My girlfriend showed me a social science paper that she was required to
read for class. Without going into too many details that would turn
this into a personal attack, it was written by a European professor
whose English was utterly illiterate. The paper gave a table that
summarized the scores from surveys, with respondents split into four
categories.

- The table talked about the "F test with 4 df", apparently referring
to the test for overall differences. Is it allowed to just drop the
denominator df like that, or is the writer just illiterate in
statistics as well?

- The table just gave F-statistics without saying their p-values. I'm
guessing again that that's just poor statistics?

- The paper also said stuff like "2 < 1,3,4" in comparing the group
means. I don't recall seeing that sort of notation in my categorical
data analysis course; it would seem to bring up a multiple testing
issue as well. Is this an accepted convention, or more illiteracy?

My girlfriend showed me a social science paper that she was
required to read for class. it was written by a European professor
whose English was utterly illiterate.

- The table talked about the "F test with 4 df",

- The table just gave F-statistics without saying their p-values.
I'm guessing again that that's just poor statistics?

- The paper also said stuff like "2 < 1,3,4" in comparing the group
means. I don't recall seeing that sort of notation in my categorical
data analysis course; it would seem to bring up a multiple testing
issue as well. Is this an accepted convention, or more illiteracy?

There are many in THIS group, from
Portuguese to American posters who do not know the meaning
of p-values. The lean on computer manuals that make errors
and assume that a "canned program sold for money" must be
correct, or something to that fallacious effect.

"Reef Fish" <large_nassua_grouper@yahoo.com> wrote in
news:1167337286.578317.202960@42g2000cwt.googlegroups.com:

There are many in THIS group, from
Portuguese to American posters who do not know the meaning
of p-values. The lean on computer manuals that make errors
and assume that a "canned program sold for money" must be
correct, or something to that fallacious effect.


There are many in this group who disagree with Reef Fish's posting from
September. Every textbook consulted disagreed.

One person cited SPSS, I cited a function in R (not sold for money). I
also cited textbooks:
copying from
news:Xns9832DF4EC07dwtttttt@216.196.97.136

--------definitons from math stats texts------------
Cox and Hinkley says a "level of significance" p_obs is defined as:
p_obs= Pr(T >= t_obs;H0)

Kalbfleisch uses significance level as
SL == Pr(D>= D_obs|H0)

DeGroot defined p-value is sup(Pr(T >= t|theta))

"Reef Fish" <large_nassua_grouper@yahoo.com> wrote in message
news:1167361171.466224.211680@k21g2000cwa.googlegroups.com...

David Winsemius wrote:
"Reef Fish" <large_nassua_grouper@yahoo.com> wrote in
news:1167337286.578317.202960@42g2000cwt.googlegroups.com:

There are many in THIS group, from
Portuguese to American posters who do not know the meaning
of p-values. The lean on computer manuals that make errors
and assume that a "canned program sold for money" must be
correct, or something to that fallacious effect.


There are many in this group who disagree with Reef Fish's posting from
September. Every textbook consulted disagreed.
+++++++++++++++++++++++++++++++++++++++++++
First of all I will get chewed out for entering this discussion.

I agree with Bob on his interpretation.

Every testbook you consulted is as Bob said made an error of ommission.

The cemtral issue is properly defining the complete domain of test results.
Just making a statement only about the hypothesis defines only a part of
this domain. The reader then has to assume what is the rest of the domain
envisioned by the proposer.

Most textbooks assume that the reader knows what the domain is, so just
giving H0 they assume that the reader knows what Ha is. Bob has however
pointed out in other messages, that this is not necessarily true. To argue
why textbooks do this is beyond this sequence.

DAH
++++++++++++++++++++++++++++++++++++++++

Not the p-VALUE either. Where is the Alternative Hypothesis?

-- Reef Fish Bob.

The error of omission was in NOT stating the Alternative Hypothesis,
which is critical in the DEFINITION of the p-value.

But the examples carefully selected by David Winsemius were much
worse than that! Below are the reasons for my statement, spelled out:

On Dec 30, 6:24 am, "Reef Fish" <large_nassua_grou...@yahoo.com> wrote:

The error of omission was in NOT stating the Alternative Hypothesis,
which is critical in the DEFINITION of the p-value.


As I explained earlier this year, the difference is in whether
or not you approach this in a Fisherian way or a pure NP way.

