Calculation of incidence rates with confidence intervals on a single cohort

[Martin Wissing]

I have a quite simple problem that I do not succeed to resolve.
I have a cohort of about 500 patients followed at a kidney stone clinic with a variable indicating the time of follow up and two additional variables with the number of renal colics and stones these patients developed during the follow up period.
I want to calculate incidence rates of both colics and stones with 95% confidence intervals.
When I use the STATA ir command this is only possible when entering an exposure variable but not for the overall population. Therefore, it is not possible to obtain the incidence rates with 95% CI for the overall cohort.
Calculation of incidence rates is also possible with the stptime and strate commands after st setting the data. However, stset considers all colic or stone events >0 and <. as a single failure event. Incidence rates are therefore calculated as if patients with colics or stones had only a single event during follow up. This underestimates true incidence rates as many patients had several events. Is there a possibility in STATA to calculate IR while taking all events into account.

Any suggestion would be welcome.

Martin

[Martin Wissing]

Theoretically it is possible to find a way around the problem by entering colics or stones into a Poisson model without covariates. This provides the incidence rate as _cons with 95%CI but I do not know whether this is the correct way to address the issue.

[Clyde Schechter]

Have you tried -stset-ing our data as multiple failure data (be sure to especify -exit(time .)-) and then fit -streg- with -distribution(exponential)-. (Exponential survival model is equivalent to a Poisson event count model, but depending on how you set up the Poisson regression, it might not be properly reflecting recurring events for the same person.)

[Martin Wissing]

Dear Clyde,
unfortunately the patient files have been reviewed without noting the time of each individual stone event. I have only the cumulative number of events over the follow up period.
I cannot make assumption whether the rate at which these events have occurred varies over the follow up period. I therefore cannot stset them as recurring events (which would require one event at a time with an interval whose end-date corresponds to the occurrence of the event).
The analysis seems quite straightforward to me. It would be sufficient that the ir command provides the confidence interval of the rate for the overall cohort as it does for the populations defined by the group variable that has to be added.
Thanks for your response.
Martin

[Clyde Schechter]

I see. Well, the -ir- command is fairly simple and is not really suitable for repeated events. I think you will have to do this using a Poisson regression model. If your data set consists of a single observation for each patient, with variables reporting the number of stone events and the duration of the period of observation in which these occurred, then it's a straightforward application of -poisson-, with the duration of the period of observation as the -exposure()- option.

[Federico Tedeschi]

Is it acceptable to use a lognormal approximation, so that

1. the squared root of the inverse of the number of contacts is the standard error of the logarithm: something like

   gen selogir=sqrt(1/(Events))

2. the confidence interval is based on the logarithm and then exponentiated?