[R] Strange behavior with poisosn and glm

Rolf Turner r.turner at auckland.ac.nz
Tue Mar 2 20:52:07 CET 2010


On 2/03/2010, at 9:02 PM, Noah Silverman wrote:

> Hi,
> 
> I'm just learning about poison links for the glm function.
> 
> One of the data sets I'm playing with has several of the variables as 
> factors (i.e. month, group, etc.)
> 
> When I call the glm function with a formula that has a factor variable, 
> R automatically converts the variable to a series of variables with 
> unique names and binary values.
> 
> For example, with this pseudo data:
> 
> y        v1        month
> 2        1            january
> 3        1.4        februrary
> 1.5    6.3        february
> 1.2    4.5        january
> 5.5    4.0        march
> 
> I use this call:
> 
> m <- glm(y ~ v1 + month, family="poisson")
> 
> R gives me back a model with variables of
> Intercept
> v1
> monthJanuary
> monthFebruary
> monthMarch

	No it didn't!!!  You are kidding the troops/being economical with the truth.

	If you had used the data that you show, it would've ``given you a model with
	variables'':

	Intercept
	v1
	monthfebruray
	monthjanuary
	monthmarch

	No caps in the month name and note the miss-spelling of ``february''.

	You actually have ***four*** levels for the month factor:

		january februrary february march

	If you had spelt ``februrary'' correctly you would have got variables

	Intercept
	v1
	monthjanuary
	monthmarch

	The first level, february would have been omitted, under the default contrasts
	(contr.treatment).  You need k-1 dummy variables to specify a factor with k levels.

> I'm concerned that this might be doing some strange things to my model.

	No, you are doing strange things.

	Notice also that the Poisson distribution is a distribution of ***counts***.
	Non-negative integers.  Whole numbers.  Values like 1.5 and 1.2 make no immediate
	sense in terms of the Poisson distribution.  The Poisson likelihood can be evaluated
	with non-integer responses, but the glm() function will quite rightly worry about
	non-integer values and give you a warning.  (Which you didn't mention.)

	If you really have non-integer valued responses you shouldn't be using the Poisson
	family; the quasi family *might* be appropriate --- if you know what you're doing.

> Can anyone offer some enlightenment?

	I hope you feel enlightened.

		cheers,

			Rolf Turner
######################################################################
Attention: 
This e-mail message is privileged and confidential. If you are not the 
intended recipient please delete the message and notify the sender. 
Any views or opinions presented are solely those of the author.

This e-mail has been scanned and cleared by MailMarshal 
www.marshalsoftware.com
######################################################################



More information about the R-help mailing list