Wednesday, August 11, 2010

A Rough Comment on Levitt...

This comment makes three observations about Donohue and Levitt’s paper on abortion and crime (Quarterly Journal of Economics 119(1) (2001), 249–275). First, there is a coding mistake in the concluding regressions, which identify abortion’s effect on crime by comparing the experiences of different age cohorts within the same state and year. Second, correcting this error and using a more appropriate per capita specification for the crime variable generates much weaker results. Third, earlier tests in the paper, which exploit cross-state rather than withinstate variation, are not robust to allowing for differential state trends based on statewide crime rates that predate the period when abortion could have had a causal effect on crime.
Although this may look like "shocking" revelations to some, if you take essentially any major result in economics over the past 30-40 years and dig around with their data, you're quite likely to discover that the central finding is fraudulent.


  1. Fraudulent?

    I understand how any complex experiment or observation will, on close enough inspection, will show errors, I heard this with a high energy physics experiment (they were closely inspecting results that fell in a "notch" and excluded those results), but actually fraudulent?

    Whiskey Tango Foxtrot?

  2. Strong language perhaps, but it just happens time and again... The key logic doesn't wash. Including an obvious control reverses the results. There's a coding mistake. Crime wasn't entered "per capita". Key data is simply made up. I don't know if "fraudulent" is too strong a word to describe data that's simply been conjured out of one's imagination (or a convenient coding mistake), but when this happens repeatedly you just get jaded... Which isn't to say that there aren't any good papers -- there are. But high energy physics economics is not.