In a message dated 08/29/2000 1:13:20 AM US Eastern Standard Time,
[log in to unmask] writes:

Jarmo,
I hesitate to answer in the wake of Brian Ellis' dissertation on SIR
statistics.  Wow.  And people call ME obscure.  Personally, I think he made
half those terms up.  I never heard them in my graduate statistics classes.
Of course, I may have been sleeping at the time.  Fractile skew sounds like
one of my lathe chisels and I am pretty sure a platykurtic is a small duck
billed animal, perhaps from Cyprus <grin>.

I agree with Brian that you have a small sample size, but then J-STD-004 has
a small sample size in its SIR test.  I would not say that it was
statisically invalid, but you do have a much larger error and lower
statistical confidence when you use a smaller sample size.  In the IPC SIR
Task Group, we have even debated what a sample was.  Take the IPC-B-24 test
board for example.  It has four identical comb patterns.  If you had three
such boards, you would have 12 patterns.  Jim Maguire, now of Intel, would
argue that you have a sample size of three.  I would argue that you have a
sample size of 12.

SIR data, by its nature, is exponential.  Most statistical texts that I have
seen suggest that you deal with such data by transforming the data.  Dealing
with the data in its natural state, you might get values like 5.2E+9 with
standard deviations of 4E+3.  I deal with SIR data on a daily basis and I
would have trouble attaching meaning to those numbers.

Lets take your measured currents in nanoamperes with a 15 volt potential:

> 7.90; 1.71; 6.80; 1.23; 3.00; 2.63; 2.90; 1.04; 7.60; 1.54; 3.90; 1.87.

Converting to resistance you get:

1.90E+09
8.77E+09
2.21E+09
1.22E+10
5.00E+09
5.70E+09
5.17E+09
1.44E+10
1.97E+09
9.74E+09
3.85E+09
8.02E+09

Take the base 10 logarithm of these numbers and you get:

9.28
9.94
9.34
10.09
9.70
9.76
9.71
10.16
9.30
9.99
9.59
9.90

The arithmetic average of these numbers is 9.73.  The standard deviation is
0.30.  The transformed resistance data is most often referred to as LogOhms.

A third of a decade SD is not too bad, but there is probably room for
improvement in your technique.  You can get this much range in your data just
from humidity variations in your chamber.

If you took the antilog of the 9.73 value, you would have the geometric mean
of the calculated resistance values.

Now, the "correct" way of measuring the SIR values depends on the
specification you are testing to.  As an example, Bellcore (well, Telcordia
Technologies now) GR-78-CORE uses a geometric method, so the geometric mean
is "correct" for that spec.

If you are just doing an engineering evaluation and not trying to meet a
specification, then I would suggest using the LogOhm method.

Just make sure your whatzahoozits value is less than 0.48 slugs per fortnight
<grin>.

Doug Pauls
CSL

##############################################################
TechNet Mail List provided as a free service by IPC using LISTSERV 1.8c
##############################################################
To subscribe/unsubscribe, send a message to [log in to unmask] with following text in
the body:
To subscribe:   SUBSCRIBE TECHNET <your full name>
To unsubscribe:   SIGNOFF TECHNET
##############################################################
Please visit IPC web site (http://www.ipc.org/html/forum.htm) for additional
information.
If you need assistance - contact Keach Sasamori at [log in to unmask] or
847-509-9700 ext.5315
##############################################################