Multiv: Users Guide v. 2.4

MULTIV
Multivariate Exploratory Analysis, Randomization Testing and Bootstrap

Resampling
Users Guide v. 2.4
Copyright 2006 by Valrio DePatta Pillar

Universidade Federal do Rio Grande do Sul, Porto Alegre, Brazil
Introduction
MULTIV is a computer application program for performing flexible exploratory analysis,
randomization testing and bootstrap resampling with multivariate data (Figure 1). It can handle
qualitative, quantitative and mixed data types, offering several options for data transformation,
resemblance measures, ordination and clustering techniques. The methods of exploratory
multivariate analysis are well described elsewhere (see, i.a., Orlci 1978, Pielou 1984, Digby &
Kempton 1987, Orlci et al. 1987, Podani 2000, Legendre & Legendre 1998). The results are
presented in text files and, when applicable, in graphs. The versions of MULTIV for Windows do
not create graphics.
Integrated to the exploratory tools, MULTIV can do fast randomization testing in
univariate or multivariate comparisons between groups of sampling units defined by one or more
factors, the latter being especially useful in the analysis of variance due to factor and interaction
effects in experimental and survey data. The technique is described in Pillar & Orlci 1996, with
improvements after Pillar 2004. In addition, MULTIV can perform the multiresponse permutation
procedure (MRPP, Mielke & Berry 2001). Randomization tests are also available for pair-wise
comparisons of variables (see Manly 1991).
MULTIV can do bootstrap resampling to generate empirical confidence limits useful in
estimation, to evaluate group partition sharpness in cluster analysis (Pillar 1999a) and to evaluate
the significance of ordination dimensions (Pillar 1999b). Bootstrap resampling is integrated with
process sampling (Orlci & Pillar 1989) for examining sampling sufficiency (Pillar 1998). These
functions were originally implemented in program SAMPLER.
This manual gives a description of MULTIV, with examples showing analytical pathways.
The methods implemented are briefly presented without entering in details on appropriateness or
advantages under different circumstances, which can be found in the general references already
mentioned or in more specific ones throughout the text.
Hardware requirements
MULTIV is released in two versions for Macintosh and two for Windows OS:
(1) Multiv is the full version for Macintosh. It can run under Mac OS version 8.5 or later,
including Mac OS X.
(2) MultivMinor is a demo version of Multiv which does everything the full version does,
except that a maximum data size of 400 data elements (sampling units x variables) is
allowed.
(3) Multiv.exe is the Windows OS version equivalent to Multiv, but not showing graphics.
(4) MultivMi.exe is the Windows OS version equivalent to MultivMinor, but not showing
graphics.
To install MULTIV, copy a version that fits your hardware in your working volume and
folder. It is advised to place MULTIV with the data files (the application, not its alias). Otherwise
you will have to specify the full path when entering data file names. In any case, the output file
Prinda.txt will be placed in the same folder with the application.
Memory allocation is dynamic, thus there is no a priori upper limit for the number of
variables and sampling units that can be handled (except in MultivMinor). In the Macintosh, under
systems previous to MacOS X, if MULTIV runs out of memory it will give a message and exit. To
increase the memory allocated to the application, the Macintosh user should quit MULTIV, select
its icon in the Finder, choose Get Info in the File menu, and type a larger, compatible minimum
size in the "Memory Requirements" box.
MULTIV can run in the background, that is, the user can operate on other applications that
do not use the CPU as intensively while computations proceed in MULTIV, which is especially
useful in randomization testing and bootstrap resampling with a large number of iterations.
Output
The numerical results are stored in a file named Prinda.txt, which can be open with any text
editor. Under Mac OS X, scatter diagrams and dendrograms can be captured and saved by using
the utility named Grab, which comes with Mac OS X; under previous systems, graphics can be
stored as picture files by pressing the keys Shift Command 3 altogether while the graph is on the
screen. With graphs on the screen, the user must close the graph's window in order to continue
running the program. The versions for Windows OS do not produce graphical output.
MULTIV keeps track of the analytical steps performed on a data set, saving the
information on a file named by default Sessao.trk. The user can choose to save the session with a
different name (the extension .trk will be added automatically). Intermediate results are saved
automatically in the sessions file when working with small data sets. The program informs
whether the session is saved or not. If not, the user should explicitly choose the option S save
session in the main menu before quitting the program. Upon reopening, MULTIV will always
check in its folder for the file Sessao.trk; if the file is found, the session with the last data file
being used can be continued or not; if the file Sessao.trk or the session file pointed therein is not
found, a new session is started. The user can also at any time open a previously defined session
and continue the analysis.
Upon reopening, MULTIV will always check in the same folder for an existing file
Prinda.txt, which if present can be appended with the new results or not.
Data input
The input data file for MULTIV must be a text only file, containing observations of one or
more variables in a set of sampling units. You can use any text editor to create the data file,
provided you save it "as text only". The data can be arranged in matrix format, with variables as
rows and sampling units as columns, or vice-versa, or in a free format, provided entries in the data
file are separated by blank spaces, tabs, carriage returns, or a combination of these. Actually,
MULTIV will read as many data entries as the number of variables times the number of sampling
units that were informed, in the order indicated; anything written after these will be disregarded (it
is wise to use the end of the file to write useful information about the data, such as data origin,
names and order of variables and sampling units, how they are arranged, etc.).
The user interface is interactive. The user should follow options offered by a sequence of
menus shown on text-based screens. The options available are in general the ones that are
applicable to the type of data informed by the user and analytical step that has been reached. In the
Preferences menu the user can choose the language Portuguese or English. Major data entry is
from text files. Numbers and letters (case insensitive) specify menu options. Keyboard input is
processed upon pressing the Return key. Some questions may require multiple keyboard input, in
which case it is enough to separate the entries by a blank space, a tab or a carriage return.
Data set
quantitative, qualitative, binary, mixed
Description
Transformation
means, entropy, complexity
scalar, vector
Bootstrap
Resemblance
between variables or sampling units

