post-conduct entities cluster in privacy-focused jurisdictions (20 of 32 = 63% in BVI, Malta, St Kitts, Samoa, or Seychelles). At face value, the Wilson intervals do not overlap and the shift looks real.

The problem is not the descriptive counts. The descriptive counts are what they are. The problem is what the shift means, and specifically whether it reflects any change in what corporations did versus a change in which ICIJ leak happens to include each entity. Cross-tabulating alignments by ICIJ source leak and temporal category resolves the question.

The source-leak by temporal-category cross-tabulation

Three observations, each fatal to the naive reading of the jurisdictional shift.

First, the pre-conduct Big Four dominance is arithmetically forced by leak composition. Bahamas Leaks contributes 12 pre-conduct alignments, all Bahamas by construction (that leak is a scrape of the Bahamas corporate registry). Paradise Papers Barbados registry contributes 18 pre-conduct alignments, all Barbados. Paradise Papers Appleby contributes 19 pre-conduct alignments concentrated in Bermuda. Together these three sources supply 49 of the 67 pre-conduct alignments (73%) and cover exactly the three “mature Big Four” jurisdictions the shift narrative rests on. If those leaks had not been included in the ICIJ corpus, the pre-conduct Big Four share would be much smaller.

Second, the post-conduct privacy-jurisdiction concentration is similarly leak-forced. All four post-conduct Malta observations come from the Paradise Papers Malta corporate registry subset (Malta only by construction). All three post-conduct St Kitts observations come from the Paradise Papers Nevis corporate registry subset (St Kitts only by construction). These are not defendants choosing Malta or St Kitts; these are leaks that happen to cover Malta and St Kitts and happen to contain post-conduct-dated entities linked to these defendants.

Third, the two multi-jurisdiction leaks that could in principle reveal a jurisdictional shift not driven by coverage (Panama Papers and Offshore Leaks) do not show one. Panama Papers has 8 pre-conduct and 12 post-conduct alignments; Offshore Leaks has 3 of each. In neither of these leaks is there a within-leak temporal-jurisdictional pattern that could support a defendant-behavior interpretation. Whatever the multi-jurisdiction leaks show is not what the naive block-level shift describes.

Sixteen nodes service more than one defendant.

Around the 107 seed offshore entities, ICIJ’s relationships file records 480 direct service-provider connections: 373 unique officers, 40 intermediaries, 39 addresses. Sixteen of these nodes are shared: they connect to entities belonging to two or more different FCPA defendants. The list is entirely legitimate professional infrastructure. Big Four accounting firms, Bermuda law firm addresses, established Bahamas and BVI corporate-services intermediaries. Two entries in an earlier working note on this section were interpretively singled out (one intermediary, two individual officers). Neither interpretation survives closer testing. Both are documented below with the reasons they do not carry forward.

The one flag worth naming, and why it does not survive closer testing

Two individual-officer observations, unverified

Two entries in the shared-infrastructure table are individual named persons: “Amissah, M Tonesan” (3 defendants: Baker Hughes, Ingersoll-Rand, Schering-Plough) and “Jones, Michael” (2 defendants: Schering-Plough, Tyco International). Both are surfaced without independent verification. The most likely explanation for a single individual appearing on the boards of subsidiaries of multiple unrelated multinationals is that the person is a corporate secretary at one of the shared Bermuda law firms and is named on client subsidiaries as part of routine corporate-services practice. This explanation is a hypothesis. Neither individual has been verified against Bermuda corporate-services personnel records. Given the Vassiliades corporate-identity failure above, absent verification, no interpretive weight should be placed on these two entries either. They are surfaced for follow-up, not as findings.

The remaining shared nodes are what they look like: Appleby’s Bermuda office as the registered address of subsidiaries of four FCPA defendants; PwC and Deloitte listed as officers of subsidiaries of two defendants each; Chancery House in Bridgetown (a major Barbadian law firm) shared across two. These are professional services with wide market share serving multinational clients. Their appearance across FCPA defendants proves market presence, not coordination.

No specialized post-conduct service network.

Louvain community detection on the seed-entity projection produces 14 communities of size 2 or more. Six of those span multiple defendants; the other eight are within-defendant clusters (groups of Ingersoll-Rand subsidiaries connected through shared corporate secretaries, groups of Baker Hughes subsidiaries connected through shared intermediaries, and so on). The multi-defendant communities are small and temporally mixed.

The absence of post-conduct-dominant communities is itself informative. If the observed post-conduct temporal elevation had reflected a specific set of service providers who specialize in post-investigation restructuring, we would expect those providers to form dense, post-conduct-heavy communities in the projection. That is not the pattern. Post-conduct entities distribute across the community structure roughly the same way pre-conduct entities do.

This observation is worth keeping distinct from the sensitivity-tested finding. The community structure is descriptive: it says the offshore service-provider ecosystem serving these FCPA-defendant corporations is not visibly organized around post-conduct-specialized clusters. This is true independently of whether the temporal elevation reaches Bonferroni significance and independently of whether the jurisdictional pattern is a coverage artifact. What the ecosystem looks like remains what it looks like.

What the sensitivity tests resolved, and what they did not.

The earlier draft of this report presented four candidate mechanisms for the observed pattern and ranked them by prior plausibility, with leak-source coverage timing at the top as the strongest confounder to rule out. The confounder tests have now been run. Mechanism 01 was confirmed as the primary driver of the jurisdictional finding. The other three mechanisms remain untested and cannot be evaluated on the current data.

The distinction between mechanisms matters because a naive reading of this report might infer a defendant-behavior story. The data do not support that inference. What the data support is a much narrower claim: within the ICIJ Offshore Leaks corpus, alignments to FCPA-defendant corporations show a directional post-conduct elevation of about 3 to 5 percentage points and a jurisdictional footprint that mostly mirrors which leaks contribute which alignments. Both facts are true. Neither warrants the causal story that the initial pattern seemed to suggest.

What was used, and how the noise was filtered out.

What was tested and what could still change the reading.

The two most damaging confounders have been tested. Both hit. Remaining limits either reflect study scope (what could be added in future work) or reflect items still on the pre-publication list.