Elite Leverage and Costly State Behaviour

Public discourse offers two explanations when a state does something costly and destructive: ideological capture — the leadership believes something and acts on it — and elite leverage — the leadership is compromised and acts against national interest for private reasons. The working thesis was that these are not competitors but different layers.

The hypothesis under test was narrower and novel:

Leverage over elites is manufactured by the gap between a society's publicly enforced moral code and its elites' actual private conduct.

The prediction: the wider that gap, the further a state's foreign policy sits from its own public's preferences, controlling for the obvious confounds.

The answer is no. That is the finding, and it was reported under a commitment made before any model was fitted.

Blackmail is never measured, by design

Successful blackmail leaves no record. Every observable case is a failed one. Measuring observed blackmail incidence would sample on failure of the very mechanism claimed to operate, inverting any finding through selection bias.

The project therefore measures the structural conditions that generate the supply of leverage and treats leverage itself as latent and unmeasured. This was a design constraint rather than a limitation to engineer around: any proposal to add a "blackmail cases" dataset was rejected on these grounds.

What was expected, recorded in the design document before any data was loaded: "Most likely: boring controls absorb most of the variance and our mechanisms add modest marginal signal. Build for a clean null, not for a dramatic finding." The result is the one the design anticipated.

The protocol

Eleven sources — V-Dem, UCDP, COW MID, Fox RAS, Polity5, WDI, GDELT, COW contiguity, COW alliances, ATOP alliances, UNGA ideal points — each acquired with hash verification and mapped to a canonical Gleditsch–Ward entity table. Nothing joins on a country name.

Three rules did the structural work:

• Missingness is never silently imputed. Four statuses per cell. A country-year before a state exists is not a zero.
• Train and holdout are physically separate stores. A year-partitioning router sends year >= 2010 to the holdout, so the split is enforced at the data layer rather than by a filter an analyst could forget.
• A five-look budget, enforced as a mechanism rather than a promise. The holdout store requires a registered evaluation name, validates it against a closed registry, and writes the access-log line before opening the file. A read cannot occur without leaving the record that proves it occurred.

The baseline was frozen before any mechanism specification and could not be revised afterwards. The criterion was declared: the 95% cluster-bootstrap interval on the paired metric difference must exclude zero. And the commitment was explicit — if the mechanisms do not beat the baseline, that is the finding, and it is reported. No post-hoc specification search to rescue it.

Results, in the declared order

Budget fully spent, 5 of 5. Each look opened the store under its registered name, in the declared order, and is logged by name. No unregistered name was attempted. Models were fitted on training and evaluated on the holdout; the holdout selected nothing.

On both primary outcomes the mechanisms do not beat the frozen baseline. On the secondary conflict operationalization they do, by the declared criterion.

One positive out of three declared outcome tests at the 95% level is what chance alone produces. That is a base rate — not a dismissal of look 3, and not a rescue of looks 1 and 2.

Look 3 reversed its training result. On training the same specification was reliably worse (−0.00140, CI [−0.00278, −0.00021]); on the holdout it is reliably better. Both intervals exclude zero, in opposite directions.

Why look 3 reversed

A diagnostic was run that spends no look: it opens the training store only and asserts the holdout access log is byte-identical at start and end. No reported result was adjusted.

Thin event counts are ruled out — look 3 is the best-powered of the three, with 212 positive events against 37 for look 1. The defect is cluster concentration. What governs whether a cluster bootstrap on a paired difference is trustworthy is not the count of country-years, nor the nominal cluster count, but how concentrated the difference is across clusters.

A bootstrap nominally resampling 167 countries is, for look 3, behaving like one resampling about 28. Its exclusion of zero was marginal to begin with — the lower bound sits at 15% of the point estimate — and a marginal exclusion does not survive an interval too narrow by this margin.

Why this does not touch the nulls. It is not that the primary outcomes escape the problem; UCDP is concentrated worse than MID. The distinction runs the other way: understated uncertainty only threatens a conclusion that rejects. Looks 1, 2, 4 and 5 fail to reject, and widening a too-narrow interval keeps zero inside it. The defect cannot have manufactured the four nulls. It can only have manufactured the single positive.

