causal effect in panel data

[Sandra Bloem]

Hi there,

I am struggling to understand something relating to causal effects with panel data.

I have an unbalanced panel data set on the individual level (t=2008-2017).

I need to estimate three causal (or as causal as possible) effects

1. having a job -> mental health
2. finding a job -> mental health
3. losing a job -> mental health

What models should I use to estimate these 3 different effects?

One thing I don't understand is that FE only uses within-unit variation. Could FE then estimate the 1st effect? Since it's time demeaning the data, but at the same time it feels wrong.

Finding/losing a job needs a difference in employment between 2 consecutive years. Let's say an individual has 2 observations: in 2010 he is not employed, in 2014 he is. Could you use this individual in the estimation of the 2nd effect? He goes from having no job to having a job, but it's unclear whether he became employed in 2014 or for example in 2011. If he became employed in 2011, including this individual would actually estimate the effect of becoming employed in year t (2011) on mental health in year t+3 (2014). Am I seeing this correctly? And is this something you should take into account, or does it not really matter?


Thank you in advance (:

[Tom Scott]

Sandra Bloem the first this I suggest doing is showing summary statistics of your outcomes over time. A fixed effects regression gets you closer to causal attribution because you can remove the effects of time-stable covariates on your time-varying relationships and you can establish the correct temporal order (ie job variables change before mental health changes), but whether you can use it will depend on your sample size and the amount of variation over time in change in job status. It also has strong assumptions like that a change in the treatment (job status) does not impact one's future likelihood of experiencing the treatment. Because employers factor in employment history, this seems unlike to hold up. Overall, I would give up on the idea of establishing causality when using observational research designs. You should definitely do what you can to remove omitted variable bias and establish correct temporal order, but it is incorrect to think that this gets you any more than an association in the end. The first thing you should do and that will help others to give you advice is to look at your data and show the forum by means of descriptive statistics.