Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: AW: Fitted probabilities using prvalue for logit model
From
Steve Samuels <[email protected]>
To
[email protected]
Subject
Re: st: AW: Fitted probabilities using prvalue for logit model
Date
Fri, 16 Jul 2010 11:01:46 -0400
I don't know what you mean by "determine the relative importance of my
additional dummy variables relative to the benchmark model with its
explanatory variables?" But in this case -prvalue- is obviously not
working for you.
How are you measuring "importance"?
If you mean "significance", have you tested the joint significance of
the two variables with -test-? (Adding variables will always increase
the log-likelihood, so "improvement" is not a guide). If the
criterion of importance is "predictive accuracy", then compare ROC
curves for the two models with -roccomp-. Unfortunately, the ROCs for
both models will be systematically optimistic, but the differences
could still be revealing. For better accuracy, some kind of
cross-validation approach is needed.
For cross-validation approaches, see:
http://www.stata.com/statalist/archive/2008-02/msg00686.html
An unreferenced Stata program for cross-validation is contained in:
http://www.mail-archive.com/[email protected]/msg82508.html
There is also a literature on "proportional reduction in error"
approaches, including partial r-squares. See: Agrestic, Analysis of
Categorical Data, 2nd Ed (2002) Wiley, Chapter 6. Measures of
r-square based on the log-likelihood are difficult to interpret (p.
227). A Google search will turn up many references.
(By the way, -prvalue- is not an official Stata command. I presume it
is user-written. Please, as the FAQ request, give references for all
the non-Stata commands you use.)
Steve
On Fri, Jul 16, 2010 at 9:54 AM, Marc Michelsen
<[email protected]> wrote:
> Steve,
>
> of course there are four possible combinations -- however, in my set-up
> there are only three valid combinations. 1/1 is not possible.
>
> Does your statement mean that -prvalue- is not an appropriate measure to
> determine the relative importance of my additional dummy variables relative
> to the benchmark model with its explanatory variables?
>
> Marc
>
> -----Ursprüngliche Nachricht-----
> Von: [email protected]
> [mailto:[email protected]] Im Auftrag von Steve Samuels
> Gesendet: Freitag, 16. Juli 2010 15:18
> An: [email protected]
> Betreff: Re: st: AW: Fitted probabilities using prvalue for logit model
>
> There are four combinations of two dummy variables, not three, so your
> statements don't make sense. The coefficients of the variables, if you
> hold others constant, refer only to the relative associations among
> those four categories, not to any absolute levels. Those are
> determined by the values at which you fix the other covariates and by
> the constant term. .It is well-known that prediction at the means of
> covariates will not even reproduce the mean prediction, which in turn
> is the raw prevalence. It is quite possible that all four
> predictions could be lower than the crude prevalence rate. So, there's
> no reason to expect those predictions to match those of any
> "benchmark" model and a (single?) benchmark probability.
>
> Steve
>
>
>
> On Fri, Jul 16, 2010 at 6:02 AM, Marc Michelsen
> <[email protected]> wrote:
>> Dear all,
>>
>> as I didn't get an answer to my problem below, I am trying to rewrite the
>> question more precisely/generally. The reference for the approach is the
>> following: DeAngelo, H., L. DeAngelo, and R. M. Stulz. "Seasoned equity
>> offerings, market timing, and the corporate lifecycle." Journal of
> Financial
>> Economics 95 (2009), 275-295. I am referring to the table on page 284.
>>
>> I am estimating the fitted probabilities of a logit model at fixed levels
> of
>> the explanatory variables using -prvalue-. I have a benchmark model and
>> therefore also a benchmark probability of the event. Including my two
> dummy
>> variables in a second model specification (improves Peusdo-R2 and Chi2)
>> actually lowers the probability of the event. However, the probability
>> should increase if the dummy variables are coded 0 (dummy 1)/1 (dummy 2).
>> The probabilities are lower in all three possible combinations of the two
>> dummies. Although the coefficients of the logit model show the correct
> signs
>> and are statistically significant for one of the dummy variables.
>>
>> Does anybody has a view on this?
>>
>> Many thanks for considering this posting
>>
>> Marc
>>
>> -----Ursprüngliche Nachricht-----
>> Von: [email protected]
>> [mailto:[email protected]] Im Auftrag von Marc
> Michelsen
>> Gesendet: Donnerstag, 15. Juli 2010 11:12
>> An: [email protected]
>> Betreff: st: Fitted probabilities using prvalue for logit model
>>
>> Dear Statalist users,
>>
>> I am running a logit model to estimate the effect and relative importance
> of
>> market timing and rating concerns on the decision to conduct a seasoned
>> equity offering (panel data).
>>
>> Including my rating concern proxy variables in the regressions improves
> the
>> fit of the logit model (Pseudo-R2 and Chi2) compared to the standard model
>> (including only market timing and control variables). One of the two
> rating
>> concern proxies (positive rating momentum) is statistically significant at
>> 5% with a marginal effect of -1.7%. The other one (negative rating
> momentum)
>> shows a positive marginal effect but has no significant influence.
>>
>> In order to gauge the relative importance of market timing versus rating
>> concerns, I am trying to obtain predicted probabilities of conducting a
>> seasoned equity offerings (SEO) in a given year. Therefore, I am using the
>> "prvalue" command to calculate the probabilities at representative values
> of
>> the explanatory variables (control variables at sample means, good vs.
> poor
>> market timing opportunities). Neutral market timing opportunities
> translates
>> into a SEO probability of 5.2%, which is comparable to the study von
>> DeAngelo/DeAngelo/Stulz (2009) p. 284. But if I measure the probabilities
>> for positive, negative and neutral rating momentum (the other explanatory
>> variables are set equal to the former model specification), the
>> probabilities are always lower compared to the benchmark model (3.8% /
> 5.0%
>> / 4.9%). While it is reasonable to assume that positive rating momentum
>> lower the SEO probability, the results for the two other rating variables
>> are surprising.
>>
>> Obviously, this weakens my hypothesis that rating concerns are one of the
>> drivers of seasoned equity offerings.
>>
>> Does anybody have an idea why the fitted probabilities are lower in all
>> three cases although the model fit is improved if I include the respective
>> explanatory variables?
>>
>> Many thanks
>> Marc
>>
>> *
>> * For searches and help try:
>> * http://www.stata.com/help.cgi?search
>> * http://www.stata.com/support/statalist/faq
>> * http://www.ats.ucla.edu/stat/stata/
>>
>>
>> *
>> * For searches and help try:
>> * http://www.stata.com/help.cgi?search
>> * http://www.stata.com/support/statalist/faq
>> * http://www.ats.ucla.edu/stat/stata/
>>
>
>
>
> --
> Steven Samuels
> [email protected]
> 18 Cantine's Island
> Saugerties NY 12477
> USA
> Voice: 845-246-0774
> Fax: 206-202-4783
>
> *
> * For searches and help try:
> * http://www.stata.com/help.cgi?search
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
>
> *
> * For searches and help try:
> * http://www.stata.com/help.cgi?search
> * http://www.stata.com/support/statalist/faq
> * http://www.ats.ucla.edu/stat/stata/
>
--
Steven Samuels
[email protected]
18 Cantine's Island
Saugerties NY 12477
USA
Voice: 845-246-0774
Fax: 206-202-4783
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/