Both post-hoc choices survived

Two decisions were made after seeing training data, and both were carried to the holdout so they could be tested rather than asserted.

The elite/general enforcement restriction was not load-bearing. Elite and full specifications perform indistinguishably out of sample and neither beats the baseline. If anything the full specification — the one that failed the discriminant gate — did marginally better, which undercuts any reading in which the elite restriction was doing work.

The scope-threshold revision did not change conclusions. Null at the revised threshold and null at the original. The post-hoc revision did not manufacture the outcome.

Errors caught before reporting

A null is only credible if the failures caught along the way are visible. Each of these was found and fixed before the result it would have affected was reported. Several would have produced a confident wrong answer rather than a visible failure.

The Kiribati/Nauru code collision. Gleditsch–Ward code 970 is Kiribati. COW code 970 is Nauru. Under an identity crosswalk, every COW-keyed join attaches Nauru's data to Kiribati's rows and vice versa — no error, no null, no missing value to notice, just two countries holding each other's numbers for every variable, permanently. Caught by auditing the crosswalk against both source files, and now asserted by a test named for it.

Three silent country disappearances. GDELT's CAMEO actor codes retain legacy ISO spellings: ROM for Romania, WST for Samoa, TMP for East Timor. Without overrides these failed to join — not a wrong number, but thousands of events dropped and three countries absent from every model using the GDELT measure, with no error and no null to signal it.

The GDELT seasonal drift. The measure was first built from a single June window each year. Religious observances follow lunar calendars that drift against the Gregorian year, and drift asymmetrically across traditions — so a fixed window samples different points in different traditions' cycles in different years, producing a bias that trends. Corrected to four stratified quarterly windows. Two further traps surfaced during the rebuild: GDELT's daily era opens 1 April 2013, not 1 January, so that year's first window existed only as a monthly file and would have been silently lost.

The divergence-baseline residual error. The natural default when freezing the baseline was to give all three outcomes the same control set. For divergence that would have been wrong: the outcome is already the residual of UNGA position on material position. Re-entering those as controls regresses a residual on the variables it is orthogonal to by construction, producing a baseline that explains nothing and a mechanism comparison flattered by an artificially weak opponent. This is the one error on this list that would have manufactured a positive result. It was caught while the baseline was being declared and never reached committed code.

The clustered standard-error defect. Cluster-robust standard errors were initially computed with the OLS bread applied to logit residuals. This produced z-statistics near 80 on the mechanism terms — a result that would have read as overwhelming significance. Corrected to the proper weighted form, the same terms read z ≈ 1.5. Found during training iteration and fixed before any holdout look was spent.

What the null does and does not say

The finding is not that elite leverage does not exist, and not that ideological capture fails to drive state behaviour. Neither claim is testable with what was built, and neither is made.

What was tested, exactly: whether these state-level proxies, over 1990–2009 training and 2010–2014 holdout, add predictive information to material position and regime type in forecasting conflict onset or alignment divergence. On both primary outcomes, they do not.

Three limits carry the gap between that and the hypothesis. They are properties of the measures, not of the world.

1. The code-strictness measure is an indicator, not a gradient. Religious-legal establishment is rare and extreme, so the variable separates religious-law states from everyone else rather than ranking societies along a continuum. A null on it is a null on that binary distinction.
2. The enforcement measure is largely regime type by construction. Between half and two-thirds of the variable is the control it was asked to beat, so the test is close to asking a variable to beat itself.
3. Source coverage cost twenty years of panel. The declared window was 1970–2026; the moral-code source spans 1990–2014 and no mechanism specification can exceed it. The test ran on 20 training years and 5 holdout years instead of the ~57 available to the baseline.

These limits do not weaken the null within its scope; they define it. The result is reported as it came out, no specification was re-run to widen it, and the limits are stated so the finding cannot be cited for more than it supports — in either direction.

The project is finished. Its budget is spent and cannot be extended: the 2010–2014 holdout is permanently burned, five looks, logged by name, and any sixth read is an unregistered look whatever it tests.