The case for the occurence of algorithmic vote flipping

[RonRules]

As expected, Canadian charts for election 2011 flat-line.

I got the data here:
http://www.elections.ca/scripts/resval/ovr_41ge.asp?ddlEDRes_prov=35&lang=e
Province: Ontario, pollbureau10001 and pollbureau59036
(10001 was the first in the list, 59036 was the largest)
They provide .csv files, but the row/column format is opposite of what we need. I could modify the Java program to process those types of files, in day or two. I could then process ALL of Canada quickly.

I feel these charts are enough for now. Verdict: Canada flat-lines.

There is a slight slope for candidate "fuchs", but bear in mind that the largest precinct there only had 322 votes.

 

[drummergirl]

Where do you think I'm contradicting myself?

Elections follow the law of large numbers.

 

[drummergirl]

It's good to know the law of large numbers is alive and well in Canada

 

[RonRules]

I want to head to Hawaii, but I don't have any data. Can anybody track that down for me?

Thanks

 

[dsw]

I tried (by eye) to add a straight line from the origin to the endpoint for Fairfax. Correct me if I got that wrong but, when you see the overall trend, and when you take into account that the overall slope is because the number of votes is being plotted rather than percentage, is there anything particularly remarkable about the "crime" point?

To me what it looks like is that the "crime" point, is significantly lower (i.e., worse for Romney) than it's two neighbors, both of which are unremarkably consistent with the overall trend. So that you get a dip down to "crime" point, followed by going right back up to the trend. But I think what's going on here would be a lot clearer if it the y-axis were % of vote.

How much evidence do you need?

Since I have all the Virginia data in one spreadsheet, I decided to chart absentee ballots only throughout the state. Note at how Romney's line slope clearly climbs upwards at around 240 votes tallied per precincts. We've seen that magic 240 count mark in many other states/counties.

This is so clear to me.

For sake of making the chart easier to see, I omitted Fairfax county, which has a very large absentee ballot count. The missing data point is here:

FAIRFAX COUNTY:
Total: 1203
Ron Paul 379 31.5%
Romney 824 68.5%
Ratio Romney/Ron Paul 2.17

I ommitted that point from the chart because it causes most results on the chart to be on the left 1/4 of the page and leaves a lot of blank space up to Fairfax.

It certainly strengthen the case by just looking at the values above. Romney earns a whooping 68.5% of the vote in that absentee ballot county.

 

[RonRules]

You're doin't wrong.

You can't make an honest curve fit with two points (origin and Fairfax) when you have 119 points available.

What my graph showed was that after 112 precincts counted, a strange thing started to happen. Creeepy Romney started creeping way too fast.

The proper way to extend that line is to do a linear, least-squares fit on those first 112 points. See where the extrapolated line should be:

[/IMG]

 

[dsw]

Elections follow the law of large numbers.

Some things are correlated and some aren't. And some things that are correlated in some cases, aren't in others. There are MANY ways you could sort the precincts in order to have a graph that doesn't flatten out, you just need to pick any sorting variable that correlates with % votes. And when that happens you've got a case where the law of large numbers doesn't apply.

Suppose someone sorts the precincts by %republican for national presidential races and finds that the graphs do not flatten out (none of them!) and so they cite the law of large numbers and the fact that "elections follow the law of large numbers" to prove that every presidential election of our generation has been fraudulently manipulated because the odds of the graph not flattening out are infinitesimal. They're wrong (at least in their reasoning, if not in their conclusion) because there's no mathematical law that says their graph should flatten out. They've sorted the precincts in an order that introduces a correlation, then tried to apply a mathematical law that only applies if there is no correlation. The math doesn't say what they think it says, and their argument, or at least the mathematical portion of it, does not prove fraud.

 

[dsw]

You left out the other points, but the extrapolated line ends at (500,300), compared to the actual of about (500,320) there. Are you just looking at the slope from the "crime" point to the next one, even though that's just how it's going to look for the next point any time the preceding point was lower than average? There's an even sharper slope a few points back from there, around 130.

Show the rest of the points, and plot it as a percentage rather than the number of votes. Does it still look suspicious?

You're doin't wrong.

