Fixed effect (xtreg) and control variables

[Ashley Lee]

Hello.
I am working with fixed effect model on the impact of highspeed rail on local economic activities using night light intensity as an economic development proxy.

(light = light intensity at township level; treat= binary variable indicating treatment status (1=treat, 0=control); t=binary variable indicating pre/post period; other control variables= at county level)

I have run the regression with and without control variables, and the results are as such.

xtreg light i.treat##i.t i.year i.treat##c.year gdppc consgds gvtrev gvtexp fixasset indentp pop hh avwage gdp area, fe vce(cl id)

As you can see, since the control variables have many missing values due to data limitation, the observations halved with full control variables.
My thesis advisor advised me to play with control variables and figure out which ones to use for the main specification, and I am having difficult time finding out what to do to find out 'good' controls. I have tried including only one control variable in each regression to see their standard errors but I am not sure what else to look for.

Below are the summary of missing values and the regression result with only one control variable (and all of the controls showed statistically significant result when run individually). Any suggestions dealing with halved observations? Thank you in advance.

[Carlo Lazzaro]

Ashley:
my guess is that the most fruitful approach is to deal with missing values first (see -ipolate- or, better, the -mi- suite in Stata .pdf manual). and then go -xtreg,fe-.
Perhaps you can present your analysis (one regression model only) with both complete cases and inouted values in your (master?) thesis.
Conversely, I would shy away from fiddling with so many regression models and focus on the one the gives the fairest and truest view of the data generating process.
Eventually, an unsolicited amateur's thought: isn't possible that an affluent local economy influences highspeed railway, in turn?