Performing a log-linear analysis when cells aren't independent (because same participant contributes to multiple cells)
Locked
Performing a log-linear analysis when cells aren't independent (because same participant contributes to multiple cells)
 
|
Hi everyone, I'm wondering how to do a log-linear analysis in which cells aren't independent. In my psychological study on learning, I manipulated independent variable A (two conditions A1 & A2) and independent variable B (two conditions B1 and B2), and look at the effect on the dependent measure: whether people discovered pattern X, pattern Y, or both X and Y (FOr this test I'm not looking at how many people didn't discover a pattern). The data looks like this (fictional numbers):
I'd like to test whether A and B interact in influencing whether people discover X versus discover Y. Since a small minority of people discover both, I'd like to create two cells X' and Y' (as below). I'd like to just do a loglinear analysis of A x B x (X', Y'), but the problem is that the same observations contribute to multiple cells, and so they're not independent – which as far as I know is an assumption of loglinear tests. Any suggestions for how I can do this analysis? Is there anything like a within-subjects or repeated-measures equivalent for log-linear analysis? (btw there are a couple reasons I prefer not to just do the A X B X (X, Y, X &Y) analysis).
Thank you, Joseph Joseph Williams PhD Student in Cognitive Psychology, UC Berkeley
|
Re: Performing a log-linear analysis when cells aren't independent (because same participant contributes to multiple cells)
|
|
I believe you want to try the generalized linear mixed models in SPSS. This has a function for loglinear variables in the analysis. This will allow for a
multi-level loglinear model. Unfortunately I’ve never done this before, so I really can’t be of much help. Matthew J Poes Research Data Specialist Center for Prevention Research and Development University of Illinois 510 Devonshire Dr. Champaign, IL 61820 Phone: 217-265-4576 email:
[hidden email] From: SPSSX(r) Discussion [mailto:[hidden email]]
On Behalf Of Joseph Williams Hi everyone, I'm wondering how to do a log-linear analysis in which cells aren't independent. In my psychological study on learning, I manipulated independent variable A (two conditions A1 & A2) and independent variable B (two conditions B1 and B2), and look
at the effect on the dependent measure: whether people discovered pattern X, pattern Y, or both X and Y (FOr this test I'm not looking at how many people didn't discover a pattern). The data looks like this (fictional numbers):
I'd like to test whether A and B interact in influencing whether people discover X versus discover Y. Since a small minority of people discover both, I'd like to create two cells X' and Y' (as below). I'd like to just do a loglinear analysis of
A x B x (X', Y'), but the problem is that the same observations contribute to multiple cells, and so they're not independent – which as far as I know is an assumption of loglinear tests. Any suggestions for how I can do this analysis? Is there anything like a within-subjects or repeated-measures equivalent for log-linear analysis? (btw there are a couple reasons I prefer not to just do the A X B X (X, Y, X &Y) analysis).
Thank you, Joseph Joseph Williams PhD Student in Cognitive Psychology, UC Berkeley |
Re: Performing a log-linear analysis when cells aren't independent (because same participant contributes to multiple cells)
|
|
In reply to this post by Joseph Williams-2
Hi Joseph,
You can fit this kind of data with a generalized linear mixed model (GENLINMIXED) or generalized estimating equations (GENLIN). See http://publib.boulder.ibm.com/infocenter/spssstat/v20r0m0/topic/com.ibm.spss.statistics.help/idh_glmm.htm and http://publib.boulder.ibm.com/infocenter/spssstat/v20r0m0/topic/com.ibm.spss.statistics.help/idh_idd_gee_repeated.htm, and the related topics. Alex From: Joseph Williams <[hidden email]> To: [hidden email] Date: 05/16/2012 01:55 PM Subject: Performing a log-linear analysis when cells aren't independent (because same participant contributes to multiple cells) Sent by: "SPSSX(r) Discussion" <[hidden email]> Hi everyone,
I'm wondering how to do a log-linear analysis in which cells aren't independent. In my psychological study on learning, I manipulated independent variable A (two conditions A1 & A2) and independent variable B (two conditions B1 and B2), and look at the effect on the dependent measure: whether people discovered pattern X, pattern Y, or both X and Y (FOr this test I'm not looking at how many people didn't discover a pattern). The data looks like this (fictional numbers):
I'd like to test whether A and B interact in influencing whether people discover X versus discover Y. Since a small minority of people discover both, I'd like to create two cells X' and Y' (as below). I'd like to just do a loglinear analysis of A x B x (X', Y'), but the problem is that the same observations contribute to multiple cells, and so they're not independent – which as far as I know is an assumption of loglinear tests.
Any suggestions for how I can do this analysis? Is there anything like a within-subjects or repeated-measures equivalent for log-linear analysis? (btw there are a couple reasons I prefer not to just do the A X B X (X, Y, X &Y) analysis).
Thank you,
Joseph
|
Re: Performing a log-linear analysis when cells aren't independent (because same participant contributes to multiple cells)
|
|
In reply to this post by Joseph Williams-2
I probably can't help with what program to use, unless the
answer is to retreat to repeated measures. But I would like to figure out what the example is specifying. If my guess-work here is wrong, please forgive the guess, and please correct my error -- As I follow it, you have each subject in each of the 4 conditions. You are tabulating how many detected X-alone, Y-alone, and X+Y. Since the X+Y is small, I *think* that you have 51+61+5 identifications in the first column. Since the overlap (X+Y) is small, the total N must be hundreds. If the not-real numbers are true-to-life in showing similar Ns across a row, it *seems* that each person, with exceptions mostly in the X+Y row, performed the same in all 4 conditions. - For practical purposes of analyzing the 4 cells, every person can be *dropped* who does the same in all 4 cells. That implies that your effective N is more like 30 than it is 300. - The fictional numbers, as it happens, suggests a simple linear model in proportions, in the remaining numbers. So an ANOVA would give a fine fit and valid tests. -- Rich Ulrich Date: Wed, 16 May 2012 11:53:18 -0700 From: [hidden email] Subject: Performing a log-linear analysis when cells aren't independent (because same participant contributes to multiple cells) To: [hidden email] Hi everyone, I'm wondering how to do a log-linear analysis in which cells aren't independent. In my psychological study on learning, I manipulated independent variable A (two conditions A1 & A2) and independent variable B (two conditions B1 and B2), and look at the effect on the dependent measure: whether people discovered pattern X, pattern Y, or both X and Y (FOr this test I'm not looking at how many people didn't discover a pattern). The data looks like this (fictional numbers):
I'd like to test whether A and B interact in influencing whether people discover X versus discover Y. Since a small minority of people discover both, I'd like to create two cells X' and Y' (as below). I'd like to just do a loglinear analysis of A x B x (X', Y'), but the problem is that the same observations contribute to multiple cells, and so they're not independent – which as far as I know is an assumption of loglinear tests. Any suggestions for how I can do this analysis? Is there anything like a within-subjects or repeated-measures equivalent for log-linear analysis? (btw there are a couple reasons I prefer not to just do the A X B X (X, Y, X &Y) analysis).
Thank you, Joseph Joseph Williams PhD Student in Cognitive Psychology, UC Berkeley
|
| Free forum by Nabble | Edit this page |