Saturday, November 7, 2020

Evidence Suggesting Voter Fraud in Milwaukee

 I posted a version of this on twitter, but a) the writing format there is so ugly, and b) who knows how long that thread might last. So here it is for the record. 

I’ve been looking at the vote counts within Milwaukee, and there’s suspicious patterns in the data that need explaining. Proving fraud is difficult, but there’s a lot of irregularities here that point in that direction. First, the tl;dr, then the main analysis.

1. Democrat votes started increasing massively relative to Republicans after Tuesday night counts. This can’t be accounted for by explanations like heavily Democratic wards reporting later. When we look at the changes *within wards*, 96.6% of them favored the Democrats.

2. Democrats also improved massively against third party candidates, whereas Republicans and third party candidates showed similar changes to each other. Since there’s little incentive to manipulate third party counts, this implies that the big change after Tuesday night is in Democrat votes, not in Republican ones.  

3. When we compare different down ballot races, we find that Democrat increases within each ward were larger in races where the Democrat candidate was initially behind in the overall race on Tuesday night – that is, relatively more Democrat votes appeared in races where they were more likely to alter the outcome.

4. This result is easy to explain by fraud, but is much more complicated to explain by other explanations like Democrats mostly voting by mail. Most such theories predict all Democrat candidates should benefit in equal proportions within a ward, not that more votes come in exactly where they’re most needed.  

Ward-level vote counts are from the Milwaukee County Clerk at 7pm last night  and the archived version from the count as it stood on election night . 

This idea came from Spotted Toad, who’s been doing great work on this too. I’m looking at Presidential, Congress, State Senate and Assembly races. One way to look at what happened is to compare the percentage increase in votes for Republican Candidates versus Democrat candidates within each ward after election night.

For instance, suppose the Democrat candidate vote total went up 200% from initial counting to Thursday night. How much did the Republican vote total go up? If the distribution of votes before and after is the same, the percentage gains for each group should be similar, regardless of who was ahead.

This is very different from the normal reason where candidate totals in the entire state might change as counting goes on, as different reports come in from other parts of the city. That just shows that wards differ from each other. Rather, we’re testing whether the *same ward * should continue to find the same distribution of votes before and after Tuesday night. 

In other words, if the before and after distributions were the same, as votes come from the same pool, you’d expect that half the time, the Republicans got a slightly unlucky draw in the early votes, and end up improving their position (regardless of whether they ultimately win or lose). And roughly half the time, the Democrats should increase their votes by more. 

What actually happens? The Democrat candidate vote increases relative to the Republican candidate a crazy fraction of the time. The variable in question is percentage increase in Democrat vote totals for that ward (that is, the percentage change from Tuesday night to Thursday night), minus percentage increase in Republican vote totals. 

So a value above zero means that Democrat totals went up more than Republicans in that ward/race. A value of 500 means that the Democrats went up 500% in excess of the republicans (e.g. D votes grew 600%, R votes grew 100%). Here’s a graph of the histogram. 

You see an enormously right skewed distribution –tons of large gains for Democrats, very few gains for Republicans. Not only do Democrats very often increase more than Republicans, but when they do, it’s often by a colossal amount. 

Out of the 1217 ward/race combinations with non-missing early votes for both parties, 1037 saw relative increases for the Democrats, 37 saw relative increases for Republicans, and 143 were ties. Excluding the ties, the D “win” fraction here is 96.6%.  A remarkable feat!

Depending on how you assign ties, if this were a 50/50 coin (i.e. D and R were equally likely to gain relative to the other), the probability or p-value for this is between 10^-147 and a number Excel just lists as “0”.

So, this proves incontrovertibly that something about the count skews crazily towards the Democrats after 2am Wednesday. But it doesn’t prove what it is. Maybe they counted different types of ballots or something, but only starting at 4am. 

However, there’s one thing we can test – from which party’s votes is the weirdness coming from? We can answer things by looking at vote changes for other candidates – third party races, write-in candidates etc. 

We can be virtually certain that nobody is bothering to manipulate the vote totals for fringe, no-hope write-in candidates. These form a great placebo group – what might you expect the changes to look like for a group where nobody is manipulating the totals?

