Bayesian Assessment
Priors stated before the data, posteriors after, and an explicit account of how much moved and why.
Why bother
Without a stated prior, any result can be presented as confirming. Writing priors down first makes it possible to check whether belief actually tracked evidence, or merely accompanied it.
Priors and posteriors
| Proposition | Prior | Posterior | Driver |
|---|---|---|---|
| Political finance is highly concentrated | 0.95 | 0.99 | Directly measured; little room to move |
| Most money is candidate-uncontrolled | 0.70 | 0.97 | 8.25:1 ratio exceeded expectation |
| Funding predicts roll-call voting | 0.65 | 0.15 | Large update against. H2 refuted |
| Money shapes outcomes via agenda control | 0.60 | 0.25 | Updated 10 Aug. H5 tested on 10,564 bills and refuted |
| IOCE transfers usefully from firms to states | 0.55 | 0.35 | Both behavioural channels failed; structural claim survives |
The one large update
Belief that funding predicts voting fell from 0.65 to 0.15. The likelihood ratio is unfavourable: if money strongly drove floor votes, correlations of 0.20–0.27 within party would be surprising. If it did not, they are close to what incidental association would produce — PACs give to members already disposed to agree with them.
The weak positive correlation is at least as consistent with money following positions as with money producing them. PACs allocate to reliable allies, so a modest association is predicted by the null. The observation does not discriminate between the two directions, and nothing in this report identifies causation.
The prior that was protected, then tested
An earlier version of this page held agenda control at 0.60, unchanged, on the grounds that no evidence bearing on it had been collected — and warned that raising it because H2 failed would be the most likely reasoning error in the project. That protection was the right call, and it has now been resolved the only way it should be: by getting the data.
The govinfo bill-status corpus for the 118th House — 10,564 bills, sponsor-matched — allowed H5 to be tested directly. Funding share correlates with bill advancement at r = 0.185 (Republican) and 0.117 (Democrat); correlation with total money raised is negative. The prior falls from 0.60 to 0.25.
Vote-switching and agenda advancement were the two observable routes from concentrated money to concentrated control. Both have now been tested against complete-population data and neither survives. The honest position is no longer "we have not looked" — it is that the two most plausible transmission mechanisms are not there in the places we can see.
What remains unmeasured is narrower and harder: bill content as drafted, committee chair power, leadership scheduling, and candidate selection before anyone files. Those are hypotheses awaiting data, and this page will not raise a prior for any of them until that data exists.
What would move these numbers
- Toward capture: bill-introduction data showing donor-aligned provisions entering text without floor debate; committee assignments predicted by fundraising; roll-call effects concentrated on low-salience votes where party discipline is weak.
- Against capture: OCDI-Gov scores near zero once geography is measured; legislative outputs tracking district opinion irrespective of funding.
Both are specified in advance so that neither can be treated as unexpected after the fact.