Marginal effect after double hurdle

[Gizem Cetin]

Hi,

I'm using Stata/IC 14.0 and run a double hurdle model in order to test time use decision of individuals.
For this model, I use this command:
churdle linear lnavperCC $varlist if (hhr_yaki==1 | hhr_yaki==2) & gender==1, select ( $varlist) ll(0) vce(robust)

After that, I know that I cannot interpret the coefficients of the double hurdle model, I can only interpret their significance level and positive/negative impact.

Then, I want to calculate marginal effects at means, but I get different significance level for some of the variables compared to the double hurdle model.

I could not figure out what's wrong. Any help will be appreciated.

Thanks.

[Enrique Pinzon (StataCorp)]

Hello Gizem,

There is nothing wrong when you obtain a different significance level from the results of the point estimates and the results of the marginal effects. The way to think about this is that they are two different statistics. For example, think about a function F(X*B) evaluated at the mean of X. Even if B is positive and significant if the mean of X is close to zero F(X*B) evaluated at the mean of X might be zero.

[Gizem Cetin]

Enrique,

Thank you for your answer. Then, how should I interpret them? Let's say I find "age" significant at double hurdle model, but not at marginal effect estimation.

[Enrique Pinzon (StataCorp)]

Hello Gizem,

I would suggest that you forget about the coefficients you estimated if what you are interested in is the effect. Think of the coefficients you estimated as auxiliary parameters. If what you are interested in F(X*B), B is just an input. margins will use the coefficients and standard errors of your estimation to obtain F(X*B) and its standard errors.

In non-linear models with multiple equations it is rarely the case that the coefficients tell us a clear-cut story or that they are of any particular interest. In your example, the variable age could be part of the corner equation (select() ), the interior equation y age ..., and potentially on the heteroskedasticity model het() . Rather than thinking about the coefficients, non-linear models are useful to ask questions about conditional expectations, conditional probabilities, arbitrary functions of the parameters (margins, expression(my_weird_function)), etc.

[Gizem Cetin]

Thank you so much, Enrique!