| Help with the F-test. mpftest.pro in particular. [message #53297] |
Fri, 06 April 2007 08:55  |
cgguido
Messages: 195
Registered: August 2005
|
Senior Member |
|
|
Hi all,
I am trying to figure out how the F-test works.
I created two fake datasets with 1e6 elements and a variance of 4
n = 1e6
s = 4
i1 = randomn(sss, n)*s
i2 = randomn(sss, n)*s
I then calculate the f value as the ration of the largest variance to
the smallest (the two sets have the same number of degrees of freedom,
n)
var = [variance(i1), variance(i2)]
vv = max(var, min=v)
f = vv/v
Then I print the probability that the variances are similar due to
pure chance (should be close to zero because the variances are *set*
to be th same)
print, mpftest(f, n, n)
--> 0.47958397
I must be doing something very wrong :-(
Can somebody help please?
Thanks,
Gianguido
|
|
|
|
|
|
| Re: Help with the F-test. mpftest.pro in particular. [message #53384 is a reply to message #53297] |
Fri, 06 April 2007 11:49  |
news.verizon.net
Messages: 47
Registered: August 2003
|
Member |
|
|
On Apr 6, 11:55 am, "Gianguido Cianci" <gianguido.cia...@gmail.com>
wrote:
> Then I print the probability that the variances are similar due to
> pure chance
This is not what mpftest computes -- instead "The function MPFTEST()
computes the probability for a value drawn from the F-distribution to
equal or exceed the given value of F."
> print, mpftest(f, n, n)
> --> 0.47958397
>
So for the null hypothesis, one would expect to obtain a value larger
than the observed value of f about half the time, i.e. the null
hypothesis (that the two variances come from the same distribution) is
consistent with the data.
For comparison, suppose the variance from the second dataset had been
1% larger
IDL> print,mpftest(f*1.01,n-1,n-1)
2.6478877e-07
then the null hypothesis could be rejected with very high
confidence. (Then sensitivity to the value of f comes because of
the large number (1e6) of degrees of freedom in your example.) --
Wayne
P.S. For those not in the know, MPFTEST is not an intrinsic IDL
function, but part of Craig Markwardt's great library
http://astrog.physics.wisc.edu/~craigm/idl/idl.html
|
|
|
|
| Re: Help with the F-test. mpftest.pro in particular. [message #53581 is a reply to message #53384] |
Wed, 18 April 2007 14:49 |
cgguido
Messages: 195
Registered: August 2005
|
Senior Member |
|
|
Thanks fo ryour help, and yes, the description is *in* the code :-(
I still don't understand what it all means and have been looking for
an understandable text on statisitcs :-(
I might have a more intelligent question once I read up on this.
Thanks,
G
On Apr 6, 2:49 pm, "Wayne Landsman" <vze49...@verizon.net> wrote:
> On Apr 6, 11:55 am, "Gianguido Cianci" <gianguido.cia...@gmail.com>
> wrote:
>
>> Then I print the probability that the variances are similar due to
>> pure chance
>
> This is not what mpftest computes -- instead "The function MPFTEST()
> computes the probability for a value drawn from the F-distribution to
> equal or exceed the given value of F."
>
>> print, mpftest(f, n, n)
>> --> 0.47958397
>
> So for the null hypothesis, one would expect to obtain a value larger
> than the observed value of f about half the time, i.e. the null
> hypothesis (that the two variances come from the same distribution) is
> consistent with the data.
>
> For comparison, suppose the variance from the second dataset had been
> 1% larger
>
> IDL> print,mpftest(f*1.01,n-1,n-1)
> 2.6478877e-07
>
> then the null hypothesis could be rejected with very high
> confidence. (Then sensitivity to the value of f comes because of
> the large number (1e6) of degrees of freedom in your example.) --
> Wayne
>
> P.S. For those not in the know, MPFTEST is not an intrinsic IDL
> function, but part of Craig Markwardt's great libraryhttp://astrog.physics.wisc.edu/~craigm/idl/idl.html
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
