[R] Generalized linear models

[David Winsemius]

On Apr 27, 2009, at 5:19 PM, mathallan wrote:

>
> I have to fit a generalized linear model in R, and I have never done  
> this
> before, so I'm in very much doubt.
>
> I have a dataset (of 4036 observations)
>
>  claims	      sum	              grp
> 1	3852	      34570293	          1
> 2	1194	      7776468	            1
> 3	3916	      26343305          1
> 4	1258	      5502915	           1
> 5	11594	    711453346	        1
> ...
>
> there are 4 groups (grp).
>
> The task is to determine the effect of sum and grp (and interactions  
> between
> them) on the claims.
>
> I have to test using different link functions and distributions
>
>
> What I think I should do is (in R)
>
>> glm(claims~sum*grp, family=gaussian(link="log"))
>
> Call:  glm(formula = claims ~ sum * grp, family = gaussian(link =  
> "log"))
>
> Coefficients:
> (Intercept)          sum          grp      sum:grp
>  1.215e+01   -4.466e-09    6.814e-02    5.294e-09
>
> Degrees of Freedom: 4035 Total (i.e. Null);  4032 Residual
> Null Deviance:      3.371e+16
> Residual Deviance: 3.355e+16    AIC: 131500
>
>
> Is this right? And how can the output be interpreted?

It is very difficult to determine "rightness" since you have omitted  
essential background information. The most glaring omission is what  
sort of data is in "sum". If this is either the number of policies or  
the dollar amount at risk then a categorical "NO" is the answer to the  
question.
>

David Winsemius, MD
Heritage Laboratories
West Hartford, CT