Thread by @TradeandMoney

It looks like my paper w/ Karsten Mau is now published in ReSTUD. I had written a thread on our paper here:

https://twitter.com/TradeandMoney/status/1280236322687459328

https://twitter.com/empiricalchina/status/1361314522980515844

I see the authors have written a partial response. However, I'm not persuaded. We've already written a thorough response-to-the-response here: dropbox.com/s/8yn8woym7qw5…

Here's a quick recap of why I'm not persuaded. Originally, I was unhappy that the authors did not plot any relevant data. A simple plot of pre-treatment trends suggested no impact at all of Chinese competition (Quota group, rising after 2002) on patents per firm.

In their response, the authors still don't plot any data, or even try to explain why a simple plot of pre-treatment trends appears to refute their results. Nor do they say how it was that they forgot to disclose that their patent data converges to zero in their sample.

In lieu of simple evidence, their MO is always to go into as-complex-as-possible econometrics, so people won't quite understand what is going on. Math serves the point not to clarify, but to obfuscate.

They can't admit they were wrong, so they take the view that, yes, there were two coding errors in our previous table (plus we described what we did incorrectly), but we also made a 4th error which flips

our results back to being significant. To get significance, though, they provide a textbook example of multi-collinearity. (Just think about this example next time you get rejected from RESTud w/out making such an obvious error.)

They add in the variable "initial Chinese imports" which is very highly correlated with "Chinese imports", especially since country*year & sectoral FEs are all controlled for. They haven't a clue what is driving their results, nor do they furnish any legitimate reason to include

initial Chinese imports in a panel regression w/ Chinese imports and year FEs. They also don't control for trends in patents, which is obviously necessary in the context of a diff-in-diff looking at our graph above.

Also, in the authors response, they talk about their "9 results". With apologies, we have shown that most of the results in their paper were not robust to slight changes in specifications, such as adding in trend controls, differing the FEs, or "un-winsorizing" half-winsored data

I can't help but wonder if someone not affiliated w/ Stanford or MIT could publish a difference-in-difference in a Top 5 w/out plotting pre-treatment trends? And in papers rife w/ other blatant errors? That is what just happened. Twice.

I'm probably going to get some more blowback for writing this, but from my vantage it looks to me like the "con" is still strong in academic "eCONomics".

So, to just fix ideas, here is the new BDvR-20 table where Chinese competition has a yuge impact on innovation in Europe. You see they get significance in column 3 by including Initial Chinese imports (large neg. sign), and 4 dig SIC FEs (which weren't mentioned in original).

Thus their new favored regression differs in four! ways from their initial paper. I can understand we all make mistakes, but four acknowledged errors in one regression is pushing it. And these four errors are not counting the lack of sectoral trends.

However, you won't be surprised to know that that "Current" and "Initial" Chinese imports are highly correlated (R-squared is .6 & t score is 350).

But, their column (3) above also includes 4-digit SIC FEs (the initial Chinese imports variable varies also at the country level), and they have country-year FEs. Added together, we get an R-squared of .76.

How much multi-collinearity is "too much"? Well, we can calculate a VIF of 4.2 for this case, which is greater than a danger zone cutoff of around 2.5 according to this article: statisticalhorizons.com/multicollinear… @johnvanreenen please explain.

As you can see, you have to wake up pretty early in the morning to sneak a slow-ball past RESTud referees and the editorial board.