search

easystats / report

:scroll: :tada: Automated reporting of objects in R
https://easystats.github.io/report/
Other
693 stars 69 forks source link

Incorrect use of `interpret_d` #204

Open mattansb opened 2 years ago

mattansb commented 2 years ago

Standardized slopes are not standardized differences! https://github.com/easystats/report/blob/a59f6b3caa759e461ad8358b56e3e26dd1ef3e4d/R/report.lm.R#L82

There is no simple 1:1 standardized interpretation of standardized slopes in multiple regression.

mattansb commented 2 years ago

Best you can do here is verbal interpretation ("A change of 1 SD in X corresponds to a [conditional] change of {} units in Y....").

bwiernik commented 2 years ago

It's the intercept though, so this is the standardized Y when all X's are at mean. I don't really see the value of interpreting the intercept for a standardized normal response. But if you do, d-like benchmarks seem reasonable? How many SDs from zero is y at the mean of X's?

mattansb commented 2 years ago

It's the intercept though [...]

Where are you seeing that?

bwiernik commented 2 years ago

Oh, apparently nowhere but in my "just woke up" fantasy land 😅

bwiernik commented 2 years ago

My general thinking here is that, multicollinearity issues aside, standardized coefficients are generally similar to either partial correlations (continuous X) or partial standardized mean differences (factor X). Applying those interpretation rules is generally reasonable, perhaps with a caveat if VIF is high?

mattansb commented 2 years ago

Are there such benchmarks for interpretation of partial correlations / partial standardized mean differences?

If so, I know at least one Frenchman who will be very happy to see this happen!

bwiernik commented 2 years ago

Not really. My recommendation would be to use the zero order r/d benchmarks and add a note about them maybe not applying to partial effect sizes?

mattansb commented 2 years ago

Ew :(

bwiernik commented 2 years ago

Not sure the opposition? These are all super rough guidelines in any case, and generally we might expect the partial effects to be "better" estimates of the true effect assuming our covariates are carefully chosen

mattansb commented 2 years ago

[...] assuming our covariates are carefully chosen

That's quite an assumption.

Even a VIF of 4.9 ("low") is an R2X of 0.8... Which is not small, and can still hinders interpretation....

mattansb commented 2 years ago

Anyway, just added interpret_vif (:

bwiernik commented 2 years ago

That's quite an assumption.

We can't automate thinking, just provide encouragement toward better approaches 😀

mattansb commented 2 years ago

Fine.... I concede! @DominiqueMakowski I think you owe @bwiernik a beer (: