R help - May 2010 - cannot update polr model if I specify "start" parameters

Hi,

I am trying to build an ordinal regression model using polr (from the 
MASS package).  In order to construct an initial model (without an error 
aborting it) in my setting, I must pass in a "start" parameter.  I
would
then like to use the "step" function to remove unnecessary variables 
from the model.  However, this fails with the error message:

 > mod1 <- step(model)
Start:  AIC=42
PathCR ~ Cluster + [[stuff omitted]]
Error in polr(formula = PathCR ~ [[stuff omitted]] :
  'start' is not of the correct length

The underlying problem appears to be that "step" calls
"drop1" which
calls "update" on the formula with an omitted term.  The
"update" fails
with the same error message:

 > update(model, ~ . - Cluster)
Error in polr(formula = PathCR ~ [[stuff omitted]] :
  'start' is not of the correct length

Since "update" extracts the initial function call from the model, it 
apparently passes the "start" parameters along to "polr" to
refit the
model.  Since one variable has been dropped, there are now too many 
parameters in the "start" parameter and the "update" fails.

Does anyone have a way around this difficulty?  (For an individual 
update, I could probably hack the "call" object inside the existing 
model, but I really don't see how to do that using 'step".)

Thanks,
    Kevin

 > sessionInfo()
R version 2.10.0 (2009-10-26)
i386-pc-mingw32

locale:
[1] LC_COLLATE=English_United States.1252
[2] LC_CTYPE=English_United States.1252  
[3] LC_MONETARY=English_United States.1252
[4] LC_NUMERIC=C                         
[5] LC_TIME=English_United States.1252   

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base    

other attached packages:
[1] MASS_7.3-4

loaded via a namespace (and not attached):
[1] tools_2.10.0

Hi Kevin,

The obvious work-around is to start with a model that can be fitted without
giving the start-parameters. If you have to specify the start parameters,
that usually means there is too much parameters or too much dependence in
your data for the algorithm to converge. Meaning that the outcome should not
be trusted, and even when step() would work, it's highly unlikely to give
you a result that actually makes sense.

In case you still want to go ahead, you can step in the other direction :
# senseless data
x <- factor(rep(1:3,10),levels=c(1,2,3),ordered=T)
a <- rnorm(30)
b <- runif (30,1,5)
c <- rpois(30,5)
d <- rbeta(30,2,3)
e <- rbinom(30,6,0.7)

mod <- polr(x~1)
step(mod,scope="x~a*b*c*d*e",direction="both")

In update, you can specify the new start values
update(mod,.~.+a+b,start=c(1,1,2,2))

mod2 <- polr(x~a*b*c*d*e,start=c(rep(1,7),rep(0,26)))
update(mod2,.~.-a-b-c-d-e,start=c(1,1,rep(0,26)))

But the best way is to start from a model that makes sense, and go from
there. Using the diagnostics allows you to adapt the model in such a way it
makes sense and behaves well. Trusting on the step() function for this, is
like giving a dictionary to a monkey and hoping he turns it into a
bestseller.

Cheers
Joris

fortune(241)

On Thu, May 6, 2010 at 5:42 PM, Kevin Coombes
<kevin.r.coombes@gmail.com>wrote:
> Hi,
>
> I am trying to build an ordinal regression model using polr (from the MASS
> package).  In order to construct an initial model (without an error
aborting
> it) in my setting, I must pass in a "start" parameter.  I would
then like to
> use the "step" function to remove unnecessary variables from the
model.
>  However, this fails with the error message:
>
> > mod1 <- step(model)
> Start:  AIC=42
> PathCR ~ Cluster + [[stuff omitted]]
> Error in polr(formula = PathCR ~ [[stuff omitted]] :
>  'start' is not of the correct length
>
> The underlying problem appears to be that "step" calls
"drop1" which calls
> "update" on the formula with an omitted term.  The
"update" fails with the
> same error message:
>
> > update(model, ~ . - Cluster)
> Error in polr(formula = PathCR ~ [[stuff omitted]] :
>  'start' is not of the correct length
>
> Since "update" extracts the initial function call from the model,
it
> apparently passes the "start" parameters along to
"polr" to refit the model.
>  Since one variable has been dropped, there are now too many parameters in
> the "start" parameter and the "update" fails.
>
> Does anyone have a way around this difficulty?  (For an individual update,
> I could probably hack the "call" object inside the existing
model, but I
> really don't see how to do that using 'step".)
>
> Thanks,
>   Kevin
>
> > sessionInfo()
> R version 2.10.0 (2009-10-26)
> i386-pc-mingw32
>
> locale:
> [1] LC_COLLATE=English_United States.1252
> [2] LC_CTYPE=English_United States.1252  [3] LC_MONETARY=English_United
> States.1252
> [4] LC_NUMERIC=C                         [5] LC_TIME=English_United
> States.1252
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods   base
> other attached packages:
> [1] MASS_7.3-4
>
> loaded via a namespace (and not attached):
> [1] tools_2.10.0
>
> ______________________________________________
> R-help@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.
>


-- 
Joris Meys
Statistical Consultant

Ghent University
Faculty of Bioscience Engineering
Department of Applied mathematics, biometrics and process control

Coupure Links 653
B-9000 Gent

tel : +32 9 264 59 87
Joris.Meys@Ugent.be
-------------------------------
Disclaimer : http://helpdesk.ugent.be/e-maildisclaimer.php

	[[alternative HTML version deleted]]