[Test-Equipment] Corrected link

[Tom Dawson]

wolfbob,


First, I'd say a detailed review of the article would be in order.  This
Tech Note was written by engineers at the National Bureau of Standards, now
NIST, for use in calibration laboratories.  I don't claim to be smart enough
to challenge their use of statistical methods, but my guess is that a
statistician reviewed their method prior to publication.

Anyone who bought a TI pocket calculator in the late '70s  has had the
opportunity to see that small populations do not provide reliable
information about larger populations, but that is not really the issue here.
Averaging is simply used to reduce the effect of random errors, like reading
the galvo, setting the potentiometer slider etc.  The procedure does include
methods to reduce the systematic errors, like thermal EMF's etc.  which
would confound any statistical analysis applied.

The process also includes repetitive measurements over time.  Thus the
population of measurements increases and statistical methods become more
useful.  In my own experience, the variations from measurement to
measurement over time reveal that certain cells have a little wider
variation over time than others and that some cells show a trend in one
direction, which might suggest replacement at some point.


So, read the paper for content and maybe some of your objections will be
addressed.


regards,

Tom




----- Original Message ----- 
From: "wolfbob" <wolfbob at csnsys.com>
To: "Tom Dawson" <wb3akd at earthlink.net>; <test-equipment at mailman.qth.net>
Sent: Tuesday, June 01, 2004 1:51 AM
Subject: Re: [Test-Equipment] Corrected link


> I found two or three gross approximations in the paper with a quick scan.
> First is the assumption that standard deviation can be applied. As I
> remember form 40 years ago, the concept for standard deviation only
applies
> to distributions that are Gaussian. They make no proof or present any
> evidence as to just what is the error distributions and why they should be
> random with a Gaussian distribution. The phenomena that is causing the
> variation could be defined and it is entirely possible that the errors are
> Gaussian, but I rather doubt it.
>
> Taking the average if a few things can be very misleading and fraught with
> statistical errors. If you run some basic probability test like Student T
> (again only applicable for Gaussian distributions as I recall) or what
ever
> to see just what confidence there is that the average is any more accurate
> than any one of the cells.
>
> A very good example that all can appreciate is that of taking the reading
of
> time from 6 watches and averaging them and expecting to get more accuracy
in
> knowing what time it is. Here is a case where it is obvious that the error
> is not random but arbitrary, a really quite different thing. It should be
> obvious the average reading of the watches is no more accurate than any
one
> of the watches as you do not have a clue as to the distribution of the
> errors or their cause, except the are not random, just different. Now if
you
> have 10,000 watches, you have another hunk of mathematics coming in to
play
> and that is the central limit theorem that says if you got enough
> independent samples of any distribution they will appear Gaussian and all
> the good math applies. Six is not a large enough to qualify under this
> caveat.
>
> Another way to see the error of this thinking is to make a simple
> assumption. One of the 5 cells is dead ass on the correct voltage and the
> other four are wrong. The average will also be wrong. just randomly
picking
> one of the 5 cells for the standard will make you right 20% of the time
> whereas the average will make you wrong 100% of the time.
>
> WBob
>
> ----- Original Message ----- 
> From: "Tom Dawson" <wb3akd at earthlink.net>
> To: <test-equipment at mailman.qth.net>
> Sent: Monday, May 31, 2004 8:59 AM
> Subject: [Test-Equipment] Corrected link
>
>
> > Sorry,
> >
> > Fumble fingered the previous post.
> >
> >
> > Follow the link to the following NBS Tech Note relating to measurements
of
> > standard cells, and maintaining the volt with a small group of saturated
> > standard cells.
> >
> >
> > http://www.radio-science.com/std_cells/tn_430.pdf
> >
> > and a dos program to do the arithmetic.
> >
> > http://www.radio-science.com/std_cells/cells.exe
> >
> >
> > regards,
> >
> > Tom
> >
> >
> > _______________________________________________
> > Test-Equipment mailing list
> > Test-Equipment at mailman.qth.net
> > http://mailman.qth.net/mailman/listinfo/test-equipment
> >
> >
>
>
>