PROC MI FCS Imputation Error

[ROLuke91]

Hi all,

I am experiencing an issue running a PROC MI FCS Regression. Below is my model:

proc mi data = data  seed=123456799 out=miFullImpute minmaxiter = 3500 minimum = 0 maximum = 136 nimpute=50 noprint;
   by VAR;
   fcs reg(Y = X1 X2 X3 X4 X5 X6);
   var X1 X2 X3 X4 X5 X6 Y;
run;

The errors I get are the following:

WARNING: An effect for variable Y is a linear combination of other effects. The

         coefficient of the effect will be set to zero in the imputation.

NOTE: The above message was for the following BY group:

      VAR=0

ERROR: An imputed variable value is not in the specified range after 3500 tries.

NOTE: The above message was for the following BY group:

      VAR=1

I'm not sure how to troubleshoot this issue. I ran a correlation for the entire model and none of X1-X6 correlate at 1.00 with Y. The content of the variables should not linearly combine to be Y as well. Not sure how to approach the "imputed variable outside the range" error, either.

It appears that X5 is the problem variable. X5 is specified to be continuous, but is a variable with only 13 potential integer values (0-12). Moreover, none of the observations have any values in X5 = 9, 10 (they represent counts). There is a significant skew to those with X5 = 8 (40/121) and 12 (34/121), with the other values (0-7, 11) having counts ranging from 1-10. I am assuming this is contributing to the problem.

When I dichotomize the variable to be <8 or >=8 and treat it as a class indicator, the "imputed variable..." error goes away, but the linear combination remains.

Can anybody provide some insight on how to troubleshoot? This particular variable is too germane to Y to be removed, so unsure how to proceed.

Thanks and Best,

Luke

[ballardw]

---

I'm not sure how to troubleshoot this issue. I ran a correlation for the entire model and none of X1-X6 correlate at 1.00 with Y. The content of the variables should not linearly combine to be Y as well.

"Should not" is no guarantee that they do not in your data.

WARNING: An effect for variable Y is a linear combination of other effects. The

         coefficient of the effect will be set to zero in the imputation.

Usually means that one (or more) of the X1 to X6 can be created from some of the other X variables. For instance if X1 = 2*X2 that would be a linear combination of other affects.

The error for VAR=1 could be that you have provided a limit with Minimum or Maximum options that cannot be met for one or more of the values. Which would be that every iteration when an attempted imputation was performed that the value was outside the range [0,136].

So you might increase Minmaxiter to see if you get acceptable values, change the restriction to allow a wider range.

Perhaps you need to look at your data a bit closer