From: Katharine Thayer (kthayer@pacifier.com)
Date: 11/06/00-09:57:41 AM Z
This argument doesn't work for me. I don't think the problem is that
people aren't doing adequate individual testing of their process because
they don't understand multicollinearity between independent variables.
Furthermore, introducing the idea of experimental models for data
analysis, even if there weren't enormous problems with such models given
the kind of data we're dealing with, seems massively beside the point
for alternative processes as practiced: each of us painstakingly making
one print at a time in our various garages, sheds, bathrooms,
custom-made climate-controlled workshops, and university darkrooms. I
don't believe there's a statistical model you could come up with that
would yield for example a gum printing process that would work for me in
my cold dank basement workshop that has a river running through it in
the winter, for someone in the dry hot southwest, and for someone in a
townhouse in Manhattan with steam heat, no matter how many variables you
entered into the equation.
But to respond to Wayde's assertion that the kind of experimental
design he points to might be useful for our purposes:
Just for starters, these models are intended for mathematical data,
preferably continuous mathematical data. What would the outcome data be
that you would test the interactions between independent variables
against? The only outcome data that we are interested in is: the print
looks the way I want the print to look. How exactly are you going to
quantify that, especially since it is subjective to each individual
printer? In addition, we've got each individual printer making prints by
the handfuls, when hundreds, even thousands of observations (prints)
would be required, all made under precisely controlled conditions, to
identify all the interactions between all the possible variables, even
if we could agree on some kind of outcome variable and even if
multicollinearity didn't invalidate the statistical model in the first
place, rendering the regression coefficients indeterminate and
uninterpretable. So what I'm saying is that even if we got a big grant
and set up an alt-process testing lab and hired full-time testers to set
up experimental designs to test all the interactions between all the
variables, using the kinds of statistical designs Wayde seems to be
advocating by his choice of URLs, (understand I'm NOT saying I think
this is a good idea) the conclusions they would come up with would
almost certainly be at least as unreliable as the results we come up
with on our own, testing one variable at a time. Sometimes, as someone
else rightly suggested, statistical analysis just results in a lot of
garbage, and these kinds of statistical models, used with insufficient
statistical understanding, have produced more garbage than all others
put together.
Second, my experience watching this list and other alt-photo lists has
almost never been that problems arise because some individual is having
trouble controlling their printing process because they haven't taken
correlated variables into account. Instead, my experience has been that
most people have been able to work out a process that works quite well
for them, choosing a paper they like, a light source that works for
them, and other ingredients that work with the things they've already
chosen. In other words, working through a process of keeping some things
constant while they nail down other things. No, the problems that
spring up again and again occur when egos get involved and people think
that because another person has come to a different conclusion as a
result of testing in their own particular situation with their own
particular environmental situation and choice of materials, the other
person is *wrong*. In this case, Wayde is correct: we need a better
understanding of the correlation between variables, not in order to set
up experimental designs for testing but to keep us from saying
ridiculous things. The person who makes absolute statements about gum
printing, for example, comes a cropper every time. I've heard a gum
printing "expert" say for instance that ammonium dichromate is no good
for gum printing because it always causes dichromate staining. This is a
ridiculous statement that can easily be proved false. What the person
really means, and should say, is "Potassium dichromate works better with
the kind of light source, the paper, the whatever....that I choose to
use, than ammonium dichromate. You may find ammonium dichromate works
perfectly well for you if you choose a different light source etc etc"
I won't share my opinion about the statistical information on the
websites in question, because my point has been that I don't think any
of this is relevant to alternative processes anyway, other than possibly
to make us more aware of the fact that we should think twice before
heralding our individual testing results as universal truth.
My 2cents.
Katharine Thayer, Ph.D. (statistics)
J. Wayde Allen wrote:
>
>
> > Firstly, explain to the backward what you mean by "highly correlated
> > variables."
>
> Say you have two variables A and B. The combination of which results in
> some result Y. As long as the value of either of these variables
> doesn't change the how they each affect Y one-variable-at-a-time testing
> is fine. One would say that the variables are independent.
>
> The problem I'm talking about occurs when the effect that variable A has
> on Y depends on B. Such variables are said to be highly correlated. In
> this case, holding one variable constant and changing only the other one
> doesn't give you any information about this interaction. In this case, if
> you hold variable B constant and vary A, what you see is only how A
> affects Y at this particular concentration of B. If B changes, your test
> data is no longer valid.
>
> > In gum printing, changing more than one variable at a time may
> > give you a good printing method, but little information about more than
> > that exact combination. If we ("we") keep all conditions the same and
> > change one variable, say the pigment, we ("we") at least see the changes
> > wrought by that pigment in the given combo....
>
> That is correct. I've never disagreed with you or anyone else on this
> point.
>
> > Pigment character changes with the environment, so it isn't clear why you
> > scorn one-variable-at-a-time. It's not crucial to a beginner, who needs to
> > find a general modus operandi-- but for fine tuning & control, also
> > trouble shooting, we (I) don't see how, or why, you dismiss it.
>
> I don't specifically scorn one-variable-at-a-time testing. It certainly
> has its place. However, this group quite often laments how these
> processes seem fickle. Many of the posts also deal with processes that
> were working and somehow went out of control. My reason for bringing this
> up is that one-variable-at-a-time testing in processes with known
> inter-variable correlation will always create this problem. I'm not
> saying it is bad, but rather just a fact of life. I think that the group
> should know about the limitations of the one-variable-at-a-time testing
> method. I also think you need to know that there are very good ways to
> get around these limitations. In other words, you don't have to live with
> the limitations of one-variable-at-a-time testing.
>
> - Wayde
> (wallen@lug.boulder.co.us)
This archive was generated by hypermail 2b30 : 12/01/00-11:46:56 AM Z CST