Poisson regression (how to reflect updated independent variables in the model)

[Therain]

Hello,

I use SAS 9.4 and I am looking at how the presence or absence of disease at baseline year would affect the number of health care use in the next 2 years. I have a data set like this:

id     baselinedisease     healthuse

1      0                              3

2      0                             15

3      1                              2

4      0                              5

5      1                              0

...

However, some people developed the disease at year 1 post baseline or year 2 post baseline, and I think this would affect their subsequent number of health care use. So I want to reflect this in the model. I have another data set showing me the presence or absence of disease for each participant in year 1 and year 2 post baseline. The current codes I am using is:

proc genmod data=have;
class disease (ref='0') / param=glm;
model healthuse = disease / type3 dist=poisson;
run;

Does anyone can help me figure out the problem? Thank you in advance!

[sld]

Would it make sense to create a factor (Occurence) with three levels?

Occurence=0 if no disease at t=0, t=1, or t=2

Occurence=1 if no disease at t=0, disease at t=1 and t=2

Occurence=2 if no disease at t=0 or t=1, disease at t=2

Each id would belong to one and only one level. 

[Therain]

Hello, thank you for the reply! I do not know if my understanding is correct, but I am thinking the new disease can only happen in those without disease at baseline. So if we create another factor and include it in the model, it may not be necessary because this factor cannot be the confounder (i.e. No new developed disease in those with disease at baseline)?

[sld]

I was thinking of replacing baselinedisease (with two levels) with Occurence (with 3 levels):

model healthuse = Occurence / type3 dist=poisson;

Whether this makes sense might depend on which two years you counted healthuse (e.g., the two years following the baseline assessment?), as well as other considerations.

[Therain]

Hello, Thank you for the clarification and sorry for the late reply. I think if we replace it with this, the research question would become comparing incident and non-incident disease groups, instead of comparing having the disease or not. Now I am considering using something like the following codes:

proc genmod data = have;
class study_id disease;
model healthuse = disease / dist=poisson link=log type3;
repeated subject = study_id / type=ar(1) corrw;
lsmeans disease/diff cl;
run;

There is also another file seems related to this issue: https://methodology.psu.edu/downloads/tvem

I will keep working on this and to see if this works.

[sld]

You could write a contrast for the levels of Occurrence that compares level 0 (non-incident disease group) to the mean of levels 1 and 2 (incident disease groups).