The reviewer said,
In a world in which I didn’t have only 20 minutes to read, analyze and write about this paper, I’d like to think through his model choices. I would feel much more comfortable on this point if he accepted the Russ Roberts Science challenge and have a section discussing the process by which he arrived at the process by which he arrived at his conclusions.
Look, I have a policy. I don’t do specification searches. If I don’t get reasonable results in the first two tries, I abandon the project. As it was in this case, I only did one pass through the data. I was testing for the idea that state or national governmental policy might affect book or market value returns, after adjusting for market sector.
He later commented,
I’d have two comments:
1. What’s the point of decomposing them, then?
2. Can’t you just attribute ALL variance of corporates to ‘historical accident’? Can there be no policy implications?
On point #2, I’d defend Merkel by saying that policy implications need a big enough sample that you can reasonably hold other factors constant. You’d need a dataset of every industry in every state over every conceivable macro-economic environment, then control for those other factors. Same applies for analyzing different countries.
The point of decomposing them is that you don’t know in advance what the result will be. I only did one pass at the data (please ask academic economists what they do), in this case, it showed that after adjusting for sectors and general economics (time), the states one was in did not matter much, as those that did well did not move to seek lower tax environs.
The piece I did last year did not attribute everything to historical accident. This year, I was surprised to find that few successful companies had not moved to lower tax/regulation jurisdictions.
I did not know what the decomposition would lead to — that was a major reason for doing it. If there had been some indication that companies in the US sought lower tax or regulation states, I would have published that, but it was not so, in aggregate. I does not matter that the result was ordinary. Once I start the problem, if I come to any understandable result, consensus or non-consensus, I publish it.
Now in truth, I don’t think the paper was one of my best efforts. I would like to have set error bounds, but I didn’t have access to good software. I also would have liked to use a better database, like the CRSP database, but that was not available. Given my lack of resources, it was the best I could do. Anyway, anyone with more constructive criticisms, I welcome them.