Permap 11.8 Tutorial

MDS Analysis Using Permap 11.
Dr. Ronald B. Heady, University of Louisiana at Lafayette (retired)

Dr. Jennifer L. Lucas, Agnes Scott College
January 24, 2010
________________________________________________________________________
TABLE OF CONTENTS
1.0 OVERVIEW
1.1 Permap
1.2 MDS Maps
2.0 GETTING STARTED

2.1 System Requirements
2.2 Installation
2.3 Data Input
2.4 Data Input, MDSX
3.0 MAIN SCREEN & PROGRAM CONTROL

3.1 Header and Objective Function Value
3.2 On-line Shortcut Controls
3.3 Parking Active Objects Control
3.4 Parking Active Attribute Control
3.5 Execution Control
3.5.1 Start, Stop, Continue
3.5.2 Jiggle
3.5.3 Field Movements (Mirror, Rotate, Move, Zoom)
3.5.4 Automatic Controls
3.6 Click & Keystroke Commands
3.7 Locking Object Positions
4.0 FILE, EDIT, VIEW

4.1 File Menu
4.1.1 Load Data From a Text File
4.1.2 Load Data From a Special Case
4.1.3 Enter New Data
4.1.4 Save Results & Settings
4.1.5 Reload Original Default Settings
4.1.6 Print Results
4.2 Edit Menu
Page 1 of 77
4.2.1 Edit Active Data File
4.2.2 Edit Help File
4.2.3 Edit Solution Files
4.3 View Menu
4.3.1 Object Representation
4.3.2 Object Identification
4.3.3 Unit Circle and X-Y Axes
5.0 ANALYSIS PARAMETERS

5.1 MDS Types and Error Bounds
5.1.1 Metric and Nonmetric Terminology
5.1.2 MDS and Error Bounds
5.1.3 Procedural Details For Each Analysis Type
5.1.3.1 Ratios
5.1.3.2 Ratio + Bounds
5.1.3.3 Intervals
5.1.3.4 Interval + Bounds
5.1.3.5 Ordinal
5.1.4 Object Centering
5.2 Objective Functions
5.2.1 Objective Function Form
5.2.2 Badness Measures
5.2.3 Distance Measures
5.2.4 Normalization Constants
5.3 Mapping Weights
5.3.1 Replacing Wij with Wij MWij
5.4 Dimensionality and Boundaries
5.4.1 Dimensionality
5.4.2 Boundaries
5.5 Dij Precision Degradation
6.0 ATTRIBUTES-TO-DISSIMILARITIES
6.1 Attribute Data Types
6.1.1 Changing Data Types
6.1.2 Examples of Down-Conversions
6.1.3 Rational for Using Down-Conversions
6.2 Attribute Functions
6.3 Using Attribute and Dissimilarity Data
6.4 Euclidean and City Block Attribute Normalization
7.0 MAP EVALUATION

7.1 Objects Evaluation Screen
7.1.1 Pareto Plot
7.1.2 On-Map Diagnostics
7.1.2.1 Objective Function Contributions
Page 2 of 77
7.1.2.2 Out-Of-Plane Tendencies
7.2 Links Evaluation Screen
7.2.1 Shepard Plot
7.2.2.1 Waern Links
7.2.2.2 Stretched Links
7.2.2.3 Compressed Links
7.3 Attributes Evaluation Screen
7.3.1 Attribute Fit Plot
7.3.2.1 Attribute Gradient Vector
7.3.2.2 Vector and Quartiles
7.3.2.3 Vector and Values
7.3.2.4 All Active Vectors
7.3.2.5 All Available Vectors
7.3.2.6 Vector Tie Lines
7.4 General Comments Concerning Map Evaluation
8.0 CONVERGENCE
8.1 Convergence Rate
8.2 Convergence Precision
9.0 SPECIAL CASES
10.0 SOLUTION TECHNIQUE
11.0 SOLUTION DIFFICULTIES
12.0 Permap HISTORY
13.0 GETTING HELP
14.0 LEGAL NOTICE
APPENDIX I - DATA FILE KEYWORDS
APPENDIX II - EXAMPLE INPUT DATA FILE
Page 3 of 77
1.0 OVERVIEW
1.1 Permap
The fundamental purpose of Permap is to uncover hidden structure that might be residing
in a complex data set. Compared to other data mining and data analysis techniques MDS is
growing increasingly popular because its mathematical basis is easier to understand and its
results are easier to interpret (Fitzgerald & Hubert, 1987).
Permap is an interactive computer program. It offers both metric (ratio and interval) and
nonmetric (ordinal, ratio + bounds, interval + bounds) MDS techniques. It solves problems in up
to eight dimensional space and allows boundary conditions to be imposed on the solution. In the
technical jargon, Permap treats "weighted, incomplete, one-mode, two-way" or "weighted,
incomplete, two-mode, two-way" data sets. Other jargon would say it handles weighted,
symmetric, incomplete, triangular or rectangular data sets. The word “weighted” means each
data point can have its own multiplier that reflects in some way the importance or reliability of
the point. The word “symmetric” means that Permap assumes that the (i, j) proximity value
equals the (j, i) proximity value, and “incomplete” means that it can handle missing data. The
one-mode, two-way and square references indicate that Permap can analyze a matrix of
proximity information between several objects, and the two-mode, two-way and rectangular
references means it can analyze objects each of which are specified by an array of attributes.
Permap can treat up to 1000 objects at a time (but see cautions in Section 11) and each
object can have up to 100 attributes. It is easy to use, Windows PC-based, visually oriented, and
allows real-time interaction with the analysis. It has been designed to have an intuitive interface
and it avoids many of the arcane alternatives that are seen in the research literature but are never
used in practice.
1.2 MDS Maps

Perceptual maps are sometimes called product maps, sociograms, sociometric maps,
psychometric maps, stimulus-response diagrams, relationship maps, concept maps, etc. A
perceptual map is a piece of paper, or any plane, with symbols on it that convey information
about perceived relationships between the objects represented by the symbols.
What is the difference between a perceptual map and any ordinary map? Not much,
although some people make a big deal about it. Usually, a perceptual map is taken to be a map
that involves object-to-object relationships that are not amenable to simple, physical
measurement.
What is an object? Objects can be anything. They can be stimuli, constructs, artifacts,
characteristics, traits, people, companies, bones, arrowheads, words, discussion topics, and so
forth. Anything that you want to study can be an object. If you are interested in how certain
objects relate to each other, and if you would like to present these relationships in the form of a
map, then MDS is the technique you need.
Page 4 of 77
The MDS algorithm uses object-to-object proximity information to construct the map.
What is a proximity? A proximity is some measure of likeness or nearness, or difference or
distance, between objects. It can be either a similarity (called a resemblance in some disciplines)
or a dissimilarity. If the proximity value gets larger when objects become more alike or closer in
some sense, then the proximity is a similarity. If the opposite is the case, the proximity is a
dissimilarity.
Proximity values can be calculated, measured, or just assigned based on someone's best
judgment. If calculated, they typically are based on some mathematical measure of association
(correlation, distance, interaction, relatedness, dependence, confusability, joint or conditional
probability, pilesort counts, and so forth) operating on a set of attributes.
What is an attribute? An attribute is some aspect of an object. It may be called a factor,

characteristic, trait, property, component, quantity, variable, dimension (not a good choice in
MDS work, but occasionally seen), parameter, and so forth. The attributes should be presented
in a form where each is normalized (standardized) to some kind of range or standard deviation,
but Permap can do the normalizing internally if so desired. An attribute in one study may be an
object in another study. It is all a matter of perspective and interest.
In this manual we present all formulas in terms of dissimilarities and refer to

dissimilarities rather than proximities or similarities, but the choice is arbitrary. Typeset
documents usually represent dissimilarities by the Greek del, ij, where the indices i and j
indicate that it is the proximity between object i and object j. However, the Greek symbols do
not translate correctly in some MS Word versions configured for languages other than English,
so here we use Dij to represent a dissimilarity.
Ideally, the dissimilarity of each pair of objects is known. However, because MDS is a
very robust procedure a few randomly missing values are of little consequence. In fact, except
when using very small data sets, MDS is exceedingly good at yielding reproducible results even
after many data points have been eliminated.
Dissimilarities do not have to be symmetrical. That is, in general there is no reason that
the dissimilarity Dij must always equal Dji (note the reversal in subscripts). In fact, it is fairly
easy to think of cases where this might be the case, such as where Dij represents the driving
distance between two stores (objects) and your town has one-way streets. In this case, your
distance (dissimilarity) matrix will be asymmetrical. Fortunately, asymmetrical dissimilarity
matrices are not too common in applied MDS work. In those cases where you do have
asymmetrical dissimilarities, it is up to you to justify using the average of the two halves of the
square matrix, ignoring one of the halves, or doing separate MDS analyses on each half and then
somehow combining the results. All of these options are "right" for certain circumstances (see
Young & Hamer, 1987, or almost any MDS text). Permap does not assist with resolving
questions involving asymmetrical dissimilarity matrices.
Permap requires that dissimilarity values be nonnegative. (As a quick aside, Pearson
correlation coefficients can be used, but they have to be shifted to the 0 to 1 range to keep all
Page 5 of 77
dissimilarities nonnegative. There are two ways of doing this. See Section 6.2 for details.)
Also, the diagonal of a dissimilarity matrix must hold zeros. If similarities are used, the diagonal
of the similarity matrix must hold identical values which are not exceeded elsewhere in the
matrix.
If you are new to MDS, you might benefit from the books by Cox and Cox (2000),
Davison (1992), and Borg and Groenen (2005). The Cox and Cox text is the most
mathematically orientated, Davison's is the most applied, and Borg and Groenen's falls
somewhere in between the other two. To varying degrees, all three texts cover theory,
application, and historical development. Young and Hammer (1987) is the best in terms of
explaining the historical roots of MDS.
2.0 GETTING STARTED
2.1 System Requirements

Permap will run on any contemporary personal computer using a Windows operating
system. It is not particularly demanding in terms of computer speed, but if the problem is large,
say more than 100 objects, then at least a 1-GHz computer is recommended. Problem solution
time is often controlled as much by the computer's available RAM (random access memory) as it
is by the computer's processor speed. With the XP or 2000 Windows operating systems you
should have at least 512 MB and preferably 1 GB of RAM. Vista needs 1 GB or more of RAM.
2.2 Installation
Obtain a copy of the file Permap.zip. This file is available on the Internet at
www.ucs.ull.edu/~rbh8900/permap.html. After downloading the Permap.zip file, open it using
any of numerous decompression utilities. Usually you will have a zip/unzip program included as
part of Microsoft suite of programs. If not, one such utility is ZipGenius available at
http://www.zipgenius.it/index_eng.htm. ZipGenius is a free program that is particularly user
friendly, downloads fast, installs easily, and can be run with no training.
After unzipping Permap, you will have Permap.exe, some Windows library files (ocx and
dll extensions), some example files, a default-values file, and a WAV file. You might want to
put a shortcut to Permap.exe on your desktop. Right click on Permap.exe (in whatever directory
you chose to put the unzipped files) and drag it to the desktop. Choose "Create Shortcut Here"
when you release the right mouse button. Permap is now ready to use. Permap can be
uninstalled simply by erasing the directory you used to hold the Permap files because no
Registry entries have been.
Start the program by double clicking the Permap shortcut, by double clicking on
Permap.exe in the directory where you put Permap, or by entering its location and name in the
Run input box's "command line" which is accessed via the Start button.
Permap can be controlled by a mouse or keyboard. The alt-key gets you to and from the
menu bar, the tab-key shifts between controls, and the enter-key activates a control.
Page 6 of 77
2.3 Data Input
Data are entered from a text file (i.e., a file stored in ASCII or ANSI format). Any
modern word processor can save information in these generic formats. For instance, if you use
WordPerfect, use File/Save As and choose ASCII DOS Text or ANSI Windows Text from the
File Type box at the bottom of the screen. If you use MS Word, use File/Save As and choose
Text Only or MS-DOS Text from the Save As Type box at the bottom of the screen.
Alternatively, it is often faster and easier to use Notepad, the simple text editor that comes with
all Windows operating systems. Notepad is designed for quick entry of short segments of
unformatted text.
Permap's data files are based on freeform data entry. This means that keyword identifiers
announce the presence of various data elements and that these data types can be present in the
file in any order. Comment lines can be placed freely throughout the data file as long as they are
not placed between a keyword and its following data. All optional information is covered by
default values. This means that if you choose not to use weights then they need not be
mentioned in the data file.
Here is a very simple data set that assumes your data are in the form of dissimilarities.
All the sentences to the right are for explanation purposes and are not part of the actual data file.
________________________________________________________________________
NOBJECTS=6 Gives the number of objects in the analysis.
DISSIMILARITYLIST Announces that dissimilarity values follow.

0 All values must be nonnegative and diagonal values
.21, 0 must be zero. The data can be separated with space(s),
.59, .68, 0 a comma, or both. DISSIMILARITYLIST is all one
.74, .79, .2, 0 word and capitalization is optional.
.88, .8, .24, .25, 0 Missing entries are shown by "NA" or "na."
.11, .1, .66, .7, .89, 0 It is acceptable to name the objects. See Appendix II.
________________________________________________________________________
Dissimilarity data can be in either a lower-left half-matrix, as shown above, or in a whole

matrix format. If a whole square matrix is entered, the upper-right triangle is ignored. Entering
a square matrix is allowed simply to facilitate data interchange with other programs such as
Excel.
You may wonder why it is necessary to enter the zeros of the diagonal when dissimilarity
data are involved. After all, if the data are dissimilarities then these diagonal values must be zero
by definition. The answer is that it is not really necessary. Permap will enter the missing
diagonal values for you if you don't enter them. However, experience shows that beginning
MDS users often get similarities and dissimilarities mixed up unless they are encouraged to enter
Page 7 of 77
the zeros whereupon they usually get it correct right from the start. Unfortunately, a previously
used convention of requiring that the diagonal values be in the data file was not much admired
by Permap users. In fact, it has been the source of so much grumbling that, starting with version
9.7 Permap was modified to enter the missing dissimilarity zeros if need be. But remember, this
does not apply to similarity data. You will get an error message if the similarity diagonal values
are missing.
If your proximity information is in the form of similarities instead of dissimilarities, then

replace the keyword DISSIMILARITYLIST with SIMILARITYLIST and be sure that the
diagonal values are all equal and are not exceeded by any other similarity value. There is no
space before the "LIST" part of the keyword and capitalization is not important except for
readability considerations.
If you want to use proximity data from a spreadsheet then first save the spreadsheet in
ASCII format. For instance, if using Excel, use File/Save As and choose either Text or CSV
from the Save As Type box at the bottom of the screen. Then, load the saved file in a text editor
and enter the keywords (e.g., DISSIMILARITYLIST), titles, and so forth. Save the edited file in
ASCII text format with a "txt" or "dat" extension, and use this new file as input to Permap.
Instead of entering proximity data, Permap will accept attribute data. In this case, each
object is represented by a set of attribute values and the proximities between objects are
calculated from the attribute data using any of several standard relationships. If you want to use
attributes to generate the proximities then use the following format (only the left-justified entries
are needed).
NOBJECTS=6 Gives the number of objects in the analysis.
NATTRIBUTES=3 Gives the number of attribute values for each object.
ATTRIBUTELIST Announces that attribute values follow.

1, 1, 2 Any range can be used for the data.
1, 2, 2 The data can be separated with space(s), a comma,
3, 3, 2 or both. ATTRIBUTELIST is all one word and
3, 4, 0 capitalization is optional. All data entries must be numbers.
3, NA, 0 Missing entries are shown by "NA" or "na."
1, 6, NA It is acceptable to name the objects. See Appendix II.
You may want to put a title on the map. You do this with the TITLE or MESSAGE
command somewhere (anywhere but between a keyword and its associated data) in the data file.
For example, enter this.
TITLE= Whatever follows "=", such as this line, is transferred to the top center of the
output page.
Page 8 of 77
A complete list of keywords is shown in Appendix I. Appendix II shows an annotated
data input file. EXAMPLE_A.txt shows a typical MDS problem expressed in terms of
similarities, and EXAMPLE_B.txt shows one using attributes. The EXAMPLE_A.txt and
EXAMPLE_B.txt files, as well as several other example problem files, are included in the
Permap.zip package and can be loaded by using the File/Load Data From a Text File or by using
the F2 shortcut key.
2.4 Data Input, MDSX

To make it easier to use input files originally made for the MDSX family of programs
(see http:// www.newmdsx.com/), Permap accepts certain alternate commands (keywords) to
those shown in Appendix I.
The following table gives equivalencies between the MDSX commands and Permap
commands. The translation between MDSX and Permap is difficult because MDSX is really a
suite of programs, some of which treat advanced MDS topics that are not covered by Permap.
Thus, there are some equivalencies that occur in special cases but are not listed below. Also, the
following list ignores MDSX control commands that are not applicable to a real-time interactive
environment. For example, input file commands COMPUTE, DIMENSIONS, ERROR LIMIT,
FINISH, INPUT MEDIUM, NUMBERED, PLOT, PRINT, and PRINT DATA are not needed in
an interactive environment because they can be issued (with mouse clicks) at any time by the
program operator. Further, the MDSX keyword INPUT FORMAT is not needed because
Permap accepts any numerical format.
Table 1
MDSX and Permap Input File Command Equivalencies
MDSX Permap Details / Comments

COMMENT no equivalent Any line that does not start with a
Permap keyword is considered to be
a comment.
RUN NAME TITLE = Permap’s title has an upper limit of

or 65 characters and the "=" is optional.
MESSAGE =
TASK NAME SUBTITLE = Permap's subtitle has an upper limit

or of 98 characters and the "=" is
SUBMESSAGE = optional.
# OF STIMULI NOBJECTS = MDSX's "#" can be "N" or "NO"

and Permap's "=" is optional.
# OF VARIABLES NATTRIBUTES = MDSX's "#" can be "N" or "NO"

(used in WOMBATS) and Permap's "=" is optional.
Page 9 of 77
READ CONFIG LOCATIONLIST Follow with a list of X, Y pairs
separated by a space or a comma.
PARAMETERS DISSIMILARITIES DISSIMILARITYLIST MDSX's input is on two separate

READ MATRIX lines.
PARAMETERS SIMILARITIES SIMILARITYLIST MDSX's input is on two separate

READ MATRIX lines.
PARAMETERS ATTRIBUTES ATTRIBUTELIST MDSX's input is on two separate

READ MATRIX lines.
3.0 MAIN SCREEN & PROGRAM CONTROL
This section gives a detailed description of Permap's main screen. It covers many fairly
obvious aspects of operating Permap and can be skipped by users that are familiar with
interactive computer programs that use a graphical interface.
The MDS map is presented in the center of the screen. Map identification information
and the objective function value are given in the header section. A group of control buttons
allow real-time manipulation of the most important problem parameters.
3.1 Header and Objective Function Value

The title can be up to 65 characters long and the subtitle can be up to 98 characters long
(a variable font is used so the exact number of characters depends on the specific characters
involved). Titles and subtitles are optional. Just below the header is a line that shows the latest
value of the objective function. This value will converge to a minimum as the iterative solution
process continues. If the Auto Repeat option is being used, then just below the current objective
function value the best-found value is displayed, followed by the number of times it has been
found, followed by the total number of times the problem has been solved for the given set of
parameters. The number of times the best-found value has been found can depend on Auto
Repeat's precision (convergence limit) setting. Please see Section 3.5.4, Automatic Controls, for
more information on this point.
3.2 On-line Shortcut Controls

The left side of the main screen has a group of shortcut buttons that allow easy control of
the basic analysis parameters or functions while the solution process is continuing. Each click on
a shortcut button causes the selected parameter or function to rotate to the next setting. For
instance, by clicking on the Distance shortcut button, you can cause Permap to measure map
distances using the Euclidean measure, then the City Block measure, then the general Minkowski
measure, and then back to the Euclidean. Holding down the shift key while clicking one of the
shortcut buttons causes the rotation to reverse direction. Each of these shortcut buttons has an
associated shortcut key for those that prefer using the keyboard instead of the mouse. See
Section 3.6 for a list of the click and keystroke commands, or observe the underlined letter in
each button's caption.
Page 10 of 77
The shortcut controls duplicate features available through the menu system. The menus
give more information about each parameter or function, and in some cases allow more options
to be controlled, while the main screen's shortcut buttons provide easier real-time control. Note
that the shortcuts only show alternatives that are consistent with the more detailed settings
specified on the menu screens. That is, if you use the Attributes screen to specify that only
ordinal level attribute-to-dissimilarity functions be used, then the Attributes shortcut button will
only rotate to the Spearman and Guttman correlation coefficients because they are the only
ordinal level attribute-to-dissimilarity functions built into Permap.
3.3 Parking Active Objects Control

The parking lots are on the right side of the main screen. Objects placed in the object’s
parking lot are removed from the active data set. This makes it easy to discover the influence of
a particular object or attribute on the overall configuration. Double click an object to move it
into its parking lot. Objects also can be dragged into the object parking lot. To return an object
to the active set just drag it out of the object parking and to the position you want it to be when
the analysis continues. Each time an object is parked or unparked, the map readjusts to represent
the new active data set.
3.4 Parking Active Attributes Control

If your data file contains attribute information, just below the functional shortcut buttons
will be a box that shows the "active attributes." These attributes are used to calculate the
dissimilarities by using any of several built-in attribute-to-dissimilarity functions.
Double clicking on an attribute moves it to the attribute parking lot and the dissimilarities
are recalculated using only the remaining active attributes. A parked (inactive) attribute can be
brought back into the active set by double clicking on it in the attribute parking lot.
3.5 Execution Control

