Guidance on HLM/multilevel modeling

trinker

ggplot2orBust
#1
I've decided I need to know multilevel modeling; no way around it in educational research. I can't take the class at the U because it's offered bi yearly (missed it this last Fall). So I need to take it else where and/or learn it on my own.

I'm hoping to rely on other people's experience here with multilevel models. I'm looking for the following related to multilevel modeling that you think would be helpful:

  1. Any online programs you know are reputable for taking HLM (an R based course would be extra awesome)
  2. Online websites
  3. Youtube videos/other videos
  4. Paper/article names or links
  5. Book names
  6. People's names associated with the methods
  7. Vocabulary I should investigate
  8. R related guides and/or example R scripts
  9. etc.

Thank you in advance for your suggestions :D
 

noetsi

Fortran must die
#2
I am taking hiearchial linear models this semester (a doctoral course in my program) although I am no expert obviously.

This is one of the elite softwares in the field. They have a free student version. It might meet your needs.
http://www.ssicentral.com/hlm/

I have a list of readings for the class which are, as best I can judge, pretty good. But I don't know how to send them to you (you could not get to the university website they are hosted on even if I was allowed to post it).

One book you can try (it is not simple) is Hiearchical Linear Models by Stephen Raudenbush and Anthony Bryk. They are names in that field in the context of education (probably generally).

There is a vast vocabulary to investigate. If you have specific questions I can try (as someone learning the method himself) to address them.
 

Lazar

Phineas Packard
#3
lme4 in r is more than adequate.

In terms of books, I would suggest Gelman and Hill as they give you everything you need and all examples are in r.

If you want to integrate latent and multilevel modeling your best choice is mplus.
 

Dason

Ambassador to the humans
#4
lme4 in r is more than adequate.

In terms of books, I would suggest Gelman and Hill as they give you everything you need and all examples are in r.

If you want to integrate latent and multilevel modeling your best choice is mplus.
Or you could go the bayesian route and use winbugs to get your samples. Does Gelman include winbugs code in his book? I would assume so.
 

spunky

Doesn't actually exist
#7
I used Mplus for several semesters and never realized it did HLM :p
in all honesty, if you learn to use Mplus well a lot of the other software becomes sorta useless... :p

but i side with Lazar: nothing beats good ol' lme4 (and Pinheiro & Bates are coming up w/ a book for it...)

as fate would have it, i am right now sitting on an HLM class in the dept of education... once i take the one in the Bio dept i'll have taken it 4 times: (theory of generalized linear models in the stats dept, multilevel models in the psych dept, hlm in the ed dept and im just missing the GEE approach from biostatistics for the circle to be complete!)
 

noetsi

Fortran must die
#8
HLM must be big in education. I am taking it as part of the Masters of Statistic in Education. Given that students naturally nest in schools I guess it makes sense.
 

trinker

ggplot2orBust
#9
Well if you think about it Noetsi a lot of educational studies are very structured as students in a class in a school in a district in a community in a state etc....
 

spunky

Doesn't actually exist
#10
HLM must be big in education. I am taking it as part of the Masters of Statistic in Education. Given that students naturally nest in schools I guess it makes sense.
it is HUUUUUGE rite now... linear mixed models (the real name of HLM/MultiLevel models) are incredibly versatile and very powerful... but it was the HLM software what made them really popular because it's very SPSSy to use...

(ps- trinker, the lme4 book is here: http://lme4.r-forge.r-project.org/) or if you can get ahold of the Splus book for fitting mixed effects models it's an invaluable resource both in terms of theory and how to code it...
 

spunky

Doesn't actually exist
#11
Latent as in the stuff I'm learning in psychometrics; the stuff extending Classical Test Theory?
yep... most tradtional multilevel models can be fit through a structural equation model approach and, as you know, structural equation models came up naturally from developments in psychometrics and extensions of CTT
 

spunky

Doesn't actually exist
#13
i need to log off 'cuz i think the prof realized i'm not really paying att'n but if you send me your email i'll send you all the stuff (plus a few R functions i've worked on) that should get you started..
 

noetsi

Fortran must die
#14
i need to log off 'cuz i think the prof realized i'm not really paying att'n but if you send me your email i'll send you all the stuff (plus a few R functions i've worked on) that should get you started..
:p As I painfully work through HLM assignment II - that made my day.
 

Lazar

Phineas Packard
#15
Or you could go the bayesian route and use winbugs to get your samples. Does Gelman include winbugs code in his book? I would assume so.

This is Gelman so yes. The winbugs route is the end game more or less of the book.

@Noetsi. Yes it does two-level models. It also allows for latent variables and latent aggregation (to control for sampling error). It also has the command TYPE=COMPLEX which allows you to run single level models but with SEs corrected for nested samples.

@Trinker. I use ml approaches in edu research alot so have several example scripts and other things.

I have just submitted an article (in second round of review at the moment) which gives an applied overview of double-latent multilevel models which example Mplus syntax. I could send it to people who are interested. Marsh et al. has a more technical example but mine is targeted at applied researchers.
 

noetsi

Fortran must die
#16
One of the problems I have working with HLM is their argument that when data is nested than OLS will inherently have unequal error variance (heteroscedacity) and observations will not be independent. Since all data is inherently nested in something, then all OLS would be badly harmed by these conditions. And clearly that is not the case.

HLM, like many recent discoveries has a habit of overselling itself and degrading other methods that in fact work fine.

I can't speak for R or mplus but HLM7 has no estimates of the level 1 model parameters. So you are estimating the parameters of the level 1 model in level 2, without knowing if those level 1 parameters actually predict in the level 1 dependent variable you are ultimately interested.

But it clearly is an interesting method.
 

Lazar

Phineas Packard
#19
One of the problems I have working with HLM is their argument that when data is nested than OLS will inherently have unequal error variance (heteroscedacity) and observations will not be independent. Since all data is inherently nested in something, then all OLS would be badly harmed by these conditions. And clearly that is not the case.

HLM, like many recent discoveries has a habit of overselling itself and degrading other methods that in fact work fine.
When we know what the structure is (not always the case granted) we can work out mathematically the bias to SEs that comes from not accounting for the multilevel structure. The issue of biased errors due to nested samples has been known for quite some time.