Understanding Multilevel Model

[ted00]

empirical Bayesian

 

[spunky]

And what if you have no prior information available to you?

well... then you just make it up and BS your way out of it

but honestly, how often do you say "OMG I have absolutely 0 idea of what to do here". if that happens in this field of knowledge usually someone more senior than you smacks you and says "go back to the literature!"

 

[ted00]

an approach is to do two analyses, one using the prior of a pessimist, the other using the prior of an optimist

 

[Dason]

well... then you just make it up and BS your way out of it

but honestly, how often do you say "OMG I have absolutely 0 idea of what to do here". if that happens in this field of knowledge usually someone more senior than you smacks you and says "go back to the literature!"

You're building a model with *a lot* of predictors and are including interactions (maybe even higher order interactions). Are you telling me you have an idea about all of the parameter values? Keep in mind it's a multiple linear regression so the coefficients might be opposite of what you would expect compared to a SLR.

You don't have access to an 'expert'

The parameter you're making a prior for is just a nuisance parameter and you don't care about it - you might not actually have much prior information about the parameter.

It's a parameter associated with a latent effect and you honestly have. no. clue.

Also keep in mind that there are different ways to be 'uninformative'.

 

[spunky]

You're building a model with *a lot* of predictors and are including interactions (maybe even higher order interactions). Are you telling me you have an idea about all of the parameter values? Keep in mind it's a multiple linear regression so the coefficients might be opposite of what you would expect compared to a SLR.

this is area-specific. you throw something like that to us and chances are our editors will just flat-out reject it.

You don't have access to an 'expert'

usually, the expert comes to you, in this province of knowledge.


i'm starting to realize that something that got lost a few posts ago was that noetsi & i were sort of debating this off in the context of the social sciences...

 

[spunky]

an approach is to do two analyses, one using the prior of a pessimist, the other using the prior of an optimist

compromise! i luv the idea!

 

[ted00]

compromise! i luv the idea!

yeah, it's hardly mine, remember reading it in a book or a paper, no idea where, but saw it somewhere...

 

[trinker]

Spunky on centering:

i thought that we had gone over this before, but i'll just quote it againf or refernce. from the Enders & Tofighi (2007) where they say CWC = Centering Within Cluster and CGM =
(a) CWC is appropriate if the Level 1 association between X and Y is of subtantive interest
(b) CGM is appropriate when one is primarily interested in a Level 2 predictor and wants to control for Level 1 covariates;
(c) either CGM or CWC can be used to examine the differential influence of a variable at Level 1 and Level 2;
(d) CWC is preferable for examining cross-level interactions and interactions that involve a pair of Level 1 variables, and CGM is appropriate for interactions between Level 2 variables

particularly because of reason (d) i find Centering Within Cluster particularly useful. especially if you need to work with interactions like say, i dunno, 'gender' and 'school SocioEconomicStauts' or other level2 contextual effects
@ trinker. here's the article i took it from. i'm almost sure i sent this refernce to you already. i think it's worthwhile to read it
http://www-bcf.usc.edu/~sping/PM-515/references/Enders 2007 centering in multilevel analysis.pdf