 |
|
| Science Forum Index » Statistics - Education Forum » Is the KS test OK for unpaired hypothesis testing?... |
|
Page 1 of 1 |
|
| Author |
Message |
| Guy... |
 Posted: Tue Aug 25, 2009 9:07 am |
|
|
|
Guest
|
I am engaged in a debate where a colleague is challenging my usage of
the unpaired student t test and the Mann-Whitney test in hypothesis
testing in unpaired situation.
He suggests the KS test is just as good and also overcomes the normal
distribution assumption (that affects the unpaired student t test). I
know his suggestion is not conventional. But, from a statistical
standpoint I am unable to rebut it outright.
From an academic perspective, can you clarify this issue? |
|
|
| Back to top |
|
|
|
| Rich Ulrich... |
| Posted: Tue Aug 25, 2009 2:03 pm |
|
|
|
Guest
|
[crossposted to sci.stat.math, where the same post occurred]
On Tue, 25 Aug 2009 12:07:40 -0700 (PDT), Guy <gaetanlion at (no spam) gmail.com>
wrote:
[quote:28ea47f44c]I am engaged in a debate where a colleague is challenging my usage of
the unpaired student t test and the Mann-Whitney test in hypothesis
testing in unpaired situation.
He suggests the KS test is just as good and also overcomes the normal
distribution assumption (that affects the unpaired student t test). I
know his suggestion is not conventional. But, from a statistical
standpoint I am unable to rebut it outright.
From an academic perspective, can you clarify this issue?
[/quote:28ea47f44c]
I have always considered KS to have a couple of drawbacks.
If two variances are sufficiently different, the KS can reject in
*both* directions at the same time. Therefore, the assumptions
you need to state in order to justify the test are more awkward
to detail, compared to the t-test or MW.
And that is probably why it is used less often, and therefore
less comfortable to audiences. KS is a great test for location
when the distributions have the same shape and range.
The MW assumes that distributions are of "similar form", which
is practically as stringent as using a t-test with unequal variances.
The MW, by testing ranks, does remove the impact of outliers --
outliers can be terrible for any test of means, unless the huge
impact of a single case on the mean is something that needs to
be accommodated. I figure that your stated strategy of using
t and MW together provides pretty good robustness.
--
Rich Ulrich |
|
|
| Back to top |
|
|
|
|
|
All times are GMT - 5 Hours
The time now is Mon Nov 30, 2009 12:22 am
|
|