Using MIXED to account for Examiner variability

[Nicola-16]

Hi all,
I have a dataset for about 440 patients (identified by Patient) and each
patient is randomly allocated to one of 2 different methods of surgery
(Group).  They are measured on a set of about 10 outcome variables (e.g.
ROM) at 4 different time points (with no missing data):
1.) baseline (the two methods should not be significantly different)
2.) 6 weeks post-surgery (observations have the lowest values during
the post-surgical recovery period, then generally improve from here
on)
3.) 3 months
4.) 6 months

One of 12 examiners measure a particular patient at a particular time
point.  It is not balanced, in that there is variation in the number
of patients measured by each examiner, and most examiners only take
measurements at 1 to 3 time points and a few of them were across
the two different methods (but at different time points).  A few
examiners appear to consistently record higher observation values (and
the biggest problem with this is at baseline), so I fitted examiner as
a random effect using the MIXED procedure:

MIXED
 ROM BY Time Group Examiner Patient
 /CRITERIA = CIN(95) MXITER(100) MXSTEP(5) SCORING(1) SINGULAR
 (0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
PCONVERGE (0.000001, ABSOLUTE)
 /FIXED = Time Group(Time)  | SSTYPE(3)
 /METHOD = REML
 /EMMEANS = TABLES(OVERALL)
 /EMMEANS = TABLES(Time)
 /EMMEANS = TABLES(Time*Group)
 /RANDOM Examiner Patient | COVTYPE(VC) .

The problem is the trend is still distorted (we expect an equal start
for the two methods at baseline and a dip at 6 weeks).  The marginal plots
of Examiner*Time in GLM shows that some Examiners are contrary to the
expected trend (e.g. higher measurements at 6 weeks post-surgery, some of
which appear to be a problem with outliers).

There are a few possible options I/we were thinking of:
1.) As two or three of the outcome variables are expected to be less
affected by the surgery (and of less interest to the study), but still
strongly affected by examiner effect, is there some way of obtaining
the examiner effect from these variables to plug into the model for
the other outcome variables?
2.) A method of time series analysis (I have never done this before,
so I have no idea what it might do)?
3.) Need for a different variance structure that account for the
heterogeneity of variance within examiner (those examiners that were
contrary to expected trend appear to be less consistent)? I tried COV(UN) as
suggested by someone in another forum (but without Patient as random effect
as SPSS ran out of memory even when I "SET WORKSPACE=614800.").

Thanks for your patience in getting this far and for any help/suggestions
you can give :)

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

[Reutter, Alex]

Nicola,

Instead of:

/RANDOM Examiner Patient | COVTYPE(VC)

Try using:

/RANDOM INTERCEPT | SUBJECT(Patient) COVTYPE(VC)
/RANDOM INTERCEPT | SUBJECT(Examiner) COVTYPE(VC)

If I'm reading your note correctly, in both syntax specifications, you're requesting an estimate of the variance term for Patient and the variance term for Examiner, but in the first specification, it's producing a 452x452 (diagonal) covariance matrix and in the second it's producing two 1x1 matrices.  Then, when you want to account for possible heterogeneity between examiners:

/RANDOM INTERCEPT | SUBJECT(Patient) COVTYPE(VC)
/RANDOM Examiner | COVTYPE(UN)

And it will produce a 1x1 matrix with the variance term for Patient and a 12x12 covariance matrix for Examiner.  That should, I think, solve your memory problems.

Cheers,
Alex


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Nicola
Sent: Monday, February 23, 2009 8:42 PM
To: [hidden email]
Subject: Using MIXED to account for Examiner variability

Hi all,
I have a dataset for about 440 patients (identified by Patient) and each
patient is randomly allocated to one of 2 different methods of surgery
(Group).  They are measured on a set of about 10 outcome variables (e.g.
ROM) at 4 different time points (with no missing data):
1.) baseline (the two methods should not be significantly different)
2.) 6 weeks post-surgery (observations have the lowest values during
the post-surgical recovery period, then generally improve from here
on)
3.) 3 months
4.) 6 months

One of 12 examiners measure a particular patient at a particular time
point.  It is not balanced, in that there is variation in the number
of patients measured by each examiner, and most examiners only take
measurements at 1 to 3 time points and a few of them were across
the two different methods (but at different time points).  A few
examiners appear to consistently record higher observation values (and
the biggest problem with this is at baseline), so I fitted examiner as
a random effect using the MIXED procedure:

MIXED
 ROM BY Time Group Examiner Patient
 /CRITERIA = CIN(95) MXITER(100) MXSTEP(5) SCORING(1) SINGULAR
 (0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
PCONVERGE (0.000001, ABSOLUTE)
 /FIXED = Time Group(Time)  | SSTYPE(3)
 /METHOD = REML
 /EMMEANS = TABLES(OVERALL)
 /EMMEANS = TABLES(Time)
 /EMMEANS = TABLES(Time*Group)
 /RANDOM Examiner Patient | COVTYPE(VC) .

The problem is the trend is still distorted (we expect an equal start
for the two methods at baseline and a dip at 6 weeks).  The marginal plots
of Examiner*Time in GLM shows that some Examiners are contrary to the
expected trend (e.g. higher measurements at 6 weeks post-surgery, some of
which appear to be a problem with outliers).

There are a few possible options I/we were thinking of:
1.) As two or three of the outcome variables are expected to be less
affected by the surgery (and of less interest to the study), but still
strongly affected by examiner effect, is there some way of obtaining
the examiner effect from these variables to plug into the model for
the other outcome variables?
2.) A method of time series analysis (I have never done this before,
so I have no idea what it might do)?
3.) Need for a different variance structure that account for the
heterogeneity of variance within examiner (those examiners that were
contrary to expected trend appear to be less consistent)? I tried COV(UN) as
suggested by someone in another forum (but without Patient as random effect
as SPSS ran out of memory even when I "SET WORKSPACE=614800.").

Thanks for your patience in getting this far and for any help/suggestions
you can give :)

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