Stratified regression

[Poushaly Talukdar]

Hello everyone. I am a young researcher working on a paper on fertility preferences and women's empowerment. The dependent variable is the ideal number of children as reported (categorical) and the independent variables are having a say in the household decisions (categorical). I want to find the effect of having a say in the household decisions on reported ideal no. of children for each state of India, separately for urban and rural setting. Which technique should I use in stata with a cross-sectional DHS dataset?

[Carlo Lazzaro]

Poushaly:
welcome to this forum.
Without further details from your side, I would go:

poisson kids i.havingasay i.rural

Both predictors are two-level categorical variables (see -fvvarlist- notation).

[Poushaly Talukdar]

Hey Carlo!
Thank you for the warm welcome.
Could you tell me why you chose Poisson regression and not a normal linear regression?

[Carlo Lazzaro]

Poushaly:
because your -kids- variable is discrete, as it takes on integers (0,1,2,3...) (aka counts).

[Poushaly Talukdar]

Thank you so much Carlo!
I went forward a step ahead and applied negative binomial regression since the variance was much greater than the mean

[Carlo Lazzaro]

Poushaly:
correct, but see https://www.statalist.org/forums/for...n-vs-nbinomial

[Jeff Wooldridge]

Use Poisson regression. And if you were to interact i.havingasay and i.rural, the equation would be fully saturated. And then linear regression will give the same estimated marginal effects as Poisson regression (or any other method). If you have other control variables, the estimates will differ. Of course, the Poisson coefficients will have a proportionate effect interpretation. But margins will give the same as the linear model estimated by OLS.

It is incorrect to think Poisson regression is deficient when there's overdispersion. It works just fine, but you need to use vce(robust) to obtain valid standard errors.