Can I omit confidence interval in multivariable logistic regression with no sample but entire population?

#1
Hi,

I have a question regarding the usefulness of a confidence interval when calculating ORs in a multivariable logistic regression model.

My data is not a sample, but the entire population of people who died in a given year (i.e. I have ~100,000 cases). What I want to do is calculate the odds of receiving a specific type of care, comparing a set of variables (men vs women, higher income vs lower income levels, etc).

When calculating ORs in SAS, obviously it presents 95%CIs as well. My question is this: is it meaningful to present these values in a paper or can I leave them out? What is the rationale to do one or the other?

I could only come up with the following rationale to leave them in: perhaps we could interpret the entire population of a given year as a random sample of the population in all other years? So it is in that sense considered a sample?

Many thanks in advance!
Arno
 
#2
When calculating ORs in SAS, obviously it presents 95%CIs as well. My question is this: is it meaningful to present these values in a paper or can I leave them out? What is the rationale to do one or the other?
Yes, I agree. If you just think of it a description of that specifik population then you can omit the confidence interval.


But on the other hand:
perhaps we could interpret the entire population of a given year as a random sample of the population in all other years? So it is in that sense considered a sample?
Yes, I agree again. If you believe that the year after and before that year, will be similar to your specific year, that is you infer something from the data, then you can include the confidence interval.

Sometimes one talks about that the specific population is a sample from a super population.
 

noetsi

Fortran must die
#3
Some argue you can just ignore statistics generally other than effect size when you have the population. Others argue it could be a sample of all possible populations perhaps at different points in time. In the later case CI matter.
 

hlsmith

Not a robit
#4
I agree with @GretaGarbo

Does a calendar year mean anything in your study, seasonality, etc. If not, 12 months may be a sample of 13 months, etc. Also, does conducting this study change the future?
 
#5
Thank you for your responses. I think it would be easiest to just leave the CI out.

@hlsmith The data is delivered to us as a retrospective cohort of all people who died in one calendar year - but no information is given on date of death. All variables with information on dates (eg. date of drug dispension) are recalculated into "number of days before death", so we have no information about seasonality etc.
 

hlsmith

Not a robit
#6
The only other comment I have is about collinearity. Terms can be collinear and you usually determine this via examining for it or appearance of large CI's. While estimates alone won't reveal it's presence. If you want the most concise or parsimonious model. You may look into the best subset of terms that aren't explaining the same phenomenon, but if not you are probably fine moving forward. I say this because if you are doing any prediction at all, estimates won't reveal the lack of confidence generated by using similar terms.

Cheers and welcome to the forum!