Thread by @JohnRLottJr

.@sndurlauf, again, responses in order. They have been responsive.
1) I am looking at small areas across the street from each other and using one measure of political interests to explain another one. The difference in the share of the in-person votes is already picking up a ...

https://twitter.com/sndurlauf/status/1345125314188472330

...lot of difference in the political interests between the two adjacent precincts that one might measure with other demographic variables. Indeed, for the Georgia data in 2020, the year that you say is impossible to really explain, that variable by itself explains 82% of the...

...variation in the difference in the share of the absentee votes. The proof for whether it was a good “hunch” is an empirical question.

2) Counties set some voting rules. So what? They set the rules in 2016, and there were no differences. In any case, counties are relatively limited in what they can do as far as voting rules, especially since Bush v Gore. For example, they can’t change the dates for early voting.

2a) "there is endogeneity to choice of county in which to live."
Are you arguing that people choose to live on the side of the street that is in Fulton county because of voting rules?

3) It isn’t just race and gender that are being accounted for. I am also accounting for small geographic areas that are literally across the street from each other and for the in-person vote shares that are going to proxy for many of these factors that you are raising. ...

...These precincts are small geographic units. Allegheny County has 1,323 precincts. Fulton County has 384 precincts. The voter turnout rate estimates that I have done at the county level account for many other factors. Again, these supposedly left out factors don’t explain ...

the fact that there is no difference between the adjacent precincts in 2016. If the cost of voting was so different between these adjacent precincts, how come these adjacent precincts didn’t show a difference in 2016?

4) "Identification assumes shifts in support for Democrats should have affected in person and absentee voting similarly and should not differ across precincts differentiated by county"
Response: Yes, as was true in 2016.

4a) "This ignores politicization of use of absentee ballots"
I am ignoring nothing. Your point seems like another repeat of previous issues without responding to my response. Why wouldn’t you expect two precincts adjacent to each other and have similar voting patterns and ...

demographics to be similar? Steve suggests that “politicization” would make the two precincts differ in that dimension. But why wouldn’t factors such as the difference in Trump’s share of the in-person vote and racial and gender demographics pick this up? Why would two ...

similar precincts respond differently to the politicization that is equally available to both?

5) "The 2020 election is unique"

Suppose it is unique? I am controlling for these adjacent precincts that are similar in geographic make up,how they vote in-person, and racial and...

...gender demographics. Two small areas next to each other might be expected to change similarly. In any case, if there is a lot of randomness that is being added into the relationship, that would make it harder to find a statistically significant result. The point here is ...

...that despite this unusual change, it is moving in a way that is consistent with the vote fraud hypothesis.
Lots of people will claim that different elections are unique 2008 election was unique and many other elections were unique. Yet, academics run regressions across...

...regressions all the time, and they don’t usually break things down to the point where they are able to use adjacent precincts as control variables.

5a) "assumptions can be wrong but some are useful. Not useful here."
Why not? This is just an assertion. These cross-sectionregressions have R2s that are between .8 and .95. It isn’t obvious that these variables are leaving a lot unexplained. Previous response also applies here.

6) "Conditioning on a county where Rs were unhappy with results induces selection bias. Not addressed."
Test: select 2 precincts split by 2 counties where fraud was not changed & see if results are the same. In fact, I did that as R counties are also compared to each other.

7) "county error correlations"
The errors may be correlated spatially and within counties. While I explain why I don’t think that it is appropriate in the paper, I did clustering by county and the levels of statistical significance increased.

7a) "theory implies parameters on voting behavior differ by race etc so additive separability ... a misspecification"
Voting behavior differs by race. By the way, it also varies by gender. But race and gender are controlled for in the regressions. And I have them interacted.

Durlauf Bottom line:"your arguments and defenses do not address identification"
Robustness is demonstrated by using many different specifications, applying them to different data sets, trying different measures of statistical significance, and applying different qualitative tests

I updated the paper to include information on using clustering by county. See here.
papers.ssrn.com/sol3/papers.cf…
Again comments are most welcome.