what is best interpretative statistic for me to use for negative binomial regression?

[Shelby Carr]

I am fitting a negative binomial regression and would like to get thoughts on the best interpretative statistic to use. The outcome is minutes of physical activity per week. I am thinking of expressing the coefficient as an incident rate ratio (where the incident is incident of physical activity minutes per week) does this sound ok? I saw a paper using similar data where they presented count rate ratios but did not describe their derivation; I assume this is count ratio but just re-named?
Hello, I am fitting a negative binomial regression and would like to get thoughts on the best interpretative statistic to use. The outcome is minutes of physical activity per week. I am thinking of expressing the coefficient as an incident rate ratio (where the incident is incident of physical activity minutes per week) does this sound ok? I saw a paper using similar data where they presented count rate ratios but did not describe their derivation; I assume this is count ratio but just re-named?

[Carlo Lazzaro]

Shelby:
welcome to this forum.
Time is, in general terms, a continuous variable.
Why not consdering .regress-?

[Joao Santos Silva]

Dear Shelby Carr,

I guess that you are using a count data model because you have a mass point at zero, right? Anyway, I would not use a NB regression because, as Carlo pointed out, this is not really count data and therefore the unit of measurement can be changed (e.g., seconds of activity per week). In those cases, the results of the NB regression depend on the scale of the dependent variable and are therefore arbitrary. I would just use Poisson regression.

Best wishes,

Joao

[Shelby Carr]

Thank you both for your comments.
My data is continuous but is not normally distributed and has lots of zero values. We opted against Poisson due to mean and variance not being the same.

[Leonardo Guizzetti]

Thank you both for your comments.
We opted against Poisson due to mean and variance not being the same.

Using the robust (sandwich) variance estimator with -poisson- relaxes this classical constraint.

[Joao Santos Silva]

Dear Shelby Carr,

Besides the useful comment by Leonardo, I note that when the variable is not a count the relation between the mean and the variance depends on the scale in which you measure the data. Hence, any potential over-dispersion concerns do not apply and I strongly suggest you use Poisson regression.

Best wishes,

Joao

[Shelby Carr]

Thank you - are you suggesting Poisson over zero-inflated Poisson?

[Joao Santos Silva]

Yes, I am, at least as a starting point.