Bad control

[Shreya Jain]

My dependent variable is Edu_hypo.
it is defined as Edu_hypo= 1 if wife_edu> hus_edu, 0 otherwise.
So if I am looking at the determinants of Edu_hypo from the wife's side, controlling for wife's education herself would be a bad control? Or is it important to control for that variable?

[George Ford]

leave it out.

if hus_edu is a categorical variable, which I suspect it is, then the results will be odd. You'll get a coefficient of 1 for any level of education equal to or exceeding a hs education and no standard errors.

sysuse auto, clear

tab rep78
g drep78 = rep78>=4

reg drep78 i.rep78

[Shreya Jain]

So mine is something like this using your example where disp is edu_hypo and assume length is wife_edu and Displacement is hus_edu :
sysuse auto, clear
gen disp= (length>displacement)
logit disp length

it does come up with an output

[George Ford]

That is not what you are asking. The disp variable is

g disp = displacement>200
logit disp displacement

In any case, you are essentially regressing y on y.

[Shreya Jain]

Dear Sir,
Maybe I was not clear earlier, but my variable is

gen edu_hypo= (wife_edu>hus_edu)

where both of them are categorical variables that show the highest education level of wife (wife_edu) and husband (hus_edu).
I want to study the determinants of edu_hypo i.e. educational hypogamy where I am trying to understand why women are marrying down in terms of education,
So I wanted to know if I should have wive_edu as one of my control variables in my regression.

[George Ford]

Oh, I see. So the dummy is wife educ > husband educ.

I think it's ok to include it. You'd think that the condition is more easily satisfied the higher the education level of the wife. Including both education levels as regressors may be a problem.

[Shreya Jain]

Dear Sir George Ford ,

I have a follow-up question regarding the inclusion of the wife’s education as a control variable when my dependent variable (educational hypogamy) is constructed using that variable [gen Edu_hypo= (wife_edu>hus_edu)]. Earlier, you mentioned that this inclusion is acceptable, but I want to understand how to justify that it is not problematic.

I have received feedback that including the wife’s education might be a bad control since it can itself be an outcome influenced by other covariates, such as parental education, socio-economic status, and age etc.

[Hemanshu Kumar]

I agree that the wife_edu variable will likely be endogenous. It's inclusion might help "explain" edu_hypo -- for instance, there is a simple mechanical way in which edu_hypo will go to zero as wife_edu becomes very low, and so the two are going to be positively correlated. But in a causal model you would want to find an instrument for wife_edu or use some other method of getting around the endogeneity issue.

[Shreya Jain]

Thank you Hemanshu Kumar.

[Carlo Lazzaro]

Shreya:
your regression specification may suffer from latent variable-led endogeneity, as some women' characteristics may influence, on average and other things being equal, the level of education and husband's education selection.

[Felix Bittmann]

I think, as long as the purpose of your study is of Exploratory nature you can throw in as many variables as you like. Then, a dominance analysis is what you could compute. If you want to estimate treatment effects, however, this is a very different story.

[Shreya Jain]

My research goal is to examine the effect of parental education on the likelihood of their children entering hypogamous marriages (where the wife is more educated than the husband). However, I am facing a challenge:

• When I do not include the wife's education (wife_edu) in the model, I observe that women’s parental education is positively associated with hypogamy, while men’s (husband’s) parental education shows a negative association.
• I understand this might be because the parental education variables are partly capturing the effect of their children’s own education, thus masking the true direct effect of parental education on hypogamy.

Could you please advise on how to properly model these relationships to isolate the effect of parental education on the likelihood of hypogamous marriage?

[Felix Bittmann]

You should also think about the at risk population. Women with low education dont have the option to enter hypogamy and should be excluded.
After that, I would have a binary indicator for hypogamy or not. I would include wife's education and wife's parental education. Nothing of the husband as this is already determined in the binary outcome. Why would you include husband's parental education?

[Shreya Jain]

Thank you for raising this important point Felix Bittmann . I agree that women with no education cannot realistically enter hypogamy, since they cannot be “less educated” than their husbands. In that sense, the correct at-risk population should exclude them, focusing only on women for whom hypogamy is a feasible outcome.

Regarding the inclusion of husband’s parental education, my initial rationale was that hypogamy reflects not just individual education levels but broader marital sorting across families. A groom’s parental education may influence his family’s bargaining power or willingness to accept a hypogamous match, independent of his own schooling. However, I recognize that because the binary outcome already incorporates the husband’s education, this inclusion needs to be carefully justified.