[R] Generalized linear models (GLM)

[David Winsemius]

  Although both are known now, there is a time element involved in  
which one, max.loss was fixed at the time of underwriting and loss was  
unknown at that time. This *is* an insurance question is it not?

Wouldn't the question be: Can one use the group variable to estimate  
the proportion of max.loss actually lost following underwriting?

I would vote "no" on all of those offered solutions. If you believe  
otherwise, you should write an explanation to a hypothetical professor  
what the output of any of the models actually means.
-- 
David Winsemius


On Apr 28, 2009, at 2:38 PM, mathallan wrote:

>
> Actually both max.loss and loss are known values (in dollars). I'm  
> very much
> doubt, what to choose.
>
>
> glm(max.loss~loss,family=gaussian(link="identity")
>
> or
>
> glm(formula = sum ~ claims * as.factor(grp), family = gaussian(link =
> "identity"))

There are no variables named claims or sum. "sum" would be a bad  
choice in variable names because of the potential confusion with the  
function of the same name.

>
>
> or
> glm(loss~max.loss,family=gaussian(link="identity")
>
> we have to look at gaussian and gamma, with link identity and log.
>
> But my problem is what is going to be between the ~
>
>
>
> David Winsemius wrote:
>>
>> I think you are off-track because max.loss does not sound like a
>> proper Y variable. Because max.loss is an amount that is known, in  
>> the
>> insurance applications I have seen it would have been modeled within
>> an offset term. Many of the examples have used number of ships or
>> buildings or the person years of exposure but I do not see that the
>> general strategy is limited to only such  considerations.
>>
>> I would also suggest that you consider links other than Gaussian,
>> perhaps negative binomial.
>>
>> The task for the analyst is then to translate output from the chosen
>> model into interpretable meaning on the scale of interest, but I
>> assume your course instructor will help with that.
>>
>> -- 
>> David Winsemius
>> On Apr 28, 2009, at 11:34 AM, mathallan wrote:
>>
>>>
>>> Hi
>>>
>>> I got a dataset
>>>
>>>      loss     max.loss   grp
>>> 1     10         50         2
>>> 2     15         33         1
>>> 3     18         49         2
>>> 4     33         38         1
>>> 5      8          50         3
>>> 6     19         29         1
>>> 7     22         51         4
>>> 8     50         50         2
>>> 9     16         38         1
>>> 10    24         30         3
>>>
>>> were loss and max.loss are monetary values (in dollar). Grp is group
>>> number.
>>>
>>> By use of GLM, I have to determine the effect of max.loss and grp  
>>> (and
>>> interactions between them) on loss. My question is how to do this.
>>>
>>> Is it something like
>>>
>>> glm(max.loss~loss,family=gaussian(link="identity")
>>>
>>> were ofcourse I can change gaussian with Gamma,... and link with
>>> log,...
>>>
>>> But am I on right track, or what should I change?
>>>
>>>
>>> Thanks
>>> -- 
>>> View this message in context:
>>> http://www.nabble.com/Generalized-linear-models-%28GLM%29-tp23279588p23279588.html
>>> Sent from the R help mailing list archive at Nabble.com.
>>>
>>> ______________________________________________
>>> 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.
>>
>> David Winsemius, MD
>> Heritage Laboratories
>> West Hartford, CT
>>
>> ______________________________________________
>> 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.
>>
>>
>
> -- 
> View this message in context: http://www.nabble.com/Generalized-linear-models-%28GLM%29-tp23279588p23283633.html
> Sent from the R help mailing list archive at Nabble.com.
>
> ______________________________________________
> 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.

David Winsemius, MD
Heritage Laboratories
West Hartford, CT