She makes some of the same arguments against it's usefulness as have been made here, plus one more that I found interesting:
You shouldn't use this when one number is dependant on another.
For example, in an election basically between two people, if 1000 people vote in a 2 person election with no undervotes, if one person gets 400, it means the other person got 600. 300/700, and so on.
If all of the precincts are the same size, the loser will lose more precincts than they will win.
That means the loser will have collected more low-numbers (first digit <5) than high numbers.
Since the Benford graph shows a distribution curve with more low numbers than high numbers, the losers graph will align with the distribution better, but the winner's, number will surely be skewed to the higher numbers.
One quote from the page
Whether Benford can be used to detect election fraud has been studied for decades. What everyone who studies this knows is that analyzing first digits absolutely DOES NOT WORK!
"You can't add percentages of disparate quantities"
Zatoichi's gaming tricks playlist on YouTube...
As we can see the data looks a little wonky ... the first thing to notice is 22 precincts reported zero votes for President Trump ... virtually a statistical impossibility, even in Detroit ... but it appears the numbers my have been manipulated .... So we begin to wonder, since President Trump won Michigan in 2016, might his campaign have been manipulating the data?
Is it a statistical impossibility? It appears to me the number of voters is not high enough, and the number of precincts not high enough to make this impossible. Also, I'd be curious about Washington DC. Where over 90% of the population voted for one candidate over the others. Would it be safe to assume a number of precincts (or whatever their breakdown) have zero votes for Pres Trump?