So let’s do the same thing as the earlier graph, but compare each part with “Miscellaneous”, which because the count is small, I aggregate together. I also limit the sample here to cases where there’s at least 5 votes for “Misc” in that ward by 2am Wednesday, to make sure that this isn’t coming from rounding (e.g. if you have only 1 vote, the minimum increase is 100%). 

What are we predicting to find? Well, if it’s the Democrat total that’s being wildly inflated, Democrats should also be increasing relative to Miscellaneous. Meanwhile, if Republicans are just being counted as normal, then their changes should look similar to the Miscellaneous Group.

And that’s basically what we find. First, Democrats vs Miscellaneous. Visually, the picture looks even more crazily skewed than the previous one. In terms of counts, Democrats improve relative to Miscellaneous in 520 ward/race observations. They tie 89 times, and Miscellaneous improves in relative terms just 3 times. That’s not a typo.


This corresponds to p-values between 10^-73 and 10^-177. The fraction of Democratic “wins” here (520/523), excluding ties, is a ludicrous 99.4%. 

So how do Republicans compare with Miscellaneous? It turns out that while they’re not exactly the same, they’re far, far more similar to each other than either is to the Democrats . Other than a few outliers (because “Miscellaneous” has very few votes in total, remember), the distribution is fairly symmetric around zero. 


In terms of counts, Republicans improve relative to Miscellaneous 179 times, Miscellaneous improves 251 times, and there are 74 ties. As a result, which p-value you get here depends enormously on how you allocate the ties. Give them to M, and it’s 10^-11. Give them to R, and it’s 0.55, or almost exactly chance (253 vs 251). 

Excluding ties, the R “win” percentage is 41.6%. So under some measures, they look slightly worse, but this ends up being affected by questions of rounding and the small vote totals for M. What’s incontrovertible is that D looks wildly, wildly different from either of them.

This is exactly what the null would predict, if votes before look like votes after. So this *does* roughly hold, but only when comparing Republicans vs Miscellaneous. This story is also inconsistent with the driver being something Trump did, like telling all his supporters to vote in-person. If so, why do changes in Miscellaneous votes look about the same? The important difference after Tuesday night, whatever you think it is, is coming on the Democrat side.

So maybe you’re wondering – are there reasons other than fraud that the ballots might be different before and after? If the ordering is random and they’re drawn from the same pool, no. But if each ward counts different types in a different order (those at 9am versus 4pm, or in-person versus mail-in), then this could happen. 

Whatever is making the vote distributions different before and after, it’s a factor that’s overwhelmingly just impacting Democrats, not Republicans. If you think it’s about in-person versus postal voting, you have to hypothesize that Republicans look kind of similar to Miscellaneous in this respect. This is possible, but not nearly as obvious. 

But there’s another more important aspect we can test here. In particular, if some of these Democrat increases are due to fraud, we would expect that the increases should be larger *when the fraud is more likely to impact the race. And since these include lots of down-ballot races like State Assembly Representatives, we have quite a lot of variation here. 

Sometimes the Democrat is way up after early counting, at which point it doesn’t matter much if they post big relative gains after that. But if the Democratic candidate is down early on, jacking up the total becomes much more important. I’m assuming that if the Party wants to rig votes, they’d also like to win as many races as possible for the least amount of rigging.

In other words, the comparison is now between two different races at the same ward. A Democrat voter comes to the ballot box or mailbox, and sees a number of races. For some, like President, it’s going to be a close call. For others, it might be a heavy favorite for the Democrat. 

The voter is a Democrat, so presumably he’s inclined to vote Democrat for both. We can compare within a given ward which of the two races showed bigger improvement for the Democrats in that particular ward after Tuesday night. 

Sure enough, the increase in Democrats relative to Republicans (the variable in our first histogram) is significantly higher when the Democratic race-wide vote share is lower during the early counting. In other words, within each ward, late vote counts break more heavily to Democrat in exactly those races where the change in votes is likely to affect the result.


How big is this effect? Well, one way to measure it is to see how many races it impacted. There were 8 races where Republicans were ahead on a two-party basis on Wednesday morning. By Thursday night, half of them had flipped to Democratic. By contrast, there were 19 races where the Democrat was ahead, and not a single one flipped to the Republicans. 

