Another proof point for tyler

" wrote:

gwhite wrote:

I'm talking about the 0.1% noise rating for the machine. I'd question a
peak noise rating for a machine. Caveat: I don't know squat about the
dope detecting industry. I'll read the links when I get some time. (I
don't pay attention to the dope stuff much.)

They just toss out the 0.1% number in a parenthesis. It's not
clear if it really is a property of the machine or just the
particular measurement they are talking about in that one
paragraph.

It should be RMS. I could not find a datasheet/specsheet on the
machine. I found something close, but stunningly a noise spec isn't
given (or I didn't see it).

http://www.beckmancoulter.com/Litera..._monograph.pdf
http://www.beckmancoulter.com/produc...ry/epicsxl.asp


In any case, everyone else talking about "peak" is referring
to the number of counts in a signal that looks like a peak, not
the peak height of the signal, ...

I'm not clear on how the counts could simply be on the peak (max) bin of
the lobe(s). The count for the total of each lobe (with the width
considered) would seem to be more sensible. That is, integrated counts
according the particular fluorescence lobe. "Where to cut off" counting
would be where adding more counts made no "significant" difference to
the total count of the lobe.

I couldn't spend more time studying this and could be way off base
regarding how things are counted and how the test is done. I am not
familiar with the field or the lingo, so it is a bit of a struggle at
the base of the learning curve.

... which is what you are referring to
with "peak noise" and the peak-vs-RMS jitter reference. (I agree
"peak noise" doesn't mean anything.)

I'm saying a peak rating to specify the machine noise does not make
sense. For random noise, the peak isn't known. Wait long enough, and
you'll get a higher peak. ("Wait longer" to get a higher p-rms doesn't
necessarily help Tyler, because "waiting longer" means a lower
probability of a false positive.)

But peaks certainly do matter for the *measurement*. That "bit error"
talked about in the article I linked is caused by the peak exceeding the
threshold (not AVG-RMS) and is exactly what can result in a false
detection. (The threshold detector itself has an uncertainty.)
Communications systems are designed around the error ratings (noise).
Peak values certainly matter, and are considered in a probablistic way.



Chung is questioning how they reliably tell that a sample has
two signals rather than one, which is a good question. Look
at the pdfs, but do like a good scientist: don't read them, just
look at the pictures. It will be clear what "peak" refers to
in the plots. Most of the plots, anyway; that's the issue.

The main instrumentation/measurement area I thought was interesting was
the "gating" area. Campbell said false levels of up to 1.1% could be
found with operator error. In my business we don't just consider
Signal/Noise ratios. We more completely care about Signal/(Noise +
Distortion + Spurious) ratios. For this realm, I suppose operator error
would be called a spurious result. We can't dismiss the operator from
the system, especially if the operators have been shown to make
significant errors. The machine's accuracy and precision is one part of
the system, not the whole system.

I think I also read Hamilton's reading was 1.3% (did I?). If so, that
1.3/1.1 ratio is uncomfortable, to say the least. I also read the
comments on gating in the Nelson paper. I sounded like the adjustments
are manual and that "mistuning" by the operator could occur. I was not
sure that each of the lobes could not be individually mistuned, which
would throw off a ratio.

More curious was Figure 1A-B in the Nelson paper. The dilution seems to
have *moved* the small fluorescence lobe! Why would the frequency
(horiz axis) of the fluorescence change? I guess I really don't
understand the machine and what it does. (I didn't read
altra_monograph.pdf.) Also, if the count is an integrated one per lobe,
the horiz count bar for the little lobe (Fig1B) is actually extended
into the large lobe! This seems odd, and maybe that is what Robert is
referring to.

For the comments above, be like a good scientist and presume I don't
know squat about the machine, its gating, and its method of counting.

Based on the little bit I've studied, I would probably have thrown the
case out for a few different reasons -- no judgement on Hamilton. That
doesn't mean I don't have my suspicions.