distances, product moments, association, ...
Cluster
Analysis
Ordination
Randomization
tests
comparing groups of
sampling units (analysis of
variance, different designs)
or comparing variables
Bootstrap
Bootstrap
Figure 1. Flow diagram for exploratory analysis and randomization testing with MULTIV,
showing options available at different analytical steps.
The options offered by the menus will be described in several sample runs that follow.
Note that keyboard input is in italics and comments preceded by // are inserted.
MULTIV
for multivariate analysis, randomization tests and bootstrapping version 2.4.0
------------------------------------------------------------------------------
------------------------------------------------------------------------------Analysis status: Data not available.
------------------------------------------------------------------------------MAIN MENU
* N specify data or open existing session
V descriptive attributes
T transform data
R resemblance measures
G specify groups of sampling units
D scatter diagrams
O ordination
C cluster analysis
P randomization tests comparing groups of sampling units
A randomization tests comparing variables
* E preferences
S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): n
Wish to specify new data (n) or to open (o) an existing session? n/o n
Enter name of file containing data matrix: okop37.txt
// The data file must be in the same folder with MULTIV. Upon specifying the file, the program
does not need to access it again afterwards, since the data matrix will be stored in the sessions
file. The sessions file as well must be in the same folder with MULTIV. The small data set in
file okop37.txt is from Orlci et al. (1987, p.37):
26 29 29 29 30 35 39
28 31 31 33 27 38 36
18 14 13 13 19 15 15
1234567
Hkub Poch Pbre
OK&O p.37
3 variables (rows)
7 sampling units (columns)
Number of sampling units:
7
Number of variables: 3
Types of variables:
(1) only quantitative variables, same measurement scale
(2) only quantitative variables, different scales
(3) only qualitative, non binary variables
6
(4) only binary variables
(5) several types
Enter option: 1
// If option 5 is chosen, the type of each variable will have to be specified later.
Data ordered by (the rows of the matrix correspond to):
(N) sampling units (u1v1,..., u1vp,..., unv1,..., unvp)
(T) variables (u1v1,..., unv1,..., u1vp,..., unvp)
Enter option: t
// The data is ordered by sampling units when in the file the first set of p entries refer to the
observations in p variables in the first sampling unit, the second set of p entries to the second
sampling unit, and so on up to the last set of p entries referring to the observations in the same p
variables in the last sampling unit. If the data is arranged in a matrix, the rows correspond to the
sampling units.
The data is ordered by variables when the first set of n entries refer to the observations in the
first variable, the second set of n entries to the observations in the second variable, and so on up
to the last set of n entries referring to the observations in the last variable. If the data is arranged
in a matrix, the rows correspond to the variables. This is the case in file okop37.txt.
Labels for sampling units and variables:
(1) given by default (1, 2, ...)
(2) read labels after data on file okop37.txt
(s. units first, labels with max. 5 alphanumeric symbols)
Enter option: 2
Include data matrix in the output file? y/n y
// Labels for sampling units and variables are given by default (sequential numbers) or may be
specified after the data on the same data file. The labels will be used in numerical results,
dendrograms and scatter diagrams. The option to include the data matrix in the output file is to
check whether the program read the data correctly.
// The current data file in the session is okop37.txt. If a new data is specified the information
given about the previous data file will be lost. To avoid this problem you can save the session
with a name of your choice before specifying a new data set, as in the following run; the session
can be recovered afterwards.
Analysis status:
Data file name: okop37.txt
Dimensions: 7 sampling units, 3 variables
Data type: (1) quantitative, same measurement scales
Scalar transformation: (0)none
Vector transformation: (0)none
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
* V descriptive attributes
* T transform data
* R resemblance measures
* G specify groups of sampling units
* D scatter diagrams
O ordination
C cluster analysis
7
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): s
Enter name for the session: okop37_Session
Analysis status session okop37_Session:
Session IS saved.
// The following run will specify file P&Op87.txt with mixed data type:
1011111111111
0100000000011
2121222221112
1311111221111
3621145543442
Data from Pillar & Orloci (1993) p.87
5 variables, several types (rows) describing 13 sampling units (columns).
Variable types:
11223
Session IS saved.
Analysis status:
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
* N specify data file or open existing session
* T transform data
O ordination
C cluster analysis
* E preferences
* S save session
8
* X exit
Enter name of file containing data matrix: P&Op87.txt
13
Types of variables:
(1) only quantitative variables, same measurement scales
(2) only quantitative variables, different units
(5) several types
Enter option: 5
Enter variable type (1:binary,2:qualitative,3:quantitative):
Variable: 1 2 3 4 5
1 1 2 2 3
Type:
// Binary variables have two states, which must be zero or 1. Qualitative variables may have 2
or more states, which must be identified by integers. Quantitative variables are integers and/or
real numbers.
Wish to correct? y/n n
Enter option: t
(2) read labels after data on file P&Op87.txt
Enter option: 1
See data on file Prinda.txt.
Wish to rename this session? y/n y
Enter name for the session: P&Op87_Session
Analysis status session P&Op87_Session:
Data file name: P&Op87.txt
Data type: (5) mixed
Variable: 1 2 3 4 5
Type:
1 1 2 2 3
Session IS saved.
// Note that now the current session is P&Op87_Session, with a new data file (P&Op87.txt).
The information given about the previous data file has been saved in session okop37_Session.
Descriptive attributes and bootstrap resampling

