A new paper by Barro and Lee:
http://www.nber.org/papers/w15902
Abstract:
Our panel data set on educational attainment has been updated for 146 countries from 1950 to 2010. The data are disaggregated by sex and by 5-year age intervals. We have improved the accuracy of estimation by using information from consistent census data, disaggregated by age group, along with new estimates of mortality rates and completion rates by age and education level. We use these new data to investigate how output relates to the stock of human capital, measured by overall years of schooling as well as by the composition of educational attainment of workers at various levels of education. We find schooling has a significantly positive effect on output. After controlling for the simultaneous determination of human capital and output, by using the 10-year lag of parents‘ education as an instrument variable (IV) for the current level of education, the estimated rate-of-return to an additional year of schooling ranges from 5% to 12%, close to typical Mincerian return estimates found in the labor literature.
The raw data is a fabulous public good. But I'm more sceptical about the empirics. Despite what the abstract says, the paper doesn't actually take the 10-year lag of parents' education as an IV. It takes the 10-year lag of everyone's education as an IV. Some of this effect will represent intergenerational effects, but most will be pure serial correlation (I was a university graduate in the last census, and I'll be one in the next census too).
ReplyDeleteI haven't read the paper. Lags as instruments in panel data is not necessarily a bad thing, tons of papers do it [influenced by Hsiao, Arellano & Bond's work amongst others] but you need to be very careful that you have got your dynamics right.
ReplyDeleteTests for serial correlation will pick up some of this as will the Sargan/Hansen test. Identification ultimately rests on theory 'though so the question is: are you confident in your model that the 10 year lag is excludable?
Well they say it is -- although I can't find the results of any over-ID tests(maybe because it's panel(?)). I thought Hansen J was standard?
ReplyDeleteFantastic dataset though.
I have run Hansen tests hundreds of time with panel data, there is nothing special. If they have only 1 instrument (& 1 endog variable) then you can't have an over-id test obviously.
ReplyDeleteAye yes -- I am mixing Hansen and Hausmann.
ReplyDelete