By Warren D. Smith, 8 Nov. 2014. warren.wds AT gmail.com This is a simpler shorter version of a larger report. There also is a questions & answers page.
A. The 9 close US senate (top) and 17 close state-governor (bottom) races in 2014, ordered by decreasing helping-Republican-signed discrepancy between official election margin, versus pre-election predictions. All discrepancy numbers (as % of #voters) taken from RealClearPolitics.com poll averages. Which races were "close" decided by Cook Political Report pre-election predictions ("toss-up" and "lean" categories). States colored pink were the ones with Republican Secretaries of State.
B. The 9 close US senate (top) and 17 close state-governor (bottom) races in 2014, ordered by decreasing absolute discrepancy between official election margin, versus pre-election predictions. States colored pink had year-2009 ratings of "inadequate" for their post-election audit procedures according to VerifiedVoting.org. (25 of the 50 US states had "inadequate" rating, usually meaning no post-election audit, not even a pretense.)
C. Some selected US senate elections with especially large discrepancies (as % of #voters) between official margins versus predictions from pre-election poll averages. ("+" sign means the discrepancy helped the Republican.) The |discrepancies| also are written as numbers of standard errors, for example for Kansas 9.9 or 4.38 standard errors under an optimistic and highly-pessimistic error-estimation method. The likelihoods of |discrepancies| at least this great, under a normal distribution model for polling errors, are given in the last line.
In A and B the pink states tend to lie mainly on the left of the ordering, to degrees hard to attribute to chance. To quantify that: if the party controlling the state truly were unrelated to the discrepancy between official results & polls, then the probability that the "mean leftwardnesses" in tables A, both would be at least this great, would be 181/58344≈0.0031 (exact result shown by computer exhaustive search). If the post-election audit quality for that state truly were unrelated to the |discrepancy|, then the probability that the "mean leftwardnesses" in tables B, both would be at least this great, would be 2867/3063060≈0.00094. Those are small. That tells us this was not a natural statistical fluke. And for various individual states such as in C (not necessarily all of them, but certainly some), it again is highly implausible that polls would be off by this much just by bad luck. Something more than bad luck is at work.
Conclusions: The 2014 midterm elections were manipulated to favor the Republican side. Our nonparametric (i.e. pretty much model-independent) tests find in close US senate & state governor elections
Finally the tests C which do depend on modelling assumptions (specifically "normal error distribution" of polling errors) find enormous confidences that the state elections with some of the most extreme discrepancies versus polls, were manipulated. In the 6 US senate races on the left side of table (C), these confidence levels range between 99.3% and 99.999999999999999999998% depending which state & election, and which error-estimation method is used for it. The ABC combined confidence-of-wrongness seems too large to be disputable.
Return to main page