Non-lmer, non-Bayesian models to support

[andrewheiss]

This is mostly just a checklist of the more important models we might want to support with fancy math. In theory, these are all supported automatically with broom (though we might want to have a general catchall way to show non-fancy formulas):

• Binary logistic (glm(..., family = binomial(link = "logit")))
• Binary probit (glm(..., family = binomial(link = "probit")))
• Ordered logistic (MASS::polr(..., method = "logistic") and ordinal::clm(..., link = "logit"))
• Ordered probit (MASS::polr(..., method = "probit") and ordinal::clm(..., link = "probit"))
• Multinomial logistic (nnet::multinom())
• Survival regression (survival::Surv())
• Multiple ANOVA (stats::manova())

[jrosen48]

Nice. Not sure what math it involves, but what do you think about stats::manova(...)?

[andrewheiss]

Sure! Just added it

[jebyrnes]

betareg and DirichletReg?
pscl:: zeroinfl?

[andrewheiss]

So with probit, how do we want to structure the LHS and RHS? Pr(Y = 1 | X) = Φ(β0 + β1X1 + etc.)? Or is that too complicated?

Once we figure that out, adding ordered probit will be trivial

[datalorax]

I like that. The only issue is we'll need to swap extract_rhs to use S3 methods and implement this as a new method, but we were probably going to have to do that at some point anyway. If it ends up being too difficult we can always put it off for the next version, but I don't think it would be too bad (or at least I hope).

[andrewheiss]

Yeah, the RHS is going to need stuff now bc φ, but it shouldn't be too bad.

Any ideas for notation for multinomial logit, survival, and multiple ANOVA?

Multinomial could be like ordered logit:

\log \frac{P(Y = \operatorname{outcome_1})}{P(Y = \operatorname{outcome\_base\_case})} = RHS \\
\log \frac{P(Y = \operatorname{outcome_2})}{P(Y = \operatorname{outcome\_base\_case})} = RHS \\
...

Survival models are complex and idk the best approach (nothing here looks simple)

Also don't know best way to do MANOVA (no clear regression-like syntax here - @jrosen48 might have ideas

[datalorax]

Yeah I like that for multinomial. I have no idea on survival and MANOVA, honestly, but could do some looking into it.

Do we still want to support ordered::clm too? I haven't looked into that at all. If the output from broom::tidy is the same or basically the same it should be easy though.

Also, should we have the goal to support all of these before the initial release? Or get the release out basically now and develop these for v0.2?

Last question - do you want to take an initial stab at probit or should I? I might have some time later tonight I could devote to it if you haven't already started (or if you just want to).

[andrewheiss]

I’m on it already, and also the clm ologit. It works with broom, so it should be easy I think probit + ordered probit and logit should be ready for initial release, then multinomial and gang can come later

…

On July 29, 2020 at 19:16:23, Daniel Anderson ***@***.***) wrote: Yeah I like that for multinomial. I have no idea on survival and MANOVA, honestly, but could do some looking into it. Do we still want to support ordered::clm too? I haven't looked into that at all. If the output from broom::tidy is the same or basically the same it should be easy though. Also, should we have the goal to support all of these before the initial release? Or get the release out basically now and develop these for v0.2? Last question - do you want to take an initial stab at probit or should I? I might have some time later tonight I could devote to it if you haven't already started (or if you just want to). — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#32 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAAR7P6AVVUDU25KNW7AHIDR6CUUPANCNFSM4HUX3FZQ> .

[datalorax]

Perfect!