You can't make an honest curve fit with two points (origin and Fairfax) when you have 119 points available.

What my graph showed was that after 112 precincts counted, a strange thing started to happen. Creeepy Romney started creeping way too fast.

The proper way to extend that line is to do a linear, least-squares fit on those first 112 points. See where the extrapolated line should be:

[/IMG]

 

[RonRules]

You left out the other points, but the extrapolated line ends at (500,300), compared to the actual of about (500,320) there. Are you just looking at the slope from the "crime" point to the next one, even though that's just how it's going to look for the next point any time the preceding point was lower than average? There's an even sharper slope a few points back from there, around 130.

We suspect that funky things start happening at around 240 votes. We have seen that many times in the past. To confirm that effect, I did a linear least-squares fit using all the 112 precincts up to precinct vote tally ~=240. (the actual X-Axis number at the crime point is 234). Obviously, I should not include points beyond that if I want to make the case that Romney departs from that line at around the 240 point.

If I add the remaining numbers to the linear fit, I will mix in the non-flipped votes with the flipped votes and produce a meaningless line.

 

[arsenius]

I don't know how much should be read into Romney's absentee ballots. Here is an article on how his campaign tried to get as many as possible:

http://mobile.slate.com/articles/ne...mastery_his_rivals_never_stood_a_chance_.html

 

[drummergirl]

Suppose someone sorts the precincts by %republican for national presidential races and finds that the graphs do not flatten out (none of them!)

Actually, if you sort them that way, they do flatten out. Over and over, year after year, election after election. That is the point.

 

[drummergirl]

I tried (by eye) to add a straight line from the origin to the endpoint for Fairfax. Correct me if I got that wrong

We don't eyeball; we use linear regression (least squares).

 

[dsw]

We suspect that funky things start happening at around 240 votes. We have seen that many times in the past. To confirm that effect, I did a linear least-squares fit using all the 112 precincts up to precinct vote tally ~=240. (the actual X-Axis number at the crime point is 234). Obviously, I should not include points beyond that if I want to make the case that Romney departs from that line at around the 240 point.

If I add the remaining numbers to the linear fit, I will mix in the non-flipped votes with the flipped votes and produce a meaningless line.

Can you plot it with percentages though?

Doing the regression you've done begs the question of whether the later results are flipped. But okay, how about just plotting percentages and drawing a horizontal line at the overall percentage. If there's a bias with large precincts that will help make that clear visually. The way you've got it drawn just makes it hard to visually separate the effect of the number of votes increasing with the size of the precinct from any percentage adjustment that may have occurred.

 

[dsw]

Actually, if you sort them that way, they do flatten out. Over and over, year after year, election after election. That is the point.

To make it a better example of someone misunderstanding the math and suspecting fraud, let's suppose this person sorts by increasing % registered Democrat, and looks at McCain vs. Obama. Elections obey the law of large numbers, so it should flatten out, they think. But it doesn't! And they think it's proof of fraud.

Clearly it doesn't flatten out. Early on McCain on this graph is going to be absolutely *smearing* Obama. He's going to start out near 100% in precincts that have negligible Democrat registration. His cumulative percentage is going to be heading downward, though, and Obama's heading up. By the time we get closer to the right and start to hit Democratic strongholds where there are few or not Republicans, Obama's graph will still be climbing (and McCain falling) because Obama consistently has a higher percentage in such precincts than he does overall, so each time you add one of those precincts you increase the cumulative average. You hit the overall average only when you hit the right hand side of the graph.

If you sort the graph that way it won't flatten out. The math doesn't say it should flatten out. And the fact that it doesn't flatten out doesn't prove fraud.

 

[drummergirl]

I want to head to Hawaii, but I don't have any data. Can anybody track that down for me?

Thanks

It appears that since the hawaii caucus was actually a state GOP run primary election, the HRP is the only source of data. At the moment all they have published are the totals and totals by congressional district (not useful for our purposes).

I put in a call to the HRP, but they were not in (probably after 5 hawaii time).

 

[drummergirl]

To make it a better example of someone misunderstanding the math and suspecting fraud, let's suppose this person sorts by increasing % registered Democrat, and looks at McCain vs. Obama.

