Instrumental variable with fixed effects gives unrealistic results

[Simon Bleijenberg]

Hi,

For my thesis I'm estimating the causal effect of smoking on happiness. I use panel data so I use fixed effects, so far no problems. Because there is a concern for endogeneity I also want to incorporate an instrumental variable, for this I use the consumer price index (CPI) of tobacco. This instrument is valid according to the literature and the tests the xtivreg2 command give seem to confirm this. But the results that I get are very unrealistic since smoking is associated with an increase in over 3 points on a happiness scale of 1 to 6. This increase is way too large and is 60 times larger than the model with just fixed effects. Running the model with just an instrumental variable and no fixed effects gives an even larger coefficient. All coefficients for the variable of interest are significant at the 10% level. The commands I used are:
IV+FE:
xtivreg2 happiness (smoking = CPITobacco) age age2 religion partner employment_status BMI BMI2 has_smoked , fe first robust
FE only:
xtreg happiness smoking age age2 religion partner employment_status BMI BMI2 has_smoked, fe robust
IV only:
ivreg2 happiness (smoking = CPITobacco) age age2 religion partner employment_status BMI BMI2 has_smoked, first robust

The output stata gives can be found in the attachment.
I hope someone can tell me what is going on here or what I did wrong.

Best regards,
Simon

[Carlo Lazzaro]

Simon:
ignoring endogeneity for a while, the main source of your concerns should be the very low within R-sq of -xtreg,fe-; I would check whether all the relevant predictors were included in the right-hand side of your regression equation.
For the future, please share what you typed and what Stata gave you back via CODE delimiters (read the FAQ on this and other posting-related topics). Thanks.
Other comments:
. smoking coefficient is simply not significant in both models (-xtreg- and -xtivreg2): hence, I would not worry about it. By the way, why you did not consider -xtivreg-?;
- there might be a quadratic relationship between BMI and happiness: have you explored it?;