Statistical Anomalies in Republican Primary Election Results (2012) [pdf](electiondefensealliance.org)
electiondefensealliance.org
Statistical Anomalies in Republican Primary Election Results (2012) [pdf]
http://electiondefensealliance.org/files/PrimaryElectionResultsAmazingStatisticalAnomalies_V2.1.pdf
6 comments
Some of these concerns are addressed on pages 19-21.
I find the hand-waving around the Maryland and Utah graphs to be great examples of the their motivated reasoning, and total blindness to the effect of conscious campaigning. (Appropriately, 'campaign' is another word missing from their 38 pages.)
Romney easily won Maryland's 2012 primary – 49% overall, presumably with the 'big precinct' advantage they find suspicious. (But, the researchers don't show us that graph or include any Maryland data in their tables.) But because the Maryland-2008 graph is 'flat' for the "same candidate… same state… same type of election", we are to believe all non-fraud density-driven vote-share effects are refuted.
But 2008 was a wildly different election! Other candidates were contending on the same axes of fundraising, organization, and compromise/moderation that Romney dominated in 2012. Another candidate (McCain) won the overall nomination, but looking at the Maryland graph, it appears both Romney and McCain barely even contested Maryland: both finished at about 6%, less than 1/5th the winner.
Do they show us McCain's or Obama's or Kerry's or Gore's or Dole's or Perot's or Clinton's slopes in any state or election? No. They just pick one small, relatively urban, and electorally irrelevant (for Republicans) state, Maryland, in one year, 2008, and imply its flatness means something.
The same problem applies to their choice of Utah-2012 as another flatness example: Romney won with 93%! No one else spent a cent there! How does that reflect anything about the effects of density and campaign expenditures in actively-contested races?
This kind of cherry-picking of outliers makes their case, especially with regard to the density-issue, look extra-flimsy.
Romney easily won Maryland's 2012 primary – 49% overall, presumably with the 'big precinct' advantage they find suspicious. (But, the researchers don't show us that graph or include any Maryland data in their tables.) But because the Maryland-2008 graph is 'flat' for the "same candidate… same state… same type of election", we are to believe all non-fraud density-driven vote-share effects are refuted.
But 2008 was a wildly different election! Other candidates were contending on the same axes of fundraising, organization, and compromise/moderation that Romney dominated in 2012. Another candidate (McCain) won the overall nomination, but looking at the Maryland graph, it appears both Romney and McCain barely even contested Maryland: both finished at about 6%, less than 1/5th the winner.
Do they show us McCain's or Obama's or Kerry's or Gore's or Dole's or Perot's or Clinton's slopes in any state or election? No. They just pick one small, relatively urban, and electorally irrelevant (for Republicans) state, Maryland, in one year, 2008, and imply its flatness means something.
The same problem applies to their choice of Utah-2012 as another flatness example: Romney won with 93%! No one else spent a cent there! How does that reflect anything about the effects of density and campaign expenditures in actively-contested races?
This kind of cherry-picking of outliers makes their case, especially with regard to the density-issue, look extra-flimsy.
Without historical context this analysis is meaningless. Is this one election the exception or the norm?
> Without historical context this analysis is meaningless. Is this one election the exception or the norm?
Why is it meaningless? It is still interesting and important to look for fraud during a single, particular election. Especially, one could argue, in the most recent one, since that presumably means bad-actors are more likely to be active in the next presidental election. I agree that it is interesting to see if voting fraud is common or widespread in primary elections in general, but studying individual events is not useless.
It occurred to me that you may have meant to say that the analysis is not robust without historical comparisons to establish that such anomalies are not just statistical fluctuations? That would something worthwhile to explore. It seemed (from very quickly skimming the pdf) that they did try various statistical tests to ensure their results were robust. But comparison with other datasets would likely provide a good test of the robustness of their findings.
Why is it meaningless? It is still interesting and important to look for fraud during a single, particular election. Especially, one could argue, in the most recent one, since that presumably means bad-actors are more likely to be active in the next presidental election. I agree that it is interesting to see if voting fraud is common or widespread in primary elections in general, but studying individual events is not useless.
It occurred to me that you may have meant to say that the analysis is not robust without historical comparisons to establish that such anomalies are not just statistical fluctuations? That would something worthwhile to explore. It seemed (from very quickly skimming the pdf) that they did try various statistical tests to ensure their results were robust. But comparison with other datasets would likely provide a good test of the robustness of their findings.
Would you run a stock-picker algorithm that never did backtesting? Real data-science is much more complex than just applying "this should work" reasoning and jumping to a really controversial conclusion with zero data to back it up.
This is a good hypothesis as it stands, and may be valuable as a predictor of fraud. However the conclusion is meaningless without sufficient evidence of predictivity to back it up.
Take historical examples of elections with known fraud and run this analysis to see if it can correctly label the fraud/no-fraud categories. _Then_ after we can agree that it works well, only then will I care about the results of any specific analysis.
This is a good hypothesis as it stands, and may be valuable as a predictor of fraud. However the conclusion is meaningless without sufficient evidence of predictivity to back it up.
