3 ms·

to give a bit more insight; it says essentially, imagine a deck of cards at a casino, shuffled by machine 50, 100, 500 times. That's what the ballots are like s

to give a bit more insight; it says essentially, imagine a deck of cards at a casino, shuffled by machine 50, 100, 500 times. That's what the ballots are like shuffled through the USPS....homogeneous....there's no cream rising to the top of the bottle, and there's no stacks of Kings or Aces lodged all in one spot because the system (mail) shuffled them. There should be no discernible patterns of any kind. One for you, two for me, two for me, two more for me, three for you, one for you, one for me, two for you and so on, all leading on his graphs, towards a very straight line. So there shouldn't be a stack of ballots that suddenly burst on to his graphs leaning heavily to one group or the other at any given time....but there is....in multiple states, all happening in the wee hours of the morning.....and all happening on the 4th. And in each state, they all (on the 4th), lean Democrat. Discernibly. And with one sole exception, only in those states.

bra-ket 6y ago
or let's say there's a batch of 100k votes and it's split 90k Biden votes and 10k Trump votes (in Democrat leaning precincts). In any random sample there should be a 9:1 ratio meaning that out of 100 randomly selected votes ~90 will be for Biden and ~10 will be for Trump. The larger the sample the more likely the actual count is closer to the 9:1 ratio. Any deviation from that ratio will be made up as the count goes on. So in the most extreme case, if 90k in a row were all Biden and there was a 90k spike, the next 10k should spike for Trump. That never happened because after the 4am Biden spike, the count went back to "normal" ratios. To use the casino analogy imagine if you're at the roulette table and it comes up red 100,000 times in a row without black or green ever being called. That is statistically improbable.
- bra-ket 6y ago
  The only other explanation would be that NYT dataset is messed up. Only an audit can prove this for the court. Let's hope they audit swing states results and do similar analysis/anomaly detection.
eesmith 6y ago
> There should be no discernible patterns of any kind. That's not a correct description of the model given. The author writes "The slight drift from D to R mail-ins occurs again and again, and is likely due to outlying rural areas having more R votes. These outlying areas take longer to ship their ballots to the polling centers." That is, the author says there should be a discernible pattern. Note that this is a hypothesis which was created only after looking at the patterns in different states. While it appears to be true of some other states, there's no strong argument for why it should be true of all states and all vote counting methods. Unfortunately, the author then double-downs by saying that since this is the expected pattern, and it wasn't seen in PA, then something's suspicious, to the point of being "likely evidence of ballot backdating or manufacturing." The IMO correct interpretation is observation that the random model, which was already tweaked in order to explain one data trend, might need more adjustments in order to explain other data trends. For example, the author uses PA and WI as examples of outliers to the drift-to-R-mode. But we know that (quoting https://eu.usatoday.com/story/news/politics/2020/11/06/how-states-prevent-election-fraud-protect-integrity-ballot-count/6186997002/ https://eu.usatoday.com/story/news/politics/2020/11/06/how-s... ): > While most states began processing mail-in ballots before Election Day, others had laws preventing election officials from doing so. For instance, election officials in the swing states of Wisconsin and Pennsylvania requested the ability to begin processing earlier and the Republican-controlled legislatures in those states refused, all but ensuring the high volume of mail-in ballots would not be counted until after Nov. 3. There's no reason to expect that a model trained on states with one vote counting model should be applicable to states with another vote counting model, much less be strong enough as to justify a claim of ballot backdating or manufacturing. Furthermore, I believe there's a reasonable argument that the "Republican-controlled legislatures in those states refused" in part to muddy the waters, and create anomalies which could be interpreted as fraudulent voting. This potential confounding issue was also not included in the "deck of cards at a casino" model.
- afiori 6y ago
  (actual question as you seem well informed) What about the slow D trend for mail ins? could that be that the same 4 Nov spike got spreaded over more days? Because one would expect that the very last to arrive ballots should trend R the same as rural areas

bra-ket 6y ago

or let's say there's a batch of 100k votes and it's split 90k Biden votes and 10k Trump votes (in Democrat leaning precincts). In any random sample there should be a 9:1 ratio meaning that out of 100 randomly selected votes ~90 will be for Biden and ~10 will be for Trump. The larger the sample the more likely the actual count is closer to the 9:1 ratio. Any deviation from that ratio will be made up as the count goes on. So in the most extreme case, if 90k in a row were all Biden and there was a 90k spike, the next 10k should spike for Trump. That never happened because after the 4am Biden spike, the count went back to "normal" ratios. To use the casino analogy imagine if you're at the roulette table and it comes up red 100,000 times in a row without black or green ever being called. That is statistically improbable.

bra-ket 6y ago
The only other explanation would be that NYT dataset is messed up. Only an audit can prove this for the court. Let's hope they audit swing states results and do similar analysis/anomaly detection.

eesmith 6y ago

> There should be no discernible patterns of any kind. That's not a correct description of the model given. The author writes "The slight drift from D to R mail-ins occurs again and again, and is likely due to outlying rural areas having more R votes. These outlying areas take longer to ship their ballots to the polling centers." That is, the author says there should be a discernible pattern. Note that this is a hypothesis which was created only after looking at the patterns in different states. While it appears to be true of some other states, there's no strong argument for why it should be true of all states and all vote counting methods. Unfortunately, the author then double-downs by saying that since this is the expected pattern, and it wasn't seen in PA, then something's suspicious, to the point of being "likely evidence of ballot backdating or manufacturing." The IMO correct interpretation is observation that the random model, which was already tweaked in order to explain one data trend, might need more adjustments in order to explain other data trends. For example, the author uses PA and WI as examples of outliers to the drift-to-R-mode. But we know that (quoting https://eu.usatoday.com/story/news/politics/2020/11/06/how-states-prevent-election-fraud-protect-integrity-ballot-count/6186997002/ https://eu.usatoday.com/story/news/politics/2020/11/06/how-s... ): > While most states began processing mail-in ballots before Election Day, others had laws preventing election officials from doing so. For instance, election officials in the swing states of Wisconsin and Pennsylvania requested the ability to begin processing earlier and the Republican-controlled legislatures in those states refused, all but ensuring the high volume of mail-in ballots would not be counted until after Nov. 3. There's no reason to expect that a model trained on states with one vote counting model should be applicable to states with another vote counting model, much less be strong enough as to justify a claim of ballot backdating or manufacturing. Furthermore, I believe there's a reasonable argument that the "Republican-controlled legislatures in those states refused" in part to muddy the waters, and create anomalies which could be interpreted as fraudulent voting. This potential confounding issue was also not included in the "deck of cards at a casino" model.

afiori 6y ago
(actual question as you seem well informed) What about the slow D trend for mail ins? could that be that the same 4 Nov spike got spreaded over more days? Because one would expect that the very last to arrive ballots should trend R the same as rural areas