I can't believe I got sucked into a doper thread.

gwhite wrote:
I can't believe I got sucked into a doper thread.

I'm not surprised at all to see you post a long and
ponderous followup to a thread no one is reading.

Bob Schwartz

gwhite wrote:

In any case, everyone else talking about "peak" is referring
to the number of counts in a signal that looks like a peak, not
the peak height of the signal, ...

I'm not clear on how the counts could simply be on the peak (max) bin of
the lobe(s). The count for the total of each lobe (with the width
considered) would seem to be more sensible.

The counts aren't on the max bin (the mode) of what you are
calling the lobe and everyone else is calling the peak. They
are of the total.

That is, integrated counts
according the particular fluorescence lobe. "Where to cut off" counting
would be where adding more counts made no "significant" difference to
the total count of the lobe.

The question is how to define this if the lobes/peaks aren't
well separated or are noisy. For example, if there is no
strong signal but just noise fluctuations, the signal you get
keeps going up as you integrate over more bins, so "where to
cut off" is not easy to define.

I'm saying a peak rating to specify the machine noise does not make
sense. For random noise, the peak isn't known. Wait long enough, and
you'll get a higher peak. ("Wait longer" to get a higher p-rms doesn't
necessarily help Tyler, because "waiting longer" means a lower
probability of a false positive.)

But peaks certainly do matter for the *measurement*. That "bit error"
talked about in the article I linked is caused by the peak exceeding the
threshold (not AVG-RMS) and is exactly what can result in a false
detection. (The threshold detector itself has an uncertainty.)
Communications systems are designed around the error ratings (noise).
Peak values certainly matter, and are considered in a probablistic way.

As long as you continue to use "peak" in a sense which is
standard for a different field, but different from everyone else
in this discussion, it's just going to get more confusing
rather than less.

Chung is questioning how they reliably tell that a sample has
two signals rather than one, which is a good question. Look
at the pdfs, but do like a good scientist: don't read them, just
look at the pictures. It will be clear what "peak" refers to
in the plots. Most of the plots, anyway; that's the issue.

" wrote:

gwhite wrote:

... so "where to
cut off" is not easy to define.

We agree.

As long as you continue to use "peak" in a sense which is
standard for a different field, but different from everyone else
in this discussion, it's just going to get more confusing
rather than less.

The original comment was:
JVD Sluis wrote:
"A noise level of 0,1% would mean that the graph can show a peak of 0,1%
without the corresponding level of RBCs being present."

A noise level rating for the machine would mean some "peak," as folks
are using it here, could be *greater* than 0.1%. The question is "how
much greater" with respect to the "valid test threshold" and how
probable it is that the threshold will be exceeded.

Noise level *ratings* aren't specified in "peaks," regardless of the
language. That's one reason why Campbell was concerned. (It's true I
wasn't clear on how others were using "peak." Not surprising given that
I didn't look at the papers till Friday, I think.)

Yes, we can use "peak" to describe the integrated energy (or power) of
the lobes, but that does not rule out using it also for describing
random processes and the /maximum/ (greater-than-average,
greater-than-mean, greater-than-root-mean-square) values implicit in
random processes across time. Most people I know just call a localized
maximum a peak -- that is what it is. The "peaks" that I'm talking
about (and Campbell) are in the time-domain. The peaks in the plots are
frequency-domain peaks. I'm accustomed to both domains, and the
transformation between the two.

For this test, the counts are positive. So without really knowing, we
could make a WAG for the noise distribution of the instrumentation
(plus/including the various lab technicians in various labs, and every
other possible error) to be Rayleigh. There is no "peak limit" to that
distribution. The question is how often (in time, then transformed to
the f-domain) a given threshold is exceeded (the probability of a false
positive) given the instrument and all the other errors that could
happen, including operator error. The "peaks" are what do that. That's
why Campbell wants the probability question answered.