R help - Feb 2011 - error in optim, within polr(): "initial value in 'vmmin' is not finite"

Hi all.  I'm just starting to explore ordinal multinomial regression.  My
dataset is 300,000 rows, with an outcome (ordinal factor from 1 to 9) and five
independent variables (all continuous).  My first stab at it was this:

pomod <- polr(Npf ~ o_stddev + o_skewness + o_kurtosis + o_acl_1e +
dispersal, rlc, Hess=TRUE)

  And that worked; I got a good model fit.  However, a variety of other things
that I've tried give me this error:

Error in optim(s0, fmin, gmin, method = "BFGS", ...) : 
  initial value in 'vmmin' is not finite

  This occurs, for example, when I try to use the method="probit"
option of polr().  It also occurs when I try a regression involving
interactions, such as:

pomod <- polr(Npf ~ o_stddev * o_skewness * o_kurtosis * o_acl_1e *
dispersal, rlc, Hess=TRUE)

  I have good reason to believe that interactions are important here, so I'd
very much like to be able to fit such models.  I have been doing that
successfully with logistic regression (considering my outcome variable to be
binary, either "1" or "2-9") using glm(), but now using
polr() it gives me this error.  I've searched Google and the R lists for
information about this error, and while I did find a couple of other people
asking about it, I didn't find any advice about what to do about it that I
can apply to my situation.

  I'd be happy to share my dataset with anyone willing to help me on this,
but 300,000 rows is a bit large to include in this email.  :->

  Thanks!

Ben Haller
McGill University

An update for the benefit of the list/posterity: I resolved this issue by
switching over to using the lrm() function of package rms.  It seems to pick
better starts, or something; in any case, it has been able to converge on a
solution for every model I've tried, although for the most complex ones I
needed to raise maxit (maximum iterations) above the default of 12 slightly. 
The lrm() function does not support interactions higher than third-order, and it
does only logistic regressions, not probit or other types, so it does have its
drawbacks; but it has solved my difficulties quite nicely.  Just in case anybody
cares.  :->

Ben Haller
McGill University


On Feb 16, 2011, at 5:41 PM, Ben Haller wrote:
>  Hi all.  I'm just starting to explore ordinal multinomial regression. 
My dataset is 300,000 rows, with an outcome (ordinal factor from 1 to 9) and
five independent variables (all continuous).  My first stab at it was this:
> 
> pomod <- polr(Npf ~ o_stddev + o_skewness + o_kurtosis + o_acl_1e +
dispersal, rlc, Hess=TRUE)
> 
>  And that worked; I got a good model fit.  However, a variety of other
things that I've tried give me this error:
> 
> Error in optim(s0, fmin, gmin, method = "BFGS", ...) : 
>  initial value in 'vmmin' is not finite
> 
>  This occurs, for example, when I try to use the method="probit"
option of polr().  It also occurs when I try a regression involving
interactions, such as:
> 
> pomod <- polr(Npf ~ o_stddev * o_skewness * o_kurtosis * o_acl_1e *
dispersal, rlc, Hess=TRUE)
> 
>  I have good reason to believe that interactions are important here, so
I'd very much like to be able to fit such models.  I have been doing that
successfully with logistic regression (considering my outcome variable to be
binary, either "1" or "2-9") using glm(), but now using
polr() it gives me this error.  I've searched Google and the R lists for
information about this error, and while I did find a couple of other people
asking about it, I didn't find any advice about what to do about it that I
can apply to my situation.
> 
>  I'd be happy to share my dataset with anyone willing to help me on
this, but 300,000 rows is a bit large to include in this email.  :->
> 
>  Thanks!
> 
> Ben Haller
> McGill University
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.