A set of execution control buttons appears just below the map in the center of the screen.
Execution of the program can be started, stopped, continued, jiggled, subjected to field
movements, automatically stopped, or automatically started by using these controls.
3.5.1 Start, Stop, Continue

When Start is clicked, the objects are placed at random positions on the map and
the iterative solution procedure is commenced. The Stop button stops the iteration and
displays the object labels if the labeling option has been selected (see Section 4.3.2). If
the solution has converged to a stationary value of the objective function but a stop signal
has not been given, then the ellipses following the objective function value continue to
blink. If you want to resume iteration after a stop command has been sent (either from
the Stop button, from Auto Stop, or from a click on an open area of the map) without
using new starting positions, then click the Continue button.
You might notice what appears to be a problem with the Continue command. Say
you are using dissimilarity data that describe a perfect circle, e.g., you are using the first
Page 11 of 77
of the Special Cases built into Permap (see Section 9.0). After finding the solution you
drag an object out of the circle and into the object parking lot. Then you click Continue
and you expect nothing to change, but you find that all of the objects move a little. Here
is what is happening. Your expectation that the map of all the remaining objects should
not change is correct, and in fact their relative positions do not change. The overall
movement is because Permap maintains the map's centroid at (0, 0), and unless the object
you removed was located precisely at (0, 0) the new centroid, i.e. the centroid after
recalculation without the parked object, will be slightly different. So, as soon as
Continue is clicked, Permap moves all objects just a bit to return the centroid to (0, 0).
MDS maps that differ only by translation, rotation, or reflection are considered to be
identical, so the map of the remaining objects does not "change" even though it moves.
3.5.2 Jiggle
Clicking the Jiggle button causes the objects to be moved to new positions that are
randomly perturbed from their existing positions. This is called "jiggling the solution."
A simple Jiggle click randomly displaces the objects a distance that averages 5% of the
map diameter. If the shift key is depressed while clicking Jiggle, the average
displacement is 10% of the map diameter. Using the control or alt keys causes average
displacements of 20% and 40%, respectively.
The Jiggle command contributes to understanding the nature of the solution. For
instance, it can provide information on the "depth of the potential well" that surrounds the
object positions. By slightly displacing the objects and watching the speed and directness
of the reconvergence you can judge if the displayed map is clearly better than all
neighboring configurations, or just one of several "almost as good" object placements.
You may be surprised to find that some local minima can be very stable, even when
subjected to large perturbations.
3.5.3 Field Movements (Mirror, Rotate, Move, Zoom)

Occasionally you will want to control the final orientation of a map in order to
make a simple comparison to previous results, or you might want to expand a map to
inspect a small, congested, area. These needs can be satisfied by mirroring, rotating,
moving, or zooming in. These operations are known as "field movements" because the
chosen operation applies to the whole field of objects. The field movement controls are
activated by clicking the Field button or right clicking the mouse on an open area.
Alternatively, you can press the F-key. To use one of the four field movement
commands, first select the desired type and then use the mouse to control the execution of
the movement. If Mirror is chosen, then clicking near an axis will cause the map to be
mirrored about that axis. If Rotate is chosen, then dragging the mouse about the center of
the map will cause the object set to rotate about the center. If Move is chosen, then
dragging the mouse in any direction will cause the object set to move in that direction.
Finally, if Zoom is chosen, then dragging the mouse away from the center of the map will
cause the object set to expand, and vice versa. The results of all of these movements
persist after the field controls are closed.
Page 12 of 77
Remember that if a MDS solution is mirrored it still is a valid solution. This is
true for any type of MDS, badness measure, distance measure, and weighting scheme.
The validity of a solution is similarly invariant to all rotations, but only if the Euclidean
distance measure is used. If City Block or any Minkowski distance measure with p not
equal to 2 is used, then only 90 degree rotations do not change the optimality of a
solution. This can be demonstrated by solving a problem using the City Block distance
measure, then rotating the solution about 45 degrees, and then continuing the solution.
The map will immediately rotate 45 degrees to again reach optimality.
3.5.4 Automatic Controls

The Auto Repeat and Auto Stop check boxes control whether or not the solution
process automatically stops, and then possibly automatically restarts, after a certain
convergence limit (as set using the Convergence menu) is reached. These controls
facilitate finding numerous solutions, each starting from a new set of random positions, to
be sure that a global minimum has been found.
If you are having trouble finding the global minima you might like to let the
computer run all night and have it record the best solution it finds. To do this, check
Auto Repeat and Auto Stop, and just let the program run. The best-found value is
displayed in brackets just below the current objective function value. Next to the best-
found value is the number of times that the best-found value has been found, followed by
the number of times the problem has been solved using the current set of parameters.
The number of times the best-found value has been found is influenced by Auto
Repeat's precision (i.e., convergence limit) setting. When the convergence setting is on
"low precision" all values within 0.0005 of the best-found value are counted, when it is
on "normal precision" this tolerance is 0.00005, and when it is on "high precision" it is
0.000005. If your goal is to quickly scan many cases, then use a low precision setting.
However, if there are several mappings (local minima) that have objective function
values that are close together it will be difficult to determine how many times the best-
found mapping actually has been found. In these cases it is important to use the high
precision setting. Also, some ordinal MDS solutions tend to pulsate and transient
objective function values can be caught when the precision value is too low. So, when
finalizing your work it is important to use the high precision setting. If in doubt, reload
(see below) the best-found solution and let Permap work on it for an extended period of
time. If the objective function value decreases, then you caught a transient configuration
with the automatic stop feature.
The complete details of the best-found solution are stored in the file
PermapBestSolnFound.txt. The PermapBestSolnFound.txt file can be reloaded via the
File menu to continue your analysis at a later date. A shortcut to reloading the last best-
found file is to press F7 and then select SaveExitReload from the menu bar.
You will find that many "real" MDS problems have several local minima and
need the Auto Restart feature to capture the global minima. The Ekman (1954) color
Page 13 of 77
comparison data that is provided in EXAMPLE_A.txt is typically susceptible (correct
answer: using ratio MDS, Euclidean distance, two dimensions, and normalized Stress, the
objective function value is 0.01721). The topic of finding global versus local minima
leads to the topic of using certain "smart" starting points to help the solution procedure
converge to a global minima. See Section 11.0, Solution Difficulties, for more
information on this important topic.
3.6 Click & Keystroke Commands

Experienced users often want the fastest possible ways of controlling program execution
even if the commands are not particularly intuitive. Here are some such control techniques using
the mouse, keystrokes, or combination keystrokes. All of these apply while the main screen has
the focus. The familiar Windows cut, paste, and select commands still work in an edit window,
even though their control keys have been usurped when the main screen is active.
Mouse shortcut controls:

Clicking an open area toggles between the Stop and Continue buttons.
Double clicking an open area is equivalent to clicking the Start button.
Double clicking an object parks the object.
Double clicking an attribute number parks the attribute.
Right clicking an open area reveals the field movement’s controls.
Control clicking an object toggles between locking and unlocking the object's position.
Keystroke shortcut controls:

F1 opens the help file
F2 opens the file selection dialog box
F3 opens the special cases screen
F4 opens the active data file
F5 opens the last-saved solution data file
F6 opens the last-saved short solution data file
F7 opens the best-saved solution data file (saved by Auto Repeat feature)
A-key changes the Attributes function

B-key changes the Badness function
C-key clicks Continue
D-key changes the Distance function
F-key reveals the Field Movements controls
E-key changes the level of Dij Precision
J-key clicks Jiggle
M-key changes the Mapping Weights function
N-key changes the number of Dimensions in the solution
P-key clicks Stop
S-key clicks Start
T-key changes the MDS Analysis Type function
X-key clicks Exit
Y-key changes the Boundary size
Page 14 of 77
esc-key clicks Exit
Alternate-key shortcut controls:

alt-A opens the Analysis Parameters menu
alt-B opens the Attributes-to-Dissimilarities screen
alt-C opens the Special Cases screen (same as F3)
alt-E opens the Edit menu
alt-F opens the File menu
alt-H opens the Help menu
alt-N opens the Convergence screen
alt-P opens the Map Evaluation menu
alt-V opens the View screen
Control-key shortcut controls:

ctrl-A opens the Attribute Evaluation screen
ctrl-B opens the Objective Functions selection screen
ctrl-D saves the existing analysis settings to file PermapDefaults.dft
ctrl-L opens the Links Evaluation screen
ctrl-M opens the Mapping Weights selection screen
ctrl-N opens the Dimensions and Boundary selection screen
ctrl-O opens the Objects Evaluation screen
ctrl-P prints the screen using the Windows PrintScreen command
ctrl-S saves the input data and the solution to file PermapSolutionLongFmt.txt
ctrl-T opens the MDS Types and Error Bounds selection screen
ctrl-Y changes the Boundary size at a slower rate than the Y-key
3.7 Locking Object Positions

You may want to fix the position of one or more objects on the map in order to make an
in-depth study of sub-optimal configurations. This situation could occur, for instance, if you are
absolutely confident of the relative positions of certain major objects, but are unsure whether or
not more arrangements of the minor objects are reasonable.
Locking an object's position can be done by holding down the control key while clicking
the object. Unlocking an object is done similarly, that is, hold down the control key and click on
the locked object. Locked objects can be dragged to a new positions. The field movement
options, mirror, rotate, move, and zoom can move locked objects because in these cases the
whole object set is being modified and relative positions do not change.
Even with some of the geometrically simple cases stored in Permap's special cases library
it is common to discover unexpected solutions when one or more objects are locked in place.
These special-situation solutions should not be dismissed out of hand. One should first ask
"could something in the environment cause one or more objects to be less mobile than the
others." If so, sub-optimal solutions may represent a physical reality.
Page 15 of 77
4.0 FILE, EDIT, VIEW
This section describes the use of Permap's system of menus. As with Section 3.0, it can
be skipped by users that are familiar with Window-based interactive computer programs.
4.1 File Menu

The File submenus provide several ways to load, save, or print data. These Input/Output
submenus use the format common to all Windows-compatible applications. The shortcut key for
this option is alt-F.
4.1.1 Load Data from Existing Files

Data are entered in either of two ways. Data can be read from a text file or
entered automatically by selecting any of several special case data sets (see Section 9.0).
The format of the data file is described in Section 2.3, Appendix I, and Appendix II. In
addition to using the menus, the F2 function key provides a shortcut that automatically
opens the input file dialog box. The F4 function key provides a shortcut that
automatically opens the active data file so that it can be reviewed or edited. The F5, F6,
and F7 function keys allow reloading the three kinds of "saved solution" files.
4.1.2 Load Data From a Special Case

The Special Cases screen can be opened using the F3 shortcut, or by following the
submenu File/Load Data From a Special Case. Due to requests from the "one click
mouse people," it has been duplicated as a top level Special Cases menu. Thus, it also
can be opened by the alt-C keystroke or by one click to the menu tab. Please refer to
Section 9.0 for its operation.
4.1.3 Enter New Data

If you want to define a new problem, often it is easier to call up an old data file,
modify it, and then use Save As (described in the next section) to save it with a new
name. This approach can be easier because it provides a "template" to guide your new
entries. However, sometimes you will be pasting data from some other media, and it will
be easier to start with a clean slate. To start from scratch, use File/Load Data From a
Text File, or ctrl-N, and a text box will open.
The default file name for the new file is "NewProblem.txt." You can type entries,
cut, paste, copy, etc. in the text box. To avoid overwriting valuable data the new data is
not automatically saved upon exiting this window, so be sure to explicitly save it when
finished. Normally, you will use the Save As option to give it some name other than
NewProblem.txt so that the original file will not be erased if a later Permap user decides
to use this option.
4.1.4 Save Results & Settings

Solution results and analysis settings can be saved to a file. There are four
options.
Page 16 of 77
The first option, File/Save Solution (Long Format), preserves all input values
needed to completely specify the problem and completely specify its solution. Because it
includes printing the distance and dissimilarity matrices this file can be very large when
the number of objects is large. The information is written to the text file
PermapSolutionLongFmt.txt. If PermapSolutionLongFmt.txt already exists, it is
overwritten. This solution file has keyword headers suitable for restarting the analysis at
the state existing when the save command was given. That is, the output file
PermapSolutionLongFmt.txt can be used as an input file. The shortcut key for this option
is ctrl-S.
The second option, File/Save Solution As … (Long Format), does the same as the
previous option except that it allows the specification of a different file name.
The third option, File/Save Solution (Short Format), writes out an abbreviated file
that fully specifies the solution, but does not repeat all the input data that define the
problem. The file name is PermapSolutionShortFmt.txt and any other copy of
PermapSolutionShortFmt.txt is overwritten. The shortcut key for this option is ctrl-R.
The fourth option, File/Save Current Settings As Defaults, saves all existing
settings to the defaults file, PermapDefaults.dft. This allows restarting Permap with all of
the existing settings. Thus, once you decide on a set of preferred settings these settings
need not be reset each time Permap is started. The shortcut key for this option is ctrl-D.
If you want to read or edit any of these saved solution files, including the one that
is automatically saved by the Auto Repeat feature, use the Edit menu or the F5, F6, or F7
function keys.
4.1.5 Reload Original Default Settings

When a program has as many parameters as Permap has it is easy to get "lost" and
not be able to return to previously used conditions. Most often, this occurs because some
obscure setting has been changed and forgotten about. Selecting File/Reload Original
Default Settings will cause Permap to reset all parameter settings to commonly used
settings and write out a new default file with these settings. If you are sharing a Permap
installation with others, it may be advisable to start each session by clicking File/Reload
Original Default Settings.
4.1.6 Printing Results

Most analysis time is spent examining maps drawn on the computer screen.
However, at some point a paper copy is usually needed. There are two approaches to
getting the MDS map onto a piece of paper.
The first approach is to use the information provided by Permap to construct and
print a map using a graphics application. This can be done, for instance, by saving the
object's coordinates using one of the File/Save commands discussed in Section 4.1.4, and
Page 17 of 77
then importing the saved object coordinates into a graphics or spreadsheet application to
prepare a presentation-quality map.
The second approach is to save the map to the Windows clipboard by using the
PrtScn key to capture the entire screen, or alt-PrtScn to capture just the active window,
and then open Paint or another graphics application and paste the clipboard contents.
After modifying the map as desired, it can be printed from within the graphics
application.
The third approach is to have the Windows operating system print the screen
directly to your attached printer. This can be done using "File/Print Screen (Std.
Windows)" or the shortcut key ctrl-P. Windows does not allow control over placement
(output will be in the upper left of the printer's page) or size (the output will be truncated
if it exceeds the dimensions of the paper). You will probably want to change the
orientation of your printed page to “landscape.” Do this with the following sequence:
Start Button → Settings → Control panel → Printers and Faxes → (select the printer you
will be using) → File → Printing Preferences → Landscape.
4.2 Edit Menu

The usual Windows editing keystrokes work when an editing window is open. That is,
ctrl-A selects all, ctrl-C copies the selected content to the clipboard, ctrl-X cuts the selected
content and copies it to the clipboard, and ctrl-V pastes the contents of the clipboard at the
position of the cursor. Note that these shortcut keys work as described only within an editing
window. For instance, if ctrl-A is pressed when the main screen is active it opens the Attributes
Evaluation screen, but when pressed inside an editing window it is the traditional "Select All"
command.
The following Edit screens are most easily opened using the F1, F4, F5, F6, and F7
shortcut keys. Because you may forget which function key goes with which edit screen, just like
we do, these shortcuts have been set up to return you to the main screen if they are pressed twice.
That is, pressing F1, etc., twice results in no net change. This makes it easier to "just hit
something" when you want to edit one of the three main data files and you don't remember which
function key is the correct one.
4.2.1 Edit Active Data File

You may edit your data file (or any other ASCII file for that matter) from inside
Permap. Enter the Edit menu and choose the Edit Active Data File selection. It will open
the data file and allow you to use all the usual Windows keystrokes. If you want to work
on another file, use the File option from within the Edit menu. The shortcut key for this
option is F4.
If you change the active data file, the changes do not automatically take affect for
the problem that is loaded. You must first save the modified file (ctrl-S), exit the Edit
menu, and reload the file that you just saved to have Permap analyze the data.
Page 18 of 77
Alternatively, you can use the SaveExitReload menu item to do this series of steps
automatically.
4.2.2 Edit Help File

The Edit menu also has the option of opening the help file (PermapHelp.txt). This
file contains several sections of this manual that contain operational details. It is
basically just a reminder list. It is made available through the Edit menu so that you can
make notes in the help file, and then save the file for future reference. Thus, with time
the help file becomes your personalized help file. The shortcut key for this option is F1.
4.2.3 Edit Solution Files

The Edit menu allows easy access to the problem solution files
(PermapSolutionLongFmt.txt, PermapSolutionShortFmt.txt, PermapBestSolnFound.txt),
i.e., the files you formed by using ctrl-S, ctrl-R, or the File/Save Solution menus, or the
file that is automatically saved by the Auto Repeat feature. Often you might want to
consult one of these files to find some numerical value of a parameter for a solution that
is under study. Or, you might want to change some particular parameter's value and then
watch what happens when the problem is rerun. The shortcut keys for these options are
F5, F6, and F7. As with the Active Data File case, changes to
PermapSolutionLongFmt.txt and PermapBestSolnFound.txt can be automatically saved
and reloaded to be the active data file by using the SaveExitReload menu item. This
cannot be done with the PermapSolutionShortFmt.txt file because it does not have all the
information necessary to fully describe the problem.
4.3 View Menu

Selecting the View menu opens the View Screen that lets you control the look of the
MDS map. The shortcut key for this option is alt-V.
4.3.1 Object Representation

You can customize the appearance of your map by selecting the object size, object
border width, and object border color. Each option is explained in a text box at the right
side of the View screen and a demonstration map at the bottom-right of the screen
illustrates the current settings. When you are analyzing large problems it is beneficial to
reduce the circle size to minimize overlaps and thus clarify the map.
4.3.2 Object Identification

User-supplied names, preferably just a few characters each, can be used to
identify the objects once the analysis is stopped. Or, the objects can be automatically
numbered in the order they are entered in the data file. As useful as these options are for
small problems, maps with many objects quickly become congested. For large problems,
the naming option should be turned off by selecting the "none" option.
Object names are displayed only when the solution (process iteration) has
stopped. You can stop the process by using the Stop key. Alternatively, you can toggle
Page 19 of 77
between the Stop and Continue states by a single mouse click on an open space on the
map.
4.3.3 Unit Circle and X-Y Axes

Some users prefer to show a unit circle on map. All metric dissimilarity matrices
generate object positions that lie within this circle. You turn on this feature with the first
entry under View/Mapping Space.
The second Mapping Space entry controls the display of an X-Y coordinate
system. The X-Y coordinate system described here is the internal system that is used to
make the calculations and place the objects on the map. While some users like to show
these coordinates, they are easily misunderstood and probably should not be shown
unless you are making a City Block or some other non-Euclidean distance analysis.
These coordinates are not related to the interpretative coordinates that the researcher
assigns to the map.
5.0 ANALYSIS PARAMETERS
A complete MDS analysis description requires specifying the type of MDS analysis, the
objective function (badness and distance measures), mapping weights, solution dimensionality,
boundary conditions, and the nature of the attributes being used (if any). The Analysis
Parameters set of submenus let you specify all of these aspects and assists in testing the affect of
changing dissimilarity precision via the Precision control button. Open this menu with alt-A or
by clicking on the Analysis Parameters menu.
5.1 MDS Analysis Types

The first option in the top box on the left side of the MDS Types and Error Bounds
selection screen allows selection of the type of MDS analysis to be made. This selection also can
be made by clicking the MDS TYPE shortcut button or by opening the MDS Types and Error
Bounds submenu using ctrl-T.
Permap can make ratio, interval, or ordinal analyses, as well as ratio and interval analyses
with error bounds. After providing some general comments on MDS terminology, each of these
methods is described in detail. The specific formulas used to match analysis type to the
objective function are given in Section 5.2.
5.1.1 Metric and Nonmetric Terminology

Traditionally, MDS analyses are classified as metric or nonmetric. These terms
are not entirely satisfactory because of their imprecision, but they persist. More
precisely, one should classify MDS analyses based on the relationship between the map's
inter-object distances dij and the dissimilarity Dij values. There should be at least three
major types of MDS analysis. First, if the algorithm seeks to find dij such that dij / Dij =
k1, where k1 is an arbitrary constant that determines the size of the map, then the MDS
algorithm is at the ratio level. Second, if the algorithm seeks to find dij such that (dij -
Page 20 of 77
dkl)/ (Dij - Dkl) = k1, then the MDS algorithm is at the interval level. Third, if the
algorithm seeks to find dij such that Dij < Dkl implies dij < dkl, then the MDS algorithm is
at the ordinal (rank order) level. Based on these considerations, an MDS analysis should
be classified as ratio, interval, or ordinal instead of metric or nonmetric.
Ratio and interval levels of information are two examples of metric information.
Ordinal information is an example of nonmetric information. The word "metric" harks
back to "measurement" and signifies that measured distances on the map are linearly
related to the Dij values. If you simply ignore the important difference between ratio and
interval levels of information, you arrive at the historical terminology where all MDS
analyses are classified as metric or nonmetric
5.1.2 MDS and Error Bounds

