« on: March 25, 2008, 01:13:26 AM »
Ah, yeah. .48 and .49 for the r's. Not good values in an absolute sense for actually generating predictions, but pretty stunningly robust in a social science context.
I wouldn't consider a multiple correlation of .48 "stunningly robust," even for cross-sectional data. This value, which implies an r-squared of .23, indicates (roughly) that only 23% of the variation in first-year grades can be attributed to ex ante differences in UGPA and LSAT scores. And LSAC admits that the 23% achieved in 2003 “is high” relative to other years. The LSAT alone explains only 13% of the variation. A microeconomic model with an r-squared of .23 would have difficulty surviving peer-review, especially if the model was built solely for prediction. (Notable exceptions certainly exist.)
I find this somewhat dismaying considering the relative weight adcoms place on numbers. But you're right. As it stands, it's the best indicator they've got...