2
$\begingroup$

I am using a double model with log transformed independent variables and have calculated average partial effects. Now I am now not sure how to interpret the coefficients; particularly those from the first stage which is supposed to be capturing the probability of participation using a probit model.

$\endgroup$
1
  • $\begingroup$ Did you leave out the word hurdle? $\endgroup$
    – dimitriy
    Commented Mar 1, 2014 at 20:32

1 Answer 1

2
$\begingroup$

There's some hand-wavy math in what follows, but hopefully the emphasis on intuition over rigor is worthwhile.

Your first stage model is essentially a probit: $$\Pr(y=1 \vert x)=\Phi(\alpha + \beta \cdot \ln x),$$ where $\Phi()$ is the standard normal CDF. You want to know the marginal effect of $x$. Taking the derivative with respect to $x$ and applying the chain rule yields $$\frac{\partial\Pr(y=1 \vert x)}{\partial x}=\frac{\partial \Phi(\alpha + \beta \cdot \ln x)}{\partial x}=\varphi(\alpha + \beta \cdot \ln x)\cdot\frac{\beta}{x},$$ where $\varphi()$ is the standard normal PDF. This is approximately equivalent to $$\frac{\Delta \Pr (y=1 \vert x)}{100 \cdot\frac{\Delta x}{x}}=\frac{\varphi(\alpha + \beta \cdot \ln x)\cdot\beta}{100}$$ The left hand side is an elasticity: it's the change in the conditional probability of a success for a 1% change in $x$. This is equal to the partial effect that I am guessing you calculated, divided by 100. Of course, you can divide by numbers other than 100, including 1, which corresponds to doubling $x$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.