Permap provides two new nonmetric kinds of analysis that involve applying error
bounds to ratio and interval level data. Thus, Permap makes traditional ratio (metric),
interval (metric), and ordinal (nonmetric) MDS analyses, and adds ratio + bounds
(nonmetric) and interval + bounds (nonmetric) analyses. Before presenting precise
definitions of the various types of MDS maps that Permap can make it is important to
understand the role of error bounds and why the traditional forms of MDS have ignored
the error bounds approach.
Every experimental measurement has some associated uncertainty. However, if

the MDS computer program does not have the ability to use the uncertainty information,
you must either make the invalid assumption that the dissimilarities should exactly
control the distances between objects (traditional metric analyses), or make the unrealistic
assumption that the dissimilarities are so poorly known that only their rank order can be
trusted (traditional nonmetric analyses). You should not be forced to choose between
these two extremes.
Metric MDS is based on the reasonable "first order approximation" that the input
data are linear in the construct of interest and are errorless. In the physical sciences the
second step usually is to relax the assumption of errorless data and to introduce error
bounds. However, the development of MDS did not follow the usual path. Early MDS
critics were fixated on the point that behavioral science measurements often are very
imprecise. The result was that metric MDS was not modified to accept error bounds.
Instead, ordinal MDS was developed and it essentially took over MDS. There are even
comments in the early MDS literature referring to the "near magical" nature of ordinal
MDS. The ordinal advocates, in effect, chose to discard useful information simply
because not all of the information was at a high level of reliability.
Here is an example of what happened. Say you have measurements, or estimates,

yielding dissimilarities of 1, 1.05, 1.1, 8, 9, 10, and 19. It is quite possible that your
experimental uncertainty is such that you cannot prove that the first three cases are truly
different from each other. Also, it is quite possible that while you are sure that cases four
through six are properly ordered, you are not at all sure that there is exactly one unit of
Page 21 of 77
difference between them. As a result, ordinal advocates argue that the analysis should
proceed as if the dissimilarities maintain the rank order of the original data, and that is all.
Their analyses would use pseudo Dij of 1, 2, 3, 4, 5, 6, and 7 in place of those shown
above. Certainly the ordinal advocates have valid concerns. Dij data are often fraught
with imprecision. But, just because some relationships are uncertain it does not follow
that the whole data set should be treated as ordinal. There is little doubt that you know
the seventh data point is far higher than just one step above the sixth, and that the
differences in the first group of three clearly are not the same as the differences in the
second group of three.
The correct way to approach most data sets is to ignore the old metric vs.
nonmetric arguments and simply face the fact that all experimental data have error.
Doing so suggests that every entry should have a best-estimate value and associated error
bounds. There are many ways to specify error bounds, but the differences are relatively
unimportant compared to the importance of avoiding the mistake of degrading to an
ordinal analysis when it is not necessary to do so. If, for example, all of the
dissimilarities mentioned above have associated uncertainties of  0.2, and an analysis is
made taking these error bounds into considerations, then the first three dissimilarities
points will be treated as being essentially the same. And, the far-different last
dissimilarity (i.e., 19) will still be recognized as being far different from the others.
Nowhere do we claim that we know the true values of each dissimilarity, but we do
preserve the significant differences that we do know to be correct.
To restate the forgoing argument in less quantitative terms, we can say that while
our measurements of some quantity may be sufficiently inaccurate that we cannot assign
precise values we may be able to say much more than just what rank order the quantities
have. Sometimes we can justifiably say that one value is "a little greater than," "much
greater than," or "very much greater than," another value instead of just saying one is
greater than the other.
5.1.3 Procedural Details for Each Analysis Type

In the following sections each type of MDS analysis is defined in terms of its
quantitative relationships. The formulas behind these definitions are given in Section 5.2.
5.1.3.1 Ratio
A ratio MDS algorithm attempts to find a set of dij that approximately
satisfies the relationship dij / Dij = k1, where k1 is an arbitrary constant that
determines the size of the map. This is the same thing as saying the goal is to find
a set of dij such that a linear relationship dij = k1 Dij is fit as well as possible. In
Permap, k1 is always unity because the dissimilarities or similarities are scaled to
a maximum of one immediately upon data entry. Therefore, k1 drops out of
Permap's internal equations, and is omitted from the formulas outside of Section
5.1.
5.1.3.2 Ratio + Bounds
Page 22 of 77
A ratio + bounds analysis attempts to find dij that satisfy dij / Dij = k1, but
the objective function is not penalized as long as dij is within the range k1 (Dij 
eij) where eij represents the error or uncertainty in Dij. Permap can handle two
types of error terms.
The first option in the second box on the left side of the Analysis Type and
Error Bounds selection screen is the one that most new users should start with. It
simply assumes that each Dij term is subject to the same uncertainty  eij. While
this approximation may not be precisely correct, it is often a good starting point.
By clicking on the up or down arrows, the bounds can be set anywhere up from 0
to1. If the error bounds are set to 0, then this option becomes identical to the
traditional ratio analysis and is eliminated from the rotation when the MDS TYPE
shortcut button is used.
The second option is best for those situations where the uncertainty is
proportional to the size of the factor being measured. This is often the case in
physical measurements. This option assigns uncertainties of plus or minus some
constant times Dij. By clicking on the up or down arrows, the percentage error
bounds can be set anywhere from 0% to plus or minus 100% of Dij. If the error
bounds are set to 0%, then this option becomes identical to the traditional ratio
analysis and is eliminated from the rotation when the MDS TYPE shortcut button
is used.
5.1.3.3 Interval
An interval MDS algorithm attempts to find a set of dij that at least
approximately satisfies the relationship (dij - dkl)/ (Dij - Dkl) = k1, where k1 is an
arbitrary constant that determines the size of the map. This is the same thing as
saying that the goal is to find dij that at least approximately satisfies the
relationship a dij = IC + k1 Dij, where IC is the Interval Constant. As was
mentioned in Section 5.1.3.1, in Permap, k1 is always unity. Thus, it is omitted
from further discussion and is omitted from the formulas outside of Section 5.1.
When making an interval MDS analysis, Permap finds the value of IC that
minimizes the objective function and displays it in the caption of the MDS TYPE
shortcut button.
The practical difference between interval and ratio MDS is that in interval
MDS uncertainty about the true zero point of the Dij data is acceptable. In certain
situations this is very important. For instance, when using Likert-like scales it can
happen that one is uncertain that test subjects are responding such that the bottom-
level response represents a true zero point for the construct under study. Special
Case 6 provides data that illustrate how switching from a ratio analysis to an
interval analysis can clarify the hidden structure in such a data set. Note that if
your Dij values come from some source that has a reliable zero-to-one range, such
as correlation coefficient data, then using interval MDS is not appropriate. You
should not introduce a new parameter (IC in this case) just to find a lower
Page 23 of 77
objective function value. Any new parameter must have sufficient "face validity"
that its critics will agree that it makes sense to add it to the model.
Interval MDS has several quirks that may be confusing to the new user.
First, all simplex cases (this includes the regular tetrahedron and the 10-object and
24-object simplex examples provided in Permap's library of special cases) have,
by definition, all non-diagonal Dij values equal. If you are allowed to shift the Dij
zero point, as interval MDS does, a valid solution can be found by reducing all
effective Dij values (i.e., IC + k1 Dij) to zero and thus allowing the objective
function to go to zero. Thus, interval analyses of simplex data will always result
in all objects being superimposed. This is not a degenerate solution. It is a true
solution. Second, adjusting the interval constant to minimize the objective
function may or may not minimize R2, the figure of merit used in the Links
Evaluation screen. Usually, minimizing the objective function and minimizing R2
go hand-in-hand, but not necessarily so. It depends on the objective function that
is being used. Try using SStress, Interval, and the Three Perpendicular Vectors
special case data. The expected "solution," that is, the analog to the Stress,
Multiscale, and Fractional cases, will not occur. Finally, there are strong bounds
on the value of the interval constant. This is an important point that deserves
expansion.
Calculation of the interval constant can be done in two ways. A strict

interpretation holds that, to make sense in terms of what a dissimilarity means, an
effective dissimilarity (IC + k1 Dij) must be bounded by zero and one. So, if your
dissimilarity data set has even one Dij = 0 entry then the interval constant cannot
be less than zero, and if it has even one Dij = 1 entry then the interval constant
cannot greater than zero. In these situation, changing from ratio to interval MDS
will make no difference at all because the interval constant can have no value
other than zero.
5.1.3.4 Interval + Bounds

Interval + bounds analyses are similar to ratio + bounds MDS analyses.
Please see Section 5.1.3.2 above. The objective function is not penalized as long
as dij is within the range IC + k1 (Dij  eij) where eij represents the error or
uncertainty in Dij.
5.1.3.5 Ordinal
An ordinal analysis seeks dij values such that if Dij < Dkl then dij < dkl.
This means that the objective function is not penalized as long as the calculated dij
are in the same rank order as the Dij.
When an ordinal analysis is specified, Permap treats ties using the so

called "primary" convention. That is, all objects of equal rank need not have the
same position on the map.
Page 24 of 77
5.1.4 Object Centering
The nonmetric forms of MDS, i.e., ratio + bounds, interval + bounds, and ordinal
may only determine the object positions to within certain limits. Because all object
positions that are "within the bounds" or "in the right order" contribute zero to the
objective function value, there may be infinitely many possible solutions. In reality,
these alternative solutions usually are tightly constrained and thus the map does not have
any variation at all. However, if the error bounds are large, or if an ordinal analysis is
being made with only a few objects, having multiple equally valid solutions can be
objectionable. In this case, one might want to ask that the objects be placed at the center
point between the error bounds or at the mid point between objects of adjacent ranks.
Centering can be achieved by adding a "junior term" to the objective function.

This term must be small enough that it does not affect an object's position when that
position is being determined by balancing conflicting dissimilarities, but it must be large
enough that it can "nudge the objects over" when they are "between the bounds." This
goal is achieved by adding a term to the objective function that is one tenth as large as the
normal terms, and is anchored to the applicable midpoint rather than being zero anywhere
inside the applicable boundaries. Note that when the centering option is used the value of
the objective function increases very slightly. This is a necessary artifact of the centering
procedure. It does not indicate that the solution is inferior in any way. It simply means
that another requirement has been placed on the solution and that the "cost" of this extra
requirement is being reflected in the value of the objective function.
Centering is controlled by the third option box on the right side of the MDS Types
and Error Bounds selection screen. Of course, if you are using regular (i.e., no bounds)
ratio or interval MDS then the centering feature has no affect.
5.2 Objective Functions

The second submenu of the Analysis Parameters menu allows specification of the
objective function, which involves specifying the form of the objective function, the badness
function, and the distance measure. The badness and distance problem parameters can be
controlled using the Badness and/or Distance shortcut buttons on the main screen. This allows
you to quickly scan a variety of solution types while watching the map as it shifts to satisfy the
new settings. The Objective Functions submenu can be opened using ctrl-B.
5.2.1 Objective Function Form

The objective function goes by many different names. For instance, some
disciplines call it an error function, loss function, criterion function, merit function, and
so forth. It is simply a mathematical formula that is maximized or minimized in the
process of solving a problem. It is the accumulation of one or more terms that get bigger
(or smaller in the case of a merit function) as the objects are moved away from a good
configuration.
In some early MDS papers the objective function was completely unspecified.
The steps used to optimize "something" were described, but exactly what was being
Page 25 of 77
optimized was never defined! This may seem almost unbelievable to new MDS
researchers, but it makes sense when you consider the progress that has been made in our
ability to quickly make a large number of calculations. Before the early 1960s
computational difficulties were so great that the precise details of making a MDS map
took precedence over the bigger picture involving the mathematics of the problem's goal
and constraints. Nowadays, just the opposite is true. This manual, for instance, reflects
the trend of emphasizing the description of the problem, assumptions, and constraints,
rather than the solution procedure. This separation can be noted by the little space given
to numerical analysis techniques (Section 10.0) compared to the great amount of space
given to the definition and description of the problem and solution (Sections 1.0 through
8.0).
Most objective functions are defined to be the sum over all object pairs of the
pair's weight factor times the square of the pair's badness measure Bij. This is called a
weighted sum-squares form. Mathematically, the sum-squares objective function
equation is:
Sum-Squares form: Objective Function = [SUM(all i, j < i) Wij Bij2 ] / NC
Ignore for now the normalizing constant, NC. It will be described in Section
5.2.4. Focusing only on the form of the above equation, one can see how the individual
badness terms are combined to form the quantity to be minimized. This separation
between the definition of Bij and how the Bij are accumulated represents a level of rigor
that is not always adhered to. Some authors implicitly assume a form for the combination
of the badness terms and then use the badness name for the name of the objective
function. This causes some confusion. For example, if one says that Stress is the criteria,
does that specify how badness is measured or how the entire objective function is
defined? Both usages can be found in the literature. Here, we takes Stress, Stress1,
SStress, Multiscale, and Fractional to define the badness measure, but it is just as
common to find authors that use these terms to specify the overall objective function.
The sum-squares form of the objective function is the most commonly used form.
This may be because it is very simple while still being an analytical function (i.e., all
derivatives exist) and an even function (i.e., f(x) = f(-x)), or perhaps it is because of the
analogy with common statistical estimators that are based on maximum likelihood
arguments. For whatever reason, it was used by the originators of the badness terms
Stress, SStress, Multiscale, and Fractional. However, you will find other objective
functions that use a "root-sum-squares" form. For instance, Kruskal introduced Stress1
and Stress2 (Kruskal's Stress2 never gained wide acceptability and is not discussed
further here) that use the root-sum-squares form. Mathematically, the root-sum-squares
objective function equation is:
Root-Sum-Squares form: Objective Function = SQRT{[SUM(all i, j < i) Wij Bij2 ] / NC}
Page 26 of 77
Even though the root-sum-squares form was used originally to define Stress1, you
will find some early papers that redefine Stress1 using the sum-squares way of combining
the badness terms. Similarly, even though the sum-squares form was used by the
originators of SStress, some papers use a root-sum-squares formula for it. Clearly, the
situation is confusing. It would be better had the early workers in the field given more
thought to maintaining consistency in their definitions, but it didn't happen. The situation
is tolerable only because it makes no difference to the configuration of the map whether
the sum-squares or root-sum-squares form is used. The only difference lies in the value
of the objective function that is reported for a particular map. The root-sum-squares form
makes Stress1 values considerably larger than Stress and all the other objective functions
which use the sum-squares form. In fact, making the values larger was Kruskal's stated
reason for using the square root in his definition.
Permap uses the definition which conforms to the originator's definition. Thus, it
is to be understood that the Stress1 badness always uses the root-sum-squares form of the
objective function and that all the rest use the sum-squares form. This means that if you
use the BADNESS FN shortcut button to cycle through the various badness functions the
objective function value will increase significantly when Stress1 is activated whereas the
other functions will have values that are of comparable magnitude.
Because MDS maps found using Stress and Stress1 are exactly the same (because
the badness function is exactly the same), one might want to eliminate one or the other
from the BADNESS FN shortcut rotation list. This can be done using the option box at
the lower left side of the Analysis Parameters / Objective Functions screen.
5.2.2 Badness Functions

A badness function is simply a definition of what it is that makes the positions of
a pair of objects be bad. Fundamentally, the object pair should be separated by a distance
that is consistent with the pair's dissimilarity. Thus, the badness Bij should involve a
measure of the mismatch between dij and Dij. There are several valid ways of defining
this mismatch.
Permap offers the five most common types of badness measure, Stress, Stress1,
SStress, Multiscale, and Fractional. Because the badness formulas for Stress and Stress1
are identical they are combined here. Remember, however, that Stress and Stress1 imply
different methods of combination to form the objective function. Each of the five
badness types must be adjusted for the type of MDS being used. We start with Ratio
MDS because it most clearly shows the underlying nature of the badness relationship
between dissimilarity and distance.
Ratio MDS
Stress & Stress1: Bij = dij - Dij
SStress: Bij = dij2 - Dij2
Page 27 of 77
Multiscale: Bij = ln(dij / Dij)
Fractional: Bij = (dij - Dij) / Dij
Ratio + Bounds MDS
Stress & Stress1: If dij > Dij + eij Then

Bij = dij - Dij - eij
ElseIf dij < Dij - eij Then
Bij = dij - Dij + eij
Else
Bij = 0
EndIf
SStress: If dij > Dij + eij Then

Bij = dij2 - (Dij + eij)2
Bij = dij2 - (Dij - eij)2
Else
Bij = 0
EndIf
Multiscale: If dij > Dij + eij Then
Bij = ln(dij / (Dij + eij))
Bij = ln(dij / (Dij - eij))
Else
Bij = 0
EndIf
Fractional: If dij > Dij + eij Then

Bij = (dij - Dij - eij) / (Dij + eij)
Bij = (dij - Dij + eij) / (Dij - eij)
Else
Bij = 0
EndIf
Interval MDS
Stress & Stress1: Bij = dij - IC - Dij
SStress: Bij = dij2 - (IC + Dij)2
Page 28 of 77
Multiscale: Bij = ln(dij / (IC + Dij))
Fractional: Bij = (dij - IC - Dij ) / (IC + Dij)
Interval + Bounds MDS
Stress & Stress1: If dij > IC + Dij + eij Then

Bij = dij - IC - Dij - eij
ElseIf dij < IC + Dij - eij Then
Bij = dij - IC - Dij + eij
Else
Bij = 0
EndIf
SStress: If dij > IC + Dij + eij Then

Bij = dij2 - (IC + Dij + eij)2
Bij = dij2 - (IC + Dij - eij)2
Else
Bij = 0
EndIf
Multiscale: If dij > IC + Dij + eij Then

Bij = ln(dij / (IC + Dij + eij))
Bij = ln(dij / (IC + Dij - eij))
Else
Bij = 0
EndIf
Fractional: If dij > IC + Dij + eij Then

Bij = (dij - IC - Dij - eij) / (IC + Dij + eij)
Bij = (dij - IC - Dij + eij) / (IC + Dij - eij)
Else
Bij = 0
EndIf
Ordinal MDS
Stress & Stress1: If Ldij < Hdij Then

If dij < Ldij Then
Bij = dij - Ldij
ElseIf dij > Hdij Then
Page 29 of 77
Bij = dij - Hdij
Else
Bij = 0
EndIf
Else
If dij > Ldij Then
Bij = dij - Hdij
ElseIf dij < Hdij Then
Bij = dij - Ldij
Else
Bij = 2dij - Ldij + Hdij
EndIf
EndIf
SStress: If Ldij < Hdij Then

If dij < Ldij Then
Bij = dij2 - Ldij2
Bij = dij2 - Hdij2
Else
Bij = 0
EndIf
Else
If dij > Ldij Then
Bij = dij2 - Hdij2
Bij = dij2 - Ldij2
Else
Bij = 2dij - Ldij2 + Hdij2
EndIf
EndIf
Multiscale: If Ldij < Hdij Then

If dij < Ldij Then
Bij = ln(dij / Ldij)
Bij = ln(dij / Hdij)
Else
Bij = 0
EndIf
Else
If dij > Ldij Then
Bij = ln(dij / Hdij)
Page 30 of 77
Bij = ln(dij / Ldij)
Else
Bij = ln(dij / Hdij) + ln(dij / Hdij)
EndIf
EndIf
Fractional: If Ldij < Hdij Then

If dij < Ldij Then
Bij = (dij - Ldij) / Ldij
Bij = (dij - Hdij) / Hdij
Else
Bij = 0
EndIf
Else
If dij > Ldij Then
Bij = (dij - Hdij) / Hdij
Bij = (dij - Ldij) / Ldij
Else
Bij = (dij -Ldij)/Ldij +(dij -Hdij)/Hdij
EndIf
EndIf
The ratio and interval badness definitions are taken from the literature. The ratio
+ bounds and interval + bounds definitions flow from the ratio and interval formulas after
being adjusted for a "zero penalty" zone centered about the target Dij. Ordinal definitions
are constructed similarly to the ratio + bounds definitions, except they no longer
reference Dij. For ordinal MDS all relationships are relative. The Dij come into play only
in determining the rank order of each i, j term. The new variable Ldij stands for the
distance between the two objects that have the next lowest dissimilarity beneath Dij.
Similarly, Hdij stands for the distance between the two objects that have then next highest
dissimilarity above Dij.
The ratio MDS Multiscale and Fractional badness measures have singularities at
Dij = 0, and the Multiscale measure has another singularity at dij = 0. In practice, 0.0001
is added to the vulnerable terms to avoid these singularities. This is a common technique
used in numerical analysis. It maintains a monotonic relationship with the unmodified
formulas, is computationally efficient, and causes insignificant (invisible under normal
viewing conditions) distortions to the map. Technically speaking, the introduction of the
0.0001 terms means that the Multiscale and Fractional measures should be called Pseudo
Page 31 of 77
Multiscale and Pseudo Fractional measures, but this refined terminology is essentially
never used in the literature and is not used here.
The choice of the badness measure is not always important. There are numerous
comments in the literature that this aspect can be ignored with the expectation that the
map will be unchanged. Because of this, some researchers simply assume that the Stress
badness formula is the proper choice. However, the conclusion that these details are not
important is not always true. A little work with Permap will show that it probably results
from a false generalization based on data sets that are well fit. The better the fit the more
the choice of the badness measure is immaterial. However, if the fit is not good, then the
choice can be very important. Once you have Permap up and running, check out the
differences in configuration for the Stacked Triangles test case when SStress is used
instead of a Stress. It is important that the affect of this basic problem parameter be
tested for each different data set. You will find that the more stressed your solution is the
more important the choice is.
Finally, it should be mentioned that the Multiscale measure really is of very little
value. When used with ordinal MDS, it is of even less value. It is included because, to
the best of our knowledge, the typical MDS researcher likes to "try everything" and "see
for themselves." So it is included more for satisfying curiosity than for use. Here is why
Multiscale is not very useful. First, it does not conform to most people's intuitive
definition of badness. That is, it is highly non-linear in its spatial relationship between
badness and displacement. Second, it leads to unstable mappings. In fact, unless your
data set has the possibility of a low-stress mapping (such as many of the special cases
data sets do) using Multiscale may produce wildly unstable results. Of course, you can
first solve the problem using Stress or SStress and then rotate to Multiscale to see what
the Multiscale mapping looks like, but direct convergence from random starting points is
difficult to achieve. Even using a smaller step size doesn't help in most cases. So, if you
are a "new user in a hurry," then just ignore Multiscale. Otherwise, enjoy, but don't get
upset if your data set doesn't have a stable solution.
5.2.3 Distance Measures

