Correlations

[From Bruce Abbott (970423.1210 EST)]

Bill Powers (970422.0757 MST) --

I didn't really understand this example when it was given, but now I do.
You have a function Y = f(X1, X2, X3, X4). There are no other variables,
known or unknown, affecting Y. Therefore Y, obviously, is perfectly
predicted by the function of the four variables, no matter what the function
is and no matter how the individual variables change.

The randomness of the four X variables is a red herring. They could vary in
any way, random or not, correlated or not, in any combination, and Y would
still be predicted perfectly by them. There is no random component in the X
variables: they are perfectly known.

Since the values of the four variables are perfectly known, this is not a
statistical problem.

Yes it is, Bill.

The statistical problem is that you, the researcher, do not _know_ how (or
even _if_) P1, P2, P3, and P4 relate to R. Discovering whatever
relationship exists is a problem of statistical inference.

Examining each Px variable separately, you discover that each correlates
approximately 0.50 with R; thus approximately 25% of the variance in Px and
R is shared. None of the Px variables is doing a particularly good job of
"explaining" the variance in R.

However, you submit these four predictors and the response variable to a
multiple correlation analysis, which searches for simultaneous weightings of
the individual predictors that will yield the best prediction of R
(definitely a statistical problem). This analysis reveals that all the
variance in R can be accounted for by a linear combination of P1, P2, P3,
and P4, and the analysis gives you the optimum weights, given the data. You
have discovered a powerful model for predicting R, even though each
individual predictor was by itself a poor predictor of R.

To make it a statistical problem you would have to say

Y(theoretical) = f(X1(actual), X2(actual), X3(actual), X4(actual)),
Y(observed) = f(X1(observed), X2(observed),
               X3(observed), X4(observed)),
where
  Xn(0bserved) = Xn(actual) + rn
  rn = random variable

Now the observed value of Y would differ from the theoretically expected
value of Y, and the prediction would not be perfect unless the observations
of Xn were perfect. In the case of Y = X1 + X2 + X3 + X4 as the theoretical
relationship, the prediction error would be the square root of the sum of
squares of the individual observation errors due to rn. The total error of
observation would be larger than any individual error of measurement.

No, this does not make it a statistical problem. It makes it a somewhat
_different_ statistical problem. Now your individual predictors are not
only poorly related to R, they are also subject to unreliable measurement.
Assume that measurement error is normally distributed, that _half_ of the
variance in the observed predictor Pxo is due to variation in Px and the
remainder to measurement error. How will this affect our ability to account
for the variation in R? Answer: we can account for half of the variance in
R using our four Pxo predictors, which now individually correlate with R
about 0.354. The correlation between our composite predictor and R will be
about 0.707.

This isn't _great_, but remember, we are assuming that the noise in the
measurement is as large as the variation in the signal being measured --
clearly a very bad measurement. With noise amounting to 10% of signal, we
could account for 96% of the variance in R using the four predictor variables.

My point in raising the example was that a combination of predictors, which
individually correlate rather poorly with the response variable, could
nevertheless do an excellent job of predicting R. My initial demo proved
that this is so. Random measurement error reduces this ability in
proportion to the size of the error, yet even then my statement holds: the
combination of predictors does better than any one of them alone. A low
simple correlation between two variables does not preclude this relationship
from being scientifically useful.

So Bruce Abbott's example is spurious.

No, it isn't. I illustrated the best case to bring home the point most
clearly. But the conclusion holds even in the presence of random
measurement error.

Regards,

Bruce

[From Bill Powers (970423.0649 MST)]
Bruce Abbott (970423.1210 EST)--

Bill Powers (970422.0757 MST) --

I didn't really understand this example when it was given, but now I do.
You have a function Y = f(X1, X2, X3, X4). There are no other variables,
known or unknown, affecting Y. Therefore Y, obviously, is perfectly
predicted by the function of the four variables, no matter what the
function is and no matter how the individual variables change.

The randomness of the four X variables is a red herring. They could vary
in any way, random or not, correlated or not, in any combination, and Y
would still be predicted perfectly by them. There is no random component
in the X variables: they are perfectly known.

Since the values of the four variables are perfectly known, this is not a
statistical problem.

Yes it is, Bill.

Wouldn't it be exactly the same problem if the four X's varied according to
independent but smooth, analytical, and perfectly predictable functions of
time? How would the problem be any different?

The statistical problem is that you, the researcher, do not _know_ how (or
even _if_) P1, P2, P3, and P4 relate to R. Discovering whatever
relationship exists is a problem of statistical inference.

That's not a _statistical_ problem, it's just a _mathematical_ problem. You
may have learned the mathematics of rotating axes and minimizing functions
in the context of statistics, but it has nothing fundamental to do with
random variables.

Examining each Px variable separately, you discover that each correlates
approximately 0.50 with R; thus approximately 25% of the variance in Px
and R is shared. None of the Px variables is doing a particularly good
job of "explaining" the variance in R.

However, you submit these four predictors and the response variable to a
multiple correlation analysis, which searches for simultaneous weightings
of the individual predictors that will yield the best prediction of R
(definitely a statistical problem). This analysis reveals that all the
variance in R can be accounted for by a linear combination of P1, P2, P3,
and P4, and the analysis gives you the optimum weights, given the data.
You have discovered a powerful model for predicting R, even though each
individual predictor was by itself a poor predictor of R.

In any function y = f(x1..x4), none of the x's does a "particularly good
job" of predicting the value of y even in the complete absence of
uncertainty. That's because y is a function of all the x's, not just one x,
and lacking any further statements we have to assume that the x's vary
independently (that is, none can be expressed as a function of the others).
However, if you have four samples of the function, with the x's being
independent of each other, you get four equations in four unknowns, which is
sufficient to solve for any non-stochastic linear function of the x's, y =
a*x1 + ..d*x4. You solve for the coefficients, and in the case you used you
find that they are all equal to 1. This is what your multiple correlational
analysis will find, too, because it is simply a method for finding the best
solution -- the best set of coefficients in the linear equation. If you set
all variances to zero, this method reduces to a way of solving simultaneous
algebraic equations.

It doesn't matter how the values of the four x's are generated, as long as
they're independent and the x's are observed exactly. This is why I said
their randomness is a red herring. This doesn't make a statistical problem
out of the situation; in order for a statistical problem to exist, the
relation of the x's to y would have to contain a random component, or the
observation of each x would have to contain a random component, or both. In
other words, in any given sample, the four x's would not add up to y, but to
y plus some random variation. In that case you _would_ need a statistical
analysis and much more than 4 samples to find the best set of coefficients
or weightings.

Even in that case it wouldn't matter how the individual x's were generated;
whether they each came from a random number generator or an independent
analytical or logical function. The x's are not the unknowns in the
statistical analysis; the coefficients or weightings are the unknowns. The
x's are _observed values_, and in your example are observed without any
uncertainty.

To make it a statistical problem you would have to say

Y(theoretical) = f(X1(actual), X2(actual), X3(actual), X4(actual)),
Y(observed) = f(X1(observed), X2(observed),
               X3(observed), X4(observed)),
where
  Xn(0bserved) = Xn(actual) + rn
  rn = random variable

...

No, this does not make it a statistical problem. It makes it a somewhat
_different_ statistical problem. Now your individual predictors are not
only poorly related to R, they are also subject to unreliable measurement.

That is the case where it would matter that there is a random component in
the predictors. In fact, you could have considered your example to be a case
in which you find that the predictors are _highly_ unpredictable; every time
you measure them, you get a different value, with the whole set of values
having a normal distribution around zero for each predictor. If you looked
at it that way, the best estimate of the value of each predictor would be
zero, and according to the underlying relationship y = x1 + x2 + x3 + x4,
you would predict y = 0 for every observation. But that is not the case you
used; you assumed that each value of a predictor was meaningful and was
measured accurately, and that for every sample, the sum of the values
predicted y exactly. So the uncertainty in measuring the predictors was
zero, as was the uncertainty in the relationship of the predictors to y. The
only "uncertainty" remaining was your own, because you had not yet solved
the equations to find the values of the coefficients.

[For the case of 50% variance in the measurements]

This isn't _great_, but remember, we are assuming that the noise in the
measurement is as large as the variation in the signal being measured --
clearly a very bad measurement.

Yes, that is the case when the measurement's correlation with the actual
values of the variables is only r^2 = 0.5 (Kennaway) or r = 0.707. A very
bad measurement, as you say.

With noise amounting to 10% of signal, we
could account for 96% of the variance in R using the four predictor >variables.

... and of course that would imply a correlation of sqrt(0.92) or about 0.96
between each predictor measurement and the true value of the predictor.

My point in raising the example was that a combination of predictors,
which individually correlate rather poorly with the response variable,
could nevertheless do an excellent job of predicting R.

And I'm saying that this is a spurious conclusion, which obviously does not
endear me to you. I'll agree that a combination of four predictors can
improve the _relative_ uncertainty in the response variable (compared to
that of the predictors) by a factor of sqrt(4), or 2. But in your example
there was no uncertainty: the coefficients were exact, and the measurements
were exact. There was no _statistical_ uncertainty to reduce; only the
psychological uncertainty of not yet having solved the equations to find the
exact answer. When there is true statistical uncertainty, there is no exact
answer: only a mean plus unpredictable fluctuations.

My initial demo proved
that this is so. Random measurement error reduces this ability in
proportion to the size of the error, yet even then my statement holds: the
combination of predictors does better than any one of them alone.

Yes, a little bit, when there is any uncertainty in the data, and in terms
of _relative_ variance. The absolute variance in the result is always
greater than the variance in any observation. For _particular_ functions
relating the x's to the y's, the result can be far more variable than the
variance in the observations in either absolute or relative terms,
especially if the x's are correlated. Consider the relationship

qi = Fe(o) + Fd(qd) + noise.

when a control system is involved. Fe(o) is highly correlated with Fd(qd),
yet the correlation of each with qi is usually less than 0.1.

A low simple correlation between two variables does not preclude this
relationship from being scientifically useful.

Well, that's the point you hope to convince us of, but it depends entirely
on where you set your standard for scientific usefulness, and on what use
you intend to make of the results. If you sell widgets, a correlation of 0.1
between color and sales is useful to know about, provided that all colors
cost the same amount. If your future depends on the next customer buying a
widget, a correlation of 0.1 is pretty useless.

Best,

Bill P.

[From Bruce Abbott (970423.2125 EST)]

Bill Powers (970423.0649 MST) --

Bruce Abbott (970423.1210 EST)

Wouldn't it be exactly the same problem if the four X's varied according to
independent but smooth, analytical, and perfectly predictable functions of
time? How would the problem be any different?

The statistical problem is that you, the researcher, do not _know_ how (or
even _if_) P1, P2, P3, and P4 relate to R. Discovering whatever
relationship exists is a problem of statistical inference.

That's not a _statistical_ problem, it's just a _mathematical_ problem. You
may have learned the mathematics of rotating axes and minimizing functions
in the context of statistics, but it has nothing fundamental to do with
random variables.

I used randomly varying predictors to preclude any suggestion that the
results obtained had something to do with the predictors varying
_non_randomly. In experiments (as opposed to correlational studies), the
values of the predictors (independent variables) are usually set to a number
of fixed levels and the dependent variable is observed at each of these
levels. The logic of analysis is basically the same, although the
statistical model changes somewhat (to what is termed a "fixed effects"
model). So, you are right, the result I demonstrated had nothing
fundamental to do with random variables. But I didn't claim that it did.
After all, if there is an underlying relationship among variables, they can
hardly be said to be related randomly (i.e., unrelated).

Multiple regression uses statistical criteria to determine the
"best-fitting" set of coefficients (in this case, least squares). It has no
way of "knowing" in advance that the independent predictors will turn out to
be capable of perfectly predicting R because R is simply their sum. The
problem is still the same problem whether the predictors are known precisely
or imprecisely. Whether you choose to call it a "mathematical" problem or a
"statistical" problem is unimportant. Furthermore, because the "true"
values of the variables being measured are almost never precisely known, the
problem of finding the mathematical function that best accounts for the
pattern of variation observed is a statistical problem by _your_ definition,
in just about any real example (in contrast to the example I cooked up).

In any function y = f(x1..x4), none of the x's does a "particularly good
job" of predicting the value of y even in the complete absence of
uncertainty. That's because y is a function of all the x's, not just one x,
and lacking any further statements we have to assume that the x's vary
independently (that is, none can be expressed as a function of the others).
However, if you have four samples of the function, with the x's being
independent of each other, you get four equations in four unknowns, which is
sufficient to solve for any non-stochastic linear function of the x's, y =
a*x1 + ..d*x4. You solve for the coefficients, and in the case you used you
find that they are all equal to 1. This is what your multiple correlational
analysis will find, too, because it is simply a method for finding the best
solution -- the best set of coefficients in the linear equation. If you set
all variances to zero, this method reduces to a way of solving simultaneous
algebraic equations.

Of course. How does this undermine my conclusion that four predictors
apparently poorly correlated with the response variable may together account
fully for the variation in R? The researcher does not know the nature of
the true relationship between the predictors and R, but has discovered that
each predictor correlates poorly with R. A scatterplot reveals a lot of
apparently random scatter around the bivariate regression of R on P1, or P2,
or P3, or P4. My claim is that simply observing such low correlations is no
guarantee that some highly predictive relationship does not exist between
the predictors and R. I haven't heard you say anything that undermines this
claim (although making assertions of "red herings" and "spurious
conclusions" sure makes it _sound_ like you have, I must admit).

My point in raising the example was that a combination of predictors,
which individually correlate rather poorly with the response variable,
could nevertheless do an excellent job of predicting R.

And I'm saying that this is a spurious conclusion, which obviously does not
endear me to you.

Oh, come on Bill, we're all just a bunch of adorable teddy bears; how could
you and I be other than endearing? Spurious! You say the funniest things,
you lovable guy you.

A low simple correlation between two variables does not preclude this
relationship from being scientifically useful.

Well, that's the point you hope to convince us of, but it depends entirely
on where you set your standard for scientific usefulness, and on what use
you intend to make of the results. If you sell widgets, a correlation of 0.1
between color and sales is useful to know about, provided that all colors
cost the same amount. If your future depends on the next customer buying a
widget, a correlation of 0.1 is pretty useless.

Be sure to let us know when you are able to make such point predictions
(about buyer behavior and so on) with the high accuracy obtained in tracking
studies; I can think of literally thousands of situations in which this
ability would be invaluable, and worth a considerable sum on the open market.

But seriously, I have already indicated a far more important "use" of
moderate correlations than in maximizing income from widget sales: model
construction and the testing of alternative models. That is what makes such
correlations scientifically useful, even though they may not be much good in
themselves for making point predictions.

Regards,

Teddy

[From Bill Powers (970424.0601 MST)]

Bruce Abbott (970423.2125 EST)--

Wouldn't it be exactly the same problem if the four X's varied according to
independent but smooth, analytical, and perfectly predictable functions of
time? How would the problem be any different?

The statistical problem is that you, the researcher, do not _know_ how (or
even _if_) P1, P2, P3, and P4 relate to R. Discovering whatever
relationship exists is a problem of statistical inference.

That's not a _statistical_ problem, it's just a _mathematical_ problem. You
may have learned the mathematics of rotating axes and minimizing functions
in the context of statistics, but it has nothing fundamental to do with
random variables.

I used randomly varying predictors to preclude any suggestion that the
results obtained had something to do with the predictors varying
_non_randomly. In experiments (as opposed to correlational studies), the
values of the predictors (independent variables) are usually set to a number
of fixed levels and the dependent variable is observed at each of these
levels. The logic of analysis is basically the same, although the
statistical model changes somewhat (to what is termed a "fixed effects"
model). So, you are right, the result I demonstrated had nothing
fundamental to do with random variables. But I didn't claim that it did.
After all, if there is an underlying relationship among variables, they can
hardly be said to be related randomly (i.e., unrelated).

I'm glad the point I was trying to make is understood now.

Multiple regression uses statistical criteria to determine the
"best-fitting" set of coefficients (in this case, least squares). It has no
way of "knowing" in advance that the independent predictors will turn out to
be capable of perfectly predicting R because R is simply their sum.

The multiple regression approach is clearly useful when you are trying to
find a model that explains how a set of independent variables relates to a
dependent variable. It's limited in that you can never be sure you have
enumerated all the independent variables, or that the relationship is
actually representable by a set of linear relationships. So we have no
disagreements here.

The
problem is still the same problem whether the predictors are known precisely
or imprecisely. Whether you choose to call it a "mathematical" problem or a
"statistical" problem is unimportant. Furthermore, because the "true"
values of the variables being measured are almost never precisely known, the
problem of finding the mathematical function that best accounts for the
pattern of variation observed is a statistical problem by _your_ definition,
in just about any real example (in contrast to the example I cooked up).

I call it a "mathematical" problem when random (unexplainable) variation
doesn't enter into the variables or the parameters. As you say, in practical
problems there is always measurement noise; there is also what Hans calls
"process noise," unpredictable variations in system parameters. The basic
methods of mathematics such as solving simultaneous equations have to be
modified when there is noise, because the solutions cannot be exact any
more; this is where statistical methods come into play.

In any function y = f(x1..x4), none of the x's does a "particularly good
job" of predicting the value of y even in the complete absence of
uncertainty. ...

Of course. How does this undermine my conclusion that four predictors
apparently poorly correlated with the response variable may together account
fully for the variation in R?

If your point is that "correlation" is irrelevant when there is an exact
relationship, it doesn't. But if that was the point, your example is
misleading: it seems to say that even if the correlation of each independent
variable is only 0.5 with the dependent variable, the four independent
variables together can predict perfectly -- even if there is random noise.
The implication of speaking of correlation is that there is some random
process involved, but in this case there is none. Your example creates the
false impression that even if there are random fluctations in the _measures_
of the variables, it is possible that by combining measures of many
contributing variables, one can get results that are almost perfect. This,
of course, is completely false. The amount of improvement in relative terms
is quite modest, and the total variance in the response measure is always
larger than the variance in any contributing variable.

The researcher does not know the nature of
the true relationship between the predictors and R, but has discovered that
each predictor correlates poorly with R. A scatterplot reveals a lot of
apparently random scatter around the bivariate regression of R on P1, or P2,
or P3, or P4. My claim is that simply observing such low correlations is no
guarantee that some highly predictive relationship does not exist between
the predictors and R. I haven't heard you say anything that undermines this
claim (although making assertions of "red herings" and "spurious
conclusions" sure makes it _sound_ like you have, I must admit).

I can hardly disagree with you on this point! You are talking about a case
in which failing to recognize a regular relationship leads to the
_impression_ that there is a low correlation, when in fact there is a very
high one. This is what I call "using the wrong model."

The most pertinent example, of course, is that of S-R investigations of the
behavior of a control system, where the controlled variable is not taken
into account. The manipulated variables are actually disturbances, and the
responses are actually actions that cancel the effects of the disturbances
on the controlled variable. Not knowing that, however, the experimenter does
not use the right measure of the disturbances: he uses qd instead of Fd(qd);
neither does he use the right measure of the action: he uses qo instead of
Fo(qo). Furthermore, he does not realize that there are many qd's that could
affect the same controlled variable, so he does not control the environment
so as to eliminate random effects on qi -- he can't, because he has not
identified qi. And of course he will probably include variables that have no
effect on qi. The net result is a low correlation between qd (the
"stimulus") and qo (the "response"), singly or in bunches.

If qi had been identified reasonably well, the experimenter could have made
sure that nothing else but qd (aside from qo) could affect it, and the true
connections between qd and qi, and qo and qi, could be determined by
physical examination of the environment. This would provide an exact model
of the environment, and now the relationship between qd and qo would become
much more nearly exact, most of the variance disappearing. In fact, multiple
disturbances could be used, and multiple linear correlation analysis would
now show that there is a very high (negative) correlation in the relationship

Fo(qo) = -[Fd1(qd1) +Fd2(qd2)+ ... Fdn(Qdn)].

It's only through identifying controlled input variables that one can
distinguish between environmental variables that ought to relate to actions
and those that ought NOT to relate to them. The key is to understand what
the controlled variable is.
.....................

A low simple correlation between two variables does not preclude this
relationship from being scientifically useful.

Well, that's the point you hope to convince us of, but it depends entirely
on where you set your standard for scientific usefulness, and on what use
you intend to make of the results. If you sell widgets, a correlation of
0.1 between color and sales is useful to know about, provided that all
colors cost the same amount. If your future depends on the next customer
buying a widget, a correlation of 0.1 is pretty useless.

Be sure to let us know when you are able to make such point predictions
(about buyer behavior and so on) with the high accuracy obtained in tracking
studies; I can think of literally thousands of situations in which this
ability would be invaluable, and worth a considerable sum on the open >market.

That's probably impossible: some people buy widgets and some do not. Color
influences only the ones who buy widgets, and it doesn't influence them all
the same way. If you find a correlation of 0.1 between color and widget
buying, and adjust your manufacturing process accordingly, you're simply
catering to people who like a certain color and who slightly outnumber the
others who prefer a different color. You haven't found any tendency in ALL
people to prefer that magic color although, regrettably, that is how such
findings are often interpreted.

But seriously, I have already indicated a far more important "use" of
moderate correlations than in maximizing income from widget sales: model
construction and the testing of alternative models. That is what makes such
correlations scientifically useful, even though they may not be much good in
themselves for making point predictions.

I agree. If you apply such statistical analyses strictly to modeling one
individual at a time, this can be very productive. You must also, of course,
use something beside the statistics, because whether you get high
correlations or low ones will depend on whether you're starting with the
right basic model. Finding the right model can't be done with statistical
analysis alone; all the analysis can do is tell you how good any particular
model is. When you rotate the axes, you're finding the best coefficients,
but that doesn't tell you that you have the best model. Also, as we point
out frequently in PCT, your model must tell you when a _high_ correlation
should _not_ be found, and you must not find a high correlation there. The
obvious example is the correlation between a disturbance and a controlled
variable, or (to a lesser extent, because reference signals can vary)
between an action and a controlled variable.

The application of correlation analysis across a population, of course,
gives you a model of the population that may have nothing to do with the
characteristics of the individuals. That's what my article in Hershberger's
_Volitional Action_ was about. Each of the 4000 individuals (control
systems) showed a high negative slope in the relationship between "reward"
and "effort," while the population showed a modest positive slope. What the
population analysis shows about the relation between reward and effort is
the opposite of the reality for EVERY individual.

Best,

Bill P.

[CHUCK TUCKER 970428]

        [Hans Blom, 970424]

        (CHUCK TUCKER 970421)

Chuck, in your TABLE SHOWING THE RELATIONSHIP BETWEEN SEVERAL
DESCRIPTIVE STATISTICS, you give a simple "translation" between r,
r2, k2, k and E which essentially shows that these measures are all
equivalent. Is that right? If so, why are they all in use and not
just one of them?

        Actually, one of the major reasons for putting together this
        very simple "translation" is that the statistics k2, k, and
        E are rarely used in research reports. Rather, most of the
        reports I read say that there is a "significant" or "strong
        relationship" between "X" (single or multiple form) and "Y."
        In wondering what is meant by this attribution I (actually
        we) ask "How strong is it?" or "What do mean by sigificant?"
        and we get the answer that the R is statistically significant
        at the .001 level even though the R=.35. The implication of
        the table is that such such a correlation accounts for only
        12% of the variance, has an unexplained variance of 88%, a
        lack of linear association of .94, and knowing that R in a
        random sample from a population you would be able to "predict"
        "Y" from "X" accurately 6% of the time. Now if you wish to
        claim "sigificance" for the R=.35 go ahead but using k2, k,
        and E it does not appear to be very significant to me.

        I have not received Kennaway's paper yet so I don't know what
        his position is on "the utility of a low correlation" but in my
        experience they have never (so far as I know) stopped any
        sociologist from publishing the "research" report rather than
        only publishing a note saying "we found the X was so poorly
        related to Y that more research must be done."

        The most dramatic example of the publishing of low correlations
        can be found in Clark McPhail's article "Civil Disorder
        Participation." AMERICAN SOCIOLOGICAL REVIEW 36:1058-1072, 1971
        which (in his own words) says: "... secondary statistical analyses
        of strength of association and statistical significance for 173
        tests (from ten studies) of relationship between deprivation,
        frustration and/or aggression in the form of participation in
        five riots: 32% were not statistically significant; 61% yielded
        low correlations (between .00 and .29); 6% yielded moderate
        correlations (between .30 and .39); and 1% yielded a moderate
        correlation of higher than .40 (McPhail, 1991:58)." These
        results were also found for studies published on the LA Watts
        riot and the Detroit and Newark riots (a study published on
        "riot participation" in the most recent issue of ASR ignored
        all of this research and concludes there is a "significant
        relationship" between "multiple variables" and "participation"
        in the same riots. The R2 was .33 for the multiple "X" and "Y.")

        Regards, Chuck

[From Bill Powers (971220.0415 MST)]

Bruce Abbott, Martin Taylor (various) --

From time to time it's been mentioned that the use of correlations to

characterize the fit of a control model to real behavior overstates the
quality of the predictions. I would be interested in any correction methods
that anyone might offer that would give a more realistic evaluation.

The main criticism that's been offered, as I understand it, is that the
number of data points we take during an experimental run is misleadingly
high, because the disturbance change from one point to the next is not
unpredictable. Thus the high correlations are explainable in part by the
fact that the disturbances are predictable over some time period.

It's not clear to me how we could change the data analysis to avoid that
problem, if it is a problem. If we "binned" the data by averaging over
larger chunks of time, it seems to me that this would _improve_ the
correlations, because the primary errors of prediction that we see are at
the high-frequency end of the spectrum. That is, the model's prediction
errors are the largest just where we would expect them to be smallest: over
short periods of time during which the disturbances are the most predictable.

I'm probably not interpreting the criticisms correctly. The point here is
that if there is a better way of evaluating the model, I would like to know
about it and try it out to see if it changes the picture substantially. Can
anyone on CSGnet provide a formula that would be more rigorously acceptable
than the simple correlation formula?

Best,

Bill P.