From: Katharine Thayer (kthayer@pacifier.com)
Date: 11/08/00-10:16:54 AM Z
Ah, shoot. I can't resist wading in again, now that it's starting to get
interesting. I like the idea of a concrete example and thank Judy for
bringing the whole discussion down to earth. And Wayde has done a fine
job of identifying some of the variables that might be tested in Judy's
example. So far, so good. The problems come in the next step and the
next one and the next one after that. What I see as the steps coming up
are:
1. convincing an individual printer that it makes sense to test
variables that aren't relevant to him/her, such as environmental
variables, which, while they are variable over the universe of
printers, are constant in each printer's situation. What good would it
do me to know how a particular ingredient works in a desert, for
example, when I am never likely to print in such a place. I don't know
about anyone else, but when I test it's not to uncover some universal
truth about gum printing or to identify all the different combinations
of variables under which a particular effect occurs, but simply to find
out what happens in my environment using the equipment and materials I
use.
2. Assuming that you've convinced us that we should do this, then the
next step is making sure we've identified all possible variables,
determining how to measure each, and running experiments using several
values of each of the variables, the more the better, and filling those
charts up with numbers.
3. Choosing a statistical model to determine the effects of each of the
variables alone and in combination with other variables. Here's where I
have a problem with Wayde's recommendation, as I've already said. I
could tell you all the statistical reasons why this is not a good idea,
but I'm sure my explanation wouldn't be useful to most of you. (However,
I do need to mention in an aside to Wayde that just because the models
are multivariate doesn't mean there aren't assumptions that should be
respected. There are some multivariate methods that aren't based on such
strict assumptions, but the ones described in the website you pointed to
aren't those.)
I of course agree with Wayde in principle, that variables acting in
combination with each other are probably responsible for most of the
effects we see. If he'd said that and stopped there, I'd have no
problem. However, to go on from there and say that adopting these
oversimplified correlation-based multivariate models as the way to find
out how each variable affects the result in combination with the others,
shows a certain... lack of familiarity with the models and their
attendant problems of analysis and interpretation, I'd guess, for which
I don't fault him. Which brings me to the last steps:
4. Analyzing the data according to the model, making sense of it in
practical terms, and validating the model by replicating it on another
set of data and by checking the predictions in practice. I will be
watching with interest to see how that all works out if this example is
carried out to the end. I wouldn't go near it, myself, but others are
welcome to go there and see what they find, with this small warning:
there be dragons on that path, in the form of statistical issues that
can make the results uninterpretable and lead to erroneous, unreliable,
and invalid conclusions. As I suggested before, the probability that the
result will be any less confusing than the present situation is very
close to zero. Good luck!
P.S. to respond quickly to the last point in Wayde's most recent post,
which came in just as I was about to click the send button: I am all for
repeatability. What I've been trying to say over and over again is that
your idea that repeatability can be achieved by the methods described in
the references you gave is, in my experience, a misplaced faith.
Katharine
J. Wayde Allen wrote:
>
> On Tue, 7 Nov 2000, Judy Seigel wrote:
>
> > Let's say I want to find out why the gelatined paper hardened with
> > glyoxal that sat on the table, but UNDER some other paper, for 6
> > months discolored, and my test strips in the file are as white as the
> > driven snow.
> >
> > My theory now is that something in the applied dichromate (if the print is
> > made promptly, as the test strips were) forestalls the discoloration.
> > That it wasn't *where* the papers were stored, but what stage they were in
> > that made the difference.
>
> OK, let's start trying to figure out what you really want to know? From
> what you've written I get:
>
> 1. What is causing the test strips to yellow?
>
> 2. Does dichromate affect the staining?
>
> 3. Does the storage location have anything to do with the staining?
>
> > To test, I would put the same paper side by side
> > on the table, one with a test strip exposed & developed on it, one blank.
>
> I'm assuming that your exposed and developed test strip is paper that has
> been sized, hardened with Glyoxal, coated with the gum emulsion, exposed
> for some length of time, and then water developed? Are you exposing
> through a step tablet? What do you mean by blank? I'm guessing this is a
> piece of the same paper that is sized and hardened with Glyoxal - correct?
>
> At any rate, what can you learn from either a positive
> or negative result? With your test, if one of the tests stain, but not
> the other, can you answer the question you set out to answer? I would
> submit that you really can't answer either question 1 or 3 that I listed
> above. You do get some indication about question 2, but it is hard to say
> that it was simply the dichromate that caused the effect. Could the
> exposure have also played a role in the staining? What about the water
> development?
>
> One particularly odd thing is that you are simply ignoring environmental
> conditions. If these have any bearing on the staining you get no
> information about it. This is one of the reasons why your tests may not
> give the same results in someone else's lab.
>
> Also you say that the strip that were under some paper on your desk
> yellowed whereas those in your file didn't. Any reason why you discounted
> the possibility that something in the paper laying on the test strips
> didn't contribute to the staining?
>
> > One variable, right?
>
> No, and what's worse you really haven't indicated what your supposed one
> variable is? What I'm counting off hand are:
>
> Paper substrate
> strips covered by paper or not
> Type or brand of gelatin sizing
> Pigment type or brand if any
> dichromate type (ammonium or potassium?)
> exposure
> development water (distilled or tap?)
> environment (lots of variables - illumination, temperature, etc.)
>
> I'm sure we can look at this longer an come up with more, but the point is
> that your simple test isn't really that simple. I'd guess that you are
> using the same paper type for all strips, that everything is sized with
> the same gelatin, that you've only chosen one dichromate, and that you
> only are using one pigment if any at all? You would be correct to say
> that these variables are controlled. Of course, by locking these down, you
> get no information about what affects they may have on the staining. You
> have also seemingly ignored several possibly key variables:
>
> other paper in contact with strips
> development water
> exposure
> environment
>
> You can only do that if you can make the case that they are not important
> to the staining process. If these do matter, then you've got a problem.
>
> That is what I understand about your test. You'll have to correct any of
> my mistakes and misunderstandings before we can look at this much deeper.
>
> - Wayde
> (wallen@lug.boulder.co.us)
This archive was generated by hypermail 2b30 : 12/01/00-11:46:56 AM Z CST