Expectation Maximzation for missing data - odd values?

[Kathryn Gardner]

Dear list,
I am working with questionnaire data and have used the EM (Expectation
Maximization) algorithm to estimate missing values in the data set. I have
noticed that some strange values have been though imputed though(i.e values
that do not fit within the possible range of scores e.g., outside the 1-5
range) and some minuses. So my questions are this:

1) are out of range values OK, or is this an error in the analysis?
2) are minus values acceptable, or is this an error in the analysis?

It doesn't seem to make sense if I have out of range OR negative values if
I am then using these item scores to compute an overall scale score.

Any advice welcome!
Kathryn

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

[John F Hall]

Kathryn

 

Definitely something fishy here.  Your imputed values should all be in the range 1-5.  Have you forgotten to declare missing values?

 

My instinct (as a survey researcher) would be to forget imputed values and use the data as is, but I'm not a purist.  If you send me examples of your questions (facsimile questionnaire in .doc or .pdf format) and also (in confidence) your SPSS saved file, I can offer better advice.

 

John Hall

----- Original Message -----

From: [hidden email]

To: [hidden email]

Sent: Friday, March 19, 2010 9:43 AM

Subject: Expectation Maximzation for missing data - odd values?

Dear list,
I am working with questionnaire data and have used the EM (Expectation
Maximization) algorithm to estimate missing values in the data set. I have
noticed that some strange values have been though imputed though(i.e values
that do not fit within the possible range of scores e.g., outside the 1-5
range) and some minuses. So my questions are this:

1) are out of range values OK, or is this an error in the analysis?
2) are minus values acceptable, or is this an error in the analysis?

It doesn't seem to make sense if I have out of range OR negative values if
I am then using these item scores to compute an overall scale score.

Any advice welcome!
Kathryn

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

[tjohnson]

Hi Kathryn

I think I have come across this before, and to my mind, it does make sense.  Now bear in mind that I am not a formally trained statistician, and I have not used the Expectation Maximisation procedure a great deal, but imagine that you are predicting a set of scores from another set of scores. 

If an individual scores lower than all other respondents on the second set of scores (those you are predicting from) but all respondents score extremely low on the first set of scores (those that are missing for this particular respondent), then there is no room to go lower than the current range of scores.  This would mean that for this respondent, for this particular set of attributes, you would expect to receive an imputed score of less than 1. 

Please do check that the procedure is working correctly.  However, I am not surprised that you are receiving scores outside your current range.  One option for using these scores is to bucket them into the current range of scores.  This would mean that someone receiving an imputed score of less than 1 would be given a new score of 1, an imputed score of more than 5 would be recoded as 5, and an imputed score of 2.43 would be rounded to a score of 2.

Good luck with your estimation!
Tim


-----Original Message-----
From: SPSSX(r) Discussion on behalf of Kathryn Gardner
Sent: Fri 19/03/2010 8:43 AM
To: [hidden email]
Subject:      Expectation Maximzation for missing data - odd values?

Dear list,
I am working with questionnaire data and have used the EM (Expectation
Maximization) algorithm to estimate missing values in the data set. I have
noticed that some strange values have been though imputed though(i.e values
that do not fit within the possible range of scores e.g., outside the 1-5
range) and some minuses. So my questions are this:

1) are out of range values OK, or is this an error in the analysis?
2) are minus values acceptable, or is this an error in the analysis?

It doesn't seem to make sense if I have out of range OR negative values if
I am then using these item scores to compute an overall scale score.

Any advice welcome!
Kathryn

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