The effect of parent's education on children's education: endogeneity problem

[Mario Bernasconi]

Dear Statalisters,

I want to measure the causal impact of parent's education on children's education and the problem is that parent's education is likely endogenous (parent's education is correlated with parent's genes and both of them influence children's education). In previous studies three strategies have been used: data on adoptees, data on twin couples, IV strategy. I think that IV is best to solve endogeneity but it only measure the Local Average Treatment Effect, for example some studies use a school reform as instrument but then you only measure the impact on those whose parent's education would have been different without the reform. I wanted an instrument that affect the schooling of all parents. My idea is to use grandparents education as instrument. Relevance for the instrument is quite trivial. In order to assure that the instrument satisfy the exclusion restriction I would use a dataset where the parents have been adopted as children, so that there is no correlation between the genes of the grandparents and the genes of the nephews.

Do you think that this strategy could work?

Thanks a lot

[daniel klein]

Sounds interesting. Why not just use adopted children? Why go through grand parents?

Best
Daniel

[Mario Bernasconi]

Thank you for your answer. It is because I have no information regarding the method of assignment of the children to families, and I fear that using only the data on adoptees there could be a bias if there is a correlation between biological parents and adoptive parents genes. Does it make sense?

Best

Mario

[daniel klein]

Would that argument not apply to grand parents and parents as well? I am not deep into this and I have no intentions to get there, but on first sight it seems that the shift in generation does not alleviate the problem.

That said, I would be very interested in the results you obtain by whatever fancy estimation technique you use, especially in comparison to the "naive" way of just running the plain correlation/regression.

Best
Daniel

[Mario Bernasconi]

mmh yes this is possible. Let's suppose then that I have a dataset with randomply assigned adoptees. I would say that IV with grandparents is still better because I can basically control for all the unobserved determinants of children's education that are correlated with parent's education (such as child rearing talent of the parents) while if I run an OLS with adopted children then genes don't raise any problem, but other variables could. Is it reasonable?

Best,
Mario

[wbuchanan]

I'm definitely not an expert on genetics, but given the amount of genetic material that humans share in common with trees, other animals, each other, etc... I would suspect that the geneology argument would be difficult to work around. More importantly, the genetic information argument assumes that the environment plays no role in gene expression. I've not seen much of the literature, but would assume you would not find very large genetic effects that can successfully parsed out from behavioral and environmental effects.

[daniel klein]

I would say that IV with grandparents is still better because I can basically control for all the unobserved determinants of children's education that are correlated with parent's education (such as child rearing talent of the parents)

Would this require the assumption that grand parents child rearing is not correlated with parents child rearing? As a sociologist I doubt that this is plausible.

Edit:

Also as a sociologist, I am starting to ask myself what exactly it is you are trying to estimate here? There must be some way through which education of parents affect their childrens' education. One plausible mechanism might be the style of child rearing. So this might not be something you want to "control" for in the first place.

Best
Daniel

[Min Zhang]

I agree with daniel klein .


Also, a number of studies find that grandparents' class or education has direct impacts on grandchildren's education over and above parental effects.