LMM or not?

[Nyougo Omae.]

Hello,



I need some help. I have a dataset of students  nested within schools. To test if 'school' is random, I ran a LMM and found the  ICCs to be extremely small (.001 to .2) and the design effects based on this range from 1 to 5 with a mode of 2. My understanding is that once the design effect is <=2 then I can assume there's no school effect. Am I right in thinking this? I just read an article in which a design effect of 1.4 is assumed large enough to warrant inclusion of the variable in the model.



Thanks.

====================To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Johnny Amora]

Hi Nyougo Omae,

Your post has not been answered yet.  I think nobody understands your question.  You may rephrase it so that you would be understood.


--- On Wed, 12/3/08, Nyougo Omae. <[hidden email]> wrote:
From: Nyougo Omae. <[hidden email]>
Subject: LMM or not?
To: [hidden email]
Date: Wednesday, 3 December, 2008, 4:53 AM

Hello,



I need some help. I have a dataset of students  nested within schools. To
test if 'school' is random, I ran a LMM and found the  ICCs to be
extremely small (.001 to .2) and the design effects based on this range from 1
to 5 with a mode of 2. My understanding is that once the design effect is <=2
then I can assume there's no school effect. Am I right in thinking this? I
just read an article in which a design effect of 1.4 is assumed large enough to
warrant inclusion of the variable in the model.



Thanks.

====================To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD



      Design your own exclusive Pingbox today! It's easy to create your personal chat space on your blogs. http://ph.messenger.yahoo.com/pingbox

====================To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Nancy Darling-2]

I spend my life in mixed models, although I'm just getting into them in
SPSS.  In my opinion, if your intraclass correlation (ICC) goes up to
.2, you cannot possibly ignore the school effect - at least for that
variable.  In general the rule of thumb I've seen is that if you have
non-independence for more than 25% of your sample (and that's very
generous) it cannot be ignored.  You have it for 100%.  In addition, if
you are publishing in the education field, any reviewer will criticize
the work on the basis of ignoring the variable and likely reject it on
those grounds.

Since school is obviously not a major interest to you as an explanatory
factor, you can use Mixed Models to take school into account and adjust
the significance tests for the non-independence of the observations.
However, you could also take some school characteristics into account as
control factors (number of kids in free lunch programs, SES of census
track of feeder area, etc.) and probably improve your power for the
individual analyses you are interested in.

If the ICCs are really small, you can just enter school alone as a
nesting factor - it won't hurt your model.  If they are high, you can
get fancier.

I have found the mixed model unit in SPSS very difficult to learn from
the documentation, however, and very, very, very easy to mis-specify.
If you mis-specify the model, you get answers that are very wrong.  I
recommend finding a good book with documentation to help you with the
syntax (especially the covariance type - this is a not a trivial
choice).  I would also run your models twice - once ignoring the nesting
and once in mixed models - to make sure that your results look
comparable to one another.

I have been told by SPSS support that you can also run these equations
using the General Estimating Equations to get population average models
that compensate for the non-independence of observations, but I have
just started exploring those options.

Nancy

Johnny Amora wrote:

> Hi Nyougo Omae,
>
> Your post has not been answered yet.  I think nobody understands your question.  You may rephrase it so that you would be understood.
>
>
> --- On Wed, 12/3/08, Nyougo Omae. <[hidden email]> wrote:
> From: Nyougo Omae. <[hidden email]>
> Subject: LMM or not?
> To: [hidden email]
> Date: Wednesday, 3 December, 2008, 4:53 AM
>
> Hello,
>
>
>
> I need some help. I have a dataset of students  nested within schools. To
> test if 'school' is random, I ran a LMM and found the  ICCs to be
> extremely small (.001 to .2) and the design effects based on this range from 1
> to 5 with a mode of 2. My understanding is that once the design effect is <=2
> then I can assume there's no school effect. Am I right in thinking this? I
> just read an article in which a design effect of 1.4 is assumed large enough to
> warrant inclusion of the variable in the model.
>
>
>
> Thanks.
>
> ====================To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>
>
>
>       Design your own exclusive Pingbox today! It's easy to create your personal chat space on your blogs. http://ph.messenger.yahoo.com/pingbox
>
> ===================
> To manage your subscription to SPSSX-L, send a message to
> [hidden email] (not to SPSSX-L), with no body text except the
> command. To leave the list, send the command
> SIGNOFF SPSSX-L
> For a list of commands to manage subscriptions, send the command
> INFO REFCARD
>

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Maguin, Eugene]

Nyougo Omae,

My opinion is that since your data have a multilevel structure, you should
analyze it with models that represent that structure, within the level 2
sample size requirements needed to get convergence and good estimation for
those models. If your level 2 sample is too small for a multilevel model
representation, I think that an alternative would be to represent students
as nested within schools by means of GLM rather than Mixed. However, on this
last suggestion, I willingly defer to more knowledgable persons.

Gene Maguin

>>I need some help. I have a dataset of students  nested within schools. To
test if 'school' is random, I ran a LMM and found the  ICCs to be extremely
small (.001 to .2) and the design effects based on this range from 1 to 5
with a mode of 2. My understanding is that once the design effect is <=2
then I can assume there's no school effect. Am I right in thinking this? I
just read an article in which a design effect of 1.4 is assumed large enough
to warrant inclusion of the variable in the model.

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD

[Arthur Burke]

This question may be better addressed to the MULTILEVEL list ...

http://www.jiscmail.ac.uk/cgi-bin/webadmin?A0=multilevel

Art
Art Burke
Northwest Regional Educational Laboratory
101 SW Main St, Suite 500
Portland, OR 97204-3213



-----Original Message-----
From: Johnny Amora [mailto:[hidden email]]
Sent: Tuesday, December 02, 2008 9:29 PM
To: [hidden email]
Subject: Re: LMM or not?

Hi Nyougo Omae,

Your post has not been answered yet.  I think nobody understands your question.  You may rephrase it so that you would be understood.


--- On Wed, 12/3/08, Nyougo Omae. <[hidden email]> wrote:
From: Nyougo Omae. <[hidden email]>
Subject: LMM or not?
To: [hidden email]
Date: Wednesday, 3 December, 2008, 4:53 AM

Hello,



I need some help. I have a dataset of students  nested within schools. To test if 'school' is random, I ran a LMM and found the  ICCs to be extremely small (.001 to .2) and the design effects based on this range from 1 to 5 with a mode of 2. My understanding is that once the design effect is <=2 then I can assume there's no school effect. Am I right in thinking this? I just read an article in which a design effect of 1.4 is assumed large enough to warrant inclusion of the variable in the model.



Thanks.

====================To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD



      Design your own exclusive Pingbox today! It's easy to create your personal chat space on your blogs. http://ph.messenger.yahoo.com/pingbox

=======
To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD

=====================
To manage your subscription to SPSSX-L, send a message to
[hidden email] (not to SPSSX-L), with no body text except the
command. To leave the list, send the command
SIGNOFF SPSSX-L
For a list of commands to manage subscriptions, send the command
INFO REFCARD