Sample size in GLMM

[George Siardos]

Dear list members,

  I would greatly appreciate if you could inform me about the limitations
set for sample sizes in performing Generalized Linear Mixed Models and
specifically a multinomial logistic mixed model. Could you suggest me any
rule of thumb for MLM to use?
   Some years ago GLMM experts suggested that for MLM regression models
the higher level sample size be at least 20, preferable 50. Is this valid
for binary and/or multinomial logistic regression?
   I really appreciate your kind reply.

Geoge C.S.

=====================
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

[Poes, Matthew Joseph]

I would need to know more about your study design to know if the following advice is appropriate, but the short answer is, Yes that is roughly right.  The longer answer is, it depends, you might need more, you might be able to get away with less.  I'd suggest a power analysis using the Optimal Design software.  It's built around RCT education studies primarily, but can be adapted to a wide range of experimental trial studies which may have similar conditions.  The newest edition allows for analyzing your design in the context of the addition of multilevel cv's, and what impact this has on power.

I'm currently the primarily analyst of a study with a little under 40 treatment/control schools, and with the CV's we are using, we are estimating the MDE to be around .15 to .2.  When you reduce that number in half, you still have power such that the MDE is around .25 or so, more than adequate in any typical education study.  There is also evidence from other studies that the MDE's generated in these power analysis are way too high, and that we may in fact have the power to detect effects as small as .1 or less with the 40 schools, and thus .15 to .2 with half as many schools.

-Matt

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of George Siardos
Sent: Friday, February 17, 2012 1:07 PM
To: [hidden email]
Subject: Sample size in GLMM

Dear list members,

  I would greatly appreciate if you could inform me about the limitations set for sample sizes in performing Generalized Linear Mixed Models and specifically a multinomial logistic mixed model. Could you suggest me any rule of thumb for MLM to use?
   Some years ago GLMM experts suggested that for MLM regression models the higher level sample size be at least 20, preferable 50. Is this valid for binary and/or multinomial logistic regression?
   I really appreciate your kind reply.

Geoge C.S.

=====================
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

[Jason Schoeneberger]

George,

There are less 'rules of thumb' about GLMMs as opposed to the linear
version...and even less so with respect to multinomial.  I know Moineddin et
al (2007) suggest at least 50 when prevalence is low, and of course,
depending on your interest in random effects even more may be needed.

My current dissertation study (the last two conditions are running now) will
address power estimates for fixed and random effects in binary GLMM across
various level-1 and level-2 sample sizes, and prevalence values.  Stay
tuned.

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of
George Siardos
Sent: Friday, February 17, 2012 2:07 PM
To: [hidden email]
Subject: Sample size in GLMM

Dear list members,

  I would greatly appreciate if you could inform me about the limitations
set for sample sizes in performing Generalized Linear Mixed Models and
specifically a multinomial logistic mixed model. Could you suggest me any
rule of thumb for MLM to use?
   Some years ago GLMM experts suggested that for MLM regression models the
higher level sample size be at least 20, preferable 50. Is this valid for
binary and/or multinomial logistic regression?
   I really appreciate your kind reply.

Geoge C.S.

=====================
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