1

I am aware that you can compute risk differences as the difference of risk between two treatment arms $( (\text{No. events}_1/ N_1) - (\text{No. events}_2 / N_2))$.

However, can I estimate the same using patient-years incidence rather than just incidence thus: events as patient-years 1 - events patient-years 2?

Thus allowing to standardize for different follow-up periods and then computing also ARD and NNTs,or should I go for a Poisson regression instead, and just use IRRs?

Alexis
  • 26,219
  • 5
  • 78
  • 131
Claudio Laudani
  • 171
  • 1
  • 9
  • 1
    When you write "patient-years incidence rates" do you mean incidence rates calculated as **incidence densities**? If so you can certainly contrast using a risk difference (with the caveat of constancy of incidence rate over the different periods of time represented by the denominators of both groups—i.e. the incidence rate for group 1 was constant across the entire time spanned to collect the denominators of both groups, and so is the incidence rate for group 2). – Alexis Aug 13 '21 at 01:32
  • I didn't know the concept of incidence density; however, if I'm not wrong, you should have got the point. Just to be sure: I'm talking about "the incidence rate was 23.4 /1000 person-years", which should be incidence densities. Many thanks! – Claudio Laudani Aug 13 '21 at 03:27

1 Answers1

1

Yes, you can calculate a risk difference using incidence densities, but there are assumptions.

Incidence rates come in two flavors: (1) incidence proportions, where the numerator contains the number of new cases occurring during a specific calendar period and the denominator contains the number of people considered 'at risk' for being a case during the same period (typically the population of at risk people at the midpoint of the period, e.g., July 1st of a calendar year, and (2) incidence densities, where the numerator is the number of new cases observed out of all observed individuals and the denominator is the sum of person-time of each individual, regardless of whether they became a case or not (assuming new cases' person-times terminate when they become cases). It sounds like you are using incidence densities.

Because the calendar period (e.g., many years) during which people are observed to calculate incidence densities typically covers a different span than the unit of time communicated in the incidence density (e.g., person-years, person-months, etc.), interpretation of incidence densities usually assumes a constant incidence rate across the full calendar period of observation. Therefore, when creating risk contrasts using incidence densities (e.g., a relative risk, a risk difference, etc.), we must likewise assume that the incidence rates of both groups being contrasted are constant across the full calendar period of observation of both groups. So, if the incidence density in group 1 was created from data spanning 2015–2019, and the incidence density of group 2 was created from data spanning 2017–2020, then risk contrasts assume the incidence rates in each group were constant from 2015–2020.

Alexis
  • 26,219
  • 5
  • 78
  • 131
  • 1
    I really appreciated your helpful answer! One more thing: If I want to calculate the SE of Risk-Contrast calculated over Incidence-densities, using the equation given here [link](https://sphweb.bumc.bu.edu/otlt/MPH-Modules/PH717-QuantCore/PH717_ComparingFrequencies/PH717_ComparingFrequencies12.html) and I have 1000 pt-years (calculated by calculating 1 pt-year and then multipliing by 1000), should I put the multiplication also in the SE, or it is not useful? – Claudio Laudani Aug 13 '21 at 17:00
  • 1
    @ClaudioLaudani Ask a new question. :D – Alexis Aug 13 '21 at 19:33
  • 1
    I did! Hope you can help me also there :D Many thanks anyway! – Claudio Laudani Aug 14 '21 at 08:39