IV using panel data and fixed effects

[Ash Lana]

Hi,

I am using panel data about women's wellbeing and influencing factors, therefore, I have been using a fixed effects panel regression (xtreg..., fe cluster(pidp)) so far.
To establish causality I am trying to use an IV regression. I assume that I need to account for the panel nature and fe, so so far have used the following regress:
xtivreg $Y1 $X1 ($Y2 = $X2), fe
Where Y1 is the dependent variable (life satisfaction)
X2 is the endogenous variable (housewife)
X1 are controls (marital status, children, region, income, year, age)
X2 is the instrument (gender employment ratio by region and year)

I have seen that there is also a regression for IV that accounts for the endogenous variable being binary, which mine is, using treatreg $Y1 $X1, treat($Y2 = $X2 $X1), which makes the first stage probit. However, this does not account for fe, please could someone help by providing guidance on which is the best method to use?

I have some other job status' as well, (unemployed, part-time, full-time), do these need to be included in the iv regressions? Or are they accounted for as '0' in the binary variable 'housewife'?

Many thanks,

Ash

[Phil Bromiley]

Welcome to Stata list. You will increase your chances of useful answer by following the FAQ on asking questions-divide Stata code in code delimiters, readable Stata output, and sample data using dataex.

As I understand it, xtivreg is still consistent even with a binary endogenous variable. Alternatively, you could do this with SEM or GSEM, or (I suspect) with user written CMP.

If you have other important factors that potentially could influence your dependent variable, then including them in the model as controls is a good idea. If you have any additional potential instruments that meet all the criteria for instruments, then an additional instruments may improve your estimates.

Note that depending on your number of panels, you might also be able to use a treatment model with dummies for panels.

[Ash Lana]

Hi Phil,

Thanks very much, I will use xtreg2 as the command. I have now encountered two new problems. Would you be able to help? I have provided additional context to help.

I am completing a study of the wellbeing women, across a sample of 4 years and 22,000 women. I wanted to analyse the effect of the job status (unemployed, housewife, self-employed, part-time work, full-time) on women's over-all life satisfaction (olsat), and to establish causality I am using IV.

My standard regression is as follows:
xtreg olsat i.jobstatus $controls, fe cluster(pidp)

The coefficient estimates I get for the job status' are jobstatus2 = .144, jobstatus3 = .289, jobstatus4 = 0.204, jobstatus5 = 0.184 (relative to the base on jobstatus1, which is unemployment), and all are significant at the 1% level.

To establish causality I am testing jobstatus2 (housewife) using IV.

I run the commands as follow:

global Y1 olsat
global Y2 jobstatus2
global X1 "mstat2 mstat3 mstat4 mstat5 mstat6 nch02 nch34 nch511 nch1215 hwhrs hqual health loghhy age age2 ethnic2 ethnic3 ethnic4 ethnic5 region2 region3 region4 region5 region6 region7 region8 region9 region10 region11 region12 wave2 wave3 wave4"
global X2 genYratio

xtivreg2 $Y1 $X1 ($Y2 = $X2), fe cluster(pidp) first endog($Y2)

I have two questions from this:
1. The endogenity test comes out as insignificant (Chi-sq(1) P-val = 0.4083), indicating that the variable is exogenous, which does not sound correct to me and does this mean I cannot use IV?
2. In the IV regressions I do not include the other jobstatus dummies, however is the something I should be doing? When I do not include them the value of jobstatus2 on olsat has a very high p-value (0.465).

Many thanks!