No degrees of freedom left for perturbations

[lea...@gmail.com]

Hi,

I used IVEware to generate 20 imputed datasets, and the log showed the following message for some variables:

  Warning: No residual variance for perturbations

1. Is this a problem and how should I handle it? 

2. Also, I have read the IVEware manual and accompanying article but would still like additional guidance on how to use the various options. What are some good resources/tutorial for using IVEware?

3. Is there any way to indicate that I want certain variables to be used to contribute information for the imputation of other variables, but for these variables themselves not to "receive" any imputations (i.e., I want the missing values for these variables to remain missing). 

Many thanks,

Mei Yi

[Trivellore Raghunathan]

I think the predictors for this variable are highly collinear. What is your sample size? How many variables? One option is to use minrsqd or maxpred commands to reduce the number of predictors.

Raghu

--
You received this message because you are subscribed to the Google Groups "IVEware" group.
To unsubscribe from this group and stop receiving emails from it, send an email to iveware+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Trivellore Raghunathan (Raghu)

Director, Survey Research Center

Research Professor, Institute for Social Research

Professor, Department of Biostatistics

University of Michigan

Phone: (734)-764-8365

Fax: (734)-763-9831

www.src.isr.umich.edu

www-personal.umich.edu/~teraghu

"A good life is filled with selfless actions full of compassion knowing well that we are all one"

[lea...@gmail.com]

Dear Raghu,

Thanks for your response. The sample size is 91 and there are 74 variables in the tdataset, out of which 69 variables have missing values. Some of the variables are measures of similar constructs, and I was planning to average them after doing the imputation and before doing my analysis. Other variables are included as "auxiliary variables" to improve the imputation but I do not plan to use in analyses. Should I keep reducing the number in maxpred until I no longer get the warnings? (I get both warnings,   Warning: No residual variance for perturbations, and   Warning: No degrees of freedom left for perturbations).

Many thanks,

Mei Yi

[Trivellore Raghunathan]

What is the missing data pattern for the constructs that you plan to average? If they are all either fully  observed or fully missing then you can average them first and impute the averages. Otherwise specify maxpred as 30.

Raghu

[lea...@gmail.com]

Dear Raghu,

The constructs I plan to average are all missing for some participants, and partially missing for other participants. I specified maxpred=30 and there were no more "Warning: No degrees of freedom left for perturbations" BUT there were till numerous (>5000) "Warning: No residual variance for perturbations." I noticed that the warnings frequently came up for subscales of a measure that I did not plan to average because each subscale represented a different construct, but all subscales would be either present or missing if the participant did not complete the measure. I averaged two conceptually similar subscales but the same warnings still came up as frequently as before. So I dropped one of the subscales and there were no more warnings, even when I did not limit maxpred. 

This solved the problem except that I will not be able to use that dropped subscale in my analyses. The dropped subscale is not part of my main analyses but it would be nice to be able to include it in some exploratory analyses. Do you know if I can do anything else to include the dropped subscale? I would be very grateful for your advice.  

Thank you,

Mei Yi