How to deal with dependent observations in FE model

[Nicole Kapelle]

Hi all,

I'd like to run a fixed effects model, dependent variable is continuous (income) and explanatory variable is union status, to examine the effect of changes in union status on income within persons. I am using household panel data. At the moment I treat each person as independent although some of my observations are nested in couples. What options do I have to deal with this dependency without dropping observations and using for instance information from just one partner? Any thoughts? Is there a clever Stata gimmick that I could use?

Thanks,
Nicole

[Carlo Lazzaro]

Nicole.
welcome to the list.
If, as it seems, you have panel data, you shoud go -xtreg, fe- (please, see -help xtreg- and related entry in Stata .pdf manual).
Under a panel data structure, treating each observation as independent seems against the panel idea itself (repeated measurements on the same unit across equally spaced time spans).

[Clyde Schechter]

Well, if you are committed to fixed effects, I don't think you have any options. Any variable which represents the pairing of individuals within couples will be constant within person (at least for the majority of the observations not engaged in new pairings or split-ups during the era of the study). So within fixed effects, I think you have to drop one member of the pair.

If, however, you are not averse to random effects modeling, this is really a three-level model: observations within persons within pairs. -mixed- will handle this nicely.

[Phil Bromiley]

Is your income data at the individual or family level? Do you have multiple years of data? If income is at the family level, then you probably need to collapse your data to families. I think we've assumed you have individual-level income and multiple years of data.

If you have separate income for husbands and wives (or whatever pair you have) and you go with xtreg (with fixed effects by family), you may want to cluster errors by family. If you use xtreg,fe you are essentially explaining within-family variation. It throws away all variables that do not vary within-panels. This might not be what you want. Random effects can be done in xtreg or with -mixed- which let's you do much fancier error structures, at the cost of additional assumptions.

[Nicole Kapelle]

Hi all,

Thank you so much for your help. I really appreciate it! I'm new to panel data analysis and think this forum is an excellent way of learning about it.

My income data is at the individual level and yes I have multiple years of data (14 waves). However, income is only measured at 4 waves. I'd like to focus on within individual variation first. So xtreg, fe seems like the way to go, but I will definitely explore using random effects modeling. I think I will drop one member of the pair for now since it seems to be the most logical solution to my dependency issue.

Thanks,
Nicole

[Carlo Lazzaro]

Nicole:
there's no need to delete observations yourself, as Stata applies listwise deletion to all obervations with missing values in any of the -depvar- and/or -indepvars-.
Obviously, this will shrink you sample size.
Without taking its outcome more than a guidance, -hausman- can give you same clues about the best specification for your data (i.e., -fe- or -re-).