Thread by @shylockh

A 🧵 on alarming changes in Pennsylvania voter rolls between October 14 and 21. Highlights:

-12k duplicate observations, potentially allowing multiple ballots by the same person

-10% of new records (6.5k) suspiciously backdated to list registration dates years in the past

1/N

Suspiciously dated records are more likely to:
-Be missing house numbers
-To have registered on Jan 1st
-To have already voted.

They could only be identified by comparing both snapshots. They are not explained by bad updates of old data or inactive voters becoming active.

2/N

Overall, I identify 18,456 highly suspicious new entries in the recent update.

It is difficult to explain all these with simple database errors, and it raises significant questions about the introduction of fake records.

3/N

This is in addition to the previously identified suspicious entries in Pennsylvania voter data

-440 voters with impossible ages
-1 million voters lacking a house number
-252 voters listing only a post office box as their address



4/N

https://x.com/shylockh/status/1849871737254498358

-7,056 former PA undergrads still registered to vote their dorm after age 24
-14 PA voters with surnames that are blank, numbers or Unicode characters
-375k PA voters who registered to vote on a federal holiday, when government employees are not working

5/N

That Pennsylvania has had many fraudulent voter registrations recently is not in doubt. Lancaster and York County caught them in the act. The only question is now many there are, and whether these counties were unusual, or just where it got caught.



6/N

https://x.com/ericldaugh/status/1849839728750711240

In understanding recent activities that look consistent with fraud, we need to start with two important facts about how horribly error-ridden the PA voter database is. Remember, bad processes make it more likely that *both* innocent and malicious errors are occurring.

7/N

Fact 1. There are no protections against either fully or partially duplicated records.

While this may seem incredible, it is also consistent with everything else we’ve learned about the database.

8/N

On October 21 2024:
3 observations are duplicates on every variable including ID number.
A further 170 are duplicates on every variable except ID.
A further 1,512 are duplicates on name, date of birth, and address
A further 10,226 are duplicates on name and date of birth.

9/N

11,908 duplicate observations is highly worrying. Will these correspond to 11,908 duplicate ballots that can be cast on election day? I don’t know, and you don’t know either. But we’re meant to take the assurance of people running one of the worst databases I’ve ever seen.

10/N

Fraudulent voting could happen either because these people vote twice, or because the duplicate entries allow someone else to vote on their behalf, introducing a fake ballot that is assigned to the duplicate record.

11/N

Even if someone manually notices, it’s not clear anything would be done. Suppose a poll worker spots two records with the same first name, last name and birthday, but one has a middle name of “Scott” (and he’s voted already), and this guy has a middle name of “S”.

12/N

This type of duplicate happens quite a lot. Will they stop the person from voting? I doubt it. If they have different addresses it’s even less likely. When you don’t keep track of duplicates, it is extremely difficult to know if it’s the same person, or a different person.

13/N

The number of registered voters is an important check on fraud, because it limits the number of additional fake ballots that can be included without raising suspicions. Start adding in tens of thousands of fake entries, and the scale of possible fraud gets way higher

14/N

This is assuming that the duplicates are there for innocent reasons, and the fraud is just opportunistic afterwards. This may be true, but there’s reasons to doubt it. Mass attempts to add fake voters have been confirmed in deliberate fraud schemes in York and Lancaster.

15/N

Fact 2. There is no consistent process for updating Pennsylvania voter records.

This too is amazing, but also expected by this point.

16/N

Voter records have a field called “ID_Number”, which you (like me) probably assume is a unique person identifier. Often it seems to work that way, but other times, it does not.

17/N

For example, between October 14 and 21, 10,109 people changed their party affiliation while keeping the same ID_number. In addition to the ID match, they had the first name, last name, date of birth, and address, so we can be almost certain it’s the same person.

18/N

But in another 414 cases, a second record gets added with a new ID, that has the same name, same address, same date of birth, and just differs in party affiliation. The old record is deleted, and a new one with a new ID is added.

19/N

In another 447 cases, a second record gets added with a new ID, but both records are kept. So now there are two different records of the same voter at the same time, with slightly different information.

The exact same change can produce three possible different outcomes.

20/N

The same chaos happens with many types of changes. Party affiliation is just a clean one to check, because we can be almost certain it’s still the same person, and changing party affiliation two weeks before the election doesn’t affect any important outcome.

21/N

The net effect of these two facts (no prevention of duplicates, no consistent process for changes) is that there are many partially duplicated records floating around, and bizarre changes between snapshots. As we’ll see, fraud tries to hide in bad data and bad processes.

22/N

For instance, between October 14 and 21, there were 65,842 new records created, based on ID_Number. You couldn’t determine this just from the data on October 21st if you downloaded it.

23/N

Based on registration dates that occurred between October 14 and 20, you’d count 51,031 new registrations (77.5% of the actual total of new records based on ID number over this period).

24/N

But then it gets weird. 22.5% of new additions (14,811 registrations) list a registration date before October 14th. You would assume that these should have been entered in the October 14th version of the data.

25/N

It doesn’t seem likely that this is due to delays in processing, or delays in handing in the form. 96% of these suspiciously early registrations list a registration date before January 1st, 2024. 88% of them are before January 1st, 2022.