This option is for the computation of sample parameters such as means and standard
deviations of variables (when variables are quantitative or binary) and entropy (when variables are
qualitative or binary). Also, relevant if the data matrix is a contingency table, an option is offered
to compute entropy and complexity measures for sampling units and for marginal totals in the
table. For methods see Orlci (1991), Anand & Orlci (1996) and references therein. A data set
with all variables measured with the same unit is taken as a contingency table by MULTIV. The
program uses Rnyi entropy functions. One of them,
s
H =
ln pi
i =1
1
measures entropy in a frequency distribution of n sampling units P = (p1 p2 ... ps) such that pi= fi/n
and p1 + p2 + ... + ps = 1 (since n = f1 + f2 + ... + fs), where f1, f2, ... fs are counts in each of s
classes in the variable being considered. The scale factor 0 is specified by the user. If =0 is
given, H0 = ln s for any frequency distribution. If =1 is given, the program actually takes lim 1
and uses Shannons entropy
s
H = pi ln pi
i=1
If the data matrix is taken as a contingency table, these functions are used to compute
entropy with the column totals and with the row totals, that is, actually for the variable defining
rows and for the variable defining columns. Mutual entropy and coherence coefficient between
these two variables are also computed (see definitions in resemblance measures).
Kolmogorov complexity (Anand & Orlci 1996) in each sampling unit is also computed
when the data matrix is taken as a contingency table. Total complexity (L) is the total code length
after Huffman coding of the frequency distribution in the sampling unit. Structural complexity is
the difference between total complexity and Rnyi entropy of a given order .
The option for generating confidence limits by bootstrap resampling is offered in case
means and standard deviations or entropy in variables are computed. The method generates
confidence limits by the percentile method (Efron & Tibshirani 1993), combined with process
sampling (Orlci & Pillar 1989, Pillar 1998). It is based on data resampling of the sample with n
sampling units, considered here as a pseudo sampling universe. At each of many iterations a
sample (bootstrap sample) of size nk n is extracted at random with replacement from the pseudo
sampling universe and the chosen parameter (mean and standard deviation, or entropy) in the
bootstrap sample are computed and stored. After performing B iterations the values are sorted
from smallest to largest. For a given probability the lower and upper confidence limits will be
respectivelly the B/2 and 1+B(1-/2) th values in the sorted list. The number of iterations is
adjusted in order to have integers identifying the position of the limits. For instance, with 1000
iterations and = 0.05, the lower and upper confidence limits will be the values at the 25th and
976th positions respectivelly. Based on the confidence interval we can conclude, with an
probability of being wrong, that the true value of the parameter is in between the limits.
Confidence intervals indicate whether a sample with size nk is sufficient for the required precision.
Process sampling involves generating confidence limits for increasing sample sizes nk n. By
examining confidence limits over increasing sample sizes we can judge how well a better precision
would compensate the effort of getting larger samples. The user can select the way process
sampling is performed, as illustrated in the following example.
10
The run below illustrates the computation of means and standard deviations, followed by bootstrap
confidence limits and process sampling:
Analysis status:
Data file name: EEASolo.txt
Data type: (2) quantitative, different units
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
* T transform data
O ordination
C cluster analysis
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): v
------------------------------------------------------------------------------DESCRIPTIVE ATTRIBUTES
Options:
(1) means and standard deviations of variables
(99)cancel and return to main menu
Enter option: 1
Get confidence limits using bootstrap resampling? y/n y
// If the answer is yes confidence intervals for the mean and standard deviation in each variable
will be generated and the information below will be requested. As specified in this example,
process sampling will start with 10 sampling units, adding 2 sampling units at each step, until
the total sample size is reached (15). At each sample size in process sampling, bootstrap
samples will be taken to generate confidence limits for the mean and standard deviation at that
sample size. A minimum of 1000 iterations for bootstrap resampling is recommended. By
specifying the same number for initialization the analysis will reproduce identical results for the
same data and options, but initialization of random numbers should normally be automatic.
Initial sample size: 10
Number of sampling units added at each sampling step: 2
Enter number of iterations in bootstrap resampling: 1000
Enter alpha probability for confidence limits: 0.05
Initialization of random number generation:
(1) automatic
(2) specify seed
Enter option: 1
Save intermediate results? y/n n
See results on file Prinda.txt.
Analysis status:
Data file name: EEASolo.txt
11
Data type: (2) quantitative, different units
Session IS saved.
Data transformation
Table 1 indicates the relevance of transformations offered by MULTIV under different data
types. Data transformation is available only when data is quantitative, binary or mixed with both.
In scalar transformations the transformed value X'ij in variable i, sampling unit j, will depend on
the non-transformed value Xij only, while in vector transformation X'ij will depend on the set of
observations in the same variable i (within variables), or in the same sampling unit j (within
sampling units) or in both (double adjustment). Vector transformations within sampling units are
only applicable when the variables are measured in the same scale (or standardized).
Scalar transformation
(1) Presence/absence. The transformed value X'ij=1 if the original value Xij>0 and X'ij=0 if
Xij=0. After this transformation, data is treated as binary, to which will apply the relevant
resemblance measures.
(2) Square root. The transformed value X'ij= |Xij| .
(3) Logarithm. The transformed value X'ij=log(|Xij+1|)
Vector transformation
The following options are available, either within variables or within sampling units:
(1) Standardizing by marginal total. Each observation Xij is divided by the variable i total
(within variables) or the sampling unit j total (within sampling units).
(2) Centering and division by N-1 (implicit in covariance). See (3) below. This
transformation does not involve any standardization.
(3) Centering. The variable i average (within variables) or sampling unit j average (within
sampling units) is subtracted from each observation Xij. Also, this transformation does not involve
any standardization.
(4) Normalization. When it is within sampling units, the observations in each unit j are
p
divided by ( Xhj )1/2, where p is the number of variables. When normalization is within
h=1
n
variables, the observations in each variable i are divided by ( Xik )1/2 where n is the number of
k=1
sampling units.
(5) Centering and (followed by) normalization in the same direction. See (3) and (4) above.
12
(6) Deviations from expected values based on marginal totals. Assumes the input data
matrix F is a contingency table in which each observation Fjh is a frequency count. The adjusted
value F'jh = Fjh/ (Fj. F.h)1/2 - (Fj. F.h)1/2 / F.., where Fj. is the variable j total , F.h is the sampling
unit h total and F.. is the grand total of the table.
(7) Standardizing by the range. Each observation Xij is divided by the range in variable i
(when within variables) or in sampling unit j (when within sampling units). The range is (SUP INF), where SUP and INF are respectively the maximum and minimum values found in the data
vector in consideration.
(8) Rescaling into classes. The range (see (7) above) observed in each variable i (within
variables) or in each sampling unit j (within sampling units) is automatically equally divided in a
specified number of classes. The transformed value X'ij is the class (1, 2, ..., c classes) to which
observation Xij belongs.
(9) Rescaling into ranks. The observations are ranked from 1 to n sampling units (if within
variables) or from 1 to p variables (if within sampling units). Tied ranks are adjusted according to
Kendall and Gibbons 1990 p.40. Computing correlations between variables (sampling units) with
the transformed data within variables (sampling units) will give the Spearman coefficient of rank
correlation.
Table 1. Data transformations offered by MULTIV that are relevant under different data types.
Quant. Quant.
Same Different
scales scales
Binary
Qualit.
Mixed
Quant.+
qualit.
Mixed
Quant.+
binary
Mixed
All
Qual.+b mixed
inary
Scalar
transformations
yes
yes
yes
no
no
yes
no
no
Vector
transformation:
within variables
within s. units
yes
yes
yes
no
yes
yes
no
no
no
no
yes
no
no
no
no
no
The following run performs data transformation in session okop37_Session, which has been
previously saved:
Enter option (valid options with *): n

