Guidance on HLM/multilevel modeling


Fortran must die
No I dont think you need the random effects to generate the CI of the fixed effect intercept. You just need to know what it is and its SE.

Maybe this helps.

You have first level equation (dont worry what the DV is).

DV = Boj +rij the fully unconditional model

At the 2nd level you have Boj = G00 +Uoj with GOO being fixed and Uoj the random effect (the variation around the means of the groups such as schools). The question is if (or actually how) the CI around Goo shows if group effects are signficant. I dont see how it can tell you this.


Doesn't actually exist
noetsi;7691What I mean is if the fixed effect intercept at the level that is explaining between group variation [B said:
can[/B] show if between group variation is occuring (or if its ci can). I dont see how it can, but the question suggest this is possible.
sorry, i should've said that this particular interpretation does not refer to the question you're asking at this precise moment, it was the previous one where somehow i think your professor said the scaling of the variables was involved? (dunno about that but meh...)


Fortran must die
lol the question was misworded it turned out (just got the correction). It now says:

What test tells us about whether we can reject the null hypothesis that the mean of mathatt is the same among all schools? How did you reach this decision?

My answer would be the test of the random effect Uoj. If its statistically signficant it tells you that variation between) group (school) is signficant.

Is that right :p


Doesn't actually exist
At the 2nd level you have Boj = G00 +Uoj with GOO being fixed and Uoj the random effect (the variation around the means of the groups such as schools). The question is if (or actually how) the CI around Goo shows if group effects are signficant. I dont see how it can tell you this.
i am going with what this poster said on SO because it seems pretty reasonable:

so apparently the varaince of the grandmean is the weighted sum of two variance components according to this person. if there are two variance components involved then one must be u0j and the other one is the varaince of the level-1 residuals/error. which means the CI for the grand mean needs an estimate of the standard error of it which is obtained (by the poster's response) from the variance component u0j which tells you that the individuals or units of analysis at level 1 must indeed change or, otherwise one would be missing that variance component to get the standard error to calculate the confidence interval of the grand mean...

phew... i missed posting here on TS!

anyways, gotta go catch the bus... c u all guys l8er!


TS Contributor
This thread should receive an award for the speed at which it reached 40+ replies.

I second the vote for lme4 (and the newer glmmADMB, which handles zero-inflated GLMMs). The glmmADMB option is great, but it is still in development. So, for zero-inflated data, there's no R option (that I know about) for GLMMs. You have to do this manually by splitting the data. The Zuur book that Jake mentions goes into this a bit. Their new book, which is specifically about zero-inflated models, supposedly goes into more detail.

That said, the Zuur book on mixed-effects models that Jake mentioned was my primary reference when I was learning GLMMs and GAMMs, and I highly recommend it. Changed my life, I tell you. I've become an annoying addict of additive models as a result. I think I should charge Zuur for the therapy sessions.

Also, there's a good website that seems to keep up with discussions and new issues about GLMMs, etc. It might be too field-specific for most people:


Fortran must die
Threads like this remind me how little I really understand statistics as I wonder how you know if between group variation is signficant (and struggle with that) and then I read about zero-inflated data and GLMM's and GAMM's - things I have never heard about and probably won't.

Oh well exepectations for running data is much more limited where I work than that in all liklihood :p


TS Contributor
Isn't life good when the fact that something isn't there (0) or isn't doing something (0, again) is at once interesting? There's data everywhere, even where there's not. Yes, that bit of zen just happened. Go ponder THAT on this here Friday evening.


I received numerous resources and suggestions tonight and am marking this thread as solved. I am grateful and will most likely delve into this seriously at the end of the semester. I still think I want to take an online class in addition to learning it on my own.


Phineas Packard
The thing to keep in mind is that the limit of mlm models as variance in l2 approaches 0 or ininity is the complete pooling (ignore the l2 varaibles completely) and non-pooling OLS approaches repsectively. Both these approaches can be useful on occasion but given these limits, for the most part, you might as well fit a multilevel model when the data structure is known (a good discussion can be found here).

I think the point though noetsi, is that in OLS we might not have indicators or knowledge of the multilevel structure or may have good reason not to think that the multilevel structure will be an influence so we can only model the data that we have. Other times we may only have l1 hypotheses so we are only interested in ensuring that the SEs are appropriate and so typical regression approaches with TYPE=COMPLEX in mplus or a custom block bootstrap in r is perfectly acceptable.