[Clyde Schechter]

Ah! Thanks for clearing that up.

[Richard Williams]

It is unfortunate that you have to use 20 year intervals. Can you at least get them to give it to you in 10 year intervals? That would be a bit more precise, and might also make it easier to detect any nonlinearities that are there.

[Ariel Karlinsky]

I'll try. the caveat is that the data is panel data across 13 years (1999-2011), so with 10 year intervals, it's possible to just recode back to continuous for all individuals and all that work would have been for naught... with 20 year intervals at least for most, the age group doesn't change across time.

[Ariel Karlinsky]

I'm happy to report that treating age_group as a continous variable seems to be OK (i.e I use "reg y x c.age_group" instead of "reg y x i.age_group"). Thanks everyone for all the help.