Ordered probit model

[marwa]

Dear all,
I am using an ordred probit model, in which my latent variable is buy, sell or no transaction.
I assigned +1 to buy, 0 for no transaction , and -1 to sell.
I ran the regression with latent variable as my dependent variable against a list od independent variables, and it went OK with no errors.
My question is that how can I interpret a negative coefficient. Does this mean that an increase in this variable leads to sell, or not to buy????
Also, how can I calculate the marginal effects in Eviews 6.
Thank you
Marwa

[trubador]

Since your dependent variable has three distinct categories, you might have estimated two threshold/limit values. The latent variable (or index function), however, can only be constructed after the estimation of coefficients of your explanatory variables (i.e. the linear combination). You can either do this by yourself or choose Make Model option under the Proc menu in the equation dialog box and solve it. If you do the latter, you will see that probability of observed values in your dependent variable for each category is computed with respect to this latent variable as well as the estimated limit points. Therefore, change in any one of the variables will lead to a change in related probabilities. Marginal effects are simply the derivative of these probabilities with respect to explanatory variables and can easily be calculated via chain rule. Suppose that the name of your calculated latent variable and estimated limit values are latent, limit_0 and limit_1, respectively. You can compute the marginal effects of your first variable (e.g. its coefficient is c(1)) as follows:

Sell: -@dnorm(limit_0 - latent)*c(1)
Hold: -(@dnorm(limit_1 - latent) - @dnorm(limit_0 - latent))*c(1)
Buy: (@dnorm(limit_1 - latent))*c(1)

You should also compute the marginal effect of each explanatory variable in its sample average and intrepret these values instead of estimated coefficient values. The value of marginal effect (say -0.75) means that a unit change in the regarding explanatory variable will lead to that amount (75%) of decrease in the probability of related category in the dependent variable (say Hold).

[EViews Glenn]

Trubador's answer is a good one. I will merely make two additional points:

1. Where you evaluate the LATENT in trubador's expressions matters since the derivatives of the probabilities are nonlinear functions (@dnorm). Often people will evaluate at mean values for X (which I'm not particularly crazy about since there is no particular reason that evaluating the nonlinear expressions at a "mean" individual is interesting). Sometimes people will evaluate at the values for XB for each individual in the sample (as implied by trubador). More generally, you should evaluate at a set of values that you find interesting for whatever reason.

2. Only the lowest and highest ordered categories have derivative signs that are unambiguous. Derivative signs for all others (in this case, the middle category) depend on the shape of the normal density in the region of the evaluation points.

[life]

Hi,

I'm running an ordered probit model as well and have come across an error message which I'm not sure how to resolve. When I add GDP, Unemployment rate and Inflation to my variable list, I get the following message: Unable to compute automatic estimates of starting values. Use user-supplied starting values.

I clicked on this option in the option menu of the formula box, but now get a series of NA's.

Any suggestions or solutions?

Thanks

[EViews Glenn]

FIrst off, what version of EViews are you using and what's the build-date (Help/About EViews)? I know that we've made some improvements in the automatic determination of starting values in recent versions of EViews that may help.

Apart from that, there is an option in estimation to use the starting values that are in the coefficient vector C. Edit the coefficient values in the C vector to be something plausible, then go to the options page inyour equation estimation dialog, select the user-supplied option.

[File]

Dear all, I am just starting to use Ordered probit model for ratings (rating transition matrix) depending on several variables (for exaple "country" and "industry" that can count from 1,...,n).

So, I need help from the very beggining, i.e. firstly, how do I prepare the input file? I actualy have several matrices depending on these conditions.

Thank You very much!

[EViews Glenn]

It's a bit hard to answer your question directly without more information about how your data are being held. It's not clear, for example, what you mean by "matrices".

Having said that, there are some general comments that may be of use...

You want a standard EViews workfile with series representing the dependent and explanatory variables. The dependent should be coded with a set of distinct values for the response. The values for the response should be ordinal, and need not be cardinal. The obvious choice is 1, 2, 3, ..., but you are not restricted to those values.

[File]

Sorry for not being very clear in presenting the problem. Here it is: What I have are observed transitions from state i to state j on quarterly basis and I have calculated empirical transfer probabilities from this. Ofcourse they differ in specific conditions. All together I end up with:

a. empirical probabilities for every transfer (from i to j)
b. "condition one": initial state (1,2...m)
c- condition two: specific characteristic of the subject (dummy taking values 0,1...n)
d. condition three: characteristic of specific date (quarter), also dummy (taking values 0...k)
e. resulting state: j that takes values 1,2...m like the initial ones.

The ordered probit should be used here to estimate probabilities of ending in state j depending on these known conditions (similar to the example in Greene, p. 739, fifth edition). I am not sure how to prepare the input file (balanced panel? For each initial state?) and how to model dependent ordered variable (probablitiy of ending up in the state j), i.e. estimating parameters and cut-off points.

Thank You.

