I have a dataset made up of observations of procedural cases, with a count outcome (variable X carries the number of times a specific finding was detected in one procedure/case). My denominator is the total number of procedures. My count outcome starts with 1 (no zeros), therefore I am using a Poisson model.

I am not sure what the exposure variable should be as the data is a retrospective cross-sectional sample of procedures with procedural findings, from each of 4 calendar years aggregated together. What would be the offset (exposure) variable, if appropriate, in this case?

Hence the question comes to interpreting the incidence rate ratios of the independent variables.

For example, specialist vs trainee (reference level) IRR =1.50 . If no exposure (offset) variable is used, does that mean that the specialist is 50% more likely compared to the trainee to detect one addtional finding count per procedure? Or : the specialist will detect on average 50% more copies of that finding per procedure?

I have seen and watched various people phrasing the interpretations of IRRs in various ways.

With or without exposure variable (if needed) , what would be the correct interpretation of IRR in this case?

  • $\begingroup$ NB a Poisson probability model permits the use of 0s. If you are excluding cases where 0 is a relevant outcome count such as in a sample of controls, you must be aware of inducing prevalent case bias. $\endgroup$ – AdamO Feb 20 '18 at 17:53