[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

Re: SV: SV: st: From probit to dprobit to interpretation

From   "Austin Nichols" <[email protected]>
To   [email protected]
Subject   Re: SV: SV: st: From probit to dprobit to interpretation
Date   Mon, 14 Jan 2008 11:46:19 -0500

I gather you have subsumed the additional regressors into the
"constant" in your example. With additional regressors, I think you
should predict at the values of the additional regressors observed in
the data, rather than at the mean. The value -3511 is one measure of
the ATT (putative effect of X on those with X=1) and twice that is one
measure of the ATE (the effect X would have on the whole sample if X=1
for everyone) in your example.  Neither is a particularly defensible
estimate in the presence of other regressors of note, IMHO, especially
with observational (non-experimental) data.  Others on Statalist and
within Statacorp have different opinions.  Are you trying to merely
replicate output from -dprobit- or do you have a specific
counterfactual hypothesis in mind? The specific hypothesis should
motivate the method.

On Jan 11, 2008 3:08 PM,  <[email protected]> wrote:
> Thank you, Austin. I am aware that tab y x will produce the same result. I was just trying to work out the logic going fra a probit to a dprobit reporting the dicrete change of a dummy. Then I wanted to find out how to apply these probabilities to do predicted counts. I have a sample of 1000000, where there are 500000 incidences where x=1. What I am curious about is whether the discrete change og 0.7 percentage point reduction could be apllied directly to the entire sample, or if it should be applied the 500000 incidences where x=1. That is, are this correct?
> >di round(5e5*normal(-2.33))
> >4952
> >di round(5e5*normal(-2.33-0.431))
> >1441
> >di round(5e5*(normal(-2.33-0.431)-normal(-2.33)))
> >-3511
> Or is this correct?
> di round(1e6*normal(-2.33))
> 9903
> di round(1e6*normal(-2.33-0.431))
> 2881
> di round(1e6*(normal(-2.33-0.431)-normal(-2.33)))
> -7022
> The way I understand this is that the discrete change should be applied to the entire sample. I have many other regressors, but these I set to equal their mean in a dprobit model. Though, the same logic should apply to the interpretation of the discrete change in the dummy variables.
> Best wishes,
> Alexander
*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index