Sorry, an error occurred while loading the content.

## Re: [Synoptic-L] Re: Statistical Significance

Expand Messages
• ... I didn t make myself clear. For 020 we only have 3 solid results (two positives and one negative) out of 18 potential results, so I was looking to see
Message 1 of 6 , Jan 23, 2002
Stephen Carlson wrote:

> At 10:57 AM 1/22/2002 -0800, David Inglis wrote:
> >The problem is that I'm convinced that there are other statements
> >we can make regarding 020 failing to correlate with (and thus being
> >independent of) other categories, but they don't fit these 'rules'. For
> >example, for 020-022 and 020-122 we have:
> >
> > 020-022 020-122
> >r = 0.0749 r = -0.07598
> >p = 0.468247 p = 0.4155
> >Sample size = 96 Sample size = 117
>
> What kind of meaningful statement are you thinking a lack of
> correlation makes? How does it different from the negative
> correlation?

I didn't make myself clear. For 020 we only have 3 'solid' results (two
positives and one negative) out of 18 potential results, so I was looking to
see which (if any) of the other 15 results involving 020 could be used with
any confidence. Because of the 'design' of the analysis both negatives and
zeros indicate a lack of correlation in the source data, and I wasn't trying
to suggest that they were different.

> >I guess
> >what I'm looking for is something that does the opposite of 'p': I want
to
> >know when the sample size is large enough to validate (not falsify) the
null
> >hypothesis for a particular 'r'. Can it be done?
>
> The problem is that the null hypothesis is that rho = 0.0. And
significance
> testing, by its very design, cannot confirm the null hypothesis. It can
only
> reject it or tell you that you do not have data to reject it.
>
> When the 'p' value is outside of the desired range, it merely means that
the
> sampled 'r' value (e.g.0.0749) is not significantly different from zero
for
> the data that we've sample. We can always sample more data, and even for
> minute differences from zero (e.g. r = 0.01), there will be a point where
> enough data will bring the 'p' value into the realm of significant
>
> To test whether a sampled 'r' is close enough to zero, that you would
consider
> to be not significantly different from zero, you need a different null
hypothesis,
> for example, e.g. |rho| > 0.05, and run the tests with that null
hypothesis.

Thanks. Perhaps now we've got Dave Gentile's extra data we won't have to