How to interpret GLM Output

#1
Hi,

Im looking at someone elses output for differentiating mortality. I was hoping someone could help me interpret the following coeffiecents etc? Im not sure how to use them to arrive at a singular result calculation:

Log Link Function:
Coef Intercept: -1.9380
Coef Age: 0.0123
Coef sum: -0.3663
Coef Pol Age: 0.1088
Coef Age Sum: 0.0048

So if i have variables:
Age: 40
Sum: 30000 (Defined as category 1? limits given to range of sum so they are grouped, so all less than 30k are the same?)
Pol Age: 3

Is the result something like exp-(-1.9380+0.0123*40+-.3663*(1? or 30000?)+0.1088*3+?
 

hlsmith

Less is more. Stay pure. Stay poor.
#5
Logistic regression would be link=log and dist = binomial, also mortality seems binomial to me not a count - but I am basing this off what you posted. Can you provide a source to these results?
 

hlsmith

Less is more. Stay pure. Stay poor.
#7
Yes, you are right - I was thinking about the model to get probabilities (e.g., risks out). Still I am curious what the DV is and the model, since it could also be something like log link dist=gamma, etc.. But likely you exponentiate the logged coefficient whether it is log bin or log Poisson, etc.. The interpretation of that output would differ though given what type of a GLM was specified.
 

fed2

Active Member
#8
yeah i think basically you need to know how the model was specified, including link, and how the coefs were defined. As it stands i don't think youcan tell if they are categorical or continuous. I thin the dist is less criticial to interpretation,
 

fed2

Active Member
#11
ummmmmmmm, no. still exp(Beta) for you. I think you have poisson here. is age categorical here really? id double check that.
 
#12
No not age that is what is whatever the variable. ok ill check so i dont use the 'odds' equation in the link above then. I just use exp(beta)? For reference i want to know the factor this returns against the inital expectation (i.e. the default mortailty table, so say prob death in general for age 40 is 0.1 and this factor then comes out at say 60% of that)
 

hlsmith

Less is more. Stay pure. Stay poor.
#13
Just provide us the source where the results are coming from, either the code or actual model output - or ideally BOTH. If you did this earlier it would have saved 10 posts.
 
#14
Logistic regression would be link=log and dist = binomial, also mortality seems binomial to me not a count - but I am basing this off what you posted. Can you provide a source to these results?
If it's modelling for the 'probability of mortality' from deaths from a known number of initially living cases it's dist = binomial, as I understand it. If it is mortality out of a known number of pre-living cases, why would one use Poisson distribution in logodds instead of binomial distribution?
 

hlsmith

Less is more. Stay pure. Stay poor.
#15
Rates and counts can be modeled with Poisson, for the former an offset needs to be defined. If counts are large enough ~> 8 the distribution can be modelled with Gaussian with identity link. GLMs have a lot of utility.