I agree with Brian, if you are using factor analysis to develop a
scale, then throw out your crossloading items. E.g. if you are trying
to develop a measure of depression & anxiety, a crossloading means
that the item measures both. That's confusing, and you need to get
rid of it.
If you are using factor analysis to explore the dimensionality of a
construct, then you should keep the cross loading, and try to
interpret it.
Usually these things go hand in hand. Your factor analysis to develop
a scale informs your theoretical understanding of the construct, so
you can't be so fixed about one or the other. I think that Stan Mulaik
in his (1972) book on factor analysis said that you should throw out
those items - then, to demonstrate that you really know what those two
factors are, you should be able to write new items, which load on the
factor you predict. If they don't, you need to do that again. Of
course, that's something of a lifetime's work, not a PhD's work.
J
2009/6/12 Brian K. Saxby <[log in to unmask]>:
> Hi Hannah,
>
> I used Paul Kline's 'Easy Guide to Factor Analysis' in my PhD research and
> it did exactly what it said on the tin. I don't recall if it dealt
> directly with questionnaires and item selection, but may help you
> understand your factors, or maybe direct you to another type of factor
> analysis rotation (i.e. orthogonal vs oblique).
>
> If your purpose is to design a questionnaire that measures only one thing,
> then crossloadings can be seen as problematic - however, if the items that
> load on the second factor make sense together, you could well be tapping
> into two different domains. But if they don't make sense it could be
> unwanted noise. There are techniques of item analysis you can use that
> will tell you whether each item is contributing meaningfully to your
> total(s), so this might be worthwhile, as it could give you a rationale
> for excluding some problematic items.
>
> The other point I'd make though is to not be too hard on your results -
> some say factor analysis is more art than science - even well-established
> and simple measures like the 16 Bond-Lader Visual Analogue Scales have
> cross-loadings of a similar magnitude to yours. I'll forward their paper
> to you for reference (rather than to the group).
>
> Hope that helps,
>
> Brian
>
>
>> I'm hoping for some advice on how to treat items that crossload in an
>> exploratory factor analysis.
>>
>> For my PhD research I'm running a factor analysis to refine a
>> questionnaire I've created. I have several items that crossload onto a
>> second factor (at above .3 on both factors) and I'm not sure what to do
>> with these. The text books I've read don't say whether I delete them,
>> ignore them ...... Can anyone advise me on how to treat these?
>>
>>
>>
>>
>>
>>
>>
>>
>> I tried just deleting these items and re-running the factor analysis
>> without them but this then found another cross loading item. Removing this
>> item too gets rid of all cross loading items and creates a solution that
>> makes sense of the data. However this slightly reduces the overall
>> Cronbach's alpha to .78 (and for each of the five factors). The amount of
>> variance explained stays roughly the same as in the original solution
>> though.
>>
>> Any advice or a push in the right direction (a text boo
>
--
Jeremy Miles
Psychology Research Methods Wiki: www.researchmethodsinpsychology.com
|