To determine the badness, Bij, you have to have the distance between two objects,
dij. To measure dij you must have a definition of how distances are measured. For
example, if you have two points (X1, Y1) and (X2, Y2) on a map, how far apart are they?
If you believe in the Euclidean measure, then the distance is the length of a straight line
between the points. If you believe in the City Block measure, then the distance is the
absolute value of X1-X2 plus the absolute value of Y1-Y2, which is equivalent to taking an
"up and over" path. If you believe in the Dominance measure where only the larger of
the two differences is considered to be important, then the distance is the maximum of the
absolute value of X1-X2 and absolute value of Y1-Y2. These and other metrics can be
represented using the general Minkowski formula where the exponent p has a value that
can vary from one to infinity.
Euclidean: dij = SQRT{(Xi-Xj)2 + (Yi-Yj)2}
Page 32 of 77
City Block: dij = |Xi-Xj| + |Yi-Yj|
Dominance: dij = MAX(|Xi-Xj|, |Yi-Yj|)
Minkowski: dij = {|Xi-Xj|p + |Yi-Yj|p}1/p
Further generalization of the Minkowski form involves applying weights to the

various Xi-Xj and Yi-Yj differences (Hartigan, 1975, p. 64). However, these extensions
are essentially never seen in the applied MDS literature. In fact, the general Minkowski
form and the Dominance form (equivalent to Minkowski p =  but often approximated by
using p = 30) is almost never seen. In practice, only the City Block (equivalent to
Minkowski p = 1) or Euclidean (equivalent to Minkowski p = 2) distance metric is used
to any significant extent. Between City Block and Euclidean, the Euclidean is
overwhelmingly the most commonly used metric. Using the City Block metric is
recommended only if you are knowledgeable about the perceptual assumptions that go
with it. See Torgerson's early work (Torgerson, 1952, 1958, 1965) for a detailed
discussion of the relationship between spatial and psychological distances, and see
Hubert, Arabie and Hesson-Mcinnis (1992) for a discussion of the mathematics of the
City Block measure. One could go so far as to say that you should always use the
Euclidean metric unless you are thoroughly grounded in the fundamentals of measuring
differences between abstract constructs. Borg and Groenen (1997, p. 14) emphasized this
position when they wrote “. . . MDS representations that employ distance functions other
than the Euclidean tend to be misleading when inspected intuitively. Therefore, they are
useless for exploratory purposes." Despite the possibility of misuse, Permap includes the
City Block and the general Minkowski distance measures to let users explore for
themselves the various distance measurement alternatives.
A final detail about the Minkowski distance measure needs explanation. The
Minkowski formula needs double precision arithmetic in order for the calculated
distances to have the number of significant figures needed for Permap to be able to
calculate the value of the objective function to many significant figures. Double
precision arithmetic is slower than single precision arithmetic and is not used in Permap.
Thus, if you use a Minkowski distance formula you may find that for some p values the
objective function is not as stable as it would be if a different distance function was used.
5.2.4 Normalization Constants

Normalization of a quantity is achieved by dividing it by a constant, NC that is
chosen to achieve three purposes. First, you want the units of measurement to cancel out.
That is, you would like to get the same results regardless of whether the original data
were expressed in centimeters or inches. Second, you want large problems to produce the
same intensive results as small problems. This means the NC must serve as a scale
factor. Third, you want values to fall in the zero-one range so that the results based on
them will be easily interpreted and do not require scale-related weights. This last goal is
the overarching goal of normalization, but it cannot always be achieved such as when the
Page 33 of 77
quantity being normalized is not bounded. Fortunately, this is not a problem for MDS
objective functions.
Normalization of the objective function does not influence the resulting MDS
map at all. It only changes the numerical value of the objective function value. Still, to
facilitate comparisons across studies it is best to normalize the objective function.
Therefore, Permap always displays normalized objective function values.
Permap uses the following NC factors:
Stress: NC = Sum (all i, j < i) Wij Dij2
Stress1: NC = Sum (all i, j < i) Wij dij2
SStress: NC = Sum (all i, j < i) Wij Dij4
Multiscale: NC = Sum (all i, j < i) Wij
Fractional: NC = Sum (all i, j < i) Wij
By convention, if Wij is not used then they are set equal to one. In this case the
sum over the lower left corner of the weights matrix reduces to N (N - 1) / 2 where N is
the number of objects in the analysis.
Finally, if you carefully examine the preceding formulas you notice that there is
an inconsistency in the Stress1 NC definition. The normalizing constant for Stress1 is
not a constant at all. So, in this case, the term "normalizing constant" should be replaced
with "normalizing factor." The imprecision is tolerated to gain the pedagogical value of
the taxonomy and to avoid arguments with Stress1 advocates.
5.3 Mapping Weights

Before discussing mapping weights, it is helpful to briefly digress into weight factors in
general. Some MDS analyses involve assigned weights Wij that multiply the corresponding Bij2
terms. Permap allows you to enter these weights through the data file. They reflect the
importance, uncertainty, saliency, flexibility, etc., associated with each object-to-object
dissimilarity value. Sometimes they are taken to be the reciprocals of the variances in the
experimental data and sometimes they are just subjective assignments. Most commonly they are
not used. If Permap finds no weight information in the data file, then all Wij are set equal to one.
Note that if the dissimilarities are well fit by the map (the minimum objective function
value is zero) then the weights are immaterial. The map will remain stationary for any random
assignment of weights because the weights are multiplying zero badness quantities. However, if
the fit is not perfect, then the weights tell how to respond to each pair's badness. You might say
they provide a method for determining "what yields the most." For example, if you are unsure of
Page 34 of 77
one object pair's dissimilarity, and it conflicts with those of one or more other pairs that you
know with some certainty, then you should give the more uncertain Bij term a lower weight.
Now let's shift our thinking away from traditional weights to a new kind of weight factor
called a mapping weight. Mapping weights are denoted MWij. They are built into Permap and
controlled by the Analysis Parameters/Mapping Weights menu or the Mapping Weights shortcut
button. These mapping weights do not replace the regular weights Wij. They multiply the Wij
weights. It is the product Wij MWij that is used in the calculations. Thus, if you decide to use
mapping weights, then everywhere Wij appears in this manual you should replace Wij with Wij
MWij. To emphasize this point, we make it a separate subsection.
5.3.1 Replacing Wij with Wij MWij

If you choose to use mapping weights, then all formulas involving Wij should
replace Wij with Wij MWij.
5.3 Mapping Weights, continued

The importance and general applicability of mapping weights is not widely recognized.
There are three distinct reasons why mapping weights might be appropriate.
1. Ignoring questions of scale and orientation the statement that zero dissimilarity exists
is meaningful whereas the statement that complete dissimilarity exists may or may not even be
capable of precise definition.
2. Percent uncertainty values, rather than absolute uncertainty values, often better
conform to measurement reality.
3. A simple Cartesian map will not always convey the essence of the underlying
relationships.
The first of these points is fundamental. It involves the issue of knowability. Let (i, j) be
(3, 4) to simplify the discussion. The information content of saying D34 = 0 is quite different
from saying D34 = 1. Even in the simple case where two objects can be described by specifying
the value of its attributes the problem remains. Saying that D34 = 0 means that all corresponding
attribute values of object 3 and object 4 are identical. However, the same type of statement is not
valid for the complete dissimilarity case. If you know the nature of object 3 and object 4 then
saying that D34 = 1 does not mean that you know the nature of object 4. To put it still another
way, test subjects might be able to explain why they consider two objects to be identical, but
they cannot explain in similar detail why they consider two other objects to be completely
different. This means that position misfits between objects with a small dissimilarity should not
be treated like position misfits for objects with large dissimilarities. Strongly dissimilar values
should be given less weight in determining the shape of the map. In other words, mapping
weights should decrease with increasing dissimilarity values. The importance of this point is
emphasized when one realizes that most objective functions do just the opposite. They give
more weight to large Dij values than small Dij values because the function involves the squaring
of Dij.
Page 35 of 77
The second point is practical. It is often true that the degree of uncertainty in a
measurement is roughly proportional to the magnitude of the property being measured. This is
not a necessary relationship, but one that is frequently observed. If it applies, then instead of
spending considerable time estimating Wij for each dissimilarity, it is easier to leave all Wij with
the default setting Wij = 1, and let Permap generate appropriate mapping weights that decrease
with increasing dissimilarity values.
The third point is fundamentally different from the first two. It does not involve using
any concept of uncertainty. Instead, it involves recognition that a "photograph like" picture of
reality is not always the best picture to explain the essence of the matter. This is why Picasso
would take the elements of a traditional picture and move them about, change their orientation,
change their color, or distort them in some manner. In doing so he would facilitate a new
understanding of the subject. Mapping weights can do this. Consider a set of eight objects that
are arranged in three-dimensional space as if they were at the corners of a long box-kite
(Permap's Special Case 7). Now use the Stress badness measure, the Euclidean distance
measure, two dimensions, and make a ratio MDS map of the eight objects. The resulting map
shows the outline of a long rectangle made up of four symbols in an almost straight line, parallel
to and remote from another four symbols in an almost straight line. This is a valid two-
dimensional picture of the three-dimensional box-kite. It faithfully shows the two remote
subgroups and some of the symmetry involved. However, it badly distorts the relationships
existing within the two subgroups. If mapping weights are applied (click the MAPPING WTs.
shortcut button twice), the MDS map immediately shifts to two squares remote from each other.
Which picture is correct? Both are, just as Picasso's art is as valid as Rembrandt's.
Another example of how the use of mapping weights can clarify hidden structure in a
data set is given in EXAMPLE_C.txt. Here, Churchill defines ten department store
characteristics to be the objects of interest. He determines the correlations between objects based
on a set of shoppers' impressions of different department stores. After loading Permap use Stress
badness measure, the Euclidean distance measure, two dimensions, and make a ratio MDS map
of the data in EXAMPLE_C.txt. You will find, as Churchill did, that there are two groups of
characteristics. One group has six members and the other four members. Next, use the
MAPPING WTs. shortcut button to rotate to using linear mapping weights. The group of six is
only slightly changed, while the group of four takes on a significantly altered internal
arrangement. Which picture is correct? As before, both are correct. It is up to the researcher to
determine if the new configuration contributes to understanding the topic.
Permap calculates mapping weights internally based on the values of the given Dij. File
entry of MWij values is not supported. The built-in mapping weight function is:
MWij = (1 - Dij) / (1 - m Dij)
where "m" is an adjustable parameter with a value less than or equal to one.
Page 36 of 77
Experience shows that most interesting behavior can be investigated using m = 1 (no
mapping weights, equivalent to MWij = 1), m = 0.95 (weak mapping weights), m = 0 (medium
mapping weights, equivalent to a linear decrease of MWij with increases in Dij), and m = -20
(strong mapping weights). Permap allows selection of these four standard mapping weight
functions by using the MAPPING WTs. shortcut button. Or, any arbitrary value of m can be
selected using the Analysis Parameters/Mapping Weights menu. This menu also shows a graph
of MWij vs. Dij and displays a table of the corresponding MWij values.
Note that the above mapping weights function goes to zero at Dij = 1 for all m not equal
to 1. This would allow independent object groups to "float away" from each other in a numerical
analysis sense. To avoid this a "floor" is introduced. The MWij floor is 0.1 for weak, 0.01 for
medium, and 0.001 for strong mapping weights. The variable case (see the Mapping Weights
screen) uses a floor equal to MWij calculated using the above equation with Dij = 0.98.
Finding alternate views of reality, as was demonstrated by the box kite example or the
department store comparison data of Churchill in EXAMPLE_C.txt, is one use of mapping
weights. Another is to check for conflicts that might exist between large and small dissimilarity
values. To demonstrate this point, load data file EXAMPLE_A.txt which contains the Ekman
(1954) color comparison data. Find the two dimensional map using ratio MDS, the Euclidean
distance metric, and the Stress badness function. You will get the standard circular figure that is
shown in many text books. Then, click the MAPPING WTs. shortcut button (use the View menu
to turn on this feature if it does not already appear on the main screen) twice to reach the
medium mapping weights” selection. This will reduce the influence (weight) of the large
dissimilarity values vis-a-vis the small ones. The map will shift significantly, thus showing that
there is a systematic conflict between the positions of the objects with large dissimilarities versus
the positions of the objects with small dissimilarities. Further analysis will show that the large
dissimilarities favor the ring shape and the small dissimilarities act to disrupt the symmetry of
the ring. If this was your research, you would next decide which picture is most appropriate for
the goals of your research. Or, perhaps you would decide that further experimentation is called
for in order to refine your estimates of the dissimilarities.
While considering the Ekman results, it is worthwhile to note the danger of reading too
much into circular patterns. The Ekman circle may be as significant as several articles claim, but
it could be that it resulted simply because it is easier to quantify differences between similar
colors than it is to quantify differences between dissimilar colors. If such were the case then
colored objects would have their near-neighbor relationships well represented while the remote
relationships would not be well represented. This form of non-linearity can result in meandering
chain-like patterns being turned into circular MDS patterns. For example, consider this
exaggerated case. A closed chain of objects has all of its nearest-neighbor dissimilarities coded
as zero and all of its non-nearest-neighbor dissimilarities coded as one. Then, as is shown by
analyzing the three different data sets in EXAMPLE_D.txt, a two-dimensional perfect circle is
found. The circle is a valid MDS solution, but it suggests much that is not necessarily true. As
the number of objects increases from three to, say, six or seven objects, the danger is even
greater. In these cases, a simplex matrix (i.e., a matrix with all non-diagonal dissimilarity values
equal) always yields a circular symmetry. This means that if your respondents are saying that
Page 37 of 77
the less similar objects are "all about equally different" you will discover a circular relationship
in your MDS map. Thus, one must be careful not to draw quick conclusions after finding objects
in a circular pattern.
5.4 Dimensionality and Boundaries

MDS analyses are not fully described until two more aspects are covered. You must
specify how many dimensions are to be used to describe the object placements, and what
boundary conditions (external constraints) are to be placed on the solution. These aspects can be
specified using the Analysis Parameters/Dimensionality and Boundaries submenu (ctrl-Y).
5.4.1 Dimensionality
Pressing the Dimensions shortcut button or using ctrl-Y to open the
Dimensionality and Boundaries screen allows you to control the dimensionality of your
MDS solution. Permap can analyze a problem using one to eight dimensions. That is,
the solution can be described in terms of each object's coordinates in any of eight solution
spaces. Permap can write out the object's coordinates to the PermapSolutionLongFmt.txt
file if you want to see the actual values. But, the important question is, since Permap can
display only one or two dimensional maps, why would you want it to make a three or
four dimensional analysis? The answer to this question may not be what you expect.
Let's take the easier of the two cases first. Why would you want to make a four
dimensional analysis? Given that four dimensional configurations cannot be visualized,
what good are they? MDS is inherently a visual tool. Providing a way to transform a set
of numbers into a picture is its reason for being. But in this case the visual nature of the
solution is not of interest. The answer is, we make the four dimensional solution not to
understand the data's hidden structure in four dimensions but to check on the validity of a
three dimensional solution. There are many ways to test the merit of a particular
solution. One can use the objective function value, a Shepard plot, a Pareto plot of
objective function components, Waern link plots, attribute gradient plots, and so forth.
Each of these is useful and should be part of a complete analysis. But the most direct
way to see if a three dimensional map is stable, and thus "good", is to remove the
restriction to three dimensions and see if any of the objects move. This is called
"relaxing the solution into a higher dimension." If a four dimensional solution gives the
same picture as a three dimensional solution, then you have strong evidence that three
dimensions are adequate for describing your data set.
Answering the similar question of why you would want to make a three
dimensional analysis follows the same logic as that for four dimensions. The only
difference is that, as opposed to four dimensional results, three dimensional results can be
visualized. So why does the above logic still apply? It applies because we need to
communicate our results to others and many times these other people, especially those
with the financial, organizational, or political clout to do something with the information,
cannot understand the meaning of a collection of points in three dimensional space. So,
in many cases, the value of making a three dimensional analysis lies more in its ability to
Page 38 of 77
help evaluate the quality of a two dimensional map rather than in the utility of the three
dimensional solution itself.
In light of the preceding two paragraphs, why would one ever want to solve
problems in more than four dimensions? An adequate reason is that it is fun to "look
around" in higher dimensions. However, a more common reason is to make a skree
analysis that is described in many MDS texts. A skree analysis involves solving a
problem using many different dimensions, then plotting each solution's Stress versus the
dimensionality of the solution space, and then taking the problem's optimal number of
dimensions, D*, to be the number of dimensions used at the plot's "elbow." The plot is
supposed to look like the skree (loose rock) and valley floor at the base of a mountain. It
is supposed to distinctly show D* where the two straight lines ("the skree line and the
valley floor line") meet. The trouble with this very reasonable approach is that it doesn't
work. It can be made to appear to work by illustrating the analysis using a hypothetical
data set that is well fit using D* dimensions. The tip-off that artificial data are being used
is noticing that Stress is very small when using D* dimensions. Unfortunately, real data
sets don't behave this way. From experience one can say that the Stress-Dimensions
curve will have a gradual bend, not a sharp elbow, and the analyst can pick D* to be
pretty much whatever number of dimensions he or she prefers. Admittedly, that is a bit
too harsh of a description of what happens with real skree analyses. To compromise, the
authors will happily soften their conclusion once they are sent a real, unmanipulated and
unprocessed data set that shows a clear break in the optimal solution's Stress vs.
dimensionality plot, AND the Stress using D* dimensions is not nearly zero.
Finding a good X-Y plot for a three or more dimensional solution is not always
easy. It is important to understand the difference between starting an analysis in, say,
three dimensions versus starting it in two and then relaxing to three. Because Permap
displays a two dimensional map, when three or more dimensions are being used it is the
projection on the X-Y plane that is displayed. (Caution, this does not mean that a two
dimensional solution is simply a projection from a higher dimension. This is a point of
confusion for some new users.) Whenever you start a new solution Permap assigns
random positions to the coordinate values of each object and then makes the analysis.
So, each different "run" produces a different "view." That is, the X-Y projection shown
is different even though the relative object positions are the same. These changing views
can be confusing at first, but they are very valuable because they give a much richer
understanding of the three dimensional shape of the object cluster.
If you prefer to always have the X-Y plane hold the maximum amount of
variation, and usually you will, then you can do two things. One, start the analysis in two
dimensions and subsequently relax it into three dimensions. Two, instruct Permap to try
to adjust the view for you. When making three or more dimensional analyses, Permap's
algorithm can be instructed to favor X-Y maps that have large cross-sections. Permap is
not always successful in this endeavor. Exercising this option is done by checking the
box "Attempt to keep the maximum ..." at the bottom right of the Dimensionality
explanation area. Beginners are advised to check the box.
Page 39 of 77
To better understand shifting views of a solution, load Special Case 3 which has a
dissimilarity matrix based on three connected perpendicular vectors in three dimensional
space. That is, the data represent the normal Euclidean axes near the origin. Use a ratio
analysis, Euclidean distance measurement, and check the "Attempt to keep the maximum
..." dimensionality option. Solve the problem in three dimensions. The map will most
probably show what you expect. That is, a figure with three-part circular symmetry.
Now hit Start numerous times. The figure will not change significantly, except for
irrelevant rotations and perhaps a few random local minima solutions. Permap does a
good job of keeping the maximum cross-section in the X-Y plane. However, if you look
very closely you may see some slight changes that are due to Permap's imperfect
orientation mechanism. Now uncheck the "Attempt to keep the maximum ..." option and
hit Start several more times. It will become immediately obvious that you are seeing the
three vectors in 3D space from widely differing perspectives. Sometimes you will be
looking down on one of the three vectors and it will effectively disappear. Sometimes
one or two of the vectors will be greatly foreshortened. Practice with this special case
will develop your ability to understand the role of viewpoint in understanding three
dimensional solutions.
Turning now to MDS analyses in one dimension, it should be noted that these
analyses are particularly difficult. Local minima abound. This is because the objects
cannot "go around" each other when seeking their best positions. If the objects cannot go
around each other then they would have to "go over" or "go through" each other to reach
their best positions, and that can make for difficult mathematics. There are two ways to
lessen this difficulty. First, the SStress badness function is more effective than any of the
other badness functions when it comes to letting objects get past blocks. If you need to
use one of the other badness functions, then first use SStress and after its solution is
found change to your preferred badness function. Second, make the analysis in two
dimensions, then use the field commands to rotate the map's longest span to the
horizontal, and then change back to one dimension and continue the solution. Neither of
these techniques will always work, but they are worth trying.
5.4.2 Boundaries
Pressing the Boundary shortcut button, or using ctrl-Y to open the Dimensionality
and Boundaries screen, allows you to apply boundary conditions (external constraints) to
your MDS solution. It is surprising that few MDS programs provide for placing even
simple boundary conditions on MDS maps despite the fact that the topic is well
developed in the literature. Lee and Bentler (1980) were one of the first that discussed
the subject and Borg and Groenen (1997, p. 182) provide a summary of work on the
topic. Applying boundary conditions, or constraints as they are sometimes called, is very
common in the work of physical scientists, but behavioral scientists have not placed
much emphasis on it. While it is true that many "conceptual space" MDS analyses do not
have obvious boundary conditions, many others do and the possibility of their existence
should be considered in every MDS analysis.
Page 40 of 77
The idea of boundary conditions is a simple one. For example, if you know all of
the person-to-person affinities of a group of people you can solve for their preferred
positions relative to one another. However, the solution implicitly assumes that there is
space available for the preferred positions and this may not be the case. Perhaps the
people are in a small room or maybe they must congregate in a room containing tables
that get in the way. Numerous similar examples are easy to describe when the objects
represent mobile physical entities. A completely different kind of boundary condition
example involves time. If you want to superimpose a time axis on a map, you might
want to constrain the map's surface to a "path" or ribbon-like configuration. That is, you
want to make a "what if" analysis where MDS gives the best possible configuration for a
set of dissimilarities and a time-flow hypothesis. You do not want to make a one
dimensional analysis because, you believe, it is possible to have multiple threads running
side-by-side so a path or multi-lane road is a more appropriate conceptual model. Maybe
you are studying certain hominid artifacts and you know that more than one hominid
species developed during the time period of interest. By constraining the map to a long,
narrow area you can quickly see what happens when you ask for a map that is compatible
with a time-flow hypothesis. A final example involves boundary conditions based on
theoretical arguments. It could be that, say, only four different object groupings are
theoretically possible. In this case you would want to have the MDS solution fit on a
map that allows objects in at most four distinctly different areas.
Permap allows studying the effect of three kinds of boundaries. First, it allows
imposing a bounding circle on the group of objects. In essence, the objects are embedded
in a space where you can control the size of the space. If a three dimensional analysis is
used then the boundary circle becomes a boundary sphere, and if a four dimensional
analysis is used then it becomes a hyper-sphere. This is necessary because if the nature
of the boundary did not change with the solution's dimensionality then objects in a higher
dimensionality could just move "off the plane" and completely avoid the constraint.
Second, Permap allows restricting the objects to a narrow path. The path becomes a tube
in three dimensions and a hyper-tube in four dimensions. Third, Permap allows
restricting the objects to a square area. The square becomes a cube in three dimensions
and a hyper-cube in four dimensions.
When applying a boundary one might think that all solutions will simply consist
of the objects being pushed up against the boundary. This is not always the case. Some
experimentation with the special cases built into Permap will reveal interesting behavior.
The implication for the MDS analyst is that when dealing with mobile objects it is
important to consider the possibility that constraints are limiting the (X, Y) positions.
This means that when constructing a MDS map sometimes it is important to be able to
apply boundary conditions that make the results fit both the Dij data and the environment
that surrounds the objects.
This discussion of boundaries might prompt you to ask “Do the constraints
imposed on certain (X, Y) values have to come from some physical surrounding barrier?”
The answer is “no.” For instance, legal rulings, family or societal mores, religious
Page 41 of 77
beliefs, or a host of other “forces” may dictate that certain members of a group be, for
instance, no closer than some given distance. Note that in all of these cases one or more
(Xi, Yi) positions are being influenced by an external force, not something that involves
changing any Dij values. Such restrictions can be viewed as forms of “boundary
conditions,” albeit not physical boundaries. This topic is touched on in the section on
“fixed locations” although in that section the language of boundaries is not used.
5.5 Dij Precision Degradation

