imputing missing likert-type values

[ziweiguan]

My questionnaire was designed based on some "latent factors", that each "latent factor" has
a number of questions (items). The answers to the questions are likert-type scales (1 to 6, strong disagree to strong agree). In the end, there are  <2% values missing. Because the likert-type scales are ordinal variables, and has no normal distribution feature, so I feel multiple imputation will not work well.

I am thinking to use the object's median values from those similar questions within the same "latent factor" to replace the missing values of that object. This makes a lot of sense I think. Is there an option to do so?
In SPSS, it has "replace missing values with median of nearby points", which takes median of values above
or below from different objects. This is not what I want.

I have another idea, that I first do EFA, and then I can do SEM with AMOS. AMOS can do imputation too. Is this a good practice for missing data imputation?

Thank you very much

[Bruce Weaver]

You appear to be suggesting that multiple imputation can only be used for normally distributed variables.  That is not correct.  The Command Syntax reference manual entry for MULTIPLE IMPUTATION says this (under SCALEMODEL Keyword):

"By default, the type of univariate model that is used [to impute values where data are missing] depends on the measurement level of the variable whose missing values are to be imputed.  Multinomial logistic regression is always used for categorical variables."  

HTH.

ziweiguan wrote

My questionnaire was designed based on some "latent factors", that each "latent factor" has
a number of questions (items). The answers to the questions are likert-type scales (1 to 6, strong disagree to strong agree). In the end, there are  <2% values missing. Because the likert-type scales are ordinal variables, and has no normal distribution feature, so I feel multiple imputation will not work well.

I am thinking to use the object's median values from those similar questions within the same "latent factor" to replace the missing values of that object. This makes a lot of sense I think. Is there an option to do so?
In SPSS, it has "replace missing values with median of nearby points", which takes median of values above
or below from different objects. This is not what I want.

I have another idea, that I first do EFA, and then I can do SEM with AMOS. AMOS can do imputation too. Is this a good practice for missing data imputation?

Thank you very much

[Art Kendall]

Likert items are rarely very discrepant from interval level of measurement.

Do you know or do you have a guess why some items have missing values?

How was the questionnaire administered?  You may want to consider how to improve the quality of administration in future work.

If you scatterplot the results of the two approaches in my reply to your other post what do you see?

Are the scales from previous work that established the psychometric qualities of the scales?  OR Did you write the scale items your self specifically for this study?

Recall that a scale can be considered a summary of several imperfect repeated measures of a construct that is used as representing what is common to the set of items.  The summary is considered more representative of the underlying construct.

[ziweiguan]

what do people usually do when they handle the missing values of likert type scales?
This is the first time I work on a survey data, and first time with likert type scales too.
Can anyone share some opinions? Thanks.

[Rich Ulrich]

I almost always take the average of the items available for the
particular factor to represent the factor, especially when they are
Likert items or close to it. 

If item means are not homogeneous or lie at the extremes, then the
scale is not "Likert" or Likert-type.  On the other hand, even though it might
be wiser not to do so, they are usually (in my experience) treated as if they
were Likert items, and they are treated as if they were equal-interval and
normal.  That is simple and ordinary.  The newer approach that assumes
otherwise uses Item Response Theory, which has a logistic underpinning.
 
On the other hand, it seems uninformed to treat a designed, Likert scale as
if the items are categorical or (merely) ranks.  This "advice" seems to be a
hang-over from the 1950s when a few (non-statistician) psychologists started
regarding the so-called non-parametric statistics as magically, always valid.

NOTE on odd distributions.
For dichotomies, it is sometimes more meaningful to count number of items
endorsed, or number not-endorsed.  For instance there is a short Dementia
scale where the best Total score is 31, but a few items may be Missing
because the subject, for instance, is blind or immobile.  It would have been
wiser if the Scale had been normed as the count of deficiencies; which is
how I used it (that is, the Missing counted as Okay).   How you deal with
unusual cases has to be determined by the purpose of the study and the
substantial meaning of the answers. 

--
Rich Ulrich

> Date: Fri, 10 Jul 2015 08:00:22 -0700

> From: [hidden email]
> Subject: Re: imputing missing likert-type values
> To: [hidden email]
>
> what do people usually do when they handle the missing values of likert type
> scales?
> This is the first time I work on a survey data, and first time with likert
> type scales too.
> Can anyone share some opinions? Thanks.
>
>

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