Instrumental variable approach with entity and time fixed effects

[Basil Rupper]

Hi Everybody.. I am quite new to Stata and Econometrics in general, so please forgive me my lack of knowledge in this field. Currently I am working on my bachelor thesis in investigating the impact of fiscal-, monetary- and trade policy on economic growth. I use a panel data set consisting of 40 countries over a time period of 26 years. In one of my models i use fixed effects (entity and time) and an instrumental variable with the following command:

xtivreg lny lnx1 lnx2 x3 i.year (lnx4 = z1), fe vce(cluster countrynum)

The problem is, if I use the instrumental variable together with time fixed effects my results (coefficients for all variables) change dramatically and all of them become highly insignificant. If i use a instrumental variable model with only entity fixed effects or alternatively a time and entity fixed effects model without an instrument the coefficients are all highly significant. I tend to think that something is wrong with the command I used above. Is it the right command to estimate an instrumental variable model with time and entity fixed effects? Or does it make sense at all to include entity and time fixed effects and an instrumental variable together in the same regression?

Thanks a lot for your help
Best wishes
Basil Rupper

[Philip Gigliotti]

Year fixed effects have the power to kill a result. That's why they are considered more robust than unit fixed effects alone. In the context of finance research over a 26 year period, year fixed effects carry a HUGE amount of information. Significant results without time effects are likely spurious and the result of macroeconomic trends like inflation, economic growth, conflict or political trends. Results without an instrumental variable are likely biased by endogeneity.

[Phil Bromiley]

Another way to represent what Philip said is that the panel fixed effect estimator takes out all the stable country effects. This is often appropriate, but can create problems with you really want to study such stable effects. If including or not including year effects massively changes the results, then it is problematic to not include year effects. I would also check that the diagnostics indicate that lnx4 is endogenous. To do this, run ivreg with i.panel as a control variable and use estat endog after. User written xtivreg2 has some tests built in that xtivreg does not (at least in earlier versions of Stata).

One thing to try is instead of year dummies, calculate the mean of the dv for all the other countries excluding the country of interest. Something like this:

bysort country year: egen tlny=total(lny)
bysort country year: egen countlny=count(lny)
g meanlny=(tliny - lny)/(count - 1)

This provides a control similar to the year dummies, but uses far fewer degrees of freedom. It has a passing resemblance to the Mundlak approach. It does create problems if the lny for other countries depends on the lny in the focal country, but whether that occurs depends on your context and measures.