Wish to specify new data (n) or to open (o) an existing session? n/o o
Enter session name (without the extension .trk): okop37_Session
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
13
*
*
*
*
T transform data
D scatter diagrams
O ordination
C cluster analysis
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): t
Session IS saved.
------------------------------------------------------------------------------DATA TRANSFORMATION
Options:
(0)restore original data (no transformation)
(1)specify transformations on original data
(99)leave as is
Enter option: 1
Scalar transformation:
(0)none
(1)presence/absence
(2)square root (sqrt(|x|))
(3)logarithm (log(|x+1|))
Enter option: 0
Vector transformation:
(0)none
(1)within sampling units
(2)within variables
Enter option: 1
Vector transformation within sampling units:
(includes double adjustment based on marginal totals)
(0)none
(1)standardizing by marginal total
(2)centering and division by N-1 (implicit in covariance)
(3)centering
(4)normalization
(5)centering and normalization (implicit in correlation)
(6)deviations expected values via marginal totals (double adjustment)
(7)standardizing by the range
(8)rescaling into classes
(9)rescaling into ranks
Enter option: 4
14
//When within variable vector transformation is selected first, a second vector transformation
(only within sampling units) can be performed on the current transformed data, as shown in the
following run:
...
Vector transformation:
(0)none
(1)within sampling units
(2)within variables
(0)none
Enter option: 2
Vector transformation within variables:
(includes double adjustment based on marginal totals)
(0)none
(3)centering
(4)normalization
Enter option: 4
Specify more transformations (only within sampling units)? y/n y
Vector transformation within sampling units:
(include double adjustment based on marginal totals)
(0)none
(3)centering
(4)normalization
Enter option: 4
See results in file Prinda.txt.
Resemblance measures
The resemblance measure options available are the ones applicable to the type of data in
hand (see Table 2). Note that some measures carry implicit transformations, which can also be
performed explicitly in the previous menus. In the following descriptions, Xi is the observation in
variable i, sampling unit .
Quantitative and /or binary data
15
Product moment. The similarity of two sampling units and is q = (XiXi) for i =
1, ..., p variables. The similarity of two variables i and k is sik = (XiXk) for = 1, ..., n
sampling units. Note that no transformation is implicit. For other product moments, such as a
centered product moment, covariance or correlation, previous data transformation can be specified.
For instance, to compute centered product moments between variables, choose the option
"centering" in vector transformation within variables; to compute covariances (between variables),
choose the option "centering and division by n-1" in vector transformation within variables; to
compute correlations between variables, choose the option "centering and normalization" in vector
transformation within variables. Correlation can also be selected directly (see below).
Correlation. Instead of choosing product moment with previous transformation, correlation
can also be selected directly, in which case centering and normalization in the proper direction is
automatic.
Absolute value function, also known as Manhattan distance or city block distance. The
distance between sampling units and is a = |Xi Xi| for i = 1, ..., p variables. The
distance between variables i and k is aik = |Xi Xk|.
Euclidean distance. Between sampling units and the distance is d = {(Xi
Xi
for i = 1, ..., p variables. Between variables i and k the distance is dik = {(Xi
Xk)2}1/2.
)2}1/2
Chord distance. This distance between sampling units is equivalent to computing

Euclidean distance with data normalized within sampling units. If the measure is between
variables, data is normalized within variables. On choosing this option, transformation is
automatic.
Legendre-Chodorowski index. This similarity coefficient (varying from 0 to 1) is for the
comparison of sampling units only. It is implemented as described in Legendre & Legendre (1998,
p. 267). The index is a composite of partial similarities computed for each variable, akin to
Gowers index, but not applicable to qualitative data. A parameter (k) is requested.
Bray-Curtis dissimilarity or percentage difference. The dissimilarity (varying from 0 to 1)
between sampling units and is b = |Xi Xi| / |Xi + Xi| for i = 1, ..., p variables. It
does not hold Euclidean metric properties, thus may not be appropriate for ordination based on
eigenanalysis nor to incremental sum of squares clustering.
Second eigenvalue of pairwise correlation. This is the function involved in hierarchical
factorial clustering of variables (Denimal 2001, Camiz et al. 2006), and should be the resemblance
choice for using that clustering method. See cluster analysis section later.
When the data file is a contingency table Mutual information. The data matrix must be a
contingency table. To compare sampling units the function is
p
I =

