Theoretic question about covariates choice in logit / probit model

[Gabriele Dini]

Dear Statalist users,

I have a theoretic question for you about the following logit / probit model.

I have data for 2014 and 2012.

The dependent variable is a dummy variable named eligible, which equals to 1 if an Italian Household received, in 2014, the "80 euro bonus" from the Italian Government, 0 othervise.

The model includes a number of variables to control for the probability of receiving the bonus (income bracket, work status), the demographic characteristics (geographical area, education, age) and general economic conditions of the household.

All these variables refers to 2012, where no household received the bonus. (The first year in which the bonus was given was 2014).

So, the model tries to answer the following question: which is, in 2012, the probability for a generic Italian household of receiving the bonus in 2014?

In addition, the model incluedes other variables such as variation in number of emplyees in the household between 2014 and 2012, variation of number of high educated people in the household between 2014 and 2012.

My question refers to the last two variables.

Why these last two variables were included in the model?
What is the meaning of their inclusion?
Is their inclusion appropriate?

[Phil Bromiley]

We normally help people implement their research ideas. It is not clear that we have any particular advantage in explaining why someone else did something, or in particular, why someone included specific variables in a model. This is more a question of your substantive area and theory rather than Stata. They also depend on what question you are trying to answer, which we simply don't know.

So, we can't answer the substantive question why specific variables should be included, or attempt to guess why someone included them.

[Carlo Lazzaro]

Gabriele:
I do share Phil's comment.
Being myself Italian too, I'm under the impression that the 2014 variables were simpy include to bridge the gap between 2012 and 2014.
That said, my guess is that any prediction resulting from such a model is weak at best.