Rich Ulrich <[EMAIL PROTECTED]> wrote:
> On 17 Oct 2003 06:05:19 -0700, [EMAIL PROTECTED] (Enda Kelly)
> wrote:

> [ snip, discussion ]

>> This was done for a number of pairs of loci. To illustrate what I meant
>> by "narrow" and "enormous", here are some examples with the p-value, the
>> median of the bootstrap values and the upper and lower limits of 95% CI:
>> 
>> Locus pair   P-value         Median          Lower           Upper
>> 1            0.002414        0.002422        0.000055        0.037401

> In terms of ratio:  Big.
> These are enormous, Median/Lower and Upper/Median.

>> 2            0.971621        0.512296        0.181935        0.850761

> ? In terms of sanity, or a dysfunctional program:  
> Is this a typographical error, or is the point-estimate 
> outside of  CI?

>> 3            0.000000        0.000000        0.000000        0.000000
>> 4            0.018936        0.016100        0.001082        0.193285

> Ratios are again more than 10-fold, for size of p.

>> 5            0.832001        0.505662        0.173286        0.857703

> ? Point estimate is within  the CI, but not by much.

>> 
>> Judging by the comments on the relationship between p-values and
>> confidence intervals, I have a suspicion that going through the
>> computationally expensive process of bootstrapping is not actually
>> teaching me anything new, but I would like to get some sort of estimate
>> of error associated with the computed p-values. Some mention has been

> It looks to me as if #1 and 4  show a huge range, when
> judged by the appropriate standard of relative size.

> But, more important, #2  and 5  are unbelievable.
> That is, I would say that they are not 'useful' results.
> Either, they were not good candidates for bootstrapping,
> or something is wrong that is more serious than that.

They are actually what you expect, I think, from a noncentral
chi-square.  If sampling from a 1-df central chi-square, for example, one would
expect 95% of samples to fall between 0.001 and 5.02, wth the "95% CI" for the 
P-values ranging from 0.05 all the way up to 0.95.  By contrast, for
ncp=10, say, the expected 95% range of P-values for the test of the null 
is 1e-6 to 0.13 (pchisq(qchisq(c(0.975,0.025),df=1,ncp=10),df=1,lower=FALSE),
around a mean of 0.0016. 

Wrt Arlequin, it is difficult to permute around missing data (not using
Windows I don't have the benefit of the updated version ;)).  SNPhap
imputes the missing alleles based on the surrounding haplotype for the
simulation IIRC.



-- 
| David Duffy (MBBS PhD)                                         ,-_|\
| email: [EMAIL PROTECTED]  ph: INT+61+7+3362-0217 fax: -0101  /     *
| Epidemiology Unit, Queensland Institute of Medical Research   \_,-._/
| 300 Herston Rd, Brisbane, Queensland 4029, Australia  GPG 4D0B994A v 
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to