i=1
=
X i >0
Xi ln
X i X ..
X . X i.
In this, p is the number of variables, X. is the marginal total for sampling unit or and
Xi. is the total of variable i in sampling units and , i.e., Xi + Xi. The function measures the
16
mutual information between rows and columns classification criteria in the p variables x 2
sampling units contingency table extracted from the complete contingency table (see Feoli et al.
1984, Pillar & Orlci 1993). To compare variables the function is also defined, but the
contingency table is then 2 variables x n sampling units.
Table 2. Resemblance measures offered by MULTIV that are relevant under different data types
for comparing sampling units.
Quant. Quant.
Same Different
scales scales
Binary
Qualit.
Mixed
Quant.+
qualit.
Mixed
Quant.+
binary
Mixed
All
Qual.+b mixed
inary
yes
yes
no
no
yes
yes
no
no
no
no
no
no
no
no
no
no
yes
yes
yes
no
no
no
no
no
no
yes
no
no
no
no
yes
no
yes
no
no
no
no
no
no
yes
no
no
no
no
no
no
yes
no
no
no
no
no
no
yes
no
no
no
no
no
no
yes
no
no
no
no
no
no
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
yes
yes
yes
yes
yes
yes
no
no
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
yes
no
no
no
no
no
no
no
no
no
no
no
yes
yes
yes
yes
no
no
no
no
yes
yes
yes
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
Resemblance
measures between
sampling units:
Product moment
Absolute value
function
Euclidean distance
Mutual information
Gower index
MDIS
Mutual entropy
Rajski's metric
Coherence
coefficient
Chi-square
Jaccard
Simple matching
Sokal-Michener
Ochiai
Sorensen
Mean square
contingency
Correlation
Chord distance
Legendre&Chodorowski
Bray&Curtis
17
Table 3. Resemblance measures offered by MULTIV that are relevant under different data types
for comparing variables.
Quant. Quant.
Same Different
scales scales
Binary
Qualit Mixed
.
Quant.+
qualit.
Mixed
Quant.+
binary
Mixed
All
Qual.+b mixed
inary
yes
yes
no
no
yes
yes
no
no
no
no
no
no
no
no
no
no
yes
yes
no
no
no
no
no
no
no
no
no
no
no
no
yes
no
no
yes
yes
yes
yes
no
no
no
yes
yes
yes
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
yes
yes
yes
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
yes
yes
no
no
yes
no
no
no
no
no
no
no
yes
yes
no
no
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
no
no
yes
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
yes
yes
no
no
yes
yes
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
no
Resemblance
measures between
variables:
Product moment
Absolute value
function
Euclidean distance
Mutual information
Gower index
MDIS
Mutual entropy
Rajski's metric
Coherence
coefficient
Chi-square
Jaccard
Simple matching
Sokal-Michener
Ochiai
Sorensen
Mean square contingency
Correlation
Chord distance
Legendre&Chodorowski
Bray&Curtis
Second eigenvalue
pairwise correlation
Qualitative and/or binary data

Resemblance measures in this category are only applicable to comparisons between
qualitative and binary variables. Most of the measures are based on information theory (see Orlci
1991). For each pair-wise comparison a contingency table is computed containing joint
frequencies of the variables' states. The contingency tables are saved in the file Tables.txt when the
option to "Include data matrix in the output file?" is set to 'y'.
Mutual entropy. The Rnyi information function of order measures the entropy shared by
two variables, given by
si

Hik
sk
ln
j=1
h=1
p jh
( p j . p.h ) 1
1
18
In this, fjh is the joint frequency of states j and h, of variables i and k, si and sk are the number of
fjh
f.h
fj.
states in variables i and k, pjh = n , pj. = n , p.h = n , fj. = sk fjh (frequency of state j, variable i),
h=1
f.h =
sh f
j=1 jh
(frequency of state h, variable k) and is a scaling factor, specified by the user, in the
interval 0 except = 1. When approaches 1 ( = 1 specified by the user), the function

reduces to
si
Hik =
sk

j=1
pjh ln
h=1
p jh
p j. p.h
In the resemblance matrix, each element in the diagonal (Hii ) is the entropy in the frequency
distribution of the states of variable i; when approaches 1, this is the Shannon-Weaver index.
The coherence coefficient, Rajski's metric and MDIS are derived from the mutual entropy.
Coherence coefficient. Is a similarity index expressing mutual entropy in relative terms,
varying from zero (no association) to 1 (strong association), defined by ik =
1 - (dik)2 , where
dik is the Rajski's metric (see below).

Rajski's metric. It is a dissimilarity, varying from zero (maximum association) to 1 (no
association), defined as

H i+k
H ik

H i+k
dik =
In this, Hi+k is the joint entropy Hi+k = Hii + Hkk - Hik , where Hii and Hkk are entropies in the
frequency distributions of the states of variables i and k.
MDIS (Kulbach's Minimum Discriminant Information Statistics). The association between
two qualitative variables i and k can be measured by
si sk
2Iik = 2f..Hik = 2 [fhj ln (fhj/fohj)]

h=1j=1
where fohj = (fh.f.j)/f.. is the expected value based on marginal totals, Hik is the mutual entropy of
order 1 and the other terms defined as for the mutual entropy. The function approximates the chisquare (below).
Chi-square. Measures the association between variables i and k as 2ik =
si
sk

h=1
j=1
( f hj f
f
ohj
ohj
)2
, being the terms as defined above.
Binary data
For each pair-wise comparison between variables or between sampling units a contingency
table is computed containing joint frequencies of the variables' 0/1 states. As for qualitative data,
the contingency tables are saved in the file Tables.txt when the option to "Include data matrix in
19
the output file?" is set to 'y'. When comparing variables (or sampling units), the terms are defined
as:
Variable k
(sampling unit )
Variable i (sampling unit ) 1

