Hi Rebecca,
there are two ways to handle missing data.
Multiple imputation (MI) 'fills in' the holes in the dataset, and then
analyzes the data. You need to do multiple runs of imputations to
reflect the uncertainty.
(Full information) maximum likelihood (FIML/ML) estimation provides
estimates of the parameters based on the information that is available
- it does not fill in, or impute, the missing data. It just uses all
of the data that you have. ML is not used for MI (although it might
be used to estimate parameters, and just to confuse you, MI is often
done using expectation-maximization - or EM).
If you run your SEM in AMOS, it will use FIML, and you don't have a
problem. (Well, you have a problem, but much less of a problem). In
some SEM programs you can get the ML estimated descriptive statistics
(but I don't think you can in AMOS), but these are weird, because they
are model dependent - that is, if you change your model, they will
change. They assume that the model is correct, otherwise they are not
right.
You can do MI (I believe) in SPSS, but I've never done it.
Don't do mean substitution. It's evil. And bad.
There's a nice Sage little green book on this, by Paul Allison, called
'Missing Data'. One thing I've heard him say (not sure if it's in the
book) is that there are no good ways to deal with missing data. There
are just ways that are bad, and ways that are not so bad. Mean
substitution (and listwise deletion and pairwise deletion) is bad.
Jeremy
On 26 May 2010 07:50, Rebecca Graber <[log in to unmask]> wrote:
> Hi everyone,
>
>
>
> I’m hoping someone can help me with a stats question. I’m debating how to
> deal with the missing data in my (rather large) dataset for a
> non-experimental (survey) study. I will eventually be doing structural
> equation modelling using AMOS, so hope to use maximum likelihood (ML)
> imputation to estimate the missing values where a participant has skipped a
> given item. But prior to this I will be doing some correlational analyses
> and wanting to run my descriptives. Is there a way of using ML in Excel or
> SPSS? Or will I have to use mean substitution in Excel/SPSS to get my
> descriptives and correlations, and then separately use ML when I run my SEM?
> Or is there another way I’m not seeing?
>
>
>
> Thanks in advance for your help!
>
>
>
> Rebecca
>
>
>
> Rebecca Graber
>
> PhD Student
>
> Institute of Psychological Sciences
>
> University of Leeds
>
> Leeds LS2 9HT
>
> [log in to unmask]
>
> 0113 343 9197
>
>
>
> http://www.psyc.leeds.ac.uk/friendship
>
>
>
> http://www.justgiving.com/psycbennevis - Support the Alzheimer's Society as
> we hike Ben Nevis!
>
>
>
>
>
>
--
Jeremy Miles
Psychology Research Methods Wiki: www.researchmethodsinpsychology.com
|