On 19 March 2013 10:16, Nina Bay <[log in to unmask]> wrote:
> Hi Jeremy,
>
>
>
> Thank you again for your precious help! You are always very clear and
> concise.
>
>
>
> I am convinced… 5 regressions it is. Now a silly question... would I have to
> use a correction for error, since these are 5 different personality traits,
> but they still account for a common personality measure?
>
I wouldn't. People will say that you should use something like
Bonferroni correction, and then you say (something like) "Ah, but
shouldn't I only use Bonferroni when the p-values are uncorrelated,
and I have correlated outcomes and so it will be (even more)
conservative than normal." And "Bonferroni correction is
over-conservative anyway,
Here's my standard rant if they are not convinced:
We respectfully disagree with the reviewer’s comments that the
‘correct’ alpha is equal to 0.05/15=0.0033. The reviewer is
suggesting that we use Bonferroni correction, perhaps more properly
known as Dunn’s procedure (see Dunn, 1961; Stigler, 1980). This was
developed for use in contrasts following analysis of variance, and is
used to ensure that the nominal type I error rate is at or below the
value of alpha that is used. In post-hoc contrasts in analysis of
variance, this is true, because the contrasts are (or very nearly are)
orthogonal. However, in a study with multiple outcomes this is not
the case, and so the Bonferroni corrected alpha can be very much below
alpha. Even if this were not the case, we would still be averse to
using this form of correction. Miles and Banyard (2007) compared
Bonferroni correction for null hypothesis testing to using a chainsaw
for brain surgery, and many other examples can be found in the
literature, for example Perneger (1998) writing in the British Medical
Journal, suggested that “Bonferroni adjustments are, at best,
unnecessary and, at worst, deleterious to sound statistical
inference.” The p-value that emerges from each of our tests, if we
were to Bonferroni correct, is testing the null hypothesis that the
null hypothesis is true, for all outcomes simultaneously. This
multivariate null hypothesis is not of interest to us. Secondly, the
value of alpha that emerges from each test is partly dependent upon
the number of tests that are performed. If researchers are encouraged
to use Bonferonni correction, then samples have to use much larger
sample sizes, alternatively, to ensure that results have some degree
of consistency or accuracy, researchers should collect fewer data, and
we have trouble with a recommendation that scientists collect fewer
data. Finally, we note that the tests used are described as being
robust to violations of normality, and this is true when the value of
alpha that is used is 0.05, however, as we move closer to the tails of
the distribution, the level of robustness to violations is
considerably reduced.
>
>
> You are right in assuming that my 5 outcomes are the big five and my
> predictor is parenting style (with 3 levels: authoritarian, moderate, and
> lenient).
>
>
>
> You also left me quite curious when mentioning that I could be interested in
> whether parents respond differently to different personality styles of their
> children, in which case parenting style is the outcome.
>
>
>
> What statistical test would I use then? Multinomial Logistic Regression?
>
Yes, it would. (I think the p-value should be similar - or even the same.)
J
|