Minitab THREECV1 Analysis

[From Bruce Abbott (950116.1730 EST)]

The THREECV1 program has been revised to save the reference positions (1st
line) and the 3600 lines of data, which now consist of the three cursor
positions and the mouse ("handle") position. What appears below is an
analysis of the first 1800 data observations using Minitab. C1, C2, and C3
are the three cursor positions; H is the handle position.

UNIVARIATE DESCRIPTIVE STATS

                N MEAN MEDIAN TRMEAN STDEV SEMEAN
C1 1800 310.63 309.00 312.07 70.04 1.65
C2 1800 345.58 328.00 344.47 63.40 1.49
C3 1800 319.31 320.00 319.35 5.28 0.12
H 1800 -9.46 -22.00 -11.54 60.18 1.42

Note the means and medians of the three cursor positions. The reference
position was 319. Which cursor do you suppose I was controlling? (Note also
the standard deviations: the controlled cursor would be the one having the
smallest SD.)

BASIC CORRELATIONS

             C1 C2 C3
C2 0.105
C3 0.177 0.039
H 0.527 0.783 0.035

The low correlations in the positions of the three cursors emerged despite the
common influence of the mouse on all three. This is due to the fact that each
cursor was being affected by an independent disturbance. The moderately high
correlations between H and C1 and between H and C2 would be 1.0 were it not
for the disturbances, which weaken the relationship. The nearly zero
correlation between H and C3 emerges because H is opposing the disturbance on
C3. Perfect opposition would keep C3 constant, yielding zero correlation with
H.

SCATTERPLOTS

I was surprised at the usuaual look of many of the scatterplots. In normal
scatterplots the points tend to form a more-or-less oval shape, and each
univariate distribution would look more-or-less normal. It is obvious from
these plots that the points are not independent, random values from a
bivariate population distribution. Such plots should set off warning bells
for anyone performing linear regression or correlation analyses of these
relationships. The "shapes" traced out by the points were so unusual that I
gave each plot a name to suggest its appearance:

     Correlation of C1 and H = 0.527: "THE PITCH FORK"
H - 3+
         - *8+5++6
         - *953+*+
       80+ 6+373**
         - *469++4
         - *+2*34
         - 3+2 5++64*
         - 2+38883++++ *2+5+9*
        0+ 5**2 +++ ++++6
         - +3+4 6 +++ *59++++++
         - ++++ *56465696432++5 568377+++4
         - 488++++425942 788+7 28+
         - 686 8992
      -80+ +32++4244++++
         - +++8+4+5+

···

-
         -
           --------+---------+---------+---------+---------+--------C1
                 200 250 300 350 400

     Correlation of C2 and H = 0.783: "THE X-CHROMOSOME"
H - 3+9
         - 3669++67+++
         - +**38+* *4922
       80+ 8**2 54542
         - +++3 23337
         - **76+6**
         - 2++8** +
         - **2+22+93* *+6++38
        0+ *847+4+6** 456* 22*448+69+6
         - +++ *3+697+98+2 334924+*
         - 795 7++++9 23++4+++32
         - 4+5+*6++5+73969
         - +47+*
      -80+ *+++++
         - ++++++
         -
         -
           +---------+---------+---------+---------+---------+------C2
         240 280 320 360 400 440

     Correlation of C3 and H = 0.035: "LOWER MICHIGAN"
H - 2* 29* *2 24 4
         - 49 ++5 95 63 *57 69 43 *35 6*
         - *4 444 35 22 325 * *4 9*3
       80+ * 4*2 ** * * 4+ 5 *
         - 342 57 33 *33 * * 333 98 6
         - 4 3** 35 23 2* ** 2 **2 *
         - *4 626 55 2* 22* 22 26 222 +5 5
         - * 23 22 22 ++8 ++ ++ ++9 73 23 7
        0+ 23 +8 622 * 5+ +++ ++ +5
         - * 65 64 +2* +7 ++ +++ 6
         - * 33 +++ ++ ++ +++ ++ 7
         - * 6+ +++ 76 56 7+6 2* 33 236 37 35 *
         - 42 22 *3 4** +5 *4 24 *
      -80+ 5*3 *3 44 2+* 3+ ++ 643 59 59 +3*
         - 3 422 4* 42 6+5 ++ ++ 85* 97 +7
         -
         -
           ----+---------+---------+---------+---------+---------+--C3
           301.0 308.0 315.0 322.0 329.0 336.0

Deviations computed as D[j,i] = C[j,i] - H[i]:

     Correlation of D1 and H = -0.364: "THE GOLF CART"
H - +
         - ++++
         - + +
       80+ +3 *+
         - + ++
         - + 8
         - 2+5 ++8
         - ++8+ 8++97 +4++
        0+ 72 2++++2 ++
         - 3+ 6 +++95 4+++++++
         - +++8 286656565+++7 2776777++++
         - 277777+989796 788877 5+
         - *9+ +8
      -80+ + 6+5++9++
         - ++++++8
         -
         -
           ------+---------+---------+---------+---------+---------+D1
               200 250 300 350 400 450

     Correlation of H and D2 = -0.258: "THE FOX"
H - +
         - 9++9+976+9++
         - 8++5 +
       80+ + 3+
         - +++95 7+
         - 35 *+78
         - 2++8 +*
         - 37452 7+3 8+++
        0+ ++986654 357* 2346669++6
         - ++998 8+9658868566* 33 7+
         - *667* 49+9++*++++ 5+++9+++
         - ++ 9++768+6786
         - ++
      -80+ +++
         - +++6+
         -
         -
           --------+---------+---------+---------+---------+--------D2
                 300 330 360 390 420

     Correlation of D3 and H = -0.996: "THE SWORD"
H - +
         - +++878
         - 899++82
       80+ 47876
         - 689+++
         - 6+556*
         - 4++++9+
         - * 6+++4
        0+ 7+++++*
         - 5+++++
         - ++++++
         - ++++++4
         - 88++9
      -80+ 2++++++
         - 4++++
         -
         -
           --------+---------+---------+---------+---------+--------D3
                 200 250 300 350 400

The above, of course, is the only legitimate relationship on which to apply a
regression analysis. It shows that the handle position was a strong, inverse
linear function of the disturbance value.

REGRESSING HANDLE POSITION ON DISTURBANCE 3

This analysis is based on the relationship depected immediately above.

The regression equation is
H = 318 - 0.995 D3

Predictor Coef Stdev t-ratio p
Constant 317.783 0.690 460.57 0.000
D3 -0.995344 0.002064 -482.16 0.000

s = 5.274 R-sq = 99.2% R-sq(adj) = 99.2%

The analysis estimates the reference at 317.783, with handle position being
equal to -0.995 times the disturbance. The actual values are 319 and -1.0.
The reference estimate is a little worse than our earlier one based on the
mean of C3 (319.31), reflecting the effect of "regression toward the mean."
Thus the mean of the C3 distribution provides a superior estimate of the
reference value than does the intercept of the regression equation.

RESIDUAL PLOT (RESIDUALS AS A FUNCTION OF TIME-ORDER)

The "moving average chart" below actually displays the averages of consecutive
blocks of 40 points over the 1800 point observation period. The residuals are
what's left over after the "influence" of the disturbance is removed from the
handle position, and represent the estimated control error. When plotted
against time-order of observation, the residuals show systematic variation.
These probably represent mostly dynamic effects such as time lags,
undershoots, and overshoots.

                 Residual Plot
         -
M -
o - +
v - +
i 7.5+
n - +
g - + + + +
         - + + ++ +
A -----------+-+----+-----+-----------------+-+-UCL=1.610
v 0.0+-----------+--+-+-+++-------+----+---++---+--X=0.0
e -------------------------+-------+------------LCL=-1.610
r - + + + + + +
a - + + +
g -++ + +
e -7.5+
         -
         - +
         -
         +---------+---------+---------+---------+---------+
         0 10 20 30 40 50
                          Sample Number

By the way, the "UCL" and "LC" boundaries are the "upper and lower control
limits." This chart is produced by Minitab as part of its suite of process
control charts. A series of consecutive points falling outside the control
limits is supposed to evoke concern that the process being monitored is "out
of control," meaning that the quantity being monitored is falling outside of
specs. Judging from the chart, I was "out of control" quite often....

NORMAL QUANTILE PLOT OF RESIDUALS (ASSESSMENT OF NORMALITY)

Despite the systematic variation of residuals over time, the normal quantile
plot below shows that the distribution of the residuals is well-represented by
a normal curve (indicated by a straight-line relationship below):
C7 -
         - *
         - 4
      2.5+ 574 *
         - +4+2+2
         - ++++7
         - ++++
         - +++++
      0.0+ ++++++
         - ++++++
         - +++++
         - 6+++++
         - +9++5
     -2.5+ 4+3
         - 4
         - *
         -
           +---------+---------+---------+---------+---------+------Resids
       -18.0 -12.0 -6.0 0.0 6.0 12.0

I did quite a bit more "playing around" with the data, but these were some of
the results I found most interesting.

The PCT analysis on these first 1800 observations yielded the same cursor-
handle correlations as reported by Minitab. The model fit gave k = 0.0991
with an RMS error of 4.0 pixels and a correlation between model and data of
0.9978.

Regards,

Bruce

[From Rick Marken (950116.2045 PST)]

Bruce Abbott (950116.1730 EST) --

Nice analyses Bruce. But it seems like you forgot the most basic
analysis of all; the one where you find the variable that is the
best predictor of the subject's response variations (H). There
are only three possibilities because there are only three variables
that the subject can actually see: C1, C2 and C3. Why not use the
multiple regression capabilities of Minitab to find the relative
contribution of each of these variables to the variance in H?

Best

Rick

Tom Bourbon [950119.1445]

[From Bruce Abbott (950116.1730 EST)]

The THREECV1 program has been revised to save the reference positions (1st
line) and the 3600 lines of data, which now consist of the three cursor
positions and the mouse ("handle") position. What appears below is an
analysis of the first 1800 data observations using Minitab. C1, C2, and C3
are the three cursor positions; H is the handle position.

Bruce, it's good to see someone else pick up some of the programming load.
And, it's good to see someone "new" using all of the tried and true
statistical procedures and measures on some control data. I remember doing
that several years ago. I thought I had pretty much accepted the ideas in
PCT (then it was called CST), but for the longest time I just couldn't admit
that there weren't some stimulus-like events or relationships lurking in the
data. I tried every sort of correlational analysis and regression analysis
I could drag out of the SPSS and BMD stat packages. Eventually, I gave it
up. It's good for someone to do that drill from time to time.

UNIVARIATE DESCRIPTIVE STATS

               N MEAN MEDIAN TRMEAN STDEV SEMEAN
C1 1800 310.63 309.00 312.07 70.04 1.65
C2 1800 345.58 328.00 344.47 63.40 1.49
C3 1800 319.31 320.00 319.35 5.28 0.12
H 1800 -9.46 -22.00 -11.54 60.18 1.42

Note the means and medians of the three cursor positions. The reference
position was 319. Which cursor do you suppose I was controlling? (Note also
the standard deviations: the controlled cursor would be the one having the
smallest SD.)

Have you calculated the variances of the three disturbances and compared
them with the variances of the three cursors? It's neat to see that when
the person controls the position of one cursor, reducing the variance of its
position, the variances of the uncontrolled cursors are usually greater
than they would have been had the person not controlled the other cursor.
Somewhere I wrote about how the variance that would have appeared in the
controlled cursor vanishes, only to show up as increased variances of the
handle and of uncontrolled variables that are affected only incidentally.
. . .

You had some pretty nifty names for the various types of scatter plots.

    Correlation of D3 and H = -0.996: "THE SWORD"
H - +
        - +++878
        - 899++82
      80+ 47876
        - 689+++
        - 6+556*
        - 4++++9+
        - * 6+++4
       0+ 7+++++*
        - 5+++++
        - ++++++
        - ++++++4
        - 88++9
     -80+ 2++++++
        - 4++++
        -
        -
          --------+---------+---------+---------+---------+--------D3
                200 250 300 350 400

The above, of course, is the only legitimate relationship on which to apply a
regression analysis. It shows that the handle position was a strong, inverse
linear function of the disturbance value.

What do you think your traditional colleagues would make of this one, and
its accompanying regression analysis shown below?

···

REGRESSING HANDLE POSITION ON DISTURBANCE 3

This analysis is based on the relationship depected immediately above.

The regression equation is
H = 318 - 0.995 D3

Predictor Coef Stdev t-ratio p
Constant 317.783 0.690 460.57 0.000
D3 -0.995344 0.002064 -482.16 0.000

s = 5.274 R-sq = 99.2% R-sq(adj) = 99.2%

============

Subject: THREECV1 Data and Minitab

[From Bruce Abbott (950117.0930 EST)]

Rick Marken (950116.2045 PST) replied to Bruce Abbott (950116.1730 EST) --

Nice analyses Bruce. But it seems like you forgot the most basic
analysis of all; the one where you find the variable that is the
best predictor of the subject's response variations (H). There
are only three possibilities because there are only three variables
that the subject can actually see: C1, C2 and C3. Why not use the
multiple regression capabilities of Minitab to find the relative
contribution of each of these variables to the variance in H?

Why, do you think it will tell us something about how subjects control? (;->

Bruce, I think it might show us something a majority of psychologists would
mistakenly identify as "causal relationships" that run from C to H -- from
S to R.
. . .

Another thing I would like to have done is to generate some
three-dimensional plots of these data (response surface), as well as
scatterplot matrices. Minitab has a new Windows-based version out that
apparently can do these, but I haven't seen it yet.

I'm curious. Do you think those fancy plots would reveal something that
isn't seen in the original plots of mouse, target and cursor positions? Or
in the simple dsecriptive statistics? I'm just asking, not judging.

Do the PCT-analysis programs you have written estimate the reference value
from the data? When the data set contains only cursor and handle positions,
(i.e., lacks the disturbance tables), the disturbance values can be
calculated exactly only if the reference values are known (In THREECV1 all
three = 319).

That sounds a little strange. Momentary cursor position is determined
entirely by handle position plus disturbance: c = h + d. The reference
signal does not enter into that calculation - the equation describes how
the *world* works, not how the control system works. I wonder what's
going on in the program.

Later,

Tom