Re: t-test gets NaN instead of 0 for significance on x86-64

[Ben Pfaff]

Thanks, I pushed it.

I did verify that it passes "make check" at least.

John Darrington <address@hidden> writes:

> It looks good to me now. (I haven't tested it).
>
> J'
>
> On Fri, Sep 04, 2009 at 09:06:47AM -0700, Ben Pfaff wrote:
>      John Darrington <address@hidden> writes:
>      
>      > On Thu, Sep 03, 2009 at 08:10:31PM -0700, Ben Pfaff wrote:
>      >      John Darrington <address@hidden> writes:
>      >      
>      >      > As a short term solution, the best I can suggest is that we 
>      >      > clamp pow2(correlation) to 1.0.
>      >      
>      >      Thanks.  I pushed out this stopgap fix.
>      >
>      > Isn't there still a problem when pairs[i].correlation is less than
>      > -1 ?  For example -1.00000001
>      
>      You are right, of course.
>      
>      Could you check that this re-fix looks correct?
>      
>      commit 46912334775c83a23902bd5c7d72cd4ad8d23c95
>      Author: Ben Pfaff <address@hidden>
>      Date:   Fri Sep 4 09:05:08 2009 -0700
>      
>          T-TEST: Also fix the case where the correlation is slightly less 
> than -1.0.
>          
>          Thanks to John Darrington for pointing out the omission.
>      
>      diff --git a/src/language/stats/t-test.q b/src/language/stats/t-test.q
>      index dd40de9..ba7e938 100644
>      *** a/src/language/stats/t-test.q
>      --- b/src/language/stats/t-test.q
>      ***************
>      *** 1432,1447 ****
>        
>              double df = pairs[i].n -2;
>        
>      !       /* pairs[i].correlation is a correlation, so mathematically it 
> will
>      !          always be in the range [-1.0, 1.0].  Inaccurate calculations 
> sometimes
>      !          cause it to be slightly greater than 1.0, however, which 
> makes the
>      !          sqrt() below to come out as NaN instead of 0.  So force it to 
> be 1.0
>      !          or less. */
>      !       double corr = MIN (1.0, pairs[i].correlation);
>      ! 
>              double correlation_t =
>               pairs[i].correlation * sqrt (df) /
>      !        sqrt (1 - pow2 (corr));
>        
>        
>              /* row headings */
>      --- 1432,1444 ----
>        
>              double df = pairs[i].n -2;
>        
>      !       /* corr2 will mathematically always be in the range [0, 1.0].  
> Inaccurate
>      !          calculations sometimes cause it to be slightly greater than 
> 1.0, so
>      !          force it into the correct range to avoid NaN from sqrt(). */
>      !       double corr2 = MIN (1.0, pow2 (pairs[i].correlation));
>              double correlation_t =
>               pairs[i].correlation * sqrt (df) /
>      !        sqrt (1 - corr2);
>        
>        
>              /* row headings */
>      
>      -- 
>      Ben Pfaff 
>      http://benpfaff.org
>      
>      
>      _______________________________________________
>      pspp-dev mailing list
>      address@hidden
>      http://lists.gnu.org/mailman/listinfo/pspp-dev

-- 
Ben Pfaff <address@hidden>
Author and maintainer of Pintos