As a variant on this problem I'm considering use of
t-test and Mann-Whitney with data where the samples
have some uncertainty associated with the labels.
Thus I have a set of samples, each having a number
of features. The samples could be assumed to belong
to two classes, 1 and 0, say. I can then run a
t-test, Mann-Whitney or other tests to find the most
distinguishing features separating the two classes.
Unfortunately, there is an intrinsic uncertainty
associated with the class labels, thus sample A
might have a probability of 0.9 of association with 1,
but 0.1 that it should have class label 0. I can
think of various ways of adapting statistical tests
for this purpose but wondered what is the most
sound approach. Colin C.
--On 17 December 2008 09:28 +0000 "Henderson, Robin"
<[log in to unmask]> wrote:
> Dear Allstat
>
> Sincere thanks to everyone who responded. I deliberately simulated data
> from two normal populations with the same variance with n1 = 5 and n2 =
> 10. The ratio of the larger IQR to the smaller IQR exceeded 2 in around
> one in three cases. Thus, in a scenario where a Mann-Whitney test would
> be appropriate (albeit inferior to a t-test), the guideline would
> "advise" against its use. This led me doubt its validity and post the
> query. Responses are given below.
>
> Stephen Senn
> I suspect that the guideline is not very good. However, it is
> probably to do with heteroscedasticity and you have programmed
> homoscedasticty into your simulation.
>
> It is well-known that the two-sample t-test is not robust if the
> population variances are different unless the samples sizes are the
> same. If the smaller sample has the larger variance there is a
> problem. I suspect that what is being done here is to give some
> guidance to deal with possible differences in variances but I also
> suspect that it is a pretty useless rule of thumb.
>
> It might be worth looking at Gerald van Belle's
> http://www.vanbelle.org/ book to see if there is a mention .
>
> Nick Longford
> I have, and ignored it, for the good of the science.
>
> Paul Wilson
> I have never heard of this restriction on the Mann-Whitney.
> Please let me know what the general opinion of your respondents is. I
> would hazard a guess that whoever penned the "guideline" was trying to
> be clever and "transfer" the guideline that a t-test should not be used
> if the standard deviation of one sample is more than twice that of the
> other, but I could be wrong!
>
> Allan White
> From a theoretical perspective, I can see why this recommendation was
> made. The Mann-Whitney test is a member of the class of permutation
> tests. This class of tests has the property that, under the null
> hypothesis, all the rearrangements of the data performed by the test must
> be equally likely. This condition is met if, and only if, the data from
> the different groups is drawn from the same error distribution. (I mean
> by this that the distributions of scores within each group/condition
> should not differ in any respect, except that of location). In the
> scenario where the spread differs obviously (and substantially) between
> the groups, this condition is clearly violated and the test degenerates
> into one that merely tests for significant differences (OF ANY SORT)
> between the groups/conditions. Thus, under these circumstances, the test
> ceases to be a simple test for differences in location (medians).
>
> Roger Newson
> I have not seen this guideline as such. However, I have long been aware
> that the Mann-Whitney U-statistic, and the associated confidence
> interval for the Hodges-Lehmann median difference, are robust to
> non-Normality and non-robust to unequal variability. I have developed a
> package (somersd) in the Stata statistical language to calculate
> confidence intervals for rank statistics that are robust to unequal
> variability. The theory is written up in Newson (2002), Newson (2006a)
> and Newson (2006b), and also in some manuals distributed with the
> package. All of these can be downloaded from my website (see my
> signature below). If you have Stata, then you can download the package
> by typing in Stata
>
> ssc describe somersd
> ssc install somersd, replace
>
> I have done some simulations, and submitted the results for publication
> in Computational Statistics and Data Analysis, on the performance, under
> a wide range of scenarios, of various confidence intervals for median
> differences (my package, the Lehmann formula, and the equal-variance and
> unequal-variance t-tests. The message of these simulations is that the
> method implemented in the somersd package is robust to non-Normality and
> to unequal variability, at the price of being non-robust to tiny sample
> numbers, under which conditions the confidence intervals may extend from
> minus infinity to plus infinity. This is because, under those
> conditions, if we are not allowed to assume Normality and/or equal
> variability, then the median difference really could be anywhere.
>
> Chris Lloyd
> Top of my head - MW is appropriate for a SHIFT model. For testing, the
> null hypothesis is identical distributions i.e. shift zero. I did some
> simulations years ago that convinced me that it was not robust to scale
> differences - the size of the test can be pretty badly compromised,
> especially when one population is contaminated with skew (holding the
> medians equal).
>
> But the twice-scale rule. Never heard of it.
>
> Lisa Yelland
> I have not come across this guideline. However I have read a paper* which
> looks at the performance of the Mann-Whitney test in a range of scenarios
> using simulations and shows that it performs poorly when the variances in
> the two groups are unequal.
>
> I am curious about the simulation you did. Firstly, if the data are
> normal then the t-test would be preferable to the Mann-Whitney test.
> Secondly, the guideline you mentioned relates to unequal inter-quartile
> ranges and yet you simulated data with equal variances (perhaps this was
> a typo?). I think it would be more relevant to simulate non-normal data
> with a range of differences in the inter-quartile range between groups to
> judge whether the guideline is appropriate (depending on how much time
> you want to spend investigating this!).
>
> * Skovlund E & Fenstad G (2001). Should we always choose a nonparametric
> test when comparing two apparently nonnormal distributions?. Journal of
> Clinical Epidemiology; 54:86-92.
>
> Martin Bland
> This is nonsense, as the Mann Whitney test can be used for ordinal
> data. The notion of interquartile range involves subtraction, so can
> apply only to interval data. However, there is a condition that if you
> want to use the test as testing the null hypothesis that the medians are
> the same, the two distributions must differ only in location. This
> clearly applies to interval data only as ordinal data do not have a
> shape. Under these circumstances the test also tests the null
> hypothesis that the means are equal. As the standard deviations must be
> identical, you might at well do a t test and get a confidence interval.
>
> Dorothy Middleton
> In theory, the bootstrap is the only technique that should be used to
> compare the means of two populations that have quite different variances
> (that is, the Behrens-Fisher problem). Student's t, a permutation test
> using the original observations, and the permutation using ranks
> (Mann-Whitney, Wilcoxon) are all likely to yield inexact significance
> levels. Still simulations have shown that permutation tests are almost
> exact even when the variance of one population is twice that of the
> other. See http://statisticsonline.info/application.htm.
>
>
> Best Wishes
>
> Robin
>
> G Robin Henderson
> Audit Coordinator
> Scottish National Stroke Audit
> Royal Infirmary of Edinburgh
> 0131 242 6934
>
>
>
> *****************************************************************
> The information contained in this message may be confidential or
> legally privileged and is intended for the addressee only. If you
> have received this message in error or there are any problems
> please notify the originator immediately. The unauthorised use,
> disclosure, copying or alteration of this message is
> strictly forbidden.
> *****************************************************************
----------------------
ICG Campbell, Engineering Mathematics
[log in to unmask]
|