On 11 November 2011 06:39, [log in to unmask] <[log in to unmask]> wrote:
> I find this question really interesting and some of the answers even more
> so.
>
> Jeremy: how much would you say that the shape of the distributions matter at
> all in this case, given what the central limit theoram implies about using
> parametric tests, even with only 50 or so participants
> http://www.graphpad.com/www/book/choose.htm
>
In the case that was described in the original question, very little.
The distribution of the questions wasn't going to be analyzed, but the
total scale was.
> One other thing: has anyone talked about transforming the data statistically
> e.g. log transformation? This often doesn't work for me too well with small
> distributions, but is always worth a try.
>
In the case above, the transformation won't have much effect, because
it's not a truly continuous variable, it's an ordered categorical
variable. Transformations can be useful - but if you do a log
transform on an item in a scale in order to get the item statistics,
you're going to have to do that when you score the scale. Which is
going to make it very hard to score.
The other problem with transformations is that you destroy linearity
and additivity, and you're often more interested in those. You want
to say something like "People who take pill A live longer than people
who take pill B". I ask "How much longer?" You say "I don't know,
'cos I did a complicated transformation on my outcome variable, and
that question isn't one I can answer".
> May I throw in a related question? Has anyone ever tried a Box-Cox
> transformation for really skewed distributions? If so, did it work and do
> you know of user-friendly free software that can do this?
>
>
Just do it by hand. Fiddle about with the lambda parameter until it
looks nice. I've never found B-C transformations massively useful.
Jeremy
|