I have a medical longitudinal retrospective dataset, records between the observation period of 2000 and end 2016. For many reasons not every medical record spans that entire time-frame, e.g. the patient may have died, or they may have transferred in to the study half way through or transferred out at some stage.
A particular event (or exposure) is seen as a clinical event e.g., going to the doctor and saying or being told that you have a particular disease, e.g., a chest infection. That patient will also have a categorical variable to indicate whether they are a smoker or not.
I wish to count the frequency of chest infections per patient and distribute them over whether they smoke or not. I can imagine this would be a box plot with UQ and LQ being defined, frequency of disease on the Y, and a Smoke YES and NO on the X. This would be very easy to do. The problem I have though is that I am not sure how I deal with medical records of varying length. Surely there is bias if a smoker vs. non-smoker both have twenty chest infections, but there is a four year medical record difference?
Thanks
You are about to discover the concept of incidence density! This is the rate of events per unit time. In your case, the rate of events per 100 or 1000 person-years.
You can tackle the problem using Poisson regression, with length of observation set as the exposure time variable.
Alternatively, you can treat the data as time-to-event data with repeated events. This has the advantage that the probability of chest infections probably rises with age. You can use age as the time variable in the analysis, with subjects entering at the age they were first seen and exiting at the age of last follow up. In order to avoid immortal time bias, you need to declare them to be at risk from some point, say from age 18.
This is pretty straightforward in Stata, if you have it.
You are about to discover the concept of incidence density! This is the rate of events per unit time. In your case, the rate of events per 100 or 1000 person-years.
You can tackle the problem using Poisson regression, with length of observation set as the exposure time variable.
Alternatively, you can treat the data as time-to-event data with repeated events. This has the advantage that the probability of chest infections probably rises with age. You can use age as the time variable in the analysis, with subjects entering at the age they were first seen and exiting at the age of last follow up. In order to avoid immortal time bias, you need to declare them to be at risk from some point, say from age 18.
This is pretty straightforward in Stata, if you have it.
Thank you, Ronań. This information has been extremely helpful. Something new, every day!
Unfortunately I don't have Stata, rather I am using R to perform all of my analysis.
Well the good news is that almost anything is possible in R. The bad news, as you know, is that it may not be easy.
You might check out this page, and pay particular attention to the sandwich package, which you will certainly need for your standard errors.
https://stats.idre.ucla.edu/r/dae/poisson-regression/
Hello professors, what a great explanation Ronán, thanks a lot! I have a question ¿Could he also calculate an accumulated incidence by the Kaplan Meier method?
Hi Anthony
See this link for a slide explaining the incidence density
https://www.slideshare.net/AndrewMertens1/154-measures-incidenceincidence-density
In R many packages can do that. If you will treat the data as time to event data , you can use the package 'survival'.
If you want to see if smoking affects the number of events (chest infection, etc), the Poisson regression would be your best bet. If the time-to-event intervals are different, an interval-censored Poisson regression would be more useful. Please take a look at the following publication:
https://baylor-ir.tdl.org/baylor-ir/bitstream/handle/2104/8176/sydeaka_watson_phd.pdf;sequence=1
Good luck!
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics