cluster/village fix effects

[Chanda Moon]

Dear Stata users,

I am trying to estimate the impact of water quality on child health in 3 Asian countries. I am using 2SLS as we assume X is endogenous. Code I am using is

ivreg2 chdiarrhea (hh_ecoli= imp_wtr ) child_agemo child_male HH51 mo_age mo_edu hh_edu hh_sex rural i.windex5 dist_slop_avg dist_rain dist_pop dist_ temp dist_road i.country#i.year, cluster(cluster)

Do i need to ask do I need to add more fixed effects as my IV is district level and I use country# year fix effects.
If I try to add i.district or i. province, i have colliniarty and also for i.district my IV drops.
Moreover, i try to use i.cluster (village) in OLS bit i have colliniarity issue and 1 country and year is dropped. In ivreg2 command adding i.cluster, state keeps on working and does not generate any result.

Can i proceed with current model without adding other fix effects as i have added other district level characteristics as controls (e.g temperature, rainfall, economic status, population, area, slope etc)



Thank you!

[Maarten Buis]

There is no statistical police that will put you in jail if you get something wrong. So in a trivial sense the answer to your question is: yes you can.

Moreover, "wrong" is not well defined. A model always involves tradeoffs, and what tradeoff is acceptable and what tradeoff is not acceptable depends on all the details of what you want to do with your model. Even if we had all the details, people can have legitimate different views on what is acceptable and what not. Remember the definition of a model is that it is a simplification of reality, and simplification is just another word for "wrong in some useful way". So a certain amount of "wrongness" is something you have to accept from a model. The question is how much wrongness is too much, and there is no general answer to that question. (Well, there is, but not one that is both general and useful...)

That sounds depressing, does it not? It is not that bad. The trick is that you should not expect a single study to find "the Answer", instead you should think of a study as part of a much larger effort by a sub-discipline. One researcher approaches the problem this way, another that way, and so on. Each study flawed in its own way. But eventually the collective effort of all those researchers should make progress towards the answer. So yes, you should be worried about doing something wrong, but you should channel that worry by adding a discussion of possible limitations to your study. That worry should not prevent you from making a potentially useful contributions to your field. Often "better" is the enemy of "good".

On a more practical level: if you have fixed effects on the village level then you already absorbed all variance at the country level. So by including village fixed effects, you already (implicitly) included province and country fixed effects. So the fact that country fixed effects are dropped when you included village fixed effects is not a problem.