26/N

A bit over half of these suspicious registration dates could possibly be due to copied records with new IDs, under the process described above. 56% (8,287 records) match to a first name, last name, and date of birth present in the October 14th database.

27/N

It’s worth bearing in mind – we don’t know the process by which some records get assigned new IDs and others don’t. It might be innocent mistakes, but it might not, too. There are reasons to be suspicious here.

28/N

The rate of these cases (new ID, suspicious registration date, name match) as a fraction of new records is significantly lower in Lancaster County (-1.4%) and York County (-5.1%), where fake voter registrations were caught and prevented. p-values are 0.013 and <0.0000001.

29/N

The rate is much higher, by 11.1%, in Montgomery County. The p-value is <0.0000001. Montgomery is the second highest county, behind only Cameron (which only had 18 new registrations). This is important, because Montgomery had considerable evidence of voter fraud in 2020.

30/N

Evidence of fraud in Montgomery County during the 2020 election was shown using multiple independent methodologies, and these accusations were never addressed or rebutted by county officials. These articles are highly censored on search engines even now.

31/N

Montgomery County was home to “The Most Suspicious Vote Update In America”, where vote totals in multiple categories *decreased* between two updates, in ways strongly indicative of fraud, and very difficult to reconcile with innocent explanations



32/Nrevolver.news/2020/11/explos…

It also had the third most suspicious distribution of voter birthdays among Pennsylvania Counties, across multiple different measures of implausibility, and these implausible birthdays correlated with Biden vote share.



33/Nrevolver.news/2020/12/pennsy…

If you’re curious about the 2020 election, here’s my summary of voter fraud research in December 2020. A lot more came out since then, all pointing in the same direction. It has a lot of musings about how hard it is to identify fraud in general.



34/Nshylockholmes.blogspot.com/2020/12/last-t…

Back to the current case. Maybe the correlation with known fraud patterns across counties is just a coincidence. But if we remove the suspiciously dated cases that match to names and birthdays in the October 14 data, the remaining records are even harder to understand.

35/N

6,548 records have both implausible registration dates, and do not map to any name and birth date on October 14th. These are 10% of all new registrations in the last eligible week. These “doubly suspicious” records are concerning on a number of other dimensions too.

36/N

First, these doubly suspicious observations are also 2.1% more likely to be missing house numbers than other new registrations. The base rate is 12.4%, and the p-value is less than 0.000001.

37/N

Second, these doubly suspicious observations are 2.8% more likely to have been registered to vote on New Years Day, relative to other new registrations. The base rate here is only 0.13% so this is a massive increase, and the p value is less than 0.00000001

38/N

Third, these doubly suspicious entries are also significantly more likely to have already voted in 2024. They’re 0.3% more likely relative to all new registrations, and 0.5% more likely relative to suspicious registration dates with name matches

39/N

Are these just voter registrations that were previously inactive, and recently got made active? A few things suggestion this is unlikely to be the explanation.

40/N

Both the October 14th and October 21st files list inactive voters in the database – 803k on October 14, and 789k on October 21st. So being inactive doesn’t mean you get deleted.

41/N

In addition, we observe voters switching from inactive on October 14th to active on October 21st – 10,416 cases in total. So the database can handle changes in status from active to inactive without creating new records, or new IDs.

42/N

Finally, it’s useful to think about why old registration dates might be useful to a fraud operation. There are good reasons to think that large additions of new, last minute registrations might be subject to more scrutiny and challenges

43/N

If your registrations are fake, you don’t want that. Once they go into the system, if they have an old registration date, then the only way to even know that they were added recently is to have both versions of the voter file, and compare the differences.

44/N

Maybe you have some guess as to what’s going on. Maybe the database is broken is some other, even more esoteric way that I haven’t managed to think of. Believe me, I tried to think of every innocent one I could, and I'm still stumped.

45/N

But think about what this implies. At some point, it is very little defense of the voting system to say that the opportunities for innocent errors are so colossal, and the guard rails preventing them so nonexistent, that it is hard to definitively prove malicious errors.

46/N

The system for managing Pennsylvania voter data is a travesty, a farce, a disgrace. It would be comedic, if the consequences for the state and the country weren’t so serious. Everybody involved should be fired. In the private sector, they would have been, years ago.

47/N

If I were designing a database setup to facilitate fraud, it would be hard to improve on the setup maintained by the Pennsylvania Department of State. I don’t know which records are genuine. They don’t know which records are genuine. Nobody knows which records are genuine.

48/N

It is no surprise that we see as many overt attempts at fraud as we do. Pennsylvania has a system that encourages and facilitates this. There are strong reasons to suspect that the errors we note here are the tip of the iceberg.

49/N

It is absolutely impossible to look at this system and have any confidence in the integrity of elections in the state.

/End

If you liked this thread, please re-tweet the first one



And if you have election data that I can analyze (especially voter rolls or recorded votes in MI, WI, GA, AZ), or if you want to test ideas on the PA data, my DMs are open.

https://x.com/shylockh/status/1851808846525653017