Re: M6 problem survey

["Patrick G. Sobalvarro" <pgs@sobalvarro.org>]

At 04:54 PM 9/5/97 -0400, ABreull@aol.com wrote:
> You would not have wanted to live at Tschernobyl,
>although - at least the offical - chances for hat type of catastrophy are
>really low, with a lot of zeros between the point and the first cipher behind
>the point. But, you might decide for an exciting partner,  even if  chances
>are against you at .90.

I'm afraid I don't understand what this has to do with the question of
whether the survey is biased.

>Further, you can't use students t, since the 95 confidence intervall has to
>be asymmetrical (the formula is given by Marascuillo/ McSweeny, 1978, and you
>can easily program it, just 20 minutes C or Fortran - at least, if you don't
>use windows 95).

I really don't think there's any reason for you to beat your chest about
this.  Certainly it's true that the t-interval calculation assumes that the
quantity we are sampling has a t-distribution in the actual population,
which can't be true with discrete quantities and a finite population, etc.
Yes, of course, the actual population distribution is in fact binomial.
But the t-interval is the one most commonly used in cases like this because
it's typically close enough in practice.  Certainly it is pretty damned
close for the quantities we are estimating here, as you can determine for
yourself by plotting t- and binomial distributions with these parameters
and comparing them.  You can probably do this in less than twenty minutes,
whether or not you use Windows 95, although I am not in a position to make
any guarantees.  Incidentally, as long as we're all waving our dicks around
here, I was calculating 90% confidence intervals, not 95% confidence
intervals; and the names of the authors of "Nonparametric and
Distribution-Free Methods for the Social Sciences" are Marascuilo and
McSweeney, not Marascuillo and McSweeny.  

>And, you don't generalize from the sample as you did, but from the
>hypothesis. And, the small and "non representative" sample might be correct
>still (hennce the bias neglectable), at last, if the confunded variable does
>nnot produce a so called "semi-disordinal interaction", and, and, and ...

Allow me to congratulate you on being the owner of a book about statistics.
 However, you aren't making any sense here.  We weren't testing a
hypothesis.  We didn't start with a hypothesized value for the mean -- we
were generating confidence intervals for a sample made up of Bernoulli
tests.  As to the issue of bias, my last message showed very clearly that
Fernando's sample was not representative (half of all surveyed M6's were
made in 95-96, which we know not to be true of the population; but these
were precisely the M6's he found to be faulty, so there is an interaction),
and I also made the point in my original message that it is difficult to
correct for his survey's bias without more data -- in particular, we have
no obvious means of estimating our nonsampling error here.

>I still think, that Ferdinand's survey was a good idea. Because it was a
>hypothesis-generating study. So, he doesn't need all the generalization stuff
>you are excited about, but offers a hypothesis, which might be investigated
>with all th necessary hypothesis-testing stuff, type I and type II error,
>sample size, effect size, and so on.

I think you are suggesting here that we treat the means found with
Fernando's data as hypotheses whose correctness  we can test in further
studies.  I think that is a capital idea!  However, we need a volunteer to
do the work of proper experimental design and data collection.  Fernando
has already done the first survey, and I have done the initial analysis of
his results.  Why don't you undertake the next study as your own
contribution to the group?

- -Patrick