Chapter Contents
Chapter Contents
Previous
Previous
Next
Next
The MDS Procedure

Example 40.1: Jacobowitz Body Parts Data from Children and Adults

Jacobowitz (1975) collected conditional rank-order data regarding perceived similarity of parts of the body from children of ages 6, 8, and 10 years and from college sophomores. The following analysis includes data from 15 children (6-year-olds) and 15 sophomores. The method of data collection and some results of an analysis are also described by Young (1987, pp. 4-10). A portion of the data is included below. See "Mds Documentation Examples" in the SAS/STAT Sample Program Library for the complete data set.

   data body;
      title  'Jacobowitz Body Parts Data from 6 Yr Olds and 
              Adults';
      title2 'First 15 subjects (obs 1-225) are children';
      title3 'Second 15 subjects (obs 226-450) are adults';
      input  (cheek face mouth head ear body arm elbow hand
          palm finger leg knee foot toe) (2.);
      if _n_ <= 225 then subject='C'; else subject='A';
      datalines;
    0 2 1 3 410 5 9 6 7 811121314
    2 012 113 3 81011 9 7 4 5 614
    3 2 0 1 4 9 511 6 7 810131214
    2 1 3 0 4 9 5 611 7 810121314
   10 111 2 0 6 3 4 51213 7 814 9
   1412 9 613 0 8 7 51011 1 4 2 3
   1214111013 5 0 4 1 3 2 6 9 7 8
    5 714 8 6 9 1 0 2 3 410111213
   1311121014 9 3 4 0 1 2 6 5 7 8
    8 6 7 9 4 5 310 1 0 212111314
   14 513 6 912 3 4 1 2 0 7 81011
   14121311 9 7 4 6 5 310 0 8 1 2
   1211141013 4 5 8 6 7 9 1 0 2 3
   1214101311 9 4 5 8 6 7 2 3 0 1
   13 8 91114 3 6 5 71012 2 4 1 0
    0 4 2 311 91412 1 713 8 6 510
    7 011 9 1 2 8 3131410 612 4 5
   1011 0 3 7 1 813 212 6 914 5 4
    4 811 0 1 2 3 9 71310 51412 6
   14 61110 0 1 2 4 9 8 5 713 312
   11 61412 3 013 2 1 9 5 410 7 8
   1412 8 3 1 7 013 2 6 5 911 410
    610 914 3 412 0 813 5 711 1 2
   131412 1 4 6 2 7 010 9 311 8 5
    6 9 711 3 5 112 2 013 810 414
   10 6 9 81113 712 2 1 014 5 3 4
    6 9 710 4 21214 1 813 0 3 511
   13 2 8 3 1 5 9141211 7 6 010 4
   1411 8 413 212 9 5 310 6 7 0 1
   14 3 4 713 6 2 8 91110 112 5 0

      ... 405 lines omitted ...

    0 1 2 4 3141210 6 5 81311 7 9
    2 0 3 1 414 810 5 612 911 713
    2 1 0 4 31412 9 511 71310 6 8
    2 1 4 0 314 810 61112 7 9 513
    1 3 2 4 01412 9 8 7 5131011 6
   131011 114 0 3 7 512 8 2 6 4 9
   13 911101214 0 1 3 6 5 2 4 7 8
   1012 813 914 1 0 611 4 3 2 5 7
   12 911101314 5 4 0 1 3 8 6 2 7
    91211131014 5 4 1 0 2 6 7 3 8
   10121113 914 6 5 1 3 0 8 7 4 2
   121011 91314 2 5 6 8 7 0 1 3 4
   1112 9131014 3 2 7 8 4 1 0 5 6
   121011 91314 5 8 1 6 7 3 4 0 2
   10121113 914 8 7 4 6 2 3 5 1 0
   ;

The data are analyzed as row conditional (CONDITION=ROW) at the ordinal level of measurement (LEVEL=ORDINAL) using the weighted Euclidean model (COEF=DIAGONAL) in three dimensions (DIMENSION=3). The final estimates are displayed (PFINAL). The estimates (OUT=OUT) and fitted values (OUTRES=RES) are saved in output data sets. The following statements produce Output 40.1.1:

   options ps=60;
   proc mds data=body
      condition=row
      level=ordinal
      coef=diagonal
      dimension=3
      pfinal
      out=out
      outres=res
      ;
   subject subject;
   title5 'Nonmetric Weighted MDS';
   run;

Output 40.1.1: Iteration History and Final Estimates for Body Parts Data

Jacobowitz Body Parts Data from 6 Yr Olds and Adults
First 15 subjects (obs 1-225) are children
Second 15 subjects (obs 226-450) are adults
Nonmetric Weighted MDS

Multidimensional Scaling: Data=WORK.BODY
Shape=SQUARE Condition=ROW Level=ORDINAL
Coef=DIAGONAL Dimension=3 Formula=1 Fit=1

Mconverge=0.01 Gconverge=0.01 Maxiter=100 Over=2 Ridge=0.0001 Alternate=MATRIX

Iteration Type Badness-
of-Fit
Criterion
Change in
Criterion
Convergence Measures
Monotone Gradient
0 Initial 0.4091 . . .
1 Monotone 0.2053 0.2038 0.3012 0.3190
2 Gau-New 0.1937 0.0116 . .
3 Monotone 0.1862 0.007533 0.0410 0.2314
4 Gau-New 0.1847 0.001496 . .
5 Monotone 0.1779 0.006754 0.0372 0.1516
6 Gau-New 0.1773 0.000654 . .
7 Monotone 0.1758 0.001488 0.0187 0.0975
8 Gau-New 0.1755 0.000253 . .
9 Monotone 0.1751 0.000478 0.0104 0.0767
10 Gau-New 0.1750 0.000113 . .
11 Monotone 0.1748 0.000199 0.006467 0.0582
12 Gau-New 0.1747 0.0000592 . 0.0384
13 Gau-New 0.1747 0.0000184 . 0.009952

Convergence criteria are satisfied.

Configuration
  Dim1 Dim2 Dim3
cheek 1.46 0.77 -0.30
face 1.40 -0.56 0.43
mouth 1.25 -0.97 -0.51
head 1.31 0.39 0.97
ear 1.23 0.03 -1.39
body 0.24 -0.18 2.24
arm -0.70 1.23 1.08
elbow -0.73 0.40 0.29
hand -0.71 1.50 -0.26
palm -0.24 1.59 -0.96
finger -0.74 0.57 -1.35
leg -0.69 -1.44 0.88
knee -1.44 -0.64 0.42
foot -0.78 -1.51 -0.36
toe -0.87 -1.20 -1.19

Jacobowitz Body Parts Data from 6 Yr Olds and Adults
First 15 subjects (obs 1-225) are children
Second 15 subjects (obs 226-450) are adults
Nonmetric Weighted MDS

Multidimensional Scaling: Data=WORK.BODY
Shape=SQUARE Condition=ROW Level=ORDINAL
Coef=DIAGONAL Dimension=3 Formula=1 Fit=1

Dimension Coefficients
subject 1 2 3
C 1.18 1.02 0.76
C 0.94 1.05 1.01
C 0.94 1.01 1.05
C 1.08 1.10 0.79
C 0.98 1.05 0.97
C 1.28 0.85 0.79
C 0.90 1.03 1.06
C 0.95 1.04 1.01
C 1.18 1.06 0.69
C 0.95 0.99 1.05
C 1.07 1.13 0.76
C 0.98 0.94 1.08
C 1.34 0.82 0.73
C 0.95 0.99 1.05
C 1.03 1.09 0.86
A 1.30 0.82 0.80
A 1.05 0.64 1.22
A 1.28 0.85 0.81
A 0.95 0.88 1.15
A 1.34 0.75 0.80
A 1.27 0.81 0.85
A 1.50 0.37 0.78
A 1.38 0.82 0.64
A 1.44 0.56 0.78
A 1.53 0.73 0.36
A 1.18 0.95 0.84
A 0.85 0.86 1.24
A 1.29 0.62 0.97
A 1.20 0.58 1.10
A 1.25 0.61 1.04

Jacobowitz Body Parts Data from 6 Yr Olds and Adults
First 15 subjects (obs 1-225) are children
Second 15 subjects (obs 226-450) are adults
Nonmetric Weighted MDS

Multidimensional Scaling: Data=WORK.BODY
Shape=SQUARE Condition=ROW Level=ORDINAL
Coef=DIAGONAL Dimension=3 Formula=1 Fit=1

subject Number of Nonmissing
Data
Weight Badness-of-Fit Criterion Distance Correlation Uncorrected Distance
Correlation
C 210 0.03 0.16 0.85 0.99
C 210 0.03 0.25 0.51 0.97
C 210 0.03 0.23 0.58 0.97
C 210 0.03 0.16 0.85 0.99
C 210 0.03 0.21 0.69 0.98
C 210 0.03 0.15 0.89 0.99
C 210 0.03 0.26 0.40 0.96
C 210 0.03 0.25 0.47 0.97
C 210 0.03 0.13 0.91 0.99
C 210 0.03 0.24 0.55 0.97
C 210 0.03 0.15 0.87 0.99
C 210 0.03 0.23 0.59 0.97
C 210 0.03 0.15 0.90 0.99
C 210 0.03 0.21 0.69 0.98
C 210 0.03 0.19 0.75 0.98
A 210 0.03 0.12 0.94 0.99
A 210 0.03 0.17 0.84 0.98
A 210 0.03 0.12 0.93 0.99
A 210 0.03 0.21 0.68 0.98
A 210 0.03 0.14 0.91 0.99
A 210 0.03 0.16 0.87 0.99
A 210 0.03 0.09 0.98 1.00
A 210 0.03 0.11 0.95 0.99
A 210 0.03 0.11 0.96 0.99
A 210 0.03 0.13 0.95 0.99
A 210 0.03 0.13 0.90 0.99
A 210 0.03 0.21 0.72 0.98
A 210 0.03 0.12 0.93 0.99
A 210 0.03 0.10 0.95 0.99
A 210 0.03 0.11 0.94 0.99
- All - 6300 1.00 0.17 0.84 0.98


The OUTRES= data set is used to produce a plot showing the overall fit of the model, with the transformed data on the vertical axis and the distances from the model on the horizontal axis. If the model fits perfectly, all points lie on a diagonal line from lower left to upper right. The vertical departure of each point from this diagonal line represents the residual of the corresponding observation. The HAXIS and VAXIS options in the PLOT statement specify that the horizontal and vertical axes use the definitions AXIS2 and AXIS1, respectively. The identical ORDER= options in each of these AXIS statements make physical distances on each axis comparable.

The following statements produce Output 40.1.2:

   title1 'Plot of Over-All Fit';
   axis1 label=(angle=90 rotate=0) minor=none 
         order=(0 to 5 by 1);
   axis2 minor=none  order=(0 to 5 by 1);

   proc gplot data=res;
      plot fitdata*fitdist/vaxis=axis1 haxis=axis2
                           frame cframe=ligr;
   run;

Output 40.1.2: Plot of Over-All Fit for Body Parts Data
mdse1d.gif (8646 bytes)

The OUT= data set is used to plot the configuration and dimension coefficients using the %PLOTIT macro. Again it is necessary to use the VTOH= option to make the axes commensurable. The configuration is plotted by selecting observations having _TYPE_=`CONFIG' with a WHERE statement and by using the _NAME_ variable to identify each body part on the plot.

The dimension coefficients are plotted by selecting observations having _TYPE_=`DIAGCOEF' and by using the Subject variable to distinguish children from adults on the plot. The following statements produce Output 40.1.3:

   title1 'Plot of configuration';
   %plotit(data=out(where=(_type_='CONFIG')), datatype=mds,
           labelvar=_name_, vtoh=1.75);
   %plotit(data=out(where=(_type_='CONFIG')), datatype=mds,
           plotvars=dim3 dim1, labelvar=_name_, vtoh=1.75);
   run;

   title1 'Plot of Dimension Coefficients for Each Subject';
   %plotit(data=out(where=(_type_='DIAGCOEF')), symtype=vector,
           symbols='',datatype=mds, place=0, labelvar=subject, 
           vechead=, ls=100, plotopts=hzero vzero);

   %plotit(data=out(where=(_type_='DIAGCOEF')),
           plotvars=dim3 dim1, symtype=vector, symbols='',
           datatype=mds, place=0, labelvar=subject, ls=100,
           vechead=, plotopts=hzero vzero);
   run;

The configuration displayed in Output 40.1.3 has a tripodal shape with Body at the apex. The three legs of the tripod can be distinguished in the plot of dimension 2 by dimension 1, which shows three distinct clusters with Body in the center. Dimension 1 separates head parts from arm and leg parts. Dimension 2 separates arm parts from leg parts. The plot of dimension 3 by dimension 1 shows the tripod from the side. Dimension 3 distinguishes the more inclusive body parts (at the top) from the less inclusive body parts (at the bottom).

Output 40.1.3: Plot of Configuration for Body Parts Data
mdse1e.gif (3719 bytes)

mdse1f.gif (3408 bytes)

The plots of dimension coefficients in Output 40.1.4 show that children differ from adults primarily in the emphasis given to dimension 2. Children give about the same weight (approximately 1) to each dimension. Adults are much more variable than children, but all have coefficients less than 1.0 for dimension 2, with an average of

about 0.7. Referring back to the configuration plot, you can see that adults consider arm parts to be more similar to leg parts than do children. Many adults also give a high weight to dimension 1, indicating that they consider head parts to be more dissimilar from arm and leg parts than do children. Dimension 3 shows considerable variability for both children and adults.

Output 40.1.4: Plot of Dimension Coefficients for Body Parts Data
mdse1g.gif (6499 bytes)

mdse1h.gif (7032 bytes)

Chapter Contents
Chapter Contents
Previous
Previous
Next
Next
Top
Top

Copyright © 1999 by SAS Institute Inc., Cary, NC, USA. All rights reserved.