Before discussing data precision, please note that if you are using an attribute matrix to
generate your dissimilarity matrix you should read Section 6.1 on attribute data level
degradation. This is because if you have attributes then they are the independent variables and
you should do "cleaning," "segregating" and “data precision evaluation” work using the
independent variables, not the dependent variables (dissimilarities in this particular case).
The topic of precision degradation (a subtopic of the much broader “data discretization”
topic) is important. This is because it is important to know if your Dij data's overall precision is
adequate to support your results. In other words, every MDS analysis should include a
sensitivity analysis.
Before we get into the details of Dij precision degradation, we must answer the question,
why are we addressing this subject again? Aren't the error bounds of Section 5.1 sufficient to
handle the matter? The answer is that, yes, using those techniques can do the job from a certain
perspective. But, the feedback we have received from many Permap users clearly shows that
these approaches are not being used as much as they should be. Also, MDS research articles
appear every month that contain no sensitivity analyses at all. This is an important flaw.
Because we believe making a sensitivity analysis is so important we have changed Permap to
make it easier. Starting with version 10.0, an on-line button has been added to Permap's main
screen to allow one-click degradation of the Dij data. This button is labeled PRECISION and can
be activated with a mouse click or by pressing the "E" key. (If the PRECISION button is not
present on the main screen then use the View menu to activate it.) With just a few clicks you can
degrade the Dij data's precision all the way back to its (0, 1) roots. If you do this while watching
your map, you can immediately draw conclusions about the sensitivity of your map to the
precision of your Dij data.
In order to be clear about Dij degradation, we must first review some definitions.
"Precision" refers to the degree of refinement with which a measurement is reported. Precision
can be characterized in several ways. One is to observe the number of significant figures that are
reported, another is to give an uncertainty measure, another is to use the implied rule of  5 in
the next (missing) decimal place to the right, and so forth. "Accuracy" refers to the degree of
conformity between a measurement and a true value. The true value may be known or unknown.
That is immaterial. The concept of accuracy defined in terms of a true value is still useful.
Finally, the first digit of an actual measurement that differs from a true measurement marks the
start of the "false figures." Anytime the precision of a measurement exceeds its accuracy there
are false figures involved. We may not know which figures are false, but we can take steps to
make sure that they do not invalidate our results. This is particularly important because in
Page 42 of 77
various of the “soft sciences” MDS is frequently applied to Dij data of low accuracy and high
precision. Clearly false figures are often, some would say essentially always, present.
How does one decide if a MDS map is influenced by false figures? One degrades the Dij
data until the false figures disappear. In effect, the increasing error bounds envelope the false
figures and they melt away into irrelevancy. Eventually, the error bounds become so large that
all critics agree that the true Dij values fall within the bounds. If the map has remained
unchanged during the degrading, then it must not be a result of the false figures.
But, if one does not know the true values for the Dij data then one must not know which
figures are false, so how can one know when to stop degrading the data? The answer is to
degrade the precision "beyond reason," or to degrade it "with your critics at your side."
Fortunately, MDS is a very robust procedure. That is, it produces results that are not sensitive to
small errors in its input data. Because of this, you will probably find that your MDS map is little
changed even after severe, "beyond reason," degradation of its Dij data.
Here is how Permap degrades the Dij data. Beneath the caption "PRECISION" on the
precision control button there is an entry that shows the current precision of the Dij data. Starting
with the maximum number of decimal places (or maximum number of digits if the input is not
already normalized to the zero/one range) in the input Dij data, Permap rounds down to the next
lesser number of decimals with each button click. After reaching the tenths-level it starts
rounding to the nearest 1/5, 1/4, 1/3, ½, and 1 intervals. At a final interval of 1, all Dij values
equal 0 or 1. Another click on PRECISION will return the precision level back to the maximum
precision consistent with the input data. It has "Max" appended to the numerical value as a
reminder that the input (maximum possible) value is in use.
For example, if D28 = 0.34567, after one click on the PRECISION button D28 will be
0.3457, then 0.346, then 0.35, and then 0.3. Once it reaches the tenths level, Permap begins to
degrade the precision using fractional step sizes rather than decimal step sizes. That is, it
reassigns the current value to the nearest number on a scale using step sizes of 1/5, 1/4, 1/3, ½,
and then 1. This means that, continuing with the example above, D28 will change from 0.3 to
0.2, then 0.25, then 0.33, then 0.5, and then to 0. In each case the step size between possible
values increases. For example, starting from input data showing five decimals, the measurement
interval (step size) varies as follows: 0.00001, 0.0001, 0.001, 0.01, 0.1, 0.2, 0.25, 0.33, 0.50, and
1.0. In each case the uncertainty, i.e.  half the measurement interval, increases in an analogous
manner. As an aside, one might ask why the 1/3 interval is taken to be 0.33 instead of, say,
0.333. The reason is that a general rule from the theory of error analysis is that the last
significant figure in any stated answer should be of the same order of magnitude as the
uncertainty (Taylor, 1997, p. 15).
The reason for having Permap degrade the precision using fractional intervals after the
decimal places are exhausted is that these step sizes, while extreme, correspond to experimental
conditions of interest to some researchers. For instance, if a person is asked to assign a
dissimilarity to two objects by placing a mark on a line segment that goes from 0 to 1, the
dissimilarity value may be reported to the nearest 0.1 value (depending on the length of the line
Page 43 of 77
segment, the width of the marker point, and the diligence of the experimenter). Our researcher
might be quite happy with the 0.1 precision data. But another researcher might claim that some
theoretical principle says that no more than 6 detectable states are possible under the given
conditions. This second researcher would insist on recording the dissimilarity values as 0.0, 0.2,
0.4, 0.6, 0.8, or 1.0. So, the first researcher might like to know if his or her MDS map will
remain stable if the dissimilarity precision is degraded from the 0.1 interval level to the 1/5 level
(five intervals, six possible values, 0.2 interval size). Permap’s Precision Button can help answer
the question.
Finally, a note of caution is in order for certain degraded precision levels. Sometimes
you will notice a significant change in a map at a precision level of ½. The ½ level can cause
changes in the map that are even bigger than those occurring when using a measurement interval
of 1. This occurs when the Dij data are clustered in the center of the 0 to 1 interval. Under these
circumstances the 0.5  0.25 range captures most of the dissimilarities and the Dij matrix moves
toward a Dij = 0.5 simplex. This behavior is simply an artifact of a poorly conditioned Dij
matrix.
6.0 ATTRIBUTES-TO-DISSIMILARITIES
If no attribute information (an Aik matrix) is provided in the data file, then this menu item
is disabled (grayed out). If attribute information is provided, Permap uses the attributes to
produce dissimilarities and then uses the dissimilarities to produce a map. In this case each
object can have up to 100 attributes.
6.1 Attribute Data Types

The first item on the left side of the screen is the Attribute Input Data Level box.
Use it to specify the type (i.e., quality or information level) of your attribute data.
Acceptable data types are ratio (metric), interval (metric), ordinal (nonmetric), nominal
(nonmetric), binary nominal (nonmetric), and "unknown" (both metric and nonmetric).
The ordinal level is also known as rank order, and the binary nominal level is also called
dichotomous.
Ideally, all of your various different attributes should be of the same data type.
For instance, if some attribute values come from Likert data, then all of your attributes
should come from Likert data or some other ordinal data of comparable precision. This is
the ideal case. In practice this rule is frequently violated. For instance, if you are using
the Pearson correlation coefficient it is flexible enough that you can feed it a mixture of
ratio, interval, and binary nominal level attribute data. Using binary data with the
Pearson formula is not theoretically correct, but the formula works just fine and Permap
happily complies with bad instructions of this kind. Other association, correlation,
distance, interaction, etc. coefficients are more specific and cannot be used except with
one type of data. Permap only goes part way toward solving the problem of mixed data
types. It facilitates using mixed types by coercing all the data types into the type that is
compatible with the selected attribute-to-dissimilarity function. These coercions are
Page 44 of 77
described in Section 6.1.1. This approach is acceptable sometimes and other times it is
clearly unacceptable. If this approach is not acceptable and if you have a mixture of data
levels in your attributes, then it is up to you to apply an appropriate attribute-to-
dissimilarity function to each different type of attribute, and then combine those single-
case dissimilarity values into an overall dissimilarity value. See Kaufman and
Rousseeuw (1990, p. 32) for a summary of methods for combining diverse attribute
information into one dissimilarity value.
When you choose an attribute data type Permap excludes using any attribute-to-
dissimilarity functions that do not fit that data type. For example, if you choose to use
interval level data then the COS(alpha) = Angular Separation function is prohibited
because the COS(alpha) function requires ratio level data. This exclusion is implemented
by disabling (graying out) all options in the Attribute-to-Dissimilarity Functions box that
are not designed to work with the specified data level, and by causing the Attributes
shortcut button on the main screen to cycle through only the allowed functions. As
another example, if you choose the ratio level option to describe your input data, then the
ratio and interval level attribute-to-dissimilarity functions are activated, the other
functions are grayed out, and the Attributes shortcut button is caused to cycle through
only the appropriate interval and ratio level functions. The same kind of behavior applies
to the other data level selections.
If you choose the "Unknown" data level option, then all attribute-to-dissimilarity
functions are allowed. Of course, you should always know the nature of your data.
However, the Unknown option has some useful aspects. First, because most MDS users
are inveterate experimenters they like to see what happens when the rules are violated.
This option is great for doing a little "looking around." Second, as will be explained
later, sometimes an expert MDS user will have a valid reason to downgrade the attribute
data to be able to use a lower level attribute-to-dissimilarity function. Or, perhaps this
would be done to force one’s attribute data to be of one consistent level. After describing
how Permap coerces data into different data types we will discuss further why an
experienced MDS researcher might want to use the Unknown option and downgrade the
attribute information.
6.1.1 Changing Data Types

If you select an attribute-to-dissimilarity function that does not agree with
the data type of one or more of your attributes, Permap converts (coerces) the
mismatched Aik data into the required format.
Up-Conversions
Up-conversion is somewhat controversial, often interesting, sometimes
provides hints about how to improve future data collection, and equally often
yields nonsense.
Up-conversions are those conversions that treat lower level data as if they
were higher level data. Sometimes this involves nothing more that ignoring
Page 45 of 77
reality and sometimes it involves valid manipulations of the data. For example,
using interval level data in a ratio level function such as COS(alpha) involves an
up-conversion of the first kind. It usually produces nonsense. On the other hand,
adding a constant to interval level data in a manner that shifts it to ratio level data
(think of converting temperatures on the interval Celsius scale to the ratio Kelvin
scale by adding 273.15) involves an up-conversion of the second kind.
Here is how Permap makes up-conversions. When moving from binary

nominal level Aik data to a higher level of data Permap just pretends that the
binary nominal level data are ratio, interval, ordinal, or nominal and proceeds to
use the formulas given in Section 6.2 to calculate the dissimilarities. That is,
either it treats the zeros and ones of the binary data as if they were the results of
measurements that yielded zeros and ones (for ratio, interval, and ordinal
formulas), or it treats them as the names "zero" and "one" (for nominal formulas).
Up-conversion from nominal level Aik data to metric data involves the
esoteric matter of ordered lists and unordered lists, and how ordered lists differ
from ordinal information (answer, only in their bound set of functions). Because
Permap uses numbers as names for nominal data, Permap makes these up-
conversions automatically. Equally automatically, the results are usually
nonsense.
Up-converting ordinal or interval level Aik data to ratio data is as easy as

upgrading binary nominal data. Permap just pretends that the numbers are more
than what they really are. If a formula is used that requires ratio level data but
lower level data is entered, nonsense often results. This is because ratio data must
have a meaningful zero and the lower levels of data may do not even
approximately satisfy this requirement.
In summary, up-conversions are occasionally useful to experienced MDS

users and probably should be avoided by beginners.
Down-Conversions
Down-converting attribute data is analogous to precision degradation of
dissimilarity data. They are much more interesting, and more useful, than up-
conversions. They produce rigorously correct results, even though the results do
not exploit the full potential of the information contained in the data. Basically,
down-conversions involve ignoring some information content of the data. The
technique possibly provides a way of strengthening your conclusions. We will
discuss this more in the following section after the nuts and bolts of Permap's
conversions are described.
Downgrading from ratio level Aik data to interval level data requires no
action on Permap's part. You just start using the interval attribute-to-dissimilarity
Page 46 of 77
formulas. All the interval level formulas work off differences in attribute values
and thus the attribute's zero point is irrelevant.
Downgrading from ratio or interval level Aik data to ordinal level is done
by first sorting the data into an array with the smallest values first. Ties are
broken arbitrarily. Then, the first data group is defined to contain the entry with
smallest value, or the first several entries where each value is less than or equal to
0.0001 more than the lowest value found in the group. Next, the previously
grouped values are ignored and the grouping process is repeated. This continues
until all entries are assigned a group. Finally, each attribute value is exchanged
for the group's identification index (rank).
In coercing ratio, interval, or ordinal level Aik data to nominal data the
original data are rounded to the nearest integer, and the integers are treated as if
they are names. The new "nominal" data will exhibit a greater coarseness than the
original data. The degree of degradation is dependent on the range of the original
data (you can increase the range by multiplying the data by a constant). If the
original data are all integers, for example, there is no information loss. However,
the map still will change because different types of attribute-to-dissimilarity
formulas are now used.
Finally, ratio, interval, ordinal, or nominal level Aik (remember that the
nominal names are represented internally by numbers) data can be coerced into
binary nominal data. There are two common ways to make the change and the
choice is made using the Conversion to Binary option box at the bottom of the
left-hand panel of the Dissimilarities-from-Attributes selection screen. The first
option is to treat any zero value as binary zero, and any non-zero value as a binary
one. The second is to treat any value less than or equal to the average value as
binary zero, and all others as binary one. The average value is the average over
all possible objects of the attribute in question. That is, it does not change if some
objects are removed from the active data set by parking them.
6.1.2 Examples of Down-Conversions

Here are several examples of down-conversion to illustrate the rules given
above. If your original attribute values are 0, .3, .6, 1.2, 55, and 11, they will
become 0, 1, 2, 3, 5, and 4 when converted to ordinal level, and become 0, 0, 1, 1,
55, and 11 when converted to nominal level. If converted to binary nominal level,
they will become 0, 1, 1, 1, 1, and 1 if the first metric-to-binary option is used,
and 0, 0, 0, 0, 1 and 0 if the second option is used.
If your original attribute values are 0, .3, .6, .2, .1, and .5, they will
become 0, 3, 5, 2, 1, and 4 when converted to ordinal level, and 0, 0, 1, 0, 0, and 1
when converted to nominal level. If converted to binary nominal level, they will
become 0, 1, 1, 1, 1, and 1 if the first metric-to-binary option is used, and 0, 1, 1,
Page 47 of 77
0, 0 and 1 if the second option is used. Clearly, the range of the metric data can
play an important role in the conversion process.
6.1.3 Rational for Using Down-Conversions

Down-conversion is a valid procedure that should be understood by all
MDS analysts. MDS is a very robust technique and sometimes down-conversion
does not significantly change the final map. In these cases, your conclusions are
strengthened because they are seen to be subject to less restrictive assumptions
about data quality than they originally were.
For instance, say your attribute data consists of the averages of the
responses of a group of tribal elders' judgments about the importance or
occurrence frequency of a list of factors associated with numerous tribal
ceremonies (objects). You have recorded all sorts of factors (attributes) and had
the tribal elders use a Likert-like scheme where the bottom category of each
factor's range was of the form "zero importance," "little importance," and so forth.
Furthermore, you have arranged the check boxes in a manner so that there is an
equal change in factor value between boxes. Based on your abundant confidence
in your survey construction expertise, you tell Permap to assume that your data
are ratio level and make the MDS map using COS(alpha) attribute-to-dissimilarity
function (most likely a mistake but that is beside the point here). However, your
research advisor says you cannot do that because you don't know that a
checkmark in the bottom box truly means that the value of the factor in question
was zero. For instance, maybe the tribal elders classified very unusual events as
being of zero importance just because they were rare events. You disagree with
your advisor. Instead of arguing the point (smart move) you reinterpret your
attribute data as interval level data and redo the analysis (i.e., just click the
Attributes shortcut button) using a less restrictive formula. Your results likely
will not change at all and as a result you are in a stronger position to support your
conclusions. Now, say your advisor further complains that you do not truly know
that the intervals between the Likert levels represented by adjacent check boxes
are equal. Again, you should reinterpret your attribute data as being ordinal level
and redo the analysis (i.e., just click the Attributes shortcut button). Your MDS
map may or may not remain stationary.
This line of reasoning can be extended to include the radical case of

degrading your supposedly ratio level attribute data all the way down to the
binary nominal level. For instance, a doctoral thesis examiner once was observed
exclaiming that a candidate's data were so unreliable that they should not be
called "measurements" at all. He stated that they were only indicators of the
existence or nonexistence of the event under study. By now you should know
what the candidate should have done before his thesis defense. He should have
reinterpreted his data as binary nominal level data and checked to see if it
mattered to the resulting MDS map. Depending on the nature of the data, and
Page 48 of 77
critically on the number of data points, the resulting map may not change at all
and the candidate could have avoided a most undesirable confrontation.
Can any real data set undergo such severe down-conversion (precision
degradation) and still yield the same MDS map? Many cannot but some can.
Sometimes a very great amount of attribute data degradation causes almost no
change in the MDS map. Whether or not the map is immune to precision
degradation depends on the dimensionality of the data's hidden structure. If there
really is a low-dimensional structure in the data, then the dissimilarity data are
highly redundant and the results will be insensitive to some measurements being
eliminated or having their precision degraded. This redundancy is expensive to
the analyst in terms of the effort needed to produce the dissimilarity matrix, but it
is beneficial when it comes to the analysis stage.
It should be clear that no magic is involved in the robustness of MDS

results. Robustness should be expected whenever the MDS map shows a low
stress configuration. The dissimilarity table has N (N - 1) / 2 values describing
object-to-object relationships. However, if the data are well fit in two
dimensions, then no more than 3N-6 (for N>2) object-to-object relationships are
needed to construct an equivalent map (translations, rotations, and reflections are
ignored). So, for example, if there are 15 objects then of the 105 dissimilarity
values only 39 are really needed. Furthermore, if you know, say, that the objects
are controlled by the value of two attributes, then even less information is
required. In this case no more than 2N-2 data points are needed to fully specify
the map and only 28 of the 105 original points are needed. Because of this great
amount of redundancy it should be no surprise that data containing two-
dimensional hidden structures can suffer significant degradation, or contain a
significant amount of random error, or have many missing values, without the
map significantly changing. That is one reason why MDS is so liked by so many
data analysts.
We have discussed only one of two ways that the effect of uncertainty in
attribute data should be evaluated. In addition to considering changes in attribute
information level (i.e., ratio, interval, …) one should consider each attribute’s
value imprecision and determine how this imprecision might affect the MDS map.
This second approach might involve assigning error bounds to each attribute
value and then using the rules of the theory of error analysis to flow these errors
through to the resulting dissimilarities and then to the object positions on the map.
This is almost never done. Actually, it is probable that it has never been done. To
get useful results one would have to assign probability distributions (using ranges
yields unreasonably pessimistic results) for the uncertainties present in each
attribute value. Obviously, this approach is very difficult for problems of any
significant size and we know of no MDS program that supports this kind of
sensitivity analysis.
Page 49 of 77
6.2 Attribute Functions
The Attribute-to-Dissimilarity Functions box lets you pick a particular attribute-
to-dissimilarity function that is consistent with your data type. The only advantage to
making a selection using this screen rather than using the main screen's Attributes
shortcut button to cycle to your preferred choice is that the text box on the right side
gives the mathematical formulas that are used. This assumes that you recognize i and j as
object subscripts, k as an attribute subscript, that |x| represents the absolute value of x,
that Rngk = Range of Ak = MAX(Ak) - MIN(Ak), that r = the product moment correlation
coefficient, and that N is the number of valid cases (which is NOBJECTS from your data
file unless you have missing data or parked objects).
1. Ratio Level Attribute Data Required

Function Formula
COS(alpha) Dij = 1 - Cosine(alpha)
Alpha is the angle between attribute vectors i and j. This is sometimes called the
Angular Separation coefficient.
2. Interval Level or Better Attribute Data Required

Function Formula
Euclidean (unnorm.) Dij = SQRT {SUM(all k) {(Aik-Ajk)2 }/ N}
Euclidean (normalized) Dij = SQRT {SUM(all k) {(Aik-Ajk)2 /
RANGE(Ak)2}/ N}
City Block (unnorm.) Dij = SUM(all k) { |Aik-Ajk| } / N
City Block (normalized) Dij = SUM(all k) { |Aik-Ajk| / RANGE(Ak)}/ N
Guttman Dij = (1 - mu2) / 2
Pearson Type 1 Dij = (1 - r) / 2
The Pearson r is the usual product moment correlation coefficient. The Guttman
mu2 monotonicity coefficient is less well known; it is sometimes used with
ordinal analyses. Occasionally it is appropriate to use a Person Type 2
transformation using Dij = 1 - |r|. This must be determined by the nature of your
data.
3. Ordinal Level or Better Attribute Data Required

Function Formula
Spearman Dij = (1 - r) / 2
The Spearman r is the same as the Pearson product moment correlation

coefficient, except it is applied to the ordinal indices (ranks) instead of directly to
the attribute data.
4. Nominal Level or Better Attribute Data Required

Function Formula
Nominal SMC Dij = 1 - A / N
Page 50 of 77
The count A is found by counting the times the attribute names (i.e., the assigned
numbers given to Permap) of objects i and j are both are the same. N is the total
number of valid cases. It equals the number of different attributes used to define
an object unless there are missing attributes.
5. Binary Level or Better Attribute Data Required

Function Formula
Jaccard Dij = 1 - A / (A + B + C)
Gower/Russell/Rao Dij = 1 - A / N
SMC/Sokal/Michene Dij = 1 - (A + D) / N
Hamman Dij = 0.5 + 0.5 (A - B - C + D) / N
Yule Dij = 1 - (AD - BC) / (AD + BC
Askin/Charles Dij = 1 - MAX{A / (A + C), A / (A + B)}
The counts A, B, C, and D are found by summing the occurrences of the four
possible cases. That is, A is incremented each time the binary attributes of objects
i and j are both are 1, B is incremented each time the first attribute is 1 and the
second attribute is 0, C is incremented each time both are 0, and D is incremented
each time the first is 0 and the second is 1. This is usually illustrated using a 2-
by-2 contingency (association) table where A and B are the first row entries, and
C and D are the second row entries. A + B + C + D = N, the total number of valid
cases mentioned above.
6.3 Using Attribute and Dissimilarity Data

The last option in the Attribute-to-Dissimilarity Functions box is not often needed
but is very useful in some special cases. This option is disabled (grayed out), and it does
not appear in the Attributes shortcut button's rotation list unless you have both
dissimilarity and attribute data in your data file. If you independently know dissimilarity
and attribute values (most unusual, count your blessings), then it is important to know if
they provide consistent results. By selecting the Use Provided Dissimilarities you tell
Permap to use the dissimilarity data rather than calculate them using the attribute-to-
dissimilarity formulas. Then, by clicking the Attributes shortcut button, you can see if
switching back to using the attributes makes any difference.
6.4 Euclidean and City Block Attribute Normalization

The interval level Euclidean attribute-to-dissimilarity formula is used frequently.
However, it and the City Block formula, come in several varieties each differing by their
normalizing factor. Note that this is quite different from all the other formulas presented
above. All of the other formulas are defined as either normalized quantities or their
normalization is inherent (e.g., the "counting" formulas).
The key question in designing Permap was whether or not to offer the
unnormalized (sometimes called "raw") forms of the Euclidean and City Block attribute-
to-dissimilarity formulas. While the unnormalized forms can lead to mischief in the
Page 51 of 77
hands of a novice, they are important to the expert. For instance, if Permap were to insist
on using only normalized forms, then experts would be prohibited from complete control
over their analysis. They might want to, say, enter a set of data where each object has
two attributes that are based on the same units of measurement but one has a range of 0.5
and the other 1.0. There may be good reasons why some experts want the differing
ranges, and they do not want their data normalized by Permap. The experts are correct in
insisting that they be able to control the treatment of their data. Thus, despite the
possibly of misuse, Permap includes the unnormalized Euclidean and City Block forms.
It is also important to recognize that the literature contains several different

techniques for normalizing the Euclidean and City Block formulas. The three most
common methods are (1) dividing by the range, (2) dividing by the average absolute
deviation, and (3) dividing by the standard deviation. All three approaches are valid and
all three save those new to MDS from using input that results in Dij values greater than
one, which in turn forces the rescaling of all Dij values back to a maximum of one and
may result in a serious imbalance between the relative influences of the various attributes.
In order to keep the Permap interface simple, and because the three normalizing
techniques produce similar results, only the range-normalized equations are used.
Control over which form, normalized or unnormalized, Permap uses is achieved

by using the option box in the middle of the left side panel on the Attributes-to-
Dissimilarities selection screen. If the Unnormalized Form option is selected, then no
changes are made to your attribute data and the unnormalized equation will be used. In
this case it is solely up to you to provide properly scaled attribute data. On the other
hand, if the Normalized Form option is selected, then the Euclidean and City Block
formulas are normalized by range and you can enter attribute data based on any desired
scale.
7.0 MAP EVALUATION
Harshman (1984) wrote an important paper addressing how to check the appropriateness
of MDS results. He discussed the importance of not making an analysis until questions about the
theoretical appropriateness of the MDS model have been considered. That is, is there reason to
believe that the relationships between objects should be properly represented by distances? If
not, then MDS may not be the appropriate tool for the analysis. Torgerson (1958, 1965) also
addressed this point. These papers should be the starting point for those new to MDS analysis.
Once you have a map, you must address the question Kruskal (Schiffman, Reynolds, &
Young, 1981) asked, "How well do we know what we know? How do we spot warning signals -
- and bring them effectively to the attention of first-time users?" While the situation has
improved since 1981, it is still true that MDS maps often are misinterpreted and artifacts are
mistaken for relationships. It is important to qualify a map before undertaking the painstaking
effort of defining and interpreting groups and assigning dimensions.
Page 52 of 77
Several map evaluation tests are given in Sections 7.1, 7.2, and 7.3. All of these tests
have to do with the validity of the two-dimensional representation of multi dimensional data.
They do not address the interpretation of the map.
7.1 Objects Evaluation Screen

If some object has a high relative contribution to the value of the objective function or to
its overall out-of-plane tendency (to be defined in the following paragraphs) then there is reason
to believe that its placement on the map does not result from a two-dimensional association with
its neighbors. This can be tested by examining a Pareto Plot of the objective function
components, or by two other ways that involve superimposing information on the map. Each
option has context-sensitive information provided in a "Notes" text box at the bottom of the Map
Evaluation/Objects Evaluation screen. The Evaluation screen can be reached by using the menu
system or by the ctrl-O shortcut.
Printing a map with the evaluation markings on it is facilitated by having these markings
persist after the evaluation screen is closed. See the Section 6.4.1 for the various printing
techniques. After closing the evaluation screen if the Start or Continue button is pressed, or an
open space is clicked, all evaluation markings are removed and Permap operation is returned to
its usual mode.
7.1.1 Pareto Plot

A Pareto plot is a bar chart showing some important quantity on the Y-axis and its
associated name, index, type, etc. on the X-axis, where the events on the X-axis are
presented in the order of decreasing Y values. To MDS users the contribution to the
objective function, or the out-of-plane tendency, is the quantity of interest.
Pareto plots are used in several disciplines. They serve to keep focus on the most
important events and give an indication of the nature of the distribution of some
important characteristic. The Pareto Principle, sometimes called the 80/20 rule, is often
stated as "80% of the results (profits, complaints, fines, etc.) come from 20% of the
sources (products, customers, accidents, etc.)." Sometimes it is stated in the inverse
format, but regardless of statement format it emphasizes that event frequency is not
uniformly distributed across causal roots.
In passing, those new to MDS might benefit from a few comments on a related
point. The Pareto Principle forms part of the foundation of MDS. Were it not true that
many situations are controlled by the "significant few" rather than the "insignificant
many" factors (another way of stating the Pareto Principle) it would make little sense to
use MDS at all. This is because nobody can meaningfully interpret or communicate the
results of a structure in four or more dimensions, and without interpretation and
communication there is little value in making the analysis. So, one can say that MDS
users have faith that their data has a significant hidden structure and that it can be
substantially described using two (three, in rare cases) dimensions. No theory says that
Page 53 of 77
this must be so, but we proceed as if it is so. Fortunately, a vast amount of experience
suggests it is so in many cases.
It is useful to identify outliers on the Pareto plot. To do this, Permap adds

average, 1% probability, and 5% probability lines. The one-tail critical probability lines
are set at the mean plus 2.33 standard deviations and at the mean plus 1.65 standard
deviations. These lines show, approximately, the values where the likelihood that an
occurrence is due to random chance is less than one percent, and five percent,
respectively. It is important to note that the classification resulting from these lines is
nothing more than a useful fiction because no evidence has been given that the
underlying distribution is normal (i.e., a Gaussian or bell curve distribution). More
information on this is given in the next section.

In addition to a Pareto plot, the Objects Evaluation screen lets you evaluate the
map's two-dimensional fit by superimposing information on the mapped objects. Two
kinds of information can be displayed on the map. The first is in terms of contributions
to the objective function and the second is in terms of out-of-plane tendencies.
7.1.2.1 Objective Function Contributions

Each object's contribution to the objective function can be displayed three
different ways. First, the range of all objects' contributions to the objective
function can be split into quarters, and each object colored according to its place
in the range. Black objects are the worst. They are in the highest quarter-range in
terms of driving up the objective function. The second highest quarter-range is
colored light gray, then gray, and then clear. Sometimes a quarter range is empty.
For instance, if all objects contribute only V1 or V2 amounts where V1 > V2, then
the objects contributing V1 are black and the rest clear.
Second, outliers can be identified. This approach is useful if you do not

care if an object's contribution is above or below the average, but instead care
whether or not an object's contribution is significantly different from those of all
other objects. Making the determination of statistical outlier status is done using
the usual one-tail rules that if a value is greater than 2.33 standard deviations
above the mean then the chance that its occurrence is due to random variations is
less than one percent, and if it is 1.65 standard deviations above the mean then the
chance is less than five percent. The "one percenters" are colored red and the
"five percenters" are colored black. It is important to note that this classification,
while very common, is just an approximation because no evidence has been given
that the underlying error distribution is normal. The use of Chebyshev
inequalities would produce more rigorous results, but they tend to be overly
conservative and are never used to our knowledge.
Third, the relative value (i.e., when all values are rescaled to a zero-to-one
range) of an object's contribution to the objective function can be superimposed
Page 54 of 77
on the object symbol. If the map is uncluttered, this is a useful technique, but the
first two techniques quickly gain superiority as the number of objects in the
analysis increases.
7.1.2.2 Out-Of-Plane Tendencies

Instead of considering an object's contribution to the objective function,
you might ask about its propensity to leave the plane of the map when making a
two-dimensional analysis. Analogously, what is its propensity to leave the line in
a one-dimension analysis or leave 3D space in a three-dimensional analysis?
Asking about out-of-plane tendencies is fundamentally different from

asking about the object's contribution to the objective function. It is possible that
objects might be highly stressed (i.e., the minimum value of the objective function
is large) even though their fit is not improved by making the map in a greater
number of dimensions. For instance, consider a tetrahedron sitting on a plane
where P denotes the object at the peak. Assume the dissimilarities between object
P and the other three objects are maintained at equal values, and all are slowly
decreased. At some point the four-object map fits perfectly (zero objective
function value) in two dimensions. The tetrahedron will have become a planar
figure. If the same three dissimilarities are further decreased then the stress will
begin to increase but the configuration of objects will not want to distort into the
third dimension. In this case the value of the objective function is not sufficient to
describe the situation. What is needed is a measure of an object's out-of-plane
tendency (OOPT).
OOPT can be calculated as follows, where we use terminology for the

two-dimensional case. The analogous formulas for higher or lower dimensions
follow immediately. Instead of writing the objective function just in terms of the
planar coordinates (X, Y), add a perpendicular coordinate, Z, to the equation.
Assume the badness measure is Stress, the distance measure is Euclidean, and that
a ratio MDS analysis is of interest. Then, for any given object, one can show that
the first partial derivative of the objective function with respect to Z is always
zero when evaluated at Z = 0. The second partial derivative is not always zero,
and therefore it controls the rate of change of the objective function at Z = 0. That
is, the second partial derivative provides a measure of the tendency of the object
to leave the X-Y plane if the two-dimensional map is slowly relaxed into a three
dimensional map. So, OOPTi can be measured by some negative constant times
the second derivative described above. A useful normalization constant is the
sum over all i and j of Wij. By convention, Wii = 0. The end result is the
following formula:
OOPTi = SUM(all j) (Wij (Dij-dij) / dij) / SUM(all j) Wij
This formula is strictly applicable for only ratio MDS using Stress and the
Euclidean distance measure. However, it is a useful approximation when dij is
Page 55 of 77
determined by any Minkowski distance measure, even including the City Block
measure. Moreover, if the Dij term is exchanged for the appropriate error or rank
bound, and OOPTi is noted to be zero if dij lies within both bounds, then it is
useful for all types of MDS analyses. So, rather than develop separate OOPT
equations for each MDS variation, we simply use the above equation. In effect,
we shift from using a derived OOPTi relationship to using the above equation as a
definition of OOPTi.
Following the approach used for each object's objective function

contribution, OOPTi is displayed in three different ways. First, by coloring the
objects according to the quarter-range of all OOPT values that OOPTi falls in.
Second, by calculating whether or not OOPTi is an outlier, as defined previously,
and highlighting the 1% and 5% outlier objects. Third, by displaying the relative
OOPTi value (i.e., OOPTi divided by the maximum absolute value OOPTi)
superimposed on the object symbol. See Section 7.1.2.1 for the procedures that
are used to display these categories.
Finally, new users might be interested in a review of what to do if some

objects have outlier OOPT values. Of course, the first step is to check all input
data for calculation or transcription errors. Next, check to be sure the
experimental conditions were appropriate for the outlier points. If these checks
fail to resolve the situation, then you are faced with a more fundamental problem.
Other than just "living with it" you have two choices. They are to segment your
data set or move to a higher dimension.
Removing offending data points is disagreeable to most researchers. The

tacit assumption is that if you just understood the underlying theory better, you
would understand how to form a new construct that would bring the outlier into
conformance with the other data. Even worse, an ethical question arises
whenever data points seem to be arbitrarily eliminated. Nevertheless, sometimes
it is exactly the right thing to do. Earlier, when describing the parking lots, an
example about dogs and cats, versus a mixture of dogs, was used. You cannot
know at the start of a study what you are going to find. And, the more refined
your study, the more likely it is that you are going to discover unanticipated
nuances. We recommend that you use the objects parking lot liberally to find a
subgroup of objects that is well represented in two dimensions. Then, repeat the
process on the objects that were parked. After both data sets are fully described,
search for a theory that recombines them into a unified whole.
If data segmentation is not successful, then you should consider going to a

higher dimensional analysis. Our advice is to go no further than three dimensions
except with the most unusual cases. While it is tempting to do a little "fishing" in
higher dimensions, it is rare that one finds experienced MDS users there. Coxon
quotes Shepard as saying "If a solution exists, it probably exists in two
dimensions; if it does not, then it certainly exists in three” (1982, p. 87). There
Page 56 of 77
have been many researchers that have discovered higher dimensional results that
were never published because of the difficulties of fully communicating their
nature. Remember, object groupings must be meaningful, reasonable, and
interpretable or they are of no value regardless of the statistical fit that has been
achieved. As Schiffman, Reynolds, and Young, (1981) point out, "dimensions
that cannot be interpreted probably do not exist."
7.2 Links Evaluation Screen

The word "link" refers to any kind of connection or relationship between two objects. In
this case, a link has two aspects: the distance between the objects, and the dissimilarity given for
the objects. So, a link analysis involves inspecting the distances between objects in relationship
to their underlying dissimilarities to determine if the map faithfully conforms to the underlying
data.
The Links screen provides a Shepard plot and three on-map analysis options that examine
the state of individual links. Each option has on-line context-sensitive information provided in a
"Notes" text box at the bottom of the Links screen. This option can be reached using the menu
system or the shortcut key ctrl-L.
Printing a map with the evaluation markings on it is facilitated by having these markings
persist after the evaluation screen is closed. See the Section 6.4.1 for the various printing
techniques. After closing the evaluation screen if the Start or Continue button is pressed, or an
open space is clicked, all evaluation markings are removed and Permap operation is returned to
its usual mode.
7.2.1 Shepard Plot

The Shepard plot is a very good way to determine the fit of an MDS map. For
metric MDS it involves nothing more than plotting Dij on the X-axis and dij on the Y-axis
(some authors reverse the orientation) and observing how the points scatter about a 45
degree line. For ordinal MDS, it involves finding a monotonically increasing line that
represents the dij vs. Dij data, and then observing how the points scatter about this
monotone line.
There are three important aspects to consider when using the Shepard plot. First,
the data points should cluster about a central 45 degree line if making a ratio or interval
MDS analysis, and about the center monotone line if making an ordinal MDS analysis.
This clustering can be evaluated by R2, the percent of the variance that is explained by
using the central line as opposed to that explained by using the average. R2 is the
coefficient of determination. Second, there should be no outliers. This is checked by
observing if any points fall outside two standard deviation bounds that are plotted
adjacent to the central line. Third, there should be no evident pattern in the distribution
of points. If these three tests are passed, then you have a good fit.
Page 57 of 77
A technical detail of an ordinal Shepard plot involves describing the procedure for
defining the central monotone line. We take the middle value to be the average of the dij
values on the lower and upper monotone hulls. The upper monotone hull's points are
found by starting with the lowest Dij value (the first X-value) and defining its
corresponding dij value to be the Y-value of the first point on the upper monotone hull.
Then the next higher Dij is considered. If its dij value is above that of the just previous Y-
value, then it is defined to be the next Y-value on the upper hull. If it is equal to or less
than the just previous Y-value, then the next Y-value is set equal to the previous Y-value.
This procedure is repeated on up to the largest Dij value, thus defining the upper
monotone hull. A reverse procedure is used to define the lower monotone hull. The
central line is obtained by simply averaging the upper and lower values. This central line
provides a clear, easily calculated, representation of the ordinal solution.
Beginning MDS users might consider this advice. Making a good analysis of a
Shepard plot is far better than comparing the objective function value to some critical
value. In fact, the "critical values" of Stress quoted in the early MDS literature are not
reliable at all. These values are fundamentally wrong, not just wrong in value. The key
distinction is that while they are rigorous for showing object placement improvement for
a given problem, they are not rigorous for showing the absolute merit of object placement
between problems. For many years this was a sensitive point. Applied users wanted
guidelines to quote to support their results and they perpetuated the use of certain critical
minimum values of Stress to confirm that a map was valid. But, the more theoretically
inclined researchers knew that the critical-values idea was fatally flawed. For instance, it
has been shown that acceptable values of Stress vary with the number of objects, the
number of dimensions, and the nature of the error in the data. These conclusions are
supported by Spence and Graef (1974), Wagenaar and Padmos (1971), Isaac and Poor
(1974), DeLeeuw and Stoop (1984), Arabie (1973), Borg and Groenen (1997, p. 38).
Arabie (1978, p. 111) stated the point most directly in saying "Another point that in the
interim has ceased to be arguable concerns the "good", "fair", etc. labels applied to Stress,
irrespective of the number of stimuli, dimensionality, Minkowski metric, and other
considerations. Those labels . . . have since been discarded by most users as unrealistic . .
. ." The bottom line is, finding the minimum value of an objective function is very
important and relative changes of the objective function between alternative mappings of
the same data are important. But, the exact value of an objective function does not mean
much. You are better advised to focus on the Shepard plot if you want to support your
claim that you have a good MDS solution.
Despite the forgoing it is likely that you will need to quote some stress values to
get your work published. It really doesn’t matter that these values are not reliable when a
journal reviewer demands that they be shown and interpreted, and this is very likely to
happen. Our advice is to focus on the Shepard plot and then simply state the value of
SStress and either Stress or Stress1. As to their interpretation, you probably should quote
Kruskal’s 1964 “rules-of-thumb” for Stress1 values are: 0 – perfect, 0.025 – excellent,
0.05 – good, 0.1 – fair, >= 0.2 poor” and then very briefly mention that these are rough
estimates that are known to not be applicable in all situations. It will do little good to
Page 58 of 77
mention that Kruskal later indicated that these Stress1 values were not reliable quality
indicators, or to mention that there have been many studies that show better critical
Stress1 values FOR CERTAIN SITUATIONS, or to mention that there has been no study
that shows a set of critical Stress1 values that apply to all situations. Our experience is
that it is counter-productive to argue these points with those not well versed in MDS.
7.2.2.1 Waern Links

