Logistic regression with nested-factor design?

[Allan Lundy, PhD]

Dear Listers,
I have a client with a design in which there are 3-7 persons in each of a couple of dozen groups.  Half the groups were presented with one manipulation, half with an alternative.  Individuals made a binary choice after discussion.  It is clear this is a nested-factor design.  It is not legit to ignore group membership because individuals influenced each other, hence individual results are likely to be more alike between 2 people in the same group than in different groups.  It would be nice to use a logistic regression, since the outcome measure is dichotomous.  But I can find nothing in SPSS on how to do this.  Until I looked at

http://www.ats.ucla.edu/stat/spss/output/logistic.htm
and found:

e.  -2 Log likelihood - This is the -2 log likelihood for the final model.  By itself, this number is not very informative.  However, it can be used to compare nested (reduced) models.

Anybody familiar with this?

Thanks in advance!
Allan

Allan Lundy, PhD
Research Consulting
[hidden email]

Business & Cell (any time): 215-820-8100
Home (8am-10pm, 7 days/week): 215-885-5313
Address:  108 Cliff Terrace, Wyncote, PA 19095
Visit my Web site at www.dissertationconsulting.net ===================== 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

[Ryan]

Hi Allan,

 

It seems that you are mixing up concepts. While it is true that the difference in log likelihood statistics can be used to compare the fit of "nested" models, this has nothing to do directly with within group dependence. Your situation is related to the latter concept.

 

SPSS does not currently offer a procedure to fit generalized linear mixed models, which on the face of it, appears to be what you desire. Do you have access to a stats program that offers a procedure to fit generalized linear mixed models?

 

Ryan

On Wed, Jun 9, 2010 at 6:28 PM, Allan Lundy, PhD <[hidden email]> wrote:

Dear Listers,
I have a client with a design in which there are 3-7 persons in each of a couple of dozen groups.  Half the groups were presented with one manipulation, half with an alternative.  Individuals made a binary choice after discussion.  It is clear this is a nested-factor design.  It is not legit to ignore group membership because individuals influenced each other, hence individual results are likely to be more alike between 2 people in the same group than in different groups.  It would be nice to use a logistic regression, since the outcome measure is dichotomous.  But I can find nothing in SPSS on how to do this.  Until I looked at

http://www.ats.ucla.edu/stat/spss/output/logistic.htm
and found:

e.  -2 Log likelihood - This is the -2 log likelihood for the final model.  By itself, this number is not very informative.  However, it can be used to compare nested (reduced) models.

Anybody familiar with this?

Thanks in advance!
Allan

Allan Lundy, PhD
Research Consulting
[hidden email]

Business & Cell (any time): 215-820-8100
Home (8am-10pm, 7 days/week): 215-885-5313
Address:  108 Cliff Terrace, Wyncote, PA 19095
Visit my Web site at www.dissertationconsulting.net

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

[Martin Holt]

----- Forwarded Message ----
From: M HOLT <[hidden email]>
To: "Allan Lundy, PhD" <[hidden email]>
Sent: Thursday, 10 June, 2010 13:08:34
Subject: Re: Logistic regression with nested-factor design?

Dear Allan,

 

So you have about 12 groups on one manipulation, and about 12 groups on one other manipulation. You say that each group within each manipulation are more likely to be similar to other other groups on that manipulation because "individuals influenced each other" hence results within manipulation are more likely to be similar than results across manipulations.

 

You say there are 3-7 persons in each group: a total of between 24x3 = 72 compared to 24x7 = 168 sample size.

 

If you were to perform logistic regression, for the least likely outcome, you would require a minimum of 10 people per factor in the test (just treating it as a straightforward logistic regression). So the number of factors you could plan to use would be between 7 and 17 per test. Many people say that 15 (not 10) are required per factor: ~ 5 to 11 depending on the actual sample size. How does that fit with you ?

 

You could do a logistic model, where you can test if the removal of one factor changes -2Log10 by a statistically significant event. (You'd also need to know the change in the number of degrees of freedom). But with so many groups, it's really like going on a fishing trip (IMHO).

