> Actually, any real statistician would cringe at a lot of the fluff that
> gets put out about rocket reliability -- for example, quoting reliability
> numbers to three or four decimal places based on a few dozen trials!  This
> is not statistics, this is numerical bullshit, pure and simple.
>
> A good rule of thumb for evaluating such numbers is:  how much would the
> number change if *one launch* had gone differently?  That gives you some
> idea of how precise, or rather imprecise, most of those figures are.


A fun and sometimes useful layman's guide for "margin of error" in a sample
is to divide 100 by  the square root of the number of samples. ie

    Margin of error % = 100 / SQRT(N)   for N samples

For eg 30 launches this gives MOE = 18% !!!

This method gives a good but rough guide, has a real basis in statistics but
is always wrong in practice :-)
(Also can be roughly done in your head which impresses (some) people).

It also happens to be essentially how they calculate the margin or error in
opinion polls!
Have a look some time - for a 1000 sample poll this would give 100 /
sqrt(1000) = 3.2%
See how that compares to the next poll MOE result you see.
You can also work backwards and work out the poll size from the quoted MOE

            Sample size  = (100 / MOE)^2


regards,


            Russell "all models are wrong, some models are useful" McMahon

______________________________________________________________


Qualifier - the above is related to population distribution shapes and areas
under distribution tails and may (WILL) give dodgy results in some (many)
real life applications. For a normal population a slightly better formula is

    MOE % = 100  x 1.98 x SQRT( p *  (1-p) / N)  %

    N = sample size
    p = real occurrence of quantity being sampled for  (0..1).
    The worst case is for p = 0.5 which reduces to the simplification given
above.

--
http://www.piclist.com hint: The PICList is archived three different
ways.  See http://www.piclist.com/#archives for details.