[Jessica Harding]

Hi Clyde,
One more quick question
In this model, should i create a new age group that specifies age at each calendar year? Or should I simply adjust for year at time of diagnosis/?
Jess

[Clyde Schechter]

I think that since you are disaggregating the results by calendar year, it makes more sense to use an age variable at calendar year. If you were doing it by years since diagnosis, then age at diagnosis would be appropriate, but that isn't what you're doing.

[Dharmenaan Palamuthusingam]

Dear Clyde (@clyde schecter),

I have similar data to Jessica and am looking to do a similar analysis. i.e. Estimate general surgical rates as a function of calendar year. After reading the above posts, I have some questions in relation to the example above that I hope you can help me understand.

1. Why do you use poisson regression when the 'lea' is not a count variable but rather a binary (yes/no)? If the patient has had multiple amputations (eg: 2), does it make more sense to create a new variable 'lea_total'?

---

2. I followed the code that has been provided above and also performed an analysis using an alternative code (vide-infra), which has produced very different results:
nb: i have created new variables --> diag_yr=year(diag_date) , lea_total=total lea per person, at_risk_followup = (time from beginning of follow-up to diag_date/death or end of follow-up)

poisson lea_total i.diag_yr, exposure(at_risk_followup) margins diag_yr, expression(predict(n)/at_risk_followup)

---

3. In relation to post #4, could you please expand on the assumptions you made? This was in relation to the code you provided in #2

gen exposure = dox1 - born poisson lea1 i.(_year age sex race), exposure(exposure) vce(robust) irr margins _year, ir Thank you
Menaan

[Clyde Schechter]

1. Why do you use poisson regression when the 'lea' is not a count variable but rather a binary (yes/no)? If the patient has had multiple amputations (eg: 2), does it make more sense to create a new variable 'lea_total'?

That was Jessica Harding's decision. It is not uncommon to use Poisson regression with a dichotomous outcome, particularly if the incidence rate is going to be low. Poisson regression is nice because it gives estimates of incidence rates, which cannot be gotten from the logistic regression that is more directly applicable to dichotomous variables. She also could have chosen to use the total number of amputations as her outcome variable, but in that case it becomes complicated to define the exposure variable in the denominator as the intervals between successive amputations will differ from each other and from the interval to the first amputation.

2. I followed the code that has been provided above and also performed an analysis using an alternative code (vide-infra), which has produced very different results:
nb: i have created new variables --> diag_yr=year(diag_date) , lea_total=total lea per person, at_risk_followup = (time from beginning of follow-up to diag_date/death or end of follow-up)

poisson lea_total i.diag_yr, exposure(at_risk_followup) margins diag_yr, expression(predict(n)/at_risk_followup)

And your question is...?

3. In relation to post #4, could you please expand on the assumptions you made? This was in relation to the code you provided in #2

It has been a long time, and I don't remember. Looking back on it quickly, I think I initially assumed that she had started out with a data set that contained one observation per person per year, or something like that. But I'm not really certain what I was thinking at that long ago time.