In addition to what Jeremy said, you need to consider your population of
participants. If they are not a random sample (and true random samples
are pretty rare in psychological research) then you can't extrapolate
from that sample to the population.
Normally (TTBOMK) we're interested in having a widely distributed range
of answers in the population, and the mere presence of skew in the
sample does not necessitate that the population distribution is skewed.
That being said, skew is often quite informative - just because an item
tends to get particular responses at the edge quite often does not
necessarily make it a bad item. It may be poor at discriminating (which
is presumably why people are advised to discard skewed items), but it
can provide a lot of useful information with regards to the people who
do not endorse it. A test item asking what 2+2 is equal to is going to
be skewed in most populations, but that does not necessarily make it a
bad item.
Finally, discarding items on the basis of tests of skew (especially with
such a small, non random sample) is probably not the best of ideas. If
you do intend to discard items use factor analysis or preferably item
response theory so that you have a better basis and justification for
the discarded items.
Hope this helps,
Richie.
On Thu, 2011-11-10 at 06:55 -0800, Jeremy Miles wrote:
> What do you mean by 'significantly skewed'. Do you mean that the
> p-value of the skew is significant, or that they are very skewed.
> (Hint: never say significant on its own, say "statistically
> significant" or use another word, so it's clear what you mean).
>
> The problem with ranking is that everyone is going to get the same
> total score, and you can't compare people. In addition, ranking more
> than a small set of items is very hard.
>
> One option might be to change your responses. I don't know what your
> items or responses are, so here's an example. Everyone (almost)
> likes their doctor. If you have the item "I like my doctor" and the
> answers are: strongly disagree, disagree, dk, agree, strongly agree,
> you'll get a massively skewed item, 'cos everyone will agree or
> strongly agree.
>
> Instead, change your responses, make your question something like:
>
> How good is your doctor.
>
> 1. OK
> 2. Good.
> 3. Great.
> 4. Absolutely fantastic.
> 5. Better than the best doctor I could imagine existing in the universe.
>
> Hope that helps,
>
> J
>
>
> On 10 November 2011 06:35, Sandi Dheensa <[log in to unmask]> wrote:
> > Hi all
> > I'm a final year PhD student in need of some advice regarding psychometrics.
> >
> > I'm trying to construct a scale that looks at men's relationship to their
> > unborn child, to see if they conceive of their relationship in a
> > biological/genetic sense, or in a more socially constructed sense (so
> > whether they're thinking about how the baby will have their DNA, and whether
> > they think that having a genetic link is really important or whether they
> > are thinking more about how to bond with the baby when it's born, and how
> > they'll help to nurture the child etc).
> >
> > There are 31 items and I've piloted it, so far, on 55 people and thought I'd
> > do an interim analysis. Unfortunately, all of the items concerning socially
> > constructed ways of enacting fatherhood are significantly skewed. According
> > to item analysis, I should discard these items...does anyone know of any
> > other ways to manage skew in scale construction? Do I really have no choice
> > but to discard the items?
> >
> > Also, I thought about maybe changing the format of the scale from likert to
> > rank-ordering, so rather than comparing scores on biological and socail
> > forms of parenting, I could get men to rank which "ways of parenting" are
> > most salient to them. Does anyone have any experience with using
> > rank-ordering surveys, and any advice about the advantages and disadvantages
> > of using such a format?
> >
> > Thank you!
> >
> > Sandi
|