[R] Problem with ordered logistic regression using polr function.

[Thomas Lumley]

On Wed, 19 Jul 2006, Debarchana Ghosh wrote:

> Hi,
> I'm trying to fit a ordered logistic regression. The response variable
> (y) has three levels (0,1,2).
> The command I've used is:
>
> /ordlog<-polr(y~x1+x2+x3+x4, data=finalbase, subset=heard, weight=wt,
> na.action=na.omit)
> /

> (There are no NA's in y but there are NA's in X's)
>
> The error I'm getting is:
> /Warning messages:
> 1: non-integer #successes in a binomial glm! in: eval(expr, envir, enclos)
> 2: design appears to be rank-deficient, so dropping some coefs in:
> polr(right.ans ~ S107 + children + work + rel + media + V013 +
> /

Those are not errors. They are warnings (and I told you yesterday that the 
first one would happen and would be harmless with probability weights). 
The second warning may indicate a real problem or not: you can tell by 
looking at which coefficients have been dropped.

Also, if that is the warning message then the polr() call you used is not 
the one you said you were using.

> Also, if I write
> /summary(ordlog)/
>
> I'm getting the following error:
>
> /Re-fitting to get Hessian
>
> Error in if (all(pr > 0)) -sum(wt * log(pr)) else Inf :
>        missing value where TRUE/FALSE needed
>
> /Could anybody point out the problem?

You can specify Hess=TRUE in the original polr() call so that refitting 
would not be necessary.  This is sensible anyway if you know you are going 
to be computing standard errors.

 	-thomas

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle