Marken's spreadsheet

[Martin Taylor 971203 16:40]

Rick Marken (971124.1100)

I've finally had a chance to play with Rick's spreadsheet, and have some
comments about the spreadsheet itself, as well as about the conclusions
he draws from it.

and in the neighbour column there are cells containing formulae
of the form" where X equals the fixed value of the left neighbour,
so that the formula is really "self + 0.5*0".

These "neighbor column" cells are the cells that compute the
perceptual variable (CV) that is controlled by each subject (control
system). An example of the formula in one of these cells is:

=($J14*$A$24+$K14+$B$24+$L14+$C$24+$M14*$D$24+$N14*E$24)-B4

                  ^ ^
That is, indeed, the formula. I assume Rick "cut and pasted" it out
of the spreadsheet. But there's a problem with it (see next comment).

The environmental variables are weighted by coefficients from
what you describe as the "big matrix of "perceptual weights"" .
The coefficients are in rows J through N of the spreadsheet; each
subject's perceptions represent a weighting of the environmental
variables by coefficients from a different row of this matrix.
In the formula above, the enviromental variables are weighted by
the coefficients in row 14 of the perceptual weights matrix
($J14*$A$24+$K14+$B$24...note the row indication associated with
columns J, K ...).

But shouldn't it read $K14*B24 and $L14*$C24, rather than $K14+$B24
and $L14+$C24 if $K14 and $L14 are the perceptual weights for two of
the "Fixed environmental variables"? I assumed this to be so, and made
these changes; all subsequent comment are based on spreadsheets
altered this way.

···

------------------

I think that the most interesting experiment you can do with the
spreadsheet right now (I plan to make improvements and eventually
write a paper about it) is type in new values for the fixed
environmental variables (cells A24 - E24) and see what happens
to the average results.

I did a different experiment first. It is normal practice in such IV-DV
experiments to assign subjects randomly to the experimental conditions.
So I did that, as well as I could, by hand. In the spreadsheet as
originally distributed, the subjects are assigned in a very specific
order, and it is this order of assignment that causes the IV-DV
correlation seen in the results.

I wanted to randomize the assignment of subjects, but I couldn't find a
way in Excel 4 to permute the rows of the perceptual weight matrix
(each row determines the characteristics of one subject, so permuting
the rows at random would reassign subjects to experimental conditions),
so I did several "hand-shuffles". (Note--if you want to do this, you have
to reset the values in the formulae in the shaded "CV" cells back to the
way they were, because Excel cleverly says "I know you don't want do what
you seem to want to do, so I will change the references to track your moves").

When I did that a few times, I found that the relationship between the
"IV" (10, 20, 30) and the DV disappeared. The average DV would sometimes
be highest in one column, sometimes in another. If this happened to be
a simulation of an ordinary behaviourist study, it would result in the
experimenter saying "there's NO effect of IV on DV here." Or, "There
is no significant relationship between the subjects' behaviour and the
experimental variable." Occasionally, of course, the "randomization"
doesn't work; 1, 2, 3, 4, 5 is a perfectly good "random" permutation
of the first 5 integers (except that I did it deliberately), but such
orderings don't happen very often. Randomization is important in group
studies, to avoid precisely the problem that occurred in Rick's
spreadsheet as it was originally distributed.
------------------

A change in the value at which the
environmental variables are held constant can change the average
results COMPLETELY. This shows dramatically that the group results
tell us NOTHING about the nature of the individuals in the group.

I guess that's right, at least. There's no relation between IV and DV,
either, so the "conventional" experimenter would come to the same
conclusion. And we (PCT-mavens) "know" that S-R studies tell us nothing
about the nature of the individuals when the S and the R are parts
of the input and output variables in a perceptual control loop.

Changing the _constant_ environment changes nothing about the
individuals but it gives a completely different piture of the
average behavior of the group.

Whoops! It changes nothing about the _internal processing_ of the
individuals, but it sure changes their output values. What you should
say is that stimulus-response measures tell you nothing about the
internal processing done by the individuals, because the outputs change
so much when the fixed (i.e., un-noticed by the "conventional" experimenter)
variables change. But we've said that in so many different ways on CSGnet
that a demo of this sort isn't going to change many minds.

If you take the group results as
an indication of what is true of the individuals, your conclusion
about those individuals will differ substantially depending on the
level at which "extraneous variables" were held constant in the
experiment.

We know that the individuals have not changed their internal performance
when the environment changes, whereas the output results have changed.

But the experimental results also show that the individuals vary
all over the lot. I haven't computed the standard deviations or the
"significance" of the differences across the columns, but by eye it
sure looks like the SD is several times the ordinary differences among
the means--hence, in conventional terms "no effect." And the "conventional"
experimenter would normally compute the SD.

What it points up is that the stimulus-response relationship can change
considerably while the CEV does not change. This is, to me, the real
arrow in the heart of "coventional" research. But the spreadsheet does
not deal with the issue it purports to address, the validity of using
group data to estimate individual data. What stays constant within the
individual is not what the measurements measure. And any conventional
experimenter would look at the spread of individual data and say that the
standard deviation of the data was huge, making the mean useless as an
estimator of individual values even of what _was_ measured.

I think a spreadsheet along the lines Rick tried might be valuable as
a demo, but this one isn't it, because the control theorist and the
conventional experimenter would come to the same conclusions.

Martin

[From Rick Marken (971203.1600)]

Martin Taylor (971203 16:40) --

But shouldn't it read $K14*B24 and $L14*$C24

Yes. Nice catch. I'll fix it.

the subjects are assigned in a very specific order, and it is
this order of assignment that causes the IV-DV correlation seen
in the results.

It's one possible order. Different results are expected with
different randomizations. I'll set up a way to reassign subjects
more easily.

When I did that a few times, I found that the relationship between
the "IV" (10, 20, 30) and the DV disappeared.

Sometimes it does. Sometimes it doesn't. You can increase the
chances of getting an average effect of IV on DV by using
more extreme values of the IV (and tuning up the values of the
"extraneous" variables). I believe it's possible to arrange things
so that you always have a chance of finding a significant effect
of IV on DV.

Randomization is important in group studies, to avoid precisely
the problem that occurred in Rick's spreadsheet as it was originally
distributed.

The only problem with the spreadsheet as originally distributed
was the problem with the formula for the perceptual variables.
When you do an experiment you only randomize once and you
see what you see. When you see something different with different
randomizations you are in the position of the psychologist who
repeats the experiment and finds no (or a different) effect.

There's no relation between IV and DV, either, so the
"conventional" experimenter would come to the same conclusion.

Only if there was no relationship between IV and DV in the
group data. This does happen sometimes. But sometimes there is
an effect. Real experiments are done once and, if the results
are significant, they are duly published as a psychological
fact. The fact that you can find subject assignments and levels
of the IV that produce a null result is expected; such results
are found all the time and not reported. The spreadsheet shows
that you can find a significant group effect of IV on DV when
every subject is doing something different (controlling a different
perception) than every other subject; no subject is like another
so the group result is not applicable to _any_ subject.

And we (PCT-mavens) "know" that S-R studies tell us nothing
about the nature of the individuals when the S and the R are parts
of the input and output variables in a perceptual control loop.

That's right. You mean you think that you have shown that this
isn't true?

But the spreadsheet does not deal with the issue it purports to
address, the validity of using group data to estimate individual
data.

We do our best. It looks like it to me like it deals with the
issue it purports to address.

