It's also a bit of a pain that it's just a summary of a video, and there's no citation to any text anyway.
From what I can tell, the first bit of evidence is from a researcher's toy language.
The second was a scan of GitHub for issues that included type errors. There's no mention of actual methodology, but I'm imagining that means 3% of Python issues in GitHub have a copy-and-pasted stacktrace that contains a type error? This is clearly not a great indicator, since it won't count issues with less description, or any bugs related to, say, a bad type coercion.
On the other hand, you're right, I've not read any research that points to static typing being a net-gain. Every static-type advocate probably remembers being burned by a dynamic-type issue, but that doesn't mean the advantages might not outweigh disadvantages - experience and intuition don't cut it here.