Back to PLS HelpMissing Data in PLS?
Posted on 10/08/07 13:57:49
Number of posts: 2
Hello all,
I'd like to perform a behavioral PLS on a PET data set. For most subjects I have 5 PET scans from 3 conditions (two of the conditions were performed twice) and a behavioral measure corresponding to each PET scan, but a few are missing a scan or behavior data point for one or more conditions. Will PLS handle a variable number of scans and/or behavior data points across subjects or must I include only subjects with complete data?
Thanks in advance,
Israel Christie
University of Pittsburgh
I'd like to perform a behavioral PLS on a PET data set. For most subjects I have 5 PET scans from 3 conditions (two of the conditions were performed twice) and a behavioral measure corresponding to each PET scan, but a few are missing a scan or behavior data point for one or more conditions. Will PLS handle a variable number of scans and/or behavior data points across subjects or must I include only subjects with complete data?
Thanks in advance,
Israel Christie
University of Pittsburgh
Replies:
Untitled Post
Posted on 10/08/07 14:47:38
Number of posts: 394
quote:
Hello all,
I'd like to perform a behavioral PLS on a PET data set. For most subjects I have 5 PET scans from 3 conditions (two of the conditions were performed twice) and a behavioral measure corresponding to each PET scan, but a few are missing a scan or behavior data point for one or more conditions. Will PLS handle a variable number of scans and/or behavior data points across subjects or must I include only subjects with complete data?
Thanks in advance,
Israel Christie
University of Pittsburgh
Hi there,Hello all,
I'd like to perform a behavioral PLS on a PET data set. For most subjects I have 5 PET scans from 3 conditions (two of the conditions were performed twice) and a behavioral measure corresponding to each PET scan, but a few are missing a scan or behavior data point for one or more conditions. Will PLS handle a variable number of scans and/or behavior data points across subjects or must I include only subjects with complete data?
Thanks in advance,
Israel Christie
University of Pittsburgh
Unfortunately, the PLS program has no mechanism to deal with missing data, so you will need to either eliminate observations with missing data or make adjustments for those missing data points:
example
1) replace with subject's mean value
2) estimate subject's data based on a linear model of the rest of group (e.g, if X is missing, predict X from Y for the rest of the group and then plug in "Y" for the subject with the missing value to estimate their value for "X")
There are more sophisticated ways to do this using ML estimators, but again, that's not part of the PLS package.
Randy
Thank & Clarification
Posted on 10/08/07 16:19:35
Number of posts: 2
Thanks for the quick reply, Randy! Are you referring specifically to missing behavioral data points or scans/conditions as well? In other words, can I include a different number of scans/conditions for different subjects? For example, can most subjects the following conditions:
A, B, B, C, C
whereas a few might have only the following:
A, B, C, C
A, B, B
B, B, C, C
Hope that makes sense.
-Israel
A, B, B, C, C
whereas a few might have only the following:
A, B, C, C
A, B, B
B, B, C, C
Hope that makes sense.
-Israel
Untitled Post
I'm Online
Posted on 10/09/07 10:55:44
Number of posts: 291
quote:
Thanks for the quick reply, Randy! Are you referring specifically to missing behavioral data points or scans/conditions as well? In other words, can I include a different number of scans/conditions for different subjects? For example, can most subjects the following conditions:
A, B, B, C, C
whereas a few might have only the following:
A, B, C, C
A, B, B
B, B, C, C
Hope that makes sense.
-Israel
No.
Thanks for the quick reply, Randy! Are you referring specifically to missing behavioral data points or scans/conditions as well? In other words, can I include a different number of scans/conditions for different subjects? For example, can most subjects the following conditions:
A, B, B, C, C
whereas a few might have only the following:
A, B, C, C
A, B, B
B, B, C, C
Hope that makes sense.
-Israel
Untitled Post
Posted on 10/09/07 11:29:27
Number of posts: 394
quote:
Thanks for the quick reply, Randy! Are you referring specifically to missing behavioral data points or scans/conditions as well? In other words, can I include a different number of scans/conditions for different subjects? For example, can most subjects the following conditions:
A, B, B, C, C
whereas a few might have only the following:
A, B, C, C
A, B, B
B, B, C, C
Hope that makes sense.
-Israel
Hi Israel;Thanks for the quick reply, Randy! Are you referring specifically to missing behavioral data points or scans/conditions as well? In other words, can I include a different number of scans/conditions for different subjects? For example, can most subjects the following conditions:
A, B, B, C, C
whereas a few might have only the following:
A, B, C, C
A, B, B
B, B, C, C
Hope that makes sense.
-Israel
The situation would apply equally to brain or behavior. The computational overhead is higher for brain, of course, but the principle is the same. The problem is the nature of the missing value. In the scenario you outline, if condition "A" is missing, it might be harder to estimate than "B" or "C" where there is a replication (case #3 in your example). Similarly, in case #3, estimating "C" may be a little less perfect that if they had one value for "C". If there only a few subjects who have missing data, and you sample size is large(ish), then the estimation scheme will probably be okay.
Randy
Login to reply to this topic.
