MetRic SCALing : MRSCAL

The MRSCAL algorithm is a metric counterpart to MINISSA. Its aim is to position a set of stimulus objects as a set of points in a space of minimum dimensionality in much the same way as MINISSA, except that the distances in this space will be a linear (or optionally a logarithmic) function of the dissimilarities between the stimuli.

Data: 2-way, 1-mode dis/similarities    Transform: Linear or Logarithmic     Model:  Euclidean (and other Minkowski) Distance

Output from MRSCAL may be input to PINDIS.

MRSCAL accepts as input the lower triangle (without diagonal) or a full symmetric matrix of (dis)similarities. The program will also accept negative values such as product moment correlations and covariances.  

The user may provide a starting configuration by means of the command READ CONFIG, using an associated INPUT FORMAT specification if the input is not free format. A coordinate for each point on each dimension is input. This may be done either by stimuli (rows) by dimensions (columns) or dimensions (rows) by stimuli (columns). In this latter case MATFORM(1) should set in the PARAMETER command.

If this is not done, however, then the program constructs an initial configuration from the original data by the Lingoes-Roskam (1973) procedure which is a good initial approximation to a solution and has desirable geometrical properties.

The user may choose the way in which the distance between the points in the configuration is defined using the MINKOWSKI parameter. The default value 2 provides for the Euclidean metric, but the user may specify any non-negative value for the parameter. Commonly used values include 1, the so-called 'city-block' metric where the distance between the two points is the sum of the differences between their co-ordinates on the axes of the space, and infinity (in MRSCAL approximated by a large number(>25)), the so-called 'dominance' metric when the largest difference on any one axis will dominate all others. (Users are warned that high MINKOWSKI values are liable to produce program failure due to overflow).

Linear and logarithmic transformations
The most common use of MRSCAL is to find a linear transformation of the data which best fits a configuration of points in the chosen dimensionality. The program will also, however, perform an analysis using logarithmic transformations of the data values. In this case the Shepard diagram will show a smooth exponential curve. The user must specify which transformation is required. If no PARAMETERS command is read and/or no specification of the transformation given, then no analysis will be performed.

Dimensionality
As a general rule solutions should be computed in a number of dimensionalities. Since a perfect fit will be obtained in n-2 dimensions the trial dimensionalities should always be in dimensionalities less than n-3. As a practical guide to the choice of trial dimensionalities it is recommended that the data compression ratio (defined by Young as the  product of stimuli x dimensions divided by the number of input data elements) should be greater than 2.

A further guide involves examining the plot of stress by dimensionality. Since MU is a measure of goodness of fit the plot will show an ascending function. The 'appropriate' dimensionality, is that at which the graph shows an 'elbow', i.e. where the addition of extra dimensions is otiose.

PARAMETERS
Keyword             Default Value            Function
DATA TYPE                        0       0: Lower-triangle matrix of similarities
                                                (high values mean high similarities
                                                 between points).
                                                 1: Lower-triangle matrix of dissimilarities
                                                (high values mean high dissimilarities
                                                 between points).
                                                2: Full-symmetric matrix of similarities
                                                (high values mean high similarities
                                                between points).
                                                3: Full-symmetric matrix of dissimilarities
                                                (high values mean high dissimilarities
                                                between points).
LINEAR TRANSFORMATION  0     0: Linear transformation is not
                                                   performed.
                                               1: Linear transformation is performed.
LOG TRANSFORMATION      0      0: Logarithmic transformation is not
                                                   performed.
                                               1: Logarithmic transformation is performed.
CRITERION                  0.00001 Sets the criterion value for terminating
                                                   the iterations.
MINKOWSKI                      2      Sets the Minkowski metric for
                                                   the analysis.
MATFORM                          0      (RELEVANT ONLY WHEN 'READ CONFIG'
                                                   IS USED)
                                               0: The input configuration is entered:
                                                   stimuli (rows) by dimensions
                                                   (columns)
                                               1: The input configuration is entered:
                                                   dimensions (rows) by stimuli
                                                   (columns)

N.B. Either LINEAR TRANSFORMATION or LOG TRANSFORMATION must be specified

NOTES
1. N OF SUBJECTS is not valid with MRSCAL.
2. N OF STIMULI may be replaced by N OF POINTS
3. a) The program expects input to be in the form of the lower
> triangle of a matrix of real (F-type) numbers.
b) The INPUT FORMAT, if specified, should read the longest,
i.e. last, row of this matrix.
4. LABELS followed by a series of labels (<= 65 char), each on a
separate line, optionally identifies the stimuli, in order and without
omissions.

PRINT options (to main output file)
Option                Form                   Description
INITIAL         p x r matrix        Initial configuration, either generated
                                             by the program or provided by the user
                                             (p = no. of stimuli, r = no. of dimensions)
FINAL            p x r matrix        Final configuration, rotated to
                                              principal components.
DISTANCES   triangle     Solution distances between points
                     with diagonal      calculated according to MINKOWSKI
                                              parameter
FITTING         Lower triangle     Fitting values : the disparities
                     with diagonal      (DHAT) values.
RESIDUALS    Lower triangle     The difference between the distances
                     with diagonal      and the disparities.

By default, only the final configuration and the final STRESS values are output.

PLOT options (to main output file)
Option                      Description
INITIAL             Up to r(r-1)/2 plots of the initial
                        configuration (r=no. of dimensions)
FINAL                Up to r(r-1)/2 plots of the final
                        configuration.
SHEPARD          The Shepard diagram of distances
                        plotted against data. Fitted values
                        are shown by *, actual data/distance
                        pairs by 0.
STRESS            Plot of STRESS values by iteration,
                        with a summary plot of stress
                        by dimensions.
POINT               Histogram of point contributions to
                        STRESS.
RESIDUALS       Histogram of residual values (logged).

By default, only the Shepard diagram and the final configuration are plotted.

PROGRAM LIMITS
Maximum no. stimuli = 300
Maximum dimensions = 8

See also

  • The NewMDSX commands in full