Well to be precise, for the iris data, the MANOVA
gives a p value of something like 1e-40, which is
unreachable by NPMANOVA for a reasonable number of
permutations. I have only tested NPMANOVA for 100,000
permutations, giving p < 1e-5, so it is indeed
possible that NPMANOVA could be orders of magnitude
less powerful than MANOVA in this case.

I would be interested to see your data set.


Regards,
Oyvind Hammer
Natural History Museum
University of Oslo
[EMAIL PROTECTED]


On Sat, 10 Nov 2007, morphmet wrote:

> Dear professors,
>
> Thanks for your prompt reply. However we think there is a little
> misunderstanding between your replies. According to Dr. Hammer, the
> classical Iris data set from Fisher provides similar results between the
> parametric and non-parametric MANOVAS. Dr. Rohlf, on the other hand,
> suggests that it is expected to have different p-values (in opposite
> directions of "significance") between the parametric and non-parametric
> alternatives.
>
> In any case we agree with the suggestion from Dr. Rohlf about the
> requirements for larger data sets in the non-parametric tests.
>
> Given that p-values are not an absolute point of reference we are also
> relying on complementary descriptions of the distance among samples, and
> complementary indexes about the robustness of p-values.
>
> Pablo Menendez
> Pablo Jarrin
>
> Quoting morphmet <[EMAIL PROTECTED]>:
>
>> As Oyvind mentions, the difference in P-values in the two methods is
>> in
>> the expected direction. It is difficult to tell whether the magnitude
>> of
>> difference is 'reasonable' without knowing sample sizes and the
>> number
>> of landmarks. A nonparametric method requires larger sample sizes.
>> In
>> the discussions concern was mentioned about the assumption of
>> normality.
>>  MANOVA is based on the assumption of homogeneity of within-group
>> covariance matrices. That can be more of a problem that modest
>> departures from normality (the central limit theorem works for
>> multivariate data too).
>>
>> In the original question it was mentioned that the data shows
>> differences mostly along the PC1 axis. That implies that the
>> explanation
>> for the differences may be one-dimensional. If not size, then
>> temperature, north-south gradient, etc. In such cases a general
>> multidimensional method such as MANOVA will have less statistical
>> power
>> than a method such as multivariate regression on some possible
>> explanatory variables.
>>
>> ------------------------
>> F. James Rohlf, Distinguished Professor
>> Ecology & Evolution, Stony Brook University
>> www: http://life.bio.sunysb.edu/ee/rohlf
>>
>>
>>> -----Original Message-----
>>> From: morphmet [mailto:[EMAIL PROTECTED]
>>> Sent: Friday, November 09, 2007 9:35 AM
>>> To: morphmet
>>> Subject: Re: MANOVA vs npMANOVA
>>>
>>> Hi, it is interesting that you get so different
>>> results between MANOVA and NPMANOVA. Would you
>>> mind sending me the data set as saved from PAST
>>> for me to look at? (I'm the author of PAST).
>>>
>>> The standard 'Iris' test data set of Fisher gives
>>> quite similar results between MANOVA and NPMANOVA
>>> of PAST. In general, the NPMANOVA is expected to
>>> be less powerful (have higher p values) than MANOVA,
>>> I believe.
>>>
>>>
>>> Oyvind Hammer
>>> Natural History Museum
>>> University of Oslo
>>>
>>>
>>> On Fri, 9 Nov 2007, morphmet wrote:
>>>
>>>> Dear colleagues,
>>>>
>>>> We would like to take advantage of Dr. Slice's comments about
>> MANOVA
>>> to
>>>> assess differences among species or biological groups.  There are
>> few
>>>> multivariate normality tests available, however it is often
>> common
>>> that
>>>> Partial Warps (and therefore Relative Warps) from morphological
>>>> structures do not fit normality. Thus, we think this is the
>> main
>>> reason
>>>> why Dr. Slice recommends the non-parametric alternatives for
>> MANOVA.
>>>> More specifically our problem is that we obtain different
>> results
>>> when
>>>> applying either of both parametric and non-parametric MANOVAS.
>>>> We are at the moment assessing the relationships between
>>> environmental
>>>> variables (remote sensing data) and morphological characters
>>> (geometric
>>>> morphometrics) in a group of Neotropical bats.  Neither of both
>> types
>>> of
>>>> variables fit normality very well.
>>>> However, what currently puzzle us is the fact that the MANOVA
>>> performed
>>>> in SPSS give us significant differences, while the np-MANOVA in
>> PAST
>>>> gives us non significant differences.   The contrast in
>> magnitude
>>>> between both p-values is extreme.  We haven?t yet looked at
>>> additional
>>>> indexes of overlap, confidence or robustness for p-values.
>>>> We will be grateful with any comments or suggestions.
>>>> Pablo Menendez
>>>> Pablo Jarrin
>>>>
>>>>
>>>>
>>>> --
>>>> Replies will be sent to the list.
>>>> For more information visit http://www.morphometrics.org
>>>>
>>>>
>>>
>>> --
>>> Replies will be sent to the list.
>>> For more information visit http://www.morphometrics.org
>>
>>
>>
>>
>> --
>> Replies will be sent to the list.
>> For more information visit http://www.morphometrics.org
>>
>
>
>
> -- 
> Replies will be sent to the list.
> For more information visit http://www.morphometrics.org
>
>



-- 
Replies will be sent to the list.
For more information visit http://www.morphometrics.org

Reply via email to