0
Total
Total
a+b
c+d
a+c
b+d
q=a+b+c+d
When comparing variables, q is the total number of sampling units; when comparing sampling
units, q is the total number of variables. The following similarity functions measure association,
varying from zero (no association) to 1 (complete association). There is positive association
between variables (or sampling units) if a > (a+b)(a+c)/q and negative if a < (a+b)(a+c)/q.
a
Jaccard. Sik (or S) = a+b+c
Simple matching. Sik (or S) = (a+d)/q. Since a+d is the total number of matchings, this
measure can also be used to compare sampling units on the basis of qualitative variables (Gower
1971), i.e., S is the number of variables in which the units match divided by the number of
variables.
Sokal-Michener. Sik (or S) = a/q
a
Ochiai. Sik (or S) = [(a+b)(a+c)]1/2
2a
Sorensen. Sik (or S) = 2a+b+c
(ad-bc)2
2
2
Mean square contingency. rik (or r) = [(a+b)(c+d)(a+c)(b+d)]
Note that r2 can be derived from the previously described correlation coefficient (squared) and chi2
square (r2 = q ).
Mixed data
Gower index. It is a similarity coefficient (varying from 0 to 1), only applicable to compare
sampling units, defined as a weighted average of p partial similarities Sh between sampling units
and (Gower 1971):
20
th h
S =
h=1
p
j
j=1
Where th and h are set according to the type of variable h:
(1) If variable h is binary, a similar rule as in the Jaccard coefficient prevails, i.e., th = 1 and
h = 1 if h is present in both sampling units, th = 0 and h = 1 if the sampling units do not
coincide in variable h, th = 0 and h = 0 if h is absent in both sampling units. MULTIV also
offers the option for counting matching absences, a similar rule as in the Simple matching
coefficient, that is, th = 1 and h = 1 if h is absent in both sampling units.
(2) If variable h is qualitative, th = 1 if sampling units , coincide for variable h, and th = 0
if they do not. In both cases h = 1.
|Xh Xh|
(3) If variable h is quantitative, th = 1 (max Xh min Xh) and h = 1, where Xh is the
value of variable h, sampling unit and Xh is the value of the same variable in sampling unit .
The following run computes Euclidean distances between sampling units:
------------------------------------------------------------------------------MAIN MENU
* T transform data
O ordination
C cluster analysis
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): r
Session IS saved.
------------------------------------------------------------------------------RESEMBLANCE MEASURES
Options:
(1)between sampling units
(2)between variables
Enter option: 1
21
Type of resemblance measure (1)between sampling units:
(1)product moment
(2)absolute value function
(3)Euclidean distance
(4)mutual information
(5)Gower index
(17)correlation
(18)chord distance
(19)Legendre-Chodorowski index
(20)Bray-Curtis dissimilarity
Enter option: 3
The following run reads the first 4 binary and qualitative variables from file P&Op87.txt and
computes a coherence coefficient between variables:
------------------------------------------------------------------------------MAIN MENU
...
Enter name of file containing data matrix: P&Op87.txt
13
Types of variables:
(1) only quantitative variables, same measurement scale
(2) only quantitative variables, different scales
(5) several types
Enter option: 5
Enter variable type (1:binary,2:qualitative,3:quantitative):
Variable: 1 2 3 4
1 1 2 2
Type:
Wish to correct? y/n n
Enter option: t
(2) read labels after data on file P&Op87.txt
Enter option: 1
See data on file Prinda.txt.
Wish to rename this session? y/n y
Enter name for the session: P&Op87_4var_Session
//In this case, since the file is ordered by variables, we could read the first four variables,
ignoring the last one.
Analysis status session P&Op87_4var_Session:
22
Variable: 1 2 3 4
Type:
1 1 2 2
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
T transform data
O ordination
C cluster analysis
P randomization testing comparing groups of sampling units
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): r
------------------------------------------------------------------------------RESEMBLANCE MEASURES
Options:
(1)between sampling units
(2)between variables
Enter option: 2
Type of resemblance measure (2)between variables:
(6)MDIS
(7)mutual entropy
(8)Rajski's metric
(9)coherence coefficient
(10)chi-square
Enter option: 9
Enter order of entropy measure (alpha in Reyni's general entropy): 12
Include resemblance matrix in the output file? y/n y
See contingency tables in file Tables.txt.
Variable: 1 2 3 4
Type:
1 1 2 2
Resemblance measure: (9)coherence coefficient, (2)between variables
Order of entropy measure (alpha in Reyni's general entropy): 12
Session IS saved.
23
//When the resemblance measure involves the construction of a contingency table for each pairwise comparison, and 'Include data matrix in the output file' is set to y, the contingency tables
are saved in a file named Tables.txt.
Ordination
Ordination is offered if a resemblance matrix is available. Three methods based on
eigenanalysis (principal coordinates analysis, principal components analysis and correspondence
analysis) are available in MULTIV.
If the selected method is principal coordinates analysis or principal components analysis,
the user may choose the option for performing bootstrapped ordination to evaluate the significance
of ordination axes. The algorithm uses bootstrap resampling and has been proposed by Pillar
(1999b). If bootstrap samples with increasing sizes are selected, the results can be used to evaluate
sampling sufficiency (Pillar 1998). The output file in Prinda.txt will contain probabilities P(io
i*). The probability P(io i*) is an indicator of the strength of the structure in an ordination, as
compared to the ordination of a null data set containing variables with the same observed
distribution but zero expected association. Setting an probability threshold will help the
interpretation of P(io i*). A small P(io i*), that is, smaller than , will indicate that the
ordination dimension in consideration is significantly more stable than that would be expected for
the same dimension in the ordination of a random data set. In this case we reject the null
hypothesis and conclude, with a probability P(io i*) of being wrong, that the given ordination
dimension is nontrivial and worthy of interpretation. Otherwise, we accept the null hypothesis and
consider the ordination dimension unstable and indistinguishable from a random data ordination.
Alternatively, we may adopt a more flexible view on the significance, and consider the resulting
P(io i*) values as representative of a gradient of reliability.
The algorithm for bootstrapped ordination involves Procrustean adjustments. The number
of dimensions involved in Procrustean adjustments is likely to affect the fitted scores and their
correlation with the reference scores. The interpretation of the probabilities should start from the
highest (the least relevant) to the first ordination dimension. In metric ordination the dimensions
are uncorrelated and meaningfully ordered by relevance. If the second dimension is significant, it
is expected that the first will also be significant. Likewise, if the third dimension is significant, the
second and first ones will be as well, and so on. Once dimension i is deemed significant, the test
may stop and all lower and more relevant i 1 dimensions are also considered significant
irrespective of the corresponding probabilities.
An insufficient sample size may cause type II error; that is, the test may not detect
significance of dimensions that would be found significant if the sample were larger. Sample size
sufficiency is indicated by stability of probabilities under increase of sample sizes (Pillar 1998). If
sample size is deemed sufficient and P(io i*) is larger than , the ordination axis is truly
irrelevant. Otherwise, we may say that a larger sample size would be needed to reach a more
confident conclusion.
Principal coordinates analysis.