In the case of continuous data, the strict inequality does not
affect the computation of a p-value, but for discrete data it
does, which is why NP suggests randomization on the
boundary.

But the examples carefully selected by David Winsemius were much
worse than that! Below are the reasons for my statement, spelled out:

David's examples (which I recognize) are both from
descriptions of Fisher-type significance testing, not NP
hypothesis testing. As I pointed out, Fisher did not
require a formal alternative hypothesis.

I know you disagree, that is fine. In reality, what is now
taught is often a hybrid of NP and Fisher theory. You
yourself have rejected it in favour of Bayesian approaches
anyway, suggesting there are multiple approaches to the
problem.

Kevin

Kevin E. Thorpe wrote:
On Dec 30, 6:24 am, "Reef Fish" <large_nassua_grou...@yahoo.com> wrote:

The error of omission was in NOT stating the Alternative Hypothesis,
which is critical in the DEFINITION of the p-value.

As I explained earlier this year, the difference is in whether
or not you approach this in a Fisherian way or a pure NP way.I consider my statistical education about as "old fashioned" as one
can be without being labeled as a dinosaur. Fisher's way was
already out-of-style long before I started statistics. It is
definitely
out of style NOW.



In the case of continuous data, the strict inequality does not
affect the computation of a p-value, but for discrete data it
does, which is why NP suggests randomization on the
boundary.
That is the nonessential part of the error. To take ALWAYS the
one-sided alternative and call it a "significance level" is an ERROR,
a major BLUNDER, no matter how you argue it.

Read this VERY carefully,

A p-value does NOT exist in the Fisherian context

It exists, and is meaningful ONLY when it works in conjection
with an Alternative Hypothesis to be able tell what "more
extreme" means.

On Dec 30, 9:41 am, "Reef Fish" <large_nassua_grou...@yahoo.com> wrote:
Kevin E. Thorpe wrote:
On Dec 30, 6:24 am, "Reef Fish" <large_nassua_grou...@yahoo.com> wrote:

The error of omission was in NOT stating the Alternative Hypothesis,
which is critical in the DEFINITION of the p-value.

As I explained earlier this year, the difference is in whether
or not you approach this in a Fisherian way or a pure NP way.I consider my statistical education about as "old fashioned" as one
can be without being labeled as a dinosaur. Fisher's way was
already out-of-style long before I started statistics. It is
definitely
out of style NOW.



In the case of continuous data, the strict inequality does not
affect the computation of a p-value, but for discrete data it
does, which is why NP suggests randomization on the
boundary.
That is the nonessential part of the error. To take ALWAYS the
one-sided alternative and call it a "significance level" is an ERROR,
a major BLUNDER, no matter how you argue it.

Let me begin this reply by reminding you that when I first
ventured in to this discussion earlier this year, I did agree
that in the NP framework, extreme is clearly a strict
inequality.

The critical region is a strict inequality

and by extension so is the calculation of a p-value from a
test statistic.

Second, I am not advocating for one approach or another
in these threads, just trying to present an opposing view,
hopefully, for the benefit of others. I know I have learned
things by thinking about these discussions.

My purpose in replying in this thread at all is to point out
that the context of the cited references from David Winsemius
is essential to the correct understanding of what the authors
were talking about.

For the benefit of others, I will briefly outline my understanding
of Fisher significance testing.

There is a NULL hypothesis.
There is a test statistic defined so that large values are
inconsistent with the NULL in an "appropriate" way. Note:
it is this that took a great deal of heat from NP. Then,
with some data in hand, you wish to know, "are my data
inconsistent with the NULL." The SIGNIFICANCE LEVEL
is then the probability of observing a test statistic as big
or bigger as the one obtained from my data, given the NULL
is true.

This may look like it is always one sided, but it's not.
If your test statistic is a likelihood ratio test compared with
a chi-square distribution, it would be two-tailed.

An obvious problem is how to define more extreme in the
absence of an explicit alternative. Many would argue
that you can't, while others would say, "it is obvious
what is intended from the context."

Read this VERY carefully,

A p-value does NOT exist in the Fisherian context

It exists, and is meaningful ONLY when it works in conjection
with an Alternative Hypothesis to be able tell what "more
extreme" means.

I think I would agree a significance level is not the same thing
as a p-value from NP testing.

I will close with a question. When you compute Fisher's
Exact Test on a 2X2 table, how do treat the observed table
in your calculation?