[EViews Glenn]

I think my advice from earlier still holds. The only thing that appears to be different is that you also want to condition on the initial state. My initial thought is that you can do this by having yet another variable which indexes the initial state, then estimating a model on each subsample of initial states.

[michelleli]

Hi,
I am using the following method to calculate the marginal effect for a ordered probit model. I have "solved" the model. However, I have a few problems.
1. The output only shows the results of the latent variables and I couldn't find limit_0 and limit_1.
2. where should I type in the command of Sell: -@dnorm(limit_0 - latent)*c(1), Hold: -(@dnorm(limit_1 - latent) - @dnorm(limit_0 - latent))*c(1), Buy: (@dnorm(limit_1 - latent))*c(1).
3. With regarding to the latent variable used to calculate marginal effect, should I use the fitted linear index value or the fitte probability of falling in each category(eg.,the probability of falling in the "hold" category ).

Please help!
Thanks

Since your dependent variable has three distinct categories, you might have estimated two threshold/limit values. The latent variable (or index function), however, can only be constructed after the estimation of coefficients of your explanatory variables (i.e. the linear combination). You can either do this by yourself or choose Make Model option under the Proc menu in the equation dialog box and solve it. If you do the latter, you will see that probability of observed values in your dependent variable for each category is computed with respect to this latent variable as well as the estimated limit points. Therefore, change in any one of the variables will lead to a change in related probabilities. Marginal effects are simply the derivative of these probabilities with respect to explanatory variables and can easily be calculated via chain rule. Suppose that the name of your calculated latent variable and estimated limit values are latent, limit_0 and limit_1, respectively. You can compute the marginal effects of your first variable (e.g. its coefficient is c(1)) as follows:

Sell: -@dnorm(limit_0 - latent)*c(1)
Hold: -(@dnorm(limit_1 - latent) - @dnorm(limit_0 - latent))*c(1)
Buy: (@dnorm(limit_1 - latent))*c(1)

You should also compute the marginal effect of each explanatory variable in its sample average and intrepret these values instead of estimated coefficient values. The value of marginal effect (say -0.75) means that a unit change in the regarding explanatory variable will lead to that amount (75%) of decrease in the probability of related category in the dependent variable (say Hold).

[EViews Glenn]

The limit_i are the limit coefficients of the ordered model.

The evaluation index is the linear index, but as in the discussion above you can evaluate at the sample values or anywhere you like. It will matter since the expression is nonlinear.

[michelleli]

The following explanation suggests to compute the marginal effect of each explanatory variable in its sample average. How can I do that in Eviews 6？


Cheers

Since your dependent variable has three distinct categories, you might have estimated two threshold/limit values. The latent variable (or index function), however, can only be constructed after the estimation of coefficients of your explanatory variables (i.e. the linear combination). You can either do this by yourself or choose Make Model option under the Proc menu in the equation dialog box and solve it. If you do the latter, you will see that probability of observed values in your dependent variable for each category is computed with respect to this latent variable as well as the estimated limit points. Therefore, change in any one of the variables will lead to a change in related probabilities. Marginal effects are simply the derivative of these probabilities with respect to explanatory variables and can easily be calculated via chain rule. Suppose that the name of your calculated latent variable and estimated limit values are latent, limit_0 and limit_1, respectively. You can compute the marginal effects of your first variable (e.g. its coefficient is c(1)) as follows:

Sell: -@dnorm(limit_0 - latent)*c(1)
Hold: -(@dnorm(limit_1 - latent) - @dnorm(limit_0 - latent))*c(1)
Buy: (@dnorm(limit_1 - latent))*c(1)

You should also compute the marginal effect of each explanatory variable in its sample average and intrepret these values instead of estimated coefficient values. The value of marginal effect (say -0.75) means that a unit change in the regarding explanatory variable will lead to that amount (75%) of decrease in the probability of related category in the dependent variable (say Hold).

[EViews Glenn]

6 and 7 should be the same in this regard.

[jleather43]

Hi, i'm fairly new to eviews and this is the first time I have attempted to run an ordered probit model. I've followed trubador's explanation as closely as I can but I am still uncertain about a couple of things.

1) When I solve the model, which is my latent variable or index function? I have new series in my workfile that represent each of the categories of my dependent variable and I also have a series that has 'i_' then the name of my dependent variable. Is it one of these?
2) I then type the equivalent formula to Sell: -@dnorm(limit_0 - latent)*c(1)
Hold: -(@dnorm(limit_1 - latent) - @dnorm(limit_0 - latent))*c(1)
Buy: (@dnorm(limit_1 - latent))*c(1)
into the command window and it generates a new series for each category, but how do I calculate the marginal effect? Is it the mean of each of these new series?

Many Thanks

[kamarinette]

My problem is this
my dependent variable takes three values​​: 1, 2 and 3. I have several variables, some of which are binary.
1) I first estimate my model and I find no significant limits point. can I still estimating the different probabilities?
2) how to calculate the marginal effects of each variable. we must use x bar, but how?

help please