Available for any resemblance matrix between sampling units (ordination of sampling
units) or variables (ordination of variables). If the resemblance matrix contains dissimilarities they
24
are squared; if they are similarities (sik), they are transformed into squared dissimilarities (ik = sii
+ skk - 2 sik). The squared dissimilarities ik are transformed into products by qik =
1
2
2 (ik - i./n - k./n + ../n ) , where n is the number of sampling units (variables). Matrix Q is
subjected to eigenanalysis. If Q has no Euclidean metric properties, meaning that negative
eigenvalues were found, a warning is shown in the results, based on the discrepancy between the
total sum of the eigenvalues and the trace of Q. The scores yik of the sampling units (variables) in
each ordination axis are in the eigenvectors of Q adjusted according to the magnitude of the
corresponding eigenvalue i, such that the sum of squared scores equals the eigenvalue, that is, yik
n
= bik ( i /
bih2 )1/2. Correlation coefficients between the original descriptors and each
h
ordination axis are computed, allowing the interpretation of variation in the ordination space.
Principal components analysis

Can be performed only when the resemblance function is a product moment or a
correlation, no restrictions regarding data transformation (a covariance matrix can be obtained by
applying the appropriate transformation). If the resemblance is between variables, the ordination is
of sampling units, and vice-versa. The resemblance matrix S is subjected to eigenanalysis. The
contribution of each component is indicated by the percentage of the corresponding eigenvalue in
relation to the total (trace of S). The normalized eigenvectors (B) are the component coefficients.
The scores of sampling units (variables) are computed by B'A, where A is the data matrix (after
vector transformations, implicit in the resemblance measure or not). The rows in A correspond to
variables when S is between variables, and vice-versa. The results also present the correlation
coefficients between the original variables (or sampling units if S is between sampling units) and
each component. The contribution of variable (sampling unit) i in ordination component k is
defined as uik = ( bki sii )2 /
uhk where p is the number of variables (sampling units).

h=1
The following run performs a principal coordinates analysis:

Resemblance measure: (3)Euclidean distance, (1)between sampling units
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
* T transform data
* O ordination
* C cluster analysis
25
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option: o
Ordination method:
(1)principal coordinates analysis
(2)principal components analysis
(3)correspondence analysis
(99)return to main menu
Enter option: 1
Save intermediate results? y/n y
//The option for bootstrapped ordination that follows is only offered if the ordination is of
sampling units and the method is principal coordinates analysis or principal components
analysis. The maximum number of ordination axes to be monitored is bounded by the rank
evaluated above, but it sets the minimum initial sample size in process sampling. Process
sampling will take bootstrap samples with sizes varying from the initial to the total number of
sampling units. The number of steps in process sampling will be a function of these choices. If
the initial sample size is equal to the total number of sampling units, there will be only one step
in process sampling.
Evaluate significance of ordination axes using bootstrap resampling? y/n y
Enter the maximum number of ordination principal axes to be monitored: 3
//A minimum of 1000 iterations in bootstrap resampling is recommended. At each step in

process sampling, this number of bootstrap samples will be taken with replacement and
subjected to the chosen ordination method. By specifying the same number for initialization the
analysis will reproduce identical results for the same data and options, but initialization of
random numbers should normally be automatic.
(1) automatic
(2) specify seed
Enter option: 1
Results saved on file Prinda.txt
//Since the option for bootstrapped ordination was set, in the Macintosh version the profiles
with the probabilities P(io i*) at increasing sample size for each ordination axis, starting
from the least important axis are shown in sequence on screen. Under Mac OS X you can use
the utility Grab to capture and save each profile. You have to close the profile window (click in
the upper left box) for the next graph, if any, to appear on screen.
26
Correspondence analysis
The method is applicable if the data set is a contingency table (or at least, in MULTIV, all
variables are in the same units). The data table must be previously transformed to deviations from
expected values based on marginal totals (see data transformation above) and the resemblance
measure must be a product moment between variables (or between sampling units). Though the
data set may be a contingency table, by convention the rows are variables when the resemblance is
comparing variables or the rows are sampling units when the resemblance is comparing sampling
units. The resemblance matrix (S) is subjected to eigenanalysis, in which the t eigenvalues are the
2
squared canonical coefficients R1 , ..., Rt . The eigenvectors of S are the scores, which after
adjustment correspond to the scores for the variables (if S compares variables):
Uhm = hm
X..
Xh.
where Uhm is the score of variable h in ordination axis m and Xh. and X.. are respectively the row
h total and the grand total in the non-transformed contingency table. The scores for sampling units
are computed by:
p XhjUhm
Vjm = X.jRm
h=1
where Vjm is the score of sampling unit j in the ordination axis m. If S is comparing sampling
units, the U scores will correspond to sampling units and the V scores to variables.
Scatter diagrams
Two dimensional scatter diagrams can be produced on screen using as axes
original variables, transformed variables or available ordination scores.
Scatter diagrams use the same scale in the horizontal and vertical axes. The
axes are not drawn through the origin. When the ordination is of sampling
units, biplots can be produced using correlations between variables and
ordination axes (see Podani 2000, p. 227). In this case the plot scale is
defined by the sampling unit scores, while the correlations are rescaled to the
maximum unit score.
------------------------------------------------------------------------------MAIN MENU
...
------------------------------------------------------------------------------Enter option (valid options with *): d
Ordination scores available: (1)principal coordinates analysis
Session IS saved.
Use as coordinates of the scatter diagram:
(1)original variables
(2)original variables transformed
27
(3)ordination scores
Enter option: 3
Type of scatter diagram:
1 plot sampling units only
2 biplot of sampling units (scores) and variables (correlations with axes)
Type option no.: 2
Labels to use in the scatter:
0 none (only squares)
1 sampling unit labels
Type option no.: 1
Connect sampling unit points in scatter? y/n n
Enter component for the horizontal axis: 1
Enter component for the vertical axis: 2
//The option to connect sampling unit points is relevant when they are spatially or temporally
ordered
//The following window is shown on screen (only in the versions for Macintosh systems):
28
//Under Mac OS X the graph can be captured by using the system utility Grab. To close the
window, click in the upper left box, after which the main menu will appear.
//Under Windows OS the diagram will not be produced, but based on file ScatterData.txt, stored
in the same folder with Prinda.txt, a similar diagram can be drawn using other tools (if you use
Excel, make sure the decimal separator is consistent with your system configuration).
Cluster analysis and dendrograms

