IV blows up coefficients

[Mathias Rechsteiner]

Dear Stata Users

Comparing a RE (xtreg, re) with and IV (xtivreg, re) model, I receive quite large coefficients in the IV case. For football I have found five Instruments.

MATH	Coeff. RE	Coeff IV
football	-0.83	-30.05
age	4.28	4.79
fear	-0.62	-0.57
parentalincome	0.004	0.004

I am wondering why the coefficient of football has such a hughe difference. According to the literature I checked the IV-Model for the following conditions (xtoverid, nois).

Endogeneity: F-Test for excluded instruments F(5, 5602) = 14.8

Weak instruments: Stock-Wogo weak ID test critical values for K1=1 and L1=5: 10% max relative bias = 10.8; 5% max relative bias = 18.3

Overidentification: Sargan-Hansen statistic: p-value = 0.34

Does anyone has an idea, why the IV model seems to be strange and probabyl wrong?

Thank you very much

[Phil Bromiley]

You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex. We don't know exactly what you ran or the standard errors on the parameters. I didn't see that xtoverid provides weak instruments tests. I don't know the significance level for the F-test (and shouldn't have to look it up to help you).

The reason for 2SLS is the it may correct biases in the xtreg estimate so getting a different parameter with 2sls than xtreg is reasonable. On the other hand, in a finite sized sample, the instrument has more noise than the original variable. So, it is not obvious which is the better estimate. Is the -30 reasonable?