How to estimate the risk of death following multiple ambulance attendances?

[Linn Gjersing]

I want to examine the risk of death following an overdose ambulance attendance by a city centre ambulance service.
During follow-up, many of the individuals have multiple attendances following the index episode.

We began inclusion at a random time (1.1.2014) and included individuals until 31.12.2018.
I want to examine the risk of death following an ambulance attendance in general, and I am therefore hesitant to use the index episode or the final episode and disregard the rest of attendances.

However, I wonder if anyone has any advice or reference on what would be the most correct manner to address this issue?
Below I have outlined the four possibilities I have considered. I think maybe no 4 is the correct manner to address this issue, but not sure.

I can estimate the study time from
1. the index episode until censoring (death or end of follow-up) and add total number of ambulance attendances as an independent variable to the regression model.
2. the final episode until censoring (death or end of follow-up) and add total number of ambulance attendances to the regression model.
3. each ambulance attendance until censoring (death or end of follow-up). This means one row for each attendance, but the person-time for each individual will not be correct.
4.each ambulance attendance and censor at the next ambulance attendance and from the final ambulance attendance censor at death or end of follow-up.

Anyone who could point me in the right direction?

[Carlo Lazzaro]

Linn:
have you taken look at -xtstreg-?

[Linn Gjersing]

Thank you, that works perfectly for the regression model.

However, I would like to estimate the mortality rate following ambulance contacts in specific days: at(2 7 30.5 91.2 183 365 )

When I am run stptime, at(2 7 30.5 91.2 183 365 ) per(1000) following
xtset id
stset nxtime, failure(died)

I get the following error message:
stptime requires that you have previously stset an id() variable

Any advice on how to address this issue? Is there a" xtstreg" option for this?

[Tor Myklebust]

There are several mays to tackle this problem. I can think of two relevant candidates:

1. You can make use of the stsplit-command to analyse the number of ambulance attendences as a time-varying covariate. After stsplit a standard survival model can be fitted.
2. You can view this as a multistate structure and estimate transition rates/probabilities. Michael Crowther has written a bunch of useful programs for this, including stmerlin, msset and predictms.

Since you have follow-up from a fixed date you might want to consider sensitivity analyses where you only analyse those who had no ambulance episodes for at least xx months. Not an expert here though!