The methods implemented are single linkage, complete linkage, average clustering
(UPGMA and WPGMA), incremental sum of squares and a topological constrained version of it,
and hierarchical factorial clustering (the latter only for clustering variables). The analysis uses the
resemblance matrix available (between variables or between sampling units). The algorithm is
agglomerative, at the beginning all groups are formed by one object (a variable or a sampling
unit):
(1) Each agglomerative step joins the two most similar groups in the matrix.
(2) The resemblance matrix is redefined according to the new group that was formed in step 1.
(3) The process is repeated until all objects are joined in one group. There will be n-1 clustering
steps, where n is the number of objects (variables or sampling units).
The clustering methods implemented differ in the criterion for redefinition of inter-group
resemblance (step 2 above). The clustering process is automatically represented in a dendrogram
on screen (Macintosh version only), which can be saved as explained for scatter diagrams. The
dendrogram presents several partition possibilities; a decision which is left to the user. The
numerical results in file Prinda.txt will also contain the possible partitions indicated by the
analysis.
The user may choose the option for performing analysis of group partition sharpness using
bootstrap resampling. The method is described in Pillar (1999a). If bootstrap samples with
increasing sizes are selected, the results can be used to evaluate sampling sufficiency (Pillar 1998).
The output file in Prinda.txt will contain probabilities P(GG*) for each partition level, from 2
groups up to the maximum specified, and for each sample size. The probability P(GG*) is the
proportion of bootstrap iterations in which G is found smaller than or equal to G*. If P(GG*) is
not larger than a specified threshold , we conclude, with a probability P(GG*) of being wrong,
that the k groups in the partition are not sharp enough to consistently reappear in resampling. That
is, we reject the null hypothesis and conclude that the groups are fuzzy. If, instead, we accept the
null hypothesis, we conclude that there is not enough evidence to refute that the groups are sharp.
Sample size influences the power of the group partition stability test. A small sample size
will likely increase the error type II, the probability of accepting the null hypothesis when it is
actually false. That is, a P(GG*) larger than a chosen threshold may either indicate the group
structure is sharp or that the sample size is too small. Sample size sufficiency can be evaluated by
examining the stability of probabilities across a series of sample sizes (process sampling). The user
chooses the initial sample size and the number of sampling units added at each sampling step. For
this, probabilities P( Gk0 Gk* ) will be found by bootstrap resampling for each sample size nk n
taken from the observed sample with n sampling units (Pillar 1998). If the null hypothesis is
accepted (P(GG*) > ), the sample is considered sufficient if the probabilities reach stability
within the range of sample sizes evaluated. If, instead, the null hypothesis is rejected (P(GG*)
29
), and the probabilities are decreasing and not yet stable, it is likely that the conclusion will not
change with larger samples.
Single linkage (nearest neighbour)

When the resemblance matrix contains dissimilarities, the dissimilarity between groups P
and Q is defined as:
dPQ = INF [ djk, for j=1, ..., n-1 and k=j+1, ..., n objects, provided j is in P and k in Q ]
where djk is an element in the resemblance matrix and INF is the lowest value in the set indicated
between brackets. When the resemblance matrix contains similarities, the similarity between
groups is defined as above, but SUP is used instead.
Complete linkage
When the resemblance matrix contains dissimilarities, the dissimilarity between groups P and Q is
defined as:
dPQ = SUP [ djk, for j=1, ..., n-1 e k=j+1, ..., n objects, provided j is in P and k in Q ]
with terms as defined for the single linkage criterion. When the resemblance matrix contains
similarities, INF is used instead.
Average clustering
The resemblance (similarity or dissimilarity) between groups P and Q is the arithmetic
average of the resemblances between the objects in one group and the ones in the other group. In
unweighted arithmetic average clustering (UPGMA), the resemblances are not weighted by group
sizes. In weighted arithmetic average clustering (WPGMA), the resemblances are weighted by the
group sizes. The implemented algorithms followed Legendre & Legendre (1998).
Incremental sum of squares

The method has been described by Ward (1963) and Orlci (1967). It works on a
dissimilarity matrix. If similarities are available, they are transformed into squared dissimilarities
(dik2 = sii + skk - 2 sik). The clustering criterion minimizes
npnq
QPQ = (np+nq) dPQ2
where dPQ2 is the squared distance between centroids of groups P and Q, and np, nq are the
number of elements in groups p and q. QPQ can be computed by difference from the total and
within sum of squares:
QPQ = QP+Q - QP - QQ
where:
30
1
QP+Q = np + nq
h
dhi2 for h=1, ..., n-1 and i=h+1, ..., n objects, provided h is in group
i
P and i in group Q.
1
QP = np dhi2 for h=1, ..., n-1 and i=h+1, ..., n objects, provided h and i are in group P
h i
1
QQ = nq dhi2 for h=1, ..., n-1 and i=h+1, ..., n objects, provided h and i are in group Q
h i
QPQ expresses the increment in the variance within group P+Q comparing to the variance existing
within groups P and Q if isolated.
The option for performing a topological constrained clustering (Grimm 1987) with this
method is also available. The user may indicate the topology is given by the neighborhoods in the
sampling unit order in the data set or specify an upper half matrix with the topology (using 1 for
adjacent pairs and 0 for otherwise). The option for evaluating group sharpness by bootstrap
resampling is not available under topological constrained clustering.
Hierarchical factorial clustering of variables

The method, initially proposed by Denimal (2001, