PROC MI FCS WARNING -LINEAR COMBINATION OF OTHER EFFECTS

[OMA]

Hello,

I am looking for help with troubleshooting these warnings that I am getting when I run proc mi FCS. I am analyzing data from a randomized clinical trial. Assessment are at 6 week interval, at each interval there are either binary flags for new diagnosis, or medication use, and also continuous variables and ordinal variables. Note that the same variables are measured at each 6 week interval. 

I have been getting the following warning "An effect for the variable X is a linear combination of other effects. The coefficient of the effect will be set to zero in the imputation". This appears for almost all variables in the imputation model. I have used proc corr to check correlations of continuous variables and none are =1. Next I took each variable out one-by-one to see which one generated the error. I have been identified 3 variables (total 6 measured at different timepoints) that generate the warning. However, I am not sure how they might have a linear combination. These variables are simply just binary flags for  the presence of specific diagnosis at weeks 6 and 12. Note that if I put W6 only or w12 only, then there is no warning. The 2X2 table is below.

DBD_W6	DBD_W12	N
.	.	11
0	.	80
0	0	351
1	.	4
1	1	17

ADHD_W6	ADHD_W12	N
.	.	11
0	.	74
0	0	330
1	.	10
1	1	38

SEVERE_W6	SEVERE_W12	N
.	.	11
0	.	84
0	0	365
0	1	2
1	1	1

Another warning is "The maximum likelihood estimates for the FCS method logistic model for variable X in an iteration process may not exist. The resulting posterior predictive distribution of the parameters used in the imputation process is based on the maximum likelihood estimates in the last maximum likelihood iteration. the likelihood=augment option can be used to derive the last maximum likelihood estimates by using an augmented dataset" This is only generated for one variable. This is an ordinal variable (likert 7 point scale) which I have specified as FCS LOGISTIC (X/order=internal). What does this warning mean? and how should I fix it? I am not sure what it means to use a likelihood=augment option.

[ballardw]

It is usually a good idea to post the code you are using along with the warning/error messages. Copy it all from the Log to be sure that the code reflects what was actually run. Paste into a code box opened using the forum's {I} icon to maintain formatting as the message windows will reformat text, sometime losing diagnostic information.

In the extremely limited example data you show, every non-missing value of DBD_W12 = DBD_W6. So those variables would be linearly related. Same with ADHD_W12 and ADHD_W6

You may want to revisit your Proc Corr. When I run a correlation using the posted values and counts for the DBD variables I get a correlation of 1 with p < 0.0001