Best

Rick

···

--
Richard S. Marken Phone or Fax: 310 474-0313
Life Learning Associates e-mail: rmarken@earthlink.net
http://home.earthlink.net/~rmarken

[From Bruce Abbott (971203.2025 EST)]

Bill Powers (971202.1710 MST) --

It's been a long day, Bruce, and I am in no condition to suffer foolish
statements gladly. You showed no such thing. You showed that IF all the
people in a group are affected similarly, the group measure will also show
the effect. But you've ignored the case in which all the people are
affected differently, yet the group still shows the effect. Since you can't
say in advance which will be the case, you simply can't draw the conclusion
you want to draw.

If all the people are affected differently, then whatever average effect is
shown by the group-based method will be weak and certainly not the sort of
thing that shows sufficient regularity (under the conditions studied) to be
worth additional effort.

Martin Taylor 971203 16:40 to Rick Marken --

But shouldn't it read $K14*B24 and $L14*$C24, rather than $K14+$B24
and $L14+$C24 if $K14 and $L14 are the perceptual weights for two of
the "Fixed environmental variables"? I assumed this to be so, and made
these changes; all subsequent comment are based on spreadsheets
altered this way.

. . . I found that the relationship between the
"IV" (10, 20, 30) and the DV disappeared. The average DV would sometimes
be highest in one column, sometimes in another. If this happened to be
a simulation of an ordinary behaviourist study, it would result in the
experimenter saying "there's NO effect of IV on DV here." Or, "There
is no significant relationship between the subjects' behaviour and the
experimental variable."

Bruce

[From Bill Powers (971203.1920 MST)]

Bruce Abbott (971203.2025 EST)--

Not quite sure what the quotations in your post are supposed to imply,
since you didn't explain.

Don't you think you could come up with a population in which there was an
effect of an IV on a DV, but in which the effects on all the individuals
were different from each other in magnitude and sign? I'd be willing to try
it, but it would mean much more coming from you.

Best,

Bill P.

[Martin Taylor 971204 10:15]

Bill Powers (971203.1920 MST) to Bruce Abbott (971203.2025 EST)--

Don't you think you could come up with a population in which there was an
effect of an IV on a DV, but in which the effects on all the individuals
were different from each other in magnitude and sign? I'd be willing to try
it, but it would mean much more coming from you.

That would be easy, I think. But isn't it the opposite of what the argument
is about. After all, in Marken's spreadsheet, the IV _does_ affect the DV
for every subject, and the individuals show DVs that differ in magnitude
and sign (though the spreadsheet itself does not show how the IV affects
the DV for any individual). The problem with the spreadsheet is that the
group averages do not seem to show any "significant" effect when the
subjects are (pseudo-)randomized across the experimental conditions.

What I think you want to ask for is a condition in which the group data
would show "significant" effects (or, my preference, narrow confidence
limits) despite variations in the individual effects that are equally
distributed around zero in magnitude and sign. I think this would be
_very_ hard to set up.

Or, in a weaker proposition that would be accepted by just about any
"conventional" psychologist, you might be asking for a condition in
which some individuals show effects opposite to others even though
the group data are "significantly different from zero." You can
easily set up that kind of a situation.

But what would be the point? Rick's spreadsheet shows quite dramatically
the impropriety of using S-R relationships to infer the properties of the
control loop, and surely that's the main issue at hand. In the
spreadsheet, all the subjects process the data the same way, using
the same parameters (except for being influenced by different "fixed"
environmental variables "unknown to the experimenter." Yet their IV-DV
relationships vary wildly.

A second issue is the impropriety of using measurements of group
data about X to infer individual values of Y. Your own study shows
this equally dramatically, using the slope data. And you have argued
well that many research papers are published though they are based on
this very mistake.

It's hardly news that individuals vary, and that any one individual is
unlikely to have exactly the average value of anuy measured property X.
So why put effort into further demonstrations of that non-controversial
fact?

Martin

[Martin Taylor 971204 08:55]

Rick Marken (971203.1600)]

Martin Taylor (971203 16:40) --

The only problem with the spreadsheet as originally distributed
was the problem with the formula for the perceptual variables.
When you do an experiment you only randomize once and you
see what you see. When you see something different with different
randomizations you are in the position of the psychologist who
repeats the experiment and finds no (or a different) effect.

True. That's why there are all these criteria for how many subjects
are enough in randomized between-subjects studies. No matter how many
subjects there are, the possibility always exists that they are assigned
to experimental conditions in a way that badly unbalances some
characteristic that the experimenter doesn't know to be important.
The larger the number of subjects, the lower the probability of that
happening.

There's no relation between IV and DV, either, so the
"conventional" experimenter would come to the same conclusion.

Only if there was no relationship between IV and DV in the
group data. This does happen sometimes. But sometimes there is
an effect. Real experiments are done once and, if the results
are significant, they are duly published as a psychological
fact.

"If the results ae significant," yes. And that notion of "significant"
always bugs me, too. However, in _none_ of the trials of your spreadsheet
would the results approach significance, at least by my "eyeball" test.
Individual result values range all over the lot, with bigger variances
for the groups having larger values of the "experimental variable."

The spreadsheet shows
that you can find a significant group effect of IV on DV when
every subject is doing something different (controlling a different
perception) than every other subject; no subject is like another
so the group result is not applicable to _any_ subject.

Maybe my eyeball test is wrong. What _is_ the "significance level" for
your originally distributed spreadsheet with the subjects carefully
ordered?

And we (PCT-mavens) "know" that S-R studies tell us nothing
about the nature of the individuals when the S and the R are parts
of the input and output variables in a perceptual control loop.

That's right. You mean you think that you have shown that this
isn't true?

I'm not clear why you would ask this question without a smiley:-),
given that the thrust of my comments was that your spreadsheet is a
good demonstration that it _is_ true. As witness this comment from
my posting, which immediately preceded your next quote:

What it points up is that the stimulus-response relationship can change
considerably while the CEV does not change. This is, to me, the real
arrow in the heart of "coventional" research.

But the spreadsheet does not deal with the issue it purports to
address, the validity of using group data to estimate individual
data.

We do our best. It looks like it to me like it deals with the
issue it purports to address.

Then I guess you missed the other main thrust of my comment. I'll put
it in an analogy: Your claim for the spreadsheet is that you can't estimate
an individual's body temperature from measurements of the body
temperature of a lot of people and finding the mean and S.D. of those
measurements; my comment is that you demonstrate the impossibility of
estimating an individual's body temperature from measurements of the
height of a lot of people and ...

To put it in the terms of the spreadsheet, you say (correctly) that
the group data for the _output_ values of the subjects cannot be used
to estimate the subject's internal processing parameters. (Analogies
to "height" and "body temperature"). You then say (incorrectly) that
this proves group measurements of any quantity X cannot be used to
estimate individual measures of quantity X. I understand this latter
to be the issue the spreadsheet purports to address (an understanding
based on _many_ messages over the past couple of weeks). The spreadsheet
does not address it. So my comments [Martin Taylor 971122 21:40]
still stand:

A demonstration that you can't use group data to predict individual
data has to use the _same_ measure for both.

and, to Marken's proposal:

I'm sure it would be easy to build a spreadsheet model where the
group relationship between IV and DV differs from the actual
relationship etween IV and DV for _every_ member of the population.

Do it, using a legitimate measure, and we will all be indebted.

It hasn't been done yet. I look forward to seeing it done.

Martin

[From Bill Powers (971204.1543 MST)]

Martin Taylor 971204 10:15 --

But what would be the point? Rick's spreadsheet shows quite dramatically
the impropriety of using S-R relationships to infer the properties of the
control loop, and surely that's the main issue at hand. In the
spreadsheet, all the subjects process the data the same way, using
the same parameters (except for being influenced by different "fixed"
environmental variables "unknown to the experimenter." Yet their IV-DV
relationships vary wildly.

I guess there's a agreement among you, Rick, and me, but not with Bruce
Abbott. Maybe you should be addressing his points directly.

Best,

Bill P.