Probit Visualization

[diggetybo]

Hello again,

After learning as much as I could about the features of the eviews model/scenario functions from the help file, I'd like to think I know what I'm doing, but I wanna be on the safe side.

I want to create a visual like the following graph:

Probit Estimation.png (31.92 KiB) Viewed 5225 times

One of the charms of this graph is we can get a visual feel for some of the dynamics of the model. For example the value of x when y = .5, aka the point of equal opportunity. In the above graph, this point occurs at roughly: x=.48.

I'm almost positive that eviews model/scenario can do this, and a mixed graph can make it look nice and pretty like the above image. The graph needs the following components:

Y Axis: dependent variable actuals (0's, and 1's)
X Axis: the desired independent variable
CND: cumulative normal distribution S curve for Y as a function of X

I'm not sure how to integrate all these components in eviews. Is a model needed, or can I use something else right after estimation of the equation? I ask because I tried it the model way with many datasets but my visuals do not look "S" shaped at all. Obviously every data set will be different, but I'm suspecting that eviews model may not be the best tool for this kind of graph. Or am I wrong?

Thank you for reading

[EViews Glenn]

Could you post your workfile along with a description of the one you want to examine on the X axis. And where do you wish to evaluate all of the other variables? At their actual values, or at some other point like means?

[diggetybo]

I was just using randomly generated data. However, if it's of any help, here is the link to the full lecture.

http://www.columbia.edu/~so33/SusDev/Lecture_9.pdf

They first talk about this graph on slide 28.

As to your other question about whether other variables are estimated at their means or not, I'm not entirely sure. If I had to guess, I'd say the professor is plotting y and an x against eachother in actuals form -- seeing as marginal effects are not talked about until slide 68 (40 slides after the graph). However, clearly it's not your typical x - y actuals scatter plot. The probit CND curve is in the mix too.

[PS], it's a lengthy pdf, I'm not expecting people to read it cover to cover or especially I dont want people to feel like they're doing my work for me. Basically, if someone from the community recognizes what's going on in the graph and is able to put it in eviews terms - that's what I'm really after. Because, like I said, I tried my way, and I'm starting to think I'm approaching this from the wrong way.

[EViews Glenn]

It's not generally a natural computation for EViews as it combines the observed data with fits for observations that are potentially not in the workfile. That is to say, you could have 15 observations (0, 1)s, and if you tried to plot the response curve, 15 would presumably not be smooth enough.

For specific data, especially with larger numbers of (0,1) values and a description of what you want to where you want to evaluate the other data, I think we could come up with something. But it would help if you had a more concrete example. As I said, it's not something that we have built-in.

[diggetybo]

Ok, thanks for looking into it.

I have included a more concrete example. I run a probit estimation called "eq01." And tried to get part of the above visualization using a graph object called "visual," these may be found in the attached workfile.

In the estimation there is a very statistically significant relationship between the y (in labor force) and the x (education). However when I did:

Code: Select all

graph visual.scat educ inlf

The pattern looks totally symmetrical and misleading. But like I said, I may have missed something. Have a look, when you get the chance.

mroz(visual).wf1

(169.08 KiB) Downloaded 240 times

[diggetybo]

I will just reiterate briefly,

After taking a break, and coming back to the problem after a few weeks I was hoping that having a fresh look would help me see something I may have been missing.

However, I'm still confused as to why my scatter graph looks the way it does. If you look inside equation "eq01" the coefficient on educ is .1309, p=.0000. However, the scatter of the actuals against educ seems a little counter intuitive. I'd expect there to be more clustering of dots around low x & low y, and high x and high y.

I'm assuming that's the method used to create the visual that I included in the first post of this thread, although that may be part of the reason it's not working, is I may be assuming too much. However that visual doesn't look too terribly complicated, just a scatter with the probit s shaped cumulative, from what I can tell.

If anyone else has made headway, feel free to share