 

Other group members will chime in, I hope, but I would recommend a hierarchical (i.e. multi-level) model. When comparing just two groups, this is equivalent to a logistic regression, but if you work out how many comparisons of two groups (A vs B), I think you'll understand why I call it a fishing trip. Cluster analysis seems to me to be the way to go....I don't know if that is OK with, say, 3 individuals in one group.

 

Professor Martin Bland has extensive references on this:

 

 

---

http://www-users.york.ac.uk/~mb55/clust/clusref.htm

 

There are about 10 articles cited herein, most available by double-click.

 

http://www-users.york.ac.uk/~mb55/clust/clustud.htm more on the design side.

 

HTH, and I would appreciate someone telling me if there is a minimum sample size per group. (I'm sure it's in these links; it's just
I've got to go soon.

 

Martin Holt

 

 

 

---

From: "Allan Lundy, PhD" <[hidden email]>
To: [hidden email]
Sent: Wednesday, 9 June, 2010 23:28:18
Subject: Logistic regression with nested-factor design?


Dear Listers,
I have a client with a design in which there are 3-7 persons in each of a couple of dozen groups.  Half the groups were presented with one manipulation, half with an alternative.  Individuals made a binary choice after discussion.  It is clear this is a nested-factor design.  It is not legit to ignore group membership because individuals influenced each other, hence individual results are likely to be more alike between 2 people in the same group than in different groups.  It would be nice to use a logistic regression, since the outcome measure is dichotomous.  But I can find nothing in SPSS on how to do this.  Until I looked at

http://www.ats.ucla.edu/stat/spss/output/logistic.htm
and found:

e.  -2 Log likelihood - This is the -2 log likelihood for the final model.  By itself, this number is not very informative.  However, it can be used to compare nested (reduced) models.

Anybody familiar with this?

Thanks in advance!
Allan

Allan Lundy, PhD
Research Consulting
[hidden email]

Business & Cell (any time): 215-820-8100
Home (8am-10pm, 7 days/week): 215-885-5313
Address:  108 Cliff Terrace, Wyncote, PA 19095
Visit my Web site at www.dissertationconsulting.net ===================== 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

[Bruce Weaver]

If I follow, the main explanatory variable in your model is a binary variable I'll call Treatment (Treatment 1 vs Treatment 2).  But subjects were not tested individually--rather they were tested in groups of 3-7 persons.  About 12 such groups were tested in each of the two treatments.  And the total N is somewhere around 120.  Is this right?  If so, I agree with others who have suggested multilevel logistic regression.  But as noted by Ryan, SPSS cannot currently run that model for you.  

What do the proportions for the binary response look like within the groups?  If they are not too close to the extremes (0 or 1), you would get a half-decent model by running a multilevel linear model (via MIXED).  In this case, the coefficient for Treatment would give you a risk difference, as opposed to the odds ratio you'd get with logistic regression.

HTH.

Allan Lundy wrote

Dear Listers,
I have a client with a design in which there are 3-7 persons in each of a
couple of dozen groups.  Half the groups were presented with one
manipulation, half with an alternative.  Individuals made a binary
choice after discussion.  It is clear this is a nested-factor
design.  It is not legit to ignore group membership because
individuals influenced each other, hence individual results are likely to
be more alike between 2 people in the same group than in different
groups.  It would be nice to use a logistic regression, since the
outcome measure is dichotomous.  But I can find nothing in SPSS on
how to do this.  Until I looked at

http://www.ats.ucla.edu/stat/spss/output/logistic.htm 
and found:
e.  -2 Log likelihood - This is the -2 log likelihood for the
final model.  By itself, this number is not very informative. 
However, it can be used to compare nested (reduced) models.
Anybody familiar with this?
Thanks in advance!
Allan


Allan Lundy, PhD
Research Consulting
[e-mail]
Business & Cell (any time): 215-820-8100
Home (8am-10pm, 7 days/week): 215-885-5313
Address:  108 Cliff Terrace, Wyncote, PA 19095
Visit my Web site at

www.dissertationconsulting.net  


=====================
To manage your subscription to SPSSX-L, send a message to
[e-mail] (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