SPSSX Discussion

MVA in SPSS v17

Classic

List

Threaded

2 messages Options

Matthew Fuller-Tyszkiewicz

MVA in SPSS v17

Hello list,

I am currently reading about various methods for handling missing data and have encountered, on several occasions, mention that EM through SPSS MVA is flawed. As these articles were referring to older versions of SPSS, I am wondering if SPSS v17 is similarly flawed?

Kind regards,

Matt.

Dr Matthew Fuller-Tyszkiewicz, Lecturer
Deakin University, Waterfront campus Australia.
Phone: 03 5227 8715 International: +61 3 5227 8714
Fax: 03 5227 8621 International: +61 3 5227 8621
Email: [hidden email]
Website: http://www.deakin.edu.au
Deakin University CRICOS Provider Code 00113B (Vic)

Important Notice: The contents of this email are intended solely for the named addressee and are confidential; any unauthorised use, reproduction or storage of the contents is expressly prohibited. If you have received this email in error, please delete it and any attachments immediately and advise the sender by return email or telephone.
Deakin University does not warrant that this email and any attachments are error or virus free.

SPSS Support

Re: MVA in SPSS v17

Hello Matt,

I’ve pasted a note below my signature that addresses a problem in the imputation of values for individual cases as implemented with the EM algorithm. As noted there, Statistics 17 had added multiple imputation to the MVA module and several other Statistics procedures are capable of pooling results from the imputed samples. If you open the Help->Contents menu in Statistics 17, click the Search tab and enter “Analyzing Multiple Imputation Data”, the topic page that is found has a link for “Procedures that Support Pooling”. Clicking this link will give you a list of procedures that pool multiple imputation results and the particular outputs of those procedures for which pooling is applied.

As noted in the resolution below, multiple imputation is the preferred method of imputing values for cases, rather than imputation of a single sample by the EM method. If there were other concerns about the EM method in MVA, such as concerns about the EM estimates of means, standard deviations, correlations, and covariates, then we would need some more detail about the flaws that were suggested by the articles. Links or full citations to the articles in question would also be helpful. Thanks,

David Matheson

SPSS Statistical Support

*********************************

Resolution number: 32730 Created on: May 20 2003 Last Reviewed on: Dec 10 2008

Problem Subject: MVA EM-imputed Variances too small

Problem Description: I am using the SPSS Missing Values Analysis (MVA) procedure to obtain a complete data set. I used the EM (Expectation-Maximization) method to estimate the covariance matrix and impute values. As a check, I used the Analyze->Correlations->Bivariate to calculate the covariance of the completed data. The variances of the completed data from the Correlations procedure were lower than the variance estimates produced by EM. What is wrong?

Resolution Subject: The values imputed by EM are the values computed in the last "Expectation" step performed by the algorithm.

Resolution Description:
The values imputed by EM are the values computed in the last "Expectation" step performed by the algorithm. As you have observed, adding them to the data will introduce a downward bias to the variances. Also, unlike the EM estimated statistics computed in MVA, the imputed data cannot make use of information on joint missingness among variables with missing values for the same case. Using the EM-imputed values without further modification is not recommended.

Beginning with Release 17, the MVA module also includes options for creation of multiple imputations and the ability in many procedures to produce combined results from analyses of the created imputations using Rubin's combining rules.

In releases prior to Release 17, two options exist for dealing with this issue in SPSS. One is to use the OMS facilities to save the results of EM estimation from MVA to a file (or otherwise capture the results) and use these results directly in appropriate models. The EM option produces estimates of means, standard deviations, correlations and covariances.

If the analyses to be performed require additional information, so that one must have raw data, the second option is to manually add random variability to the imputed values produced by the EM-based imputation in MVA.

It should be noted that even with a better set of imputations, a single set of data with imputations cannot simply be naively analyzed as a complete set of data without further adjustments. At the very least, the degrees of freedom assumed must be reduced to account for the missing data, and no specific value can in general be easily derived as the correct value. Also, use of a single set of imputed data of any quality will generally not capture all of the uncertainty involved with the fact that some of the data have been imputed. Accordingly, attention to the area has moved to the multiple imputation approach, and this is likely to be the major focus of future enhancements to the SPSS MVA module.

From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Matthew Fuller-Tyszkiewicz
Sent: Monday, June 08, 2009 6:18 PM
To: [hidden email]
Subject: MVA in SPSS v17

Hello list,

Kind regards,

Matt.

Important Notice: The contents of this email are intended solely for the named addressee and are confidential; any unauthorised use, reproduction or storage of the contents is expressly prohibited. If you have received this email in error, please delete it and any attachments immediately and advise the sender by return email or telephone.
Deakin University does not warrant that this email and any attachments are error or virus free.