Take historical examples of elections with known fraud and run this analysis to see if it can correctly label the fraud/no-fraud categories. _Then_ after we can agree that it works well, only then will I care about the results of any specific analysis.
Historical results would be a better "control" against the possibility that this is the result of something other than systematic fraud.
This is a general problem with statistical analysis: all you can do is whittle down the space of possible explanations, you can never "prove" any one of them. I don't mean to nitpick about highly improbable statistical fluctuations. You can always make a "straw man" null hypothesis so shitty that it's trivially rejected (p<0.05 whee) and then claim that this supports your crazy alternative hypothesis. A robust review process should reject this kind of "p-hacking."
Controls are one of the best tools for cutting down the space of potential alternative explanations and avoiding p-hacking (intentional or not). It's not unreasonable to ask for one here. I haven't read deeply enough to know if the existing controls are convincing or not but the opinion "they aren't convincing" isn't silly.
This is a general problem with statistical analysis: all you can do is whittle down the space of possible explanations, you can never "prove" any one of them. I don't mean to nitpick about highly improbable statistical fluctuations. You can always make a "straw man" null hypothesis so shitty that it's trivially rejected (p<0.05 whee) and then claim that this supports your crazy alternative hypothesis. A robust review process should reject this kind of "p-hacking."
Controls are one of the best tools for cutting down the space of potential alternative explanations and avoiding p-hacking (intentional or not). It's not unreasonable to ask for one here. I haven't read deeply enough to know if the existing controls are convincing or not but the opinion "they aren't convincing" isn't silly.
This statistical anomaly was also noted by former NSA analyst Michael Duniho: http://www.globalresearch.ca/why-is-mitt-romney-so-confident...
I am personally shocked by this. Why didn't this make more noise in 2012? Why am I only hearing about this now?
Since voting fraud seems to take place in larger precincts, why not split them into smaller precincts that are never bigger than a certain size? This would make the life of fraudsters harder since they would have to commit fraud in more precincts instead of a few big ones.
I am personally shocked by this. Why didn't this make more noise in 2012? Why am I only hearing about this now?
Since voting fraud seems to take place in larger precincts, why not split them into smaller precincts that are never bigger than a certain size? This would make the life of fraudsters harder since they would have to commit fraud in more precincts instead of a few big ones.
This was alleged relatively loudly by the libertarian wing of the GOP. While no one expected Ron Paul to win the primaries, most of us did expect him to do significantly better than he did. Further, I recall instances in very small precincts where more people openly claimed to have voted for Paul than were counted.
Coupled with the reported suppression of Paul's delegates to the national convention, there seems to be strong evidence of an active conspiracy within the GOP to resist a perceived ideological shift in the party. Such a shift occurred last in the wake Pat Robertson's failed bid for the GOP nomination for the presidency in 1988.
Coupled with the reported suppression of Paul's delegates to the national convention, there seems to be strong evidence of an active conspiracy within the GOP to resist a perceived ideological shift in the party. Such a shift occurred last in the wake Pat Robertson's failed bid for the GOP nomination for the presidency in 1988.
If you'd asked me before seeing this thread whether Romney did better in the primaries in areas with high turnout or lower turnout, I'd have strongly guessed "higher", because it is widely believed that dogmatic conservatives -- and also dogmatic libertarians -- are more passionate about elections than typical Romney supporters.
Similarly, when Democrats benefit from higher turnout in general elections, the reason may not be that evil ward captains are stuffing the ballot boxes.
Similarly, when Democrats benefit from higher turnout in general elections, the reason may not be that evil ward captains are stuffing the ballot boxes.
The paper doesn't explain how this may have been accomplished. That would be a very interesting analysis. Since this was a primary, is it possible that the Republican party favored Romney as a candidate and thus modified vote counts to his favor? I'm curious how this would have logistically taken place.
Sooo, this is saying that Mitt Romney probably committed election fraud throughout the 2012 primary?
The paper offers VERY strong statistical evidence of vote fraud of a specific type occurred in Romney's favor in the 2012 primary. It seems that almost without a doubt, wide spread vote-flipping (the type of fraud analyzed in this paper) was in play and had large impacts on the primary, on the order of hundreds of thousands of votes and to the extent of causing the outcomes to differ (meaning the fraud caused Romney to win in places he maybe wouldn't have). The fraudelnt votes have compounding effects on the election that go beyond the flipped votes because of momentum loss effects on candidates election campaigns who were losing (but shouldn't have). Votes were being flipped from Santorum and Ron Paul to Mitt Romney. To a lesser extent votes were also flipped from Newt Gingrich. Other types of fraud were considered beyond the scope of the paper, such as ballot injection or vote removal.
What's more likely: a nationwide and unrevealed conspiracy to flip votes in favor of Romney, or that Republican voters in larger/more populous precincts favor Romney more often than voters in smaller precincts?
Edit for clarity: by larger/more populous I mean the x-axis on the graphs in the paper, though I forgot that this is only Republican voters and Republican/Democrat split in the precinct is also a factor. The x-axis doesn't necessarily mean rural to urban.
Edit for clarity: by larger/more populous I mean the x-axis on the graphs in the paper, though I forgot that this is only Republican voters and Republican/Democrat split in the precinct is also a factor. The x-axis doesn't necessarily mean rural to urban.