I'd love to see that chart. I think if you looked at a few of those, you might learn something interesting.

 

[RonRules]

Can you plot it with percentages though?

OK, next time you do you own charts, mmmkay?

Question for the gang: Why does Romney do so well in all those small absentee precincts?

Could it be related to the story that Arsenius posted?
http://mobile.slate.com/articles/ne...mastery_his_rivals_never_stood_a_chance_.html

It would make sense to spread out these absentee ballots a few at a time in all precincts, including the very small ones.

If the Romney team fraudulently submitted about 5-6 ballots in each precincts, the early very small precincts would exhibit this kind of early peaking followed by downard slope against an honest candidate that did not stuff the ballot box.

Here's a few where Romney was doing exceptionally well early on:

County Total Paul Romney
GALAX CITY 6 1 5
CHARLES CITY COUNTY 7 1 6
GRAYSON COUNTY 7 1 6
ESSEX COUNTY 9 1 8
NOTTOWAY COUNTY 9 4 5
LEXINGTON CITY 9 0 9
SMYTH COUNTY 10 3 7
ALLEGHANY COUNTY 11 1 10
GILES COUNTY 12 2 10
PRINCE EDWARD COUNTY 12 1 11
LEE COUNTY 13 3 10

 

[dsw]

I'd love to see that chart. I think if you looked at a few of those, you might learn something interesting.

Well, here's a quick-and-dirty approximation. I took the congressional district percentages from here:
http://www.swingstateproject.com/diary/4161/
I didn't have % registered Democrat for those districts, so I used something else that illustrates the same point. I sorted them by the %gore vote, which is probably a good approximation (not that it matters, any variable with a correlation will do the same). I didn't bother with district sizes, which would move some points around a bit, and I apologize for the lack of labels but this shows the (completely unsurprising) shape of the graph.

There's noise early on, as expected, then it smooths out. And it doesn't level off. None of this is in the slightest surprising.

And so if someone looked at this and said that math requires that this flatten out, they'd be wrong; if they said that the odds of it not flattening out are infinitesimal they'd be wrong; if they said the law of large numbers applies to the sampling when done in this order, they'd be wrong; if they said this proved fraud or proved anything other than the correlation we already knew about, they'd be wrong.

 

[dsw]

Thanks, I appreciate that. The "crime" point is a lot more subtle now, but it only looked more significant before because the way it was graphed was misleading. It looks more like a depression in the middle than a boost starting at "crime" and continuing to the right ... but you address that by hypothesizing more crime on the left hand side, too. Just so. (Why would it make sense to spread out those absentee ballots, i.e., to spread out completely legitimate absentee votes that demonstrate the success of a campaign operation Romney is proud of?)

OK, next time you do you own charts, mmmkay?

Question for the gang: Why does Romney do so well in all those small absentee precincts?

Could it be related to the story that Arsenius posted?
http://mobile.slate.com/articles/ne...mastery_his_rivals_never_stood_a_chance_.html

It would make sense to spread out these absentee ballots a few at a time in all precincts, including the very small ones.

If the Romney team fraudulently submitted about 5-6 ballots in each precincts, the early very small precincts would exhibit this kind of early peaking followed by downard slope against an honest candidate that did not stuff the ballot box.

Here's a few where Romney was doing exceptionally well early on:

County Total Paul Romney
GALAX CITY 6 1 5
CHARLES CITY COUNTY 7 1 6
GRAYSON COUNTY 7 1 6
ESSEX COUNTY 9 1 8
NOTTOWAY COUNTY 9 4 5
LEXINGTON CITY 9 0 9
SMYTH COUNTY 10 3 7
ALLEGHANY COUNTY 11 1 10
GILES COUNTY 12 2 10
PRINCE EDWARD COUNTY 12 1 11
LEE COUNTY 13 3 10

 

[affa]

I tried (by eye) to add a straight line from the origin to the endpoint for Fairfax. Correct me if I got that wrong but, when you see the overall trend, and when you take into account that the overall slope is because the number of votes is being plotted rather than percentage, is there anything particularly remarkable about the "crime" point?

You 'tried by eye' and drew a line clearly mapping -above- all but one of the highest possible points on the left side of the graph? C'mon.