Waern (1973) suggested using a cutoff value for discarding dissimilarity
information to simplify the job of mapping the most important structure hidden in
a dissimilarity matrix. She was able to reduce several complex similarity matrices
to fairly simple maps by using only manual calculations. Her critical values often
fell in the 0.45 to 0.65 range. Her method didn't preserve the metric level
information in the data, but it did a good job of showing some of the hidden
structure using only the simplest of procedures. As a consequence of Waern's
work, maps that show only some of the dissimilarity relationships sometimes are
called Waern link maps. Later, Kruskal and Wish (1978) suggested highlighting
links with a similarity values greater than 0.3 to reveal "horseshoe" structures
embedded in MDS plots. Even later it became common to use a cutoff value of
about 1/3 and apply it "both ways." That is, first highlight the links in the lower
1/3 of all dissimilarity values, and then highlight the links in the upper 1/3 of all
dissimilarity values. Using the "larger and smaller thirds" rule leads to two very
simple Waern plots that test the basic plausibility of a solution.
When using the "smaller third" Waern plot, the connecting lines should
not strike out across the map. They should fill in the small spaces between
adjacent and near-neighbor objects. When using the "greater third" Waern plot,
the connecting lines should criss-cross the map, connecting only the objects that
are farthest from each other. These two tests, while admittedly elementary, still
have merit. If these Waern plots don't make sense, then further analysis is not
needed. It simply means that a low-dimensional representation of the
multidimensional data is not adequate.
Finally, some analysts like to use Waern cutoff values different from the
standard upper and lower 1/3 levels. Thus, Permap has a slider just below the
Waern option buttons that allows resetting the cutoff to any value. Using small
cutoff values can lead to unexpected behavior when highly symmetric data sets
are being analyzed. This happens because many dissimilarity values are equal.
Rather than make an arbitrary elimination of some links inside a group of links
with equal dissimilarities, Permap shows all or none of a group of links that have
equal dissimilarities. Thus, links with equal dissimilarities will appear or
disappear all at once.
7.2.2.2 Stretched Links
Page 59 of 77
A link's stretch is defined to be (dij - Dij) / Dij, its fractional displacement
from its ideal length. This quantity would be called "strain" in the engineering
disciplines, but since "strain" has special meaning to Psychologists, the more
generic terms "fractional displacement," "fractional stretch," or less appropriately
but easier to say, "stretch," are used here. Actually, as was the case with the
badness functions that have singularities, Permap adds 0.0001 to the denominator
to make the definition of stretch be (dij - Dij) / (0.0001 + Dij). Thus, Permap uses
a pseudo fractional displacement formula. The word "pseudo" is dropped for the
sake of brevity and is justified because the 0.0001 factor is inconsequential in
terms of the visual display of the information.
For ordinal MDS, the stretch definition is modified to be (dij - Ldij) /

(0.0001 + Ldij) or (dij - Hdij) / (0.0001 + Hdij), depending on whether the dij value
is below (compressed) or above (stretched) its preferred value. See Section 5.2.2
for a definition of Ldij and Hdij.
The fractional stretch display provides the option of showing either all
stretched links, i.e., those with mapped distances that are longer than Dij, or just
the links that are stretched so much that they are statistical outliers. If you choose
to show all stretched links, only those with more than a 0.5% fractional stretch are
actually shown. This is to keep from reporting links as being stretched when in
fact their "stretch" is only "numerical noise."
Outlier links are identified using the same approach outlined for objects
described in Section 7.1.2.1, except that you seek to determine if a particular
link's stretch is significantly greater than what can be explained by random
variation. This is done using the usual one-tail rules that if a value is greater than
2.33 standard deviations above the mean then the chance that its occurrence is due
to random variation is less than one percent, and if it is 1.65 standard deviations
above the mean then the chance is less than five percent. The "one percenters"
are colored red and the "five percenters" are colored black. As with the object
outliers, it is important to note that this classification is not rigorous because no
evidence has been given that the underlying distribution is normal. It is, however,
a useful approach for showing outliers.
While outlier identification is useful, it is important to remember its true

meaning. It simply means that the identified link is significantly different from all
the others. That does not mean the link is necessarily erroneous. If almost all
dissimilarities are well fit, then a link with any stretch, even stretch due to trivial
numerical noise, may appear to be an outlier. Conversely, if a set of objects is
very badly fit then the standard deviation of the stretch factors may be very large
and no particular link, even one that is greatly stretched, would be identified as an
outlier. In both cases, the identifications are correct but misleading. Finally, don't
forget that the outliers are determined by fractional deviations. The Shepard plot
(see Section 7.2.1) shows deviations, not fractional deviations. Thus, it is
Page 60 of 77
possible to find that a link is identified as an outlier even though its placement on
the Shepard plot is within the 1.65 and 2.33 standard deviation boundary lines,
and vice versa.
Another word of caution is needed if you are using something other than
Euclidean mapping distances. For instance, if you have chosen the City Block
metric then all plots and stress calculations are made in terms of the City Block
distance rather than the "apparent" distance showing on the map. This can lead to
results that are correct, but "don't look right" when the map and the Shepard plot
or map diagnostics are compared. Using the City Block metric can seldom be
justified and beginners are advised to stick to the Euclidean distance measure.
7.2.2.3 Compressed Links

This option allows showing either all compressed links, i.e., those with
mapped distances that are shorter than Dij, or just the links that are compressed so
much that they are statistical outliers. The details are identical to those described
in Section 7.2.2.2 on Stretched Links, except the quantity (dij - Dij) / (0.0001 +
Dij) is replaced by (Dij - dij) / (0.0001 + Dij).
7.3 Attributes Evaluation Screen

The Attributes menu provides an Attribute Fit plot and five ways to superimpose
attribute gradient vectors on the MDS map. Each option has on-line context-sensitive
information provided in a "Notes" text box at the bottom of the Attributes menu screen.
If no attributes are given in the problem definition, then the Attributes menu entry is
disabled (grayed out). The shortcut key for this option is ctrl-A.
Immediately below the Analysis Options box is a box that lists all available
attributes. By highlighting a member of this box you determine which attribute is to be
used for the Attribute Fit plot and for plotting the various on-map gradient vectors. If
you have more than seven attributes you will need to use the right arrow to highlight the
“missing” attribute numbers in the Attribute Selection List box.
Printing a map with the evaluation markings on it is facilitated by having these

markings persist after the evaluation screen is closed. See the Section 6.4.1 for the
various printing techniques. After closing the evaluation screen if the Start or Continue
button is pressed, or an open space is clicked, all evaluation markings are removed and
Permap operation is returned to its usual mode.
7.3.1 Attribute Fit Plot

The Attribute Fit plot is a scatter plot that shows how well one or all of the
attributes are described by an optimally oriented uni-variant gradient running
across the map. As usual, the coefficient of determination, R2, is the key measure.
A perfect fit results in R2 = 1, whereas random behavior causes R2 to approach
zero.
Page 61 of 77
Each object's attribute value is represented by a point with the actual
attribute value, after being scaled to fit the zero/one domain, plotted on the X-axis
and its projected value, using the best-fit gradient, plotted on the Y-axis. This
scatter plot gives a clear visual method of determining the usefulness of a simple
uni-variant gradient for a particular problem. While the X-axis values are scaled
to fit the graph, if the best-fit is very bad the projected values may be outside the
bounds of the Y-axis. These cases are shown with red half-circles at the upper or
lower limits of the Y-axis. See the following section for a discussion of the
principles involved in defining the gradient vector.

This section describes various biplots that superimpose a second set of
information onto an existing MDS map.
7.3.2.1 Attribute Gradient Vector

Once a map is made, if attribute values are known it is reasonable
to ask about the distribution of the attribute's values across the map. In
some cases the underlying distribution can be understood simply by
printing the attribute values on the objects or by coloring the objects
according to their attribute values. Permap offers both of these methods.
However, if there is considerable scatter in the attribute values it can be
difficult to determine the underlying pattern. In this case, the most
important piece of information concerns the direction of steepest ascent of
the attribute's values. This gradient can be shown by making a biplot
which superimposes a vector on the map that points in the direction that
the selected attribute increases most rapidly.
Finding the best possible gradient vector is easily done by using

standard least-squares ideas. Because only linear uni-variant gradient
fields are of interest, all parallel vectors are equivalent. Therefore, without
loss of generality, we place the vector through the map's center point. The
vector is defined by the angle theta, measured counter clockwise from the
positive X axis.
Assume there is a linear attribute gradient across the map. Let the
ideal attribute values be represented parametrically by the formula a + b t,
where t varies from t = 0 at the vector tail to t = 1 at the vector tip. Actual
attribute values, Ai, are known for each object i. The t-value, ti, which
corresponds to each Ai, is determined by the perpendicular line running
from the vector and passing through object i. One wants the ideal values
to fit the actual values as closely as possible, so least-squares minimization
is used to calculate the parameters a and b. That is, a and b are chosen to
minimize the sum of the squares of the differences between a + b ti and Ai
for all i.
Page 62 of 77
Having calculated a and b, one can calculate R2 for the vector
specified by angle theta. Then, theta is varied and R2 is calculated for
each value of theta. The ideal theta is the angle where R2 is a minimum.
While the vector described above gives the "best possible"

direction of a linear attribute gradient, this best possible value may be no
good at all. For instance, if attribute values are randomly assigned to the
objects there will still be a best possible gradient vector, but it is
meaningless. The R2 value is given in the text box at the lower right of the
screen. It can be used to help tell whether or not the linear attribute
gradient is meaningful.
7.3.2.2 Vector and Quartiles

In addition to showing the vector described in Section 7.3.2.1, this
option uses the selected attribute's value to color each object. If a value
falls in the upper fourth of the range of attribute values for all objects, then
the object is colored black. Objects in the next lower quarter range are
gray, then light gray, and then clear. This option is useful when the screen
is congested and reading individual values is difficult.
7.3.2.3 Vector and Values

In addition to the vector described in Section 7.3.2.1, this option
prints the value of the selected attribute on each object. This option is
useful when the map has few objects that overlap each other. Use the
View screen (Section 4.3) to set the object diameter to accommodate
printing the values.
7.3.2.4 All Active Vectors

This option shows vectors for each of the attributes that were used
to determine the dissimilarities that were used to make the map. This set
of attributes is called the “active attributes” set.
Showing all vectors at once can help determine which attributes

are redundant and which are fundamentally different from each other. For
instance, if two or more vectors line up together (co-linear either in the
same or opposite directions), then they are providing similar information
and all but one are redundant for the given set of conditions. It might be
that a composite variable (a new "construct" or a linear combination)
would provide greater insight and show less random variation because of
the averaging affect. On the other hand, if there are vectors that are ninety
degrees to each other (orthogonal), then you have a good lead on assigning
meaningful coordinates to the map. The best of all possible worlds is
when, after your very first MDS analysis, all of your attribute vectors fall
into two perpendicular groups!
Page 63 of 77
7.3.2.5 All Available Vectors
This option is much like Section 7.3.2.4, except it includes all
attributes, even those that were not used in making the map.
This option facilitates doing "grounded research" where, for

instance, you want the best possible synthesis of all available data, and
you are willing to give up generalizability in order to squeeze the last bit
of information out of your data. If you have a low stress map with a good
interpretation, and you are interested in evaluating additional attributes,
then use this option.
7.3.2.6 Vector Tie Lines

To understand the construction of the gradient vector, it can be
helpful to see the connections between the object's attribute values and the
gradient as represented by the best-fit regression values assigned to the
vector. This information can be revealed by checking the "Show Vector
Tie-Lines" box. A perfect fit would have all tie-lines running
perpendicular to the gradient vector.
It is possible that the tie-lines will fly off the map. Some tie-lines
may run almost parallel to the gradient vector. This behavior is not due to
an error in Permap. It can happen when the best possible fit is a very bad
fit (say, R2 < 0.1). While the domain of the independent regression
variable is constrained to the map, if the attribute data are poorly fit there
is no reason that the dependent regression values must be within the
boundaries of the map.
7.4 General Comments Concerning Map Evaluation

Almost all experts in the field agree that MDS results should not pass or fail depending
on some particular value of some particular statistic. The first consideration is, do your results
have face value? This means that they should "make sense" to an intelligent but non-expert
observer. The lack of slavish obedience to statistical tests works both ways. Weak statistics
should not necessarily condemn meaningful results, and strong statistics should not be used to
prop up frivolous results. All results should pass the "adequacy, interpretability, stability,
meaningfulness, reasonability, and generalizability" hurdles, and if they do then they should be
accepted.
These ideas are not new. Torgerson (1965), and Law, Snyder, Hattie, and McDonald
(1984) strongly emphasized that adequacy and interpretability should dominate the debate over
the appropriateness of a MDS result. They clearly stated that it is the researcher's responsibility
to understand and rationally probe into the nature of the data, and that the researcher must not be
enslaved to lesser indicators such as Stress and R2. In the end, if a conflict exists between
measures of statistical fit and a meaningful interpretation, the measures of statistical fit must
yield.
Page 64 of 77
8.0 CONVERGENCE
Selecting the Convergence menu allows you to control two parameters that are seldom
changed. They control operations that are going on behind the scenes.
8.1 Convergence Rate

It can happen that you get a problem whose solution is unstable from a numerical
standpoint. This seldom happens with "real" data sets containing a reasonable number of
objects, but it is fairly common for test cases with highly symmetrical or uniform dissimilarities
or for very large data sets (say, larger than 200 objects). In such cases you can increase the
stability of the solution process by changing the iteration step size. This increased stability
comes with a decrease in solution speed. Moreover, it does not always work. About all one can
say is that it is worth a try if you happen onto a "bad actor" matrix or if you really have to treat
hundreds of objects at a time (a questionable requirement). See Section 11.0, Solution
Difficulties, for additional ways to increase stability and for the difficulties associated with
making MDS analyses of very large data sets.
8.2 Convergence Precision

Convergence precision is normally not a concern because the solution algorithm
continues to run while you are studying the map and the map's precision quickly becomes more
than even the most demanding analyst wants. However, if you are using the Auto Start and Auto
Repeat options to find multiple solutions starting from random initial positions, you may want to
control the convergence precision in order to control the speed with which you scan many
solutions. This can be done by setting the convergence limits using the Convergence menu.
Permap offers three convergence limits. The first allows fast screening of data, the second
provides a balance between speed and precision and is adequate for most situations, and the third
is so demanding that it will usually exceed all reasonable requirements.
It is possible for solutions to pulsate. That is, the field of objects will slowly expand and
contract and the objection will slowly rise and fall. This happens because Permap optimizes in a
step-by-step manner and it takes a few iterations for all parts of the problem to be equally
optimized. When this happens, if the convergence precision is set too low, it is possible to
record fictitious values as being the best-found values. Be sure to use a high level of precision
when finalizing your work.
9.0 SPECIAL CASES
Data for some interesting special cases are built into Permap. They include geometrically
simple cases that illustrate certain points. They are valuable for learning about MDS and for
checking Permap's results against known solutions. They also show unintuitive results and non-
convergent behavior. The F3 function key provides a shortcut to the Special Cases selection
screen.
Page 65 of 77
The prime entry point to this screen was moved to a top level menu item at the request of
users, but this screen's content logically (according to Microsoft's menu philosophy) fits as a
submenu under the top level File menu. Thus, an entry point is placed there too. The two paths
lead to the same screen. So, the Special Cases screen can be opened using the File menu, the
Special Cases menu, alt-C, or F3. That ought to keep almost everyone happy.
Some special cases involved dissimilarity or attribute data that are generated using a
pseudo random number generator. Usually you will want a new set of random numbers
generated each time you visit the Special Cases selection screen. However, upon occasion you
may want Permap to use the same set of random numbers each time a special case is selected.
The choice is made by using the Reuse Random Numbers check box at the bottom-left of the
Special Cases election screen.
10.0 SOLUTION TECHNIQUE
How a program finds a configuration that minimizes the objective function is probably
not of particular interest to you. However, you must have confidence that the program does what
it says it does. That is, a "black box" (solution algorithm) is acceptable only if you have
confidence that the black box works right. This confidence usually comes either from knowing
that the algorithm has been used successfully for a long time or from testing it against many test
cases where the correct answer is known. However, there should be no doubt about the fact that
an algorithm is never proven to be correct by using these approaches. Only a mathematical
analysis of the algorithm can do that. Unfortunately, the mathematical analysis of algorithms is
so difficult, and there are so few mathematicians that specialize in this kind of analysis, that the
rigorous approach has become an endangered species. Fortunately, Permap passes the two
empirical tests.
For those interested in what goes on inside Permap, it uses an iterative method of solving
the MDS minimization problem. Basically, the method is a heavily damped form of Newton's
method applied to one object at a time. This approach is widely used for problems that have a
limited amount of known structure. It has proven effective for every "real" data set tested so far,
but in theory this approach does have limitations. The mathematically inclined can easily
construct Dij matrices that are unstable. In these cases the objects may precess, oscillate, or even
enter the chaotic realm.
If you have a problem that shows signs of instability, you can do several things to help
Permap find the solution. First, you can reduce the step size (Convergence menu). Sometimes
this is sufficient to tame unstable cases. Second, you can apply strong mapping weights
(Analysis Parameters/Mapping Weights menu). De-emphasizing the distant objects will often
reduce chaotic behavior. This is particularly true for data sets that have very simple geometric
arrangements (a straight line or a large matrix of objects fitting perfectly in two dimensions).
After a stable solution is found, use the Mapping Weights shortcut button to rotate to the desired
mapping weights. Third, you can use one of the more stable objective functions (e.g., SStress) to
find the solution, and then use the Badness shortcut button to rotate to using the badness function
Page 66 of 77
that you really want. Fourth, if you are making an ordinal MDS analysis, first make a ratio
analysis and then use the MDS TYPE shortcut button to rotate to an ordinal analysis. If none of
these techniques work, and you are working with a "real" data set, please email the details to Ron
Heady at ron@heady.us. It will be the first case he has seen and he will enjoy working with you
on the problem.
11.0 SOLUTION DIFFICULTIES
Local Minima
The topic of local minima was touched on in Section 3.5.4 where the automatic control
features were discussed. The topic is important. Difficulties associated with local minima were
recognized by the earliest researchers (Kruskal, 1964). They have been noted by all the major
contributors to the applied MDS literature, and continue to merit comments in current reviews
(Everitt & Dunn, 1995). Still, the topic does not get the careful attention it deserves and some
practitioners still discount its importance. Shepard (1972) noted that except in special
circumstances local minima have not been a serious problem. Several other authors have
repeated this conclusion. Even some recent reviews (Stalans, 1995) suggest that though
configurations due to local minima often occur, "they generally are not drastically different from
the global minimum solution." However, our experience is that false results due to local minima
are not only common, they are very common. Moreover, sometimes they are significantly
different from the lowest stress configuration. Borg and Groenen (1987, p. 221) back up this
conclusion. They point out that "MDS algorithms usually end up in local minimum" and then go
on to note that this tendency is even greater when nonmetric MDS is used. After using Permap,
or any other interactive MDS program, we believe that you will see that this conclusion is
correct. You might even come to wonder how the early MDS researchers, who had very limited
computational facilities, no real-time interaction, and no built-in graphical capabilities, could be
so confident that their results were indeed representative of global minima.
Most early MDS programs (and even a few currently available programs) contained
subroutines that would choose smart (sometimes called rational) starting points to combat the
difficulties associated with local minima. Torgerson (1952, 1958), published a method that
could be used to find a good set of starting points. Twenty years later Arabie (1973) and Kruskal
and Wish (1978) were arguing for an alternative approach. Both of their papers suggested that it
would be better to use random starting points. Kruskal and Wish limited their advice to when a
solution has a "suspiciously large" stress value, although they did not suggest how suspiciously
large values should be recognized. But, regardless of this advice most MDS programs continued
to use smart starting points. In our opinion, these smart starting points are a liability rather than
an asset. We believe they should be avoided. This is because even though smart starting points
often lead to finding a global minima, when they fail you do not know that they have failed. You
simply get the same wrong solution over and over. Obviously, the alternative of using random
starting positions requires far more computation. So, historically, the decision has been between
being fast and usually right, or being slow and almost certain that you are right. This choice was
difficult for many years. Fortunately, computer speeds now have increased to the point that the
Page 67 of 77
original argument is moot. As a result, the need to be sure of the nature of your solution
dominates and most MDS programs now provide for easily solving problems using many
different random starting points.
Because of the importance of the topic of smart starting points, we digress to offer new
MDS users a bit of historical perspective. Arabie's (1973) comments against using smart starting
points were not well received by many MDS researchers. For instance, Clark (1976) and Spence
and Young (1978) criticized Arabie for not providing any systematic comparisons between
multiple-start and rational-start approaches. Arabie (1978) refuted the arguments against using
random starting points and explained the difference between having theoretically correct starting
points and having state-of-the-art starting points. Nevertheless, the argument continued. In
1992, Davison (p.100) concluded that the multiple-start approach was ". . . expensive, time
consuming, and largely untested in a systematic way." Even as late as 1997, Borg & Groenen (p.
222) were still concerned about the computationally intensive aspect of the multiple-start
approach. As the body of MDS literature matured, however, Arabie's concerns proved to be
valid and the debate amongst experienced MDS users now has faded away. However, Cox and
Cox, in a draft of a journal submission (see the web site http://www.ncl.ac.uk/mds/) show that, at
least in terms of beginning MDS users, the problem is still with us. They refer to misleading
comments in SAS and SPSS manuals, and the fact that these programs do not make it easy to
make the needed randomly generated replications. They conclude that beginning MDS users are
still in danger of falling into this old trap. We agree and strongly recommend that you use
Permap's Auto Repeat and Auto Stop features and make many tests before you conclude that you
have a global minima. For first-hand experience with the problem, load the data in
EXAMPLE_E.txt. These are the data Cox and Cox used to illustrate the ease with which local
minima can be mistaken for global minima.
Degenerate Solutions
A note on degeneracy is in order. Over the years quite a few MDS articles have referred
to the problem of avoiding degenerate solutions. A degenerate solution is a rigorous solution
from a mathematical viewpoint but it is not meaningful in terms of the reality that is being
modeled. Degenerate solutions generally take the form of all objects being in a straight line or
clustered on top of each other. When first writing Permap we planned a subroutine to combat
this possibility. However, we found that it was not needed. Evidently, some of the early MDS
computer programs used either very susceptible algorithms or exceedingly unrealistic starting
points. While Permap can produce degenerate solutions, they are very rare unless you make an
effort to find them.
For teaching purposes, it is useful to be able to demonstrate the existence of degenerate