FTFA:
> Some argue that more liberal candidates are more popular in urban than in the rural precincts. To verify this hypothesis, we drew geographically random samples of precincts and computed partial correlations to filter out the population density factor. We will demonstrate in this paper that this factor has no impact on our conclusions.
> Some argue that more liberal candidates are more popular in urban than in the rural precincts. To verify this hypothesis, we drew geographically random samples of precincts and computed partial correlations to filter out the population density factor. We will demonstrate in this paper that this factor has no impact on our conclusions.
Yes, I saw that. The authors ruled out one possible explanation (urban vs rural).
Perhaps it's due to gerrymandered districts, republican vs democratic districts, large farm districts vs small towns, small towns vs large cities, or some other factor we can't think of in a few minutes.
The authors should rule out a lot more possible explanations before alleging fraud.
Perhaps it's due to gerrymandered districts, republican vs democratic districts, large farm districts vs small towns, small towns vs large cities, or some other factor we can't think of in a few minutes.
The authors should rule out a lot more possible explanations before alleging fraud.
They claim to address that, specifically:
Some argue that more liberal candidates are more popular in urban than in the rural precincts. To verify this hypothesis, we drew geographically random samples of precincts and computed partial correlations to filter out the population density factor. We will demonstrate in this paper that this factor has no impact on our conclusions.
I'll have to look at this in more detail when I have time.
Note also that just a few races have to be flipped, "momentum" is really critical. For instance, if they're correct and Romney didn't win New Hampshire, it's very easy to imagine him never getting decisive traction no matter how much money he threw into the campaign (one analysis I recently read noted he campaigned for dollars rather than votes, which of course was telling in the end).
Some argue that more liberal candidates are more popular in urban than in the rural precincts. To verify this hypothesis, we drew geographically random samples of precincts and computed partial correlations to filter out the population density factor. We will demonstrate in this paper that this factor has no impact on our conclusions.
I'll have to look at this in more detail when I have time.
Note also that just a few races have to be flipped, "momentum" is really critical. For instance, if they're correct and Romney didn't win New Hampshire, it's very easy to imagine him never getting decisive traction no matter how much money he threw into the campaign (one analysis I recently read noted he campaigned for dollars rather than votes, which of course was telling in the end).
The paper refutes the argument that Romney was more popular in urban areas (see page 20 and following).
I don't know. I remember hearing about this in 2012, but never whether it was accepted or rejected by anyone independent of these two authors, such as Andrew Gelman.
Googling now, all I seem able to find is various media articles and no independent statistical responses. The only followup seems to be http://www.kansas.com/news/politics-government/article171398... but, in a version of 'proving too much', that raises more questions for me than it answers by indicating that they've found the phenomenon to be much more general than 'Republican primaries in one year' (for starters, rereading the paper, I'm not sure I buy the foundational claim: why would there be any trend of vote bias with precinct, why wouldn't these supposed biases be simply implemented as a randomized flip on each vote, which would rig the election without leaving any noticeable sign? And think of the logistics: how can all these states and races have been rigged - what, did they all buy the exact same model of Diebold voting machines back in 2007?) This sounds like some sort of demographic or statistical artifact; they say they control for the obvious demographic and precinct-level issues, but of course, you never measure or correctly model everything.
Googling now, all I seem able to find is various media articles and no independent statistical responses. The only followup seems to be http://www.kansas.com/news/politics-government/article171398... but, in a version of 'proving too much', that raises more questions for me than it answers by indicating that they've found the phenomenon to be much more general than 'Republican primaries in one year' (for starters, rereading the paper, I'm not sure I buy the foundational claim: why would there be any trend of vote bias with precinct, why wouldn't these supposed biases be simply implemented as a randomized flip on each vote, which would rig the election without leaving any noticeable sign? And think of the logistics: how can all these states and races have been rigged - what, did they all buy the exact same model of Diebold voting machines back in 2007?) This sounds like some sort of demographic or statistical artifact; they say they control for the obvious demographic and precinct-level issues, but of course, you never measure or correctly model everything.
What if the candidate with an overwhelming fundraising/spending advantage (Romney) focused his resources in larger precincts?
The authors appear to have a pet theory ("larger margin-of-victory in larger precincts suggests fraud") and desired result ("fraud for Romney") and deploy all their reasoning powers towards that end.
Rather than just running lots of Romney share-vs-precinct-size slopes, what are the slopes for other candidates (of both parties), in other elections (including non-Presidential)?
Do Romney's slopes still stick out? Or do all "well-funded compromise frontrunners" show the same slope, where insurgents and protest candidates overperform in small precincts, farther from the influence of professional campaign/media operations?
Notable, also, is that Romney has the 'suspicious' slope even in states where he came in 3rd (Oklahoma, Alabama), or a distant 2nd (Louisiana). The slope doesn't even serve to improve his rank-order in the final results. What's the plausible mechanism, or rational risk/reward calculation, for engaging in fraud in places where your supporters are ~25% or less of the local party, and the fraud has a negligible impact on your count of nomination delegates?