And again, let’s recall what we’re observing here. It’s not that the races flipped because suddenly wards that were known to be heavy Democrat strongholds started reporting in. Rather, more votes started coming in for Democrats relative to the ratio that was coming in for that exact same ward the previous night. Moreover, within each ward, the votes also skewed more for races that the Democrats looked like they might lose. 

Importantly, this finding is surprisingly hard to explain with the commonly cited reasons for Democrats pulling ahead overall. For instance, one of the claims is that mail-in ballots are counted late, and these are more heavily Democrat. In general, this doesn’t explain why within the same ward, some races later skew Democrat more than others.

The key part is that for each voter, the decision to take a mail-in ballot is common to all races. In other words, a single voter can’t vote for some races by mail, and others in person. So if your claim is that the overall skew to Democrats is a mail ballot effect, most versions of this explanation predict that all races should be equally affected.

To simplify the logic, consider a stylized example where all Democrats and Republicans vote straight ticket. More Democrats vote by mail, and these are counted late.  This would predict that Democrats overall would improve, but the expected improvement is the same for all races, regardless of whether the Democrat is ahead or behind. 

More ballots come in Democratic, they each vote for every Democrat, so all Democrats increase in the same percentage terms. This isn’t what we find. In the data, within a ward, the important races go up more than the unimportant races.

And this prediction, that all races should be equally affected, holds for a lot of other variations too. Does the answer change if every Democrat voter has a 90% chance of voting for each Democratic candidate, if this attitude is the same between Democrats who vote in-person versus those who vote by mail? No. The increase should be the same in all races.

The answer doesn’t even change if Democrat voters in general can’t be bothered as much voting for shoo-in candidates, and only cast their votes for tight races. As long as this instinct is the same in Democrats who vote by mail and those who vote in person, there should be no difference across races in how much they break late towards Democrats.

What you need is something complicated. Democrat voters can’t be bothered voting for candidates they like but who they know are going to win anyway, AND this instinct is somehow larger in Democrat voters who vote by mail than those who vote in person, AND there has to be a larger share of mail voting by Democrats overall. 

This may sound like a confusing and complicated explanation. And it is! That’s kind of the point. We’re now a long way from the simple explanation that Democrats vote more by mail. It’s not impossible, of course, and we can’t rule it out. There are other variants on this story, but if you think this is all about mail-in ballots, there has to be some difference *within Democrat voters* who vote by mail versus in person.

In other words, the bare fact is that races swung much more towards Democrats exactly for those races where the Democrats were down on Wednesday early morning. To explain this with mail-in ballots needs a very complicated story. To explain it with fraud needs a very simple story – you commit fraud more where the fraud matters more. 

This is why the evidence suggests fraud to me, but your mileage may vary here. I’ve tried very much to stick to the facts, because I don’t have any special ability to interpret the numbers above. Whatever is going in is crying out for explanation, and the simple alternatives don’t do it. To me, it looks pretty suspicious. 

A final question worth pondering. What should our null hypothesis be here? When we say “there’s no evidence of it”, we’re claiming “no fraud” as the null hypothesis. But as I’ve argued (by metaphor), the system of vote counting is so rickety and broken that this is an incredibly difficult null to justify. 

A metaphor for the likelihood of voter fraud, for people who insist that it's a conspiracy theory, or there's no evidence of it.

Suppose Amazon wanted to know how many packages it had. Packages were kept in warehouses all over the country. The system was different in every warehouse.

Some people need to move packages around, and there's a list of who is allowed to do that in each warehouse. But if you go in and say you're that person, nobody checks. If someone else has already done that for you when you arrive, you just get another package.

Some packages get driven around by people in their own cars, some get moved around by the post office, some by volunteers or low paid government employees, and in each case they're largely unmonitored - there's no clear record of which ones left or arrived.

Packages are, by common consent, valuable for people to take. But nobody investigates closely what happens in each place, and very rarely are package thieves caught.

For what package system other than "votes" would this be considered a reliable and acceptable system?

For what important corporate outcome, if you proposed this setup as a manager, would you not be fired?

If someone told you there was no evidence of package fraud, how plausible would that claim be?

I find the possibility of voter fraud entirely plausible, and that belief has nothing to do which party you think is doing it. At a minimum, I feel strongly that this possibility needs to be investigated more seriously than it is, given the evidence above.