search

tunsmart / interactionR

An R package for full reporting of effect modification and interaction
GNU General Public License v3.0
18 stars 6 forks source link

OR and the rare outcome assumption. #14

Open mamcisaac opened 1 year ago

mamcisaac commented 1 year ago

As far as I can tell, your package labels estimates as "OR" regardless of the link function chosen in the glm model. With a log link, for example, we are estimating "RR"s, not "OR"s.

Additionally, it appears to estimate RERI as OR11 - OR01 - OR10 + 1, whereas RERI = RR11 - RR01 - RR10 + 1, which is only approximated by OR11 - OR01 - OR10 + 1 if the outcome is rare.

vanderWeele and Knol (2014) state:

The approach described above works well if the outcome is rare so that RERI_OR approximates RERI_RR. If the outcome is common, RERI_OR may not be an adequate measure of additive interaction. In such cases, for cohort data, one could estimate RERI_RR by replacing the logistic model in eq. [9] with a log-linear model, though such log-linear models with continuous covariates C may not always converge; likewise an approach for risk ratios using modified Poisson, rather than logistic regression, has also been proposed that can be used with a common outcome (Zou, 2008).

Am I misunderstanding what your package does? Does it handle different link functions differently? Does it assume logistic regression with rare outcomes?

esb32 commented 1 year ago

I am fairly certain the package runs whatever the model output contains. I ran a glm with both binomial logit link to get OR, and a poisson log link, to get RR and had two very different RERI outputs.