IV estimate is opposite sign to OLS estimate

[Kavith Thanthulage]

Hi all,

I'm using ivreg2 to analyse panel data in stata 18.0. My regression is looking at returns to education by ethnicity using a standard mincer equation and I am using a compulsory schooling law as an instrument for years of education:

ivreg2 log_hrly_wage (c.yeduc#i.ethnic = i.RoslaD#i.ethnic) i.region i.female c.exp c.exp#c.exp i.parentsoc, first robust

Where I control for region, gender, experience, experience squared and parents occupation. When I run this regression as a naive pooled OLS using regress I get positive and significant results for years of education (around 0.13) but using ivreg2 I get negative significant results of a larger magnitude (around -0.16). Just to note, this only happens when I include the variables c.exp c.exp#c.exp and parentsoc, not the other controls.


This is quite confusing for more as the instrument I've used is quite a popular one in the literature to show a positive effect of education on earnings. Have I gone terribly wrong somewhere?
Thank you all for help,

Kavith

[Carlo Lazzaro]

Kavith:
welcome to this forum.
Why not using -xtivreg- ?
As an aside, please read and act on the FAQ to post more effectively. Thanks.

[Kavith Thanthulage]

Hi Carlo, thanks for the response.

I was planning to follow up with xtivreg but wanted to build up using simpler models first.

Kind regards,

Kavith