solutions. Because Permap loads objects in a straight line, and because degenerate solutions
often take the form of a straight line, demonstrating degenerate solutions is easy. After loading a
new problem, click Continue instead of clicking Start. This causes Permap to start searching for
a solution using the existing straight line positions instead of random positions. Because a
degenerate solution is usually nearby, the objects may simply slide over into their degenerate
solution positions. If you click Jiggle the solution will immediately evolve into a meaningful
Page 68 of 77
solution. But that does not really show the fragile nature of some degenerate solutions. To show
just how unstable these degenerate solutions can be, click Stop and drag an object just very very
slightly out of the horizontal line of objects. Then click Continue, and unless you have made an
extremely small adjustment, the objects will immediately fly off to a more meaningful solution.
Interestingly, this technique works with almost any distance function, badness function, and
dissimilarity matrix. Another degenerate solution can be found by imposing extreme boundary
conditions. After solving a problem close the constraining boundary while the solution continues
to run. Do this by clicking the Boundary shortcut. If it is not present on the main screen then
use the View menu to add the it to the list of active controls. As the circle shrinks toward a zero
radius the objects are pressed together until they are superimposed upon one another. Then, with
the solution continuing to run, open up the boundary. The objects will stay in their superimposed
"solution." A jiggle will quickly show that this degenerate solution is meta stable.
Very Large Data Sets
Finally, it is important to discuss problems with a large number of objects. Everyone

wants to treat all of their objects at once. That is natural, but not always a good idea. Permap's
upper limit on the number of points was recently increased from 200 to 1000 because users were
requesting it and because newer computers have so much RAM and such high CPU speeds that
static memory can be used (to increase speed) without the operating system resorting to the slow
"memory swapping" approach to memory management. But there are two difficulties with
analyzing problems having many objects. First, large problems can become unstable because the
solution is controlled by insignificant figures. This can be helped by using a small step size
option in the Convergence menu, but when the problem gets really large even this may not be
enough. It all depends on the nature of the data. But even if the stability problem is manageable,
the "solution" almost certainly is only one of thousands (millions?) of local minima. Densely
packed objects can take on almost any number of arrangements all of which have essentially the
same objective function value. These differences must not be taken seriously even though they
seem to be significant to the untrained eye. If you use many points you must shift to looking
only for the "big blobs” and pay no attention to the fine detail. This is not a defect in Permap. It
comes from the "curse of dimensionality." MDS programs that don't reveal as much of the inner
workings as Permap have the same defect, it is just invisible. A better approach than treating all
points at once is to find high concentrations of objects (“blobs”) and analyze each separately, and
then, perhaps, replace them with archetypical objects based on the common properties of the
objects in the center of the concentration. Then, conclusions are drawn from the arrangement of
the archetypical objects. The key point is that the detailed arrangements between close objects
can, and usually is, overridden by the requirements of the relationships of the many many distant
objects.
12.0 Permap HISTORY
Permap was developed by Professor Ronald B. Heady of the University of Louisiana at

Lafayette, and Professor Jennifer L. Lucas of Agnes Scott College. An early version is described
in Heady and Lucas (1996). The current Windows-based version of Permap is a descendant of a
DOS-based version which was written in 1993-94 using a Windows-emulating program
Page 69 of 77
(VBDOS). In 2000, Permap was rewritten for the Windows operating system to further increase
its interactivity, add new features, and remove certain memory restrictions.
Professor Tony Coxon and his classes have provided valuable input that has speeded
along Permap's evolution. Many thanks to them, and to all the others that have sent in
suggestions for improvement or words of encouragement.
13.0 GETTING HELP
There are three ways to get help. The first is to press F1 or enter the Edit/Edit Help File
menu to read the brief on-line help file that was described in Section 4.2.2. This file is more of a
short reminder list than a full summary of how to use Permap. It contains some of the
operational sections of this manual. The second is to refer to this manual. The third is to email
Ron Heady at ron@heady.us. He is more than happy to help you understand Permap's
idiosyncrasies, but he cannot do much about specific hardware problems.
14.0 LEGAL NOTICE
Permap, the computer program, is copyright 1993-2010 by Ronald B. Heady. It is

declared to be academic freeware. You have permission to use Permap as much as you like for
any academic or personal use. You may make and give away as many copies of Permap as you
like as long as you do not alter the program or charge for the copies beyond the cost of the media
and a nominal handling fee. Permap is offered with no guarantee or warranty. It has no
assurance of product support.
This Operation Manual is copyright 1993-2010 by Ronald B. Heady. You may make as
many copies of this manual as you like for any academic or personal use as long as you do not
alter its content or charge for the copies beyond the cost of the media and a nominal handling
fee. The accuracy or usability of the contents of this manual is not guaranteed in any way.
Page 70 of 77
APPENDIX I
DATA FILE KEYWORDS
To be recognized, a keyword must be the first entry on a line. Keywords are usually
capitalized to make them standout, but any mixture of upper and lowercase letters is acceptable.
Keyword Purpose Example Use Comment

NOBJECTS Set number of objects NOBJECTS=... 1000 maximum
NATTRIBUTES Set number of attributes NATTRIBUTES=... 100 maximum
TITLE Set a title line to be shown on output TITLE=... 65 char. max
MESSAGE Set a message to be shown on output MESSAGE=... 65 char. max
SUBTITLE Set a subtitle line to be shown on output SUBTITLE=... 98 char. max
SUBMESSAGE Set a submessage to be shown on output SUBMESSAGE=... 98 char. max
DISSIMILARITYLIST Announce that dissimilarity data follow
SIMILARITYLIST Announce that similarity data follow
ATTRIBUTELIST Announce that attribute data follow
LOCATIONLIST Announce that initial positions follow
WEIGHTLIST Announce that weight data follow
NA A missing dissimilarity value ..., NA, ... or ..., na, ...
The following keywords are of a different nature from those shown above. Normally, they are
not used. In fact, if they are ever used it is almost certainly in a classroom. These keywords are
used to control start-up details of the program. They override default values and force Permap to
start with specific settings. Thus they provide a way for a teacher to provide a data file that also
controls the type of analysis, or, for a "save solution" file to provide all information needed to
immediately restart the analysis from the conditions existing when the file was saved.
These keywords are entered with a following equal sign and a number chosen from those shown
after each option. For instance, if the data file has STARTBadnessFunctionNum=1, then Permap
will start with the SStress badness option being selected
STARTMDSAnalysisTypeNum=0
0 Ratio, 1 Ratio + Bounds, 2 Interval, 3 Interval + Bounds, 4 Ordinal
STARTBadnessFunctionNum=0
0 Stress, 1 Stress1, 2 SStress, 3 Multiscale, 4 Fractional
STARTDistanceFunctionNum=0
0 Euclidean, 1 CityBlock, 2 Minkowski
STARTMappingWtFunctionNum=0
0 None, 1 Weak, 2 Linear, 3 Strong, 4 Adjustable
STARTAttributeFunctionNum=0
0 COS(alpha), 1 Euclidean, 2 City Block, 3 Guttman
4 Pearson, 5 Spearman, 6 Nominal SMC
7=Jaccard, 8=Gower/Russell/Rao, 9=SMC/Sokal/Michener
10=Hamman, 11=Yule, 12=Askin/Charles, 13 Dissimilarities in file
STARTAttributeDataTypeNum=0
0 Ratio, 1 Interval, 2 Nominal, 3 Binary Nominal, 4 Unknown
Page 71 of 77
STARTObjectIDMethodNum=0
0 Number, 1 Name, 2 None
STARTDimensionsNum=2
1 One Dimensional, 2 Two Dimensional, . . . , 8 Eight Dimensional
STARTConstraintRadius=1.0
A number from 0 to 1, indicating the radius of the constraining boundary
Page 72 of 77
APPENDIX II
EXAMPLE INPUT DATA
The following provides an example of a working data file. To be readable by Permap this file
must be stored in an unformatted text format. Therefore, if you want to actually run this data,
copy and paste it into a word processor and then save the file in ASCII or ANSI format. Be sure
that you do not introduce any "strange" characters into the text file (sometimes WordPerfect will
add an invisible termination character at the end of a file, and this can cause trouble).
All lines that do not start with a keyword or a number are comment lines. Comments are
disregarded by Permap.
This example uses data from Kaufman and Rousseeuw's book "Finding Groups in Data" that
gives the subjective dissimilarities between eleven sciences as seen by fourteen postgraduate
economics students from several different countries.
****************** Start Copying Just Below Here *******************
MESSAGE=Differences Between the Sciences
NOBJECTS=11
DISSIMILARITYLIST
Astr, 0
Biol, 7.87, 0
Chem, 6.50, 2.93, 0
CSci, 5.00, 6.86, 6.50, 0
Econ, 8.00, 8.14, 8.21, 4.79, 0
Geog, 4.29, 7.00, 7.64, 7.71, 5.93, 0
Hist, 8.07, 8.14, 8.71, 8.57, 5.86, 3.86, 0
Math, 3.64, 7.14, 4.43, 1.43, 3.57, 7.07, 9.07, 0
Medi, 8.21, 2.50, 2.93, 6.36, 8.43, 7.86, 8.43, 6.29, 0
Phys, 2.71, 5.21, 4.57, 4.21, 8.36, 7.29, 8.64, 2.21, 5.07, 0
Psyc, 9.36, 5.57, 7.29, 7.21, 6.86, 8.29, 7.64, 8.71, 3.79, 8.64, 0
****************** Stop Copying Just Above Here *******************
The data can be separated with space(s), a comma, or both.

If there are missing data, they should be entered as "NA" or "na."
The leading names shown above are optional, but if used they must start with a letter, must not
start with NA or na, and must not contain non-alphanumeric characters.
If you want to force Permap to start an analysis using certain object locations, use the
LOCATIONLIST keyword followed by a set of locations. When Permap writes out a solution to
Page 73 of 77
the PermapBestSolnFound.txt file, it puts these locations in the file. This means that you can use
the output file as an input file and have the analysis begin exactly where you left off. The
number of lines of data must equal the NOBJECTS setting. Leading plus signs are optional on
attribute or dissimilarity data. If desired, they can be used to improve the readability of the input
file. If you enter a location list, Permap will automatically set the number of dimensions used in
the analysis to equal the number of dimensions implied by the location list.
LOCATIONLIST
-0.5262 -0.0030
+0.2297 -0.3264
-0.0307 -0.4252
-0.2926 +0.1243
-0.1282 +0.5070
+0.1829 +0.4083
+0.4286 +0.4962
-0.3222 +0.0142
+0.2130 -0.4146
-0.3328 -0.2349
+0.5785 -0.1459
Page 74 of 77
REFERENCES
Arabie, P. (1973). Concerning Monte Carlo evaluations of nonmetric multidimensional scaling

algorithms. Psychometrika 38, 607-608.
Arabie, P. (1978). Random versus rational strategies for initial configurations in nonmetric
multidimensional scaling. Psychometrika, 43, 111-119.
Borg, I. & Groenen, P. (1997). Modern Multidimensional scaling: Theory and applications.
New York: Springer.
Clark, A. K. (1976). Re-evaluation of Monte Carlo studies in nonmetric multidimensional

scaling. Psychometrika, 41, 401-403.
Cox, T. F., & Cox, M. A. (1994). Multidimensional Scaling (Monographs on Statistics and
Applied Probability, No. 59). New York: Chapman and Hall.
Coxon, A. P. M. (1982). The users' guide to multidimensional scaling. Exeter, NH: Heinemann.
Davison, M. L. (1992). Multidimensional scaling. New York: John Wiley & Sons.
DeLeeuw, J., & Stoop, I. (1984). Upper bounds for Kruskal's stress. Psychometrika 49, 391-
402.
Ekman, G. (1954). Dimensions of colour vision. Journal of Psychology, 38, 467-474.
Everitt, R. S., & Dunn, G. (1995). Applied multivariate data analysis. London: Edward Arnold
Publishers.
Fitzgerald, L. F., & Hubert, L. J. (1987). Multidimensional scaling: Some possibilities for
counseling psychology. Journal of Counseling Psychology, 34, 469-480.
Harshman, R. A. (1984). "How can I know if it's real?, A catalog of diagnostics for use with
three-mode factor analysis and multidimensional scaling.” In H. G. Law, C. W. Snyder, J. A.
Hattie, & R. P. McDonald (Eds.), Research Methods for Multimode Data Analysis. New York:
Praeger Press.
Hartigan, J. A. (1975). Clustering algorithms. New York: John Wiley & Sons.
Heady, R. B., & Lucas, J. L. (1997). Permap: An interactive program for making perceptual
maps. Behavior Research Methods, Instruments, & Computers, 29(3), 450-455.
Hubert, L., Arabie, P., & Hesson-Mcinnis, M. (1992). Multidimensional scaling in the city-
block metric: A combinatorial approach. Journal of Classification, 9, 211-236.
Page 75 of 77
Isaac, P. D., & Poor, D. S. (1974). On the determination of appropriate dimensionality in data
with error. Psychometrika 39, 91-109.
Kaufman, L., & Rousseeuw, P. J. (1990). Finding groups in data: An introduction to cluster
analysis. New York: John Wiley & Sons.
Kruskal, J. B. (1964). Nonmetric multidimensional scaling: A numerical method.

Psychometrika, 29, 115-129.
Kruskal, J. B., & Wish, M. (1978). Multidimensional scaling: Sage university paper series on
quantitative applications in the social sciences. Beverly Hills, CA: Sage
Publications.
Law, H. G., Snyder, C. W., Hattie, J. A., & McDonald, R. P. (1984). Research methods for
multimode data analysis. New York: Praeger Press.
Lee, S-Y., & Bentler, P. M. (1980). British Journal of Mathematical and Statistical Psychology,
33, 142-150.
Schiffman, S. S., Reynolds, M. L., & Young, F. W. (1981). Introduction to multidimensional

scaling. New York: Academic Press.
Shepard, R. N., Romney, A. K., & Nerlove, S. B. (1972). Multidimensional scaling, Volume I -
Theory. New York: Seminar Press.
Spence, I., & Graef, J. (1974). The determination of the underlying dimensionality of an
empirically obtained matrix of proximities. Multivariate Behavioral Research, 9, 331-
342.
Spence, I., & Young, F. W. (1978). Monte Carlo studies in nonmetric scaling. Psychometrika
43, 115-117.
Stalans, L. J. (1995). Multidimensional scaling. In L. G. Grimm & P. R. Yarnold (Eds.),

Reading and understanding multivariate statistics. Washington, DC: American Psychological
Association, 137-168.
Taylor, J. R. (1997). An Introduction to Error Analysis. Sausalito, CA: University Science

Books.
Torgerson, W. S. (1952). Theory and methods of scaling: I. theory and method. Psychometrika,
17, 401-419.
Torgerson, W. S. (1958). Theory and methods of scaling. New York: John Wiley & Sons.
Page 76 of 77
Torgerson, W. S. (1965). Multidimensional scaling of similarity. Psychometrika, 30, 379-393.
Wagenaar, W. A., & Padmos, P. (1971). Quantitative interpretation of stress in Kruskal's

multidimensional scaling technique. British Journal of Mathematical and Statistical Psychology,
24, 101-110.
Waern, Y. (1973). Structure in similarity matrices: A graphic approach. Scandinavian Journal

of Psychology, 13, 5-16.
Young, F. W., & Hamer, R. M. (1987). Multidimensional scaling: History, theory, and
applications. Hillsdale, NJ: Lawrence Erlbaum Associates.
Page 77 of 77

Permap 11.8 Tutorial

Uploaded by

Copyright:

Available Formats

Permap 11.8 Tutorial

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Permap 11.8 Tutorial

Uploaded by

Copyright:

Available Formats

MDS Analysis Using Permap 11.

Dr. Ronald B. Heady, University of Louisiana at Lafayette (retired)

2.0 GETTING STARTED

3.0 MAIN SCREEN & PROGRAM CONTROL

4.0 FILE, EDIT, VIEW

5.0 ANALYSIS PARAMETERS

7.0 MAP EVALUATION

9.0 SPECIAL CASES

10.0 SOLUTION TECHNIQUE

11.0 SOLUTION DIFFICULTIES

12.0 Permap HISTORY

13.0 GETTING HELP

14.0 LEGAL NOTICE

APPENDIX I - DATA FILE KEYWORDS

APPENDIX II - EXAMPLE INPUT DATA FILE

1.2 MDS Maps

What is an attribute? An attribute is some aspect of an object. It may be called a factor,

In this manual we present all formulas in terms of dissimilarities and refer to

2.0 GETTING STARTED

2.1 System Requirements

NOBJECTS=6 Gives the number of objects in the analysis.

DISSIMILARITYLIST Announces that dissimilarity values follow.

Dissimilarity data can be in either a lower-left half-matrix, as shown above, or in a whole

If your proximity information is in the form of similarities instead of dissimilarities, then

NOBJECTS=6 Gives the number of objects in the analysis.

NATTRIBUTES=3 Gives the number of attribute values for each object.

ATTRIBUTELIST Announces that attribute values follow.

2.4 Data Input, MDSX

MDSX Permap Details / Comments

RUN NAME TITLE = Permap’s title has an upper limit of

TASK NAME SUBTITLE = Permap's subtitle has an upper limit

# OF STIMULI NOBJECTS = MDSX's "#" can be "N" or "NO"

# OF VARIABLES NATTRIBUTES = MDSX's "#" can be "N" or "NO"

PARAMETERS DISSIMILARITIES DISSIMILARITYLIST MDSX's input is on two separate

PARAMETERS SIMILARITIES SIMILARITYLIST MDSX's input is on two separate

PARAMETERS ATTRIBUTES ATTRIBUTELIST MDSX's input is on two separate

3.0 MAIN SCREEN & PROGRAM CONTROL

3.1 Header and Objective Function Value

3.2 On-line Shortcut Controls

3.3 Parking Active Objects Control

3.4 Parking Active Attributes Control

3.5 Execution Control

3.5.1 Start, Stop, Continue

3.5.3 Field Movements (Mirror, Rotate, Move, Zoom)

3.5.4 Automatic Controls

3.6 Click & Keystroke Commands

Mouse shortcut controls:

Keystroke shortcut controls:

A-key changes the Attributes function

Alternate-key shortcut controls:

Control-key shortcut controls:

3.7 Locking Object Positions

4.1 File Menu

4.1.1 Load Data from Existing Files

4.1.2 Load Data From a Special Case

4.1.3 Enter New Data

4.1.4 Save Results & Settings

4.1.5 Reload Original Default Settings

4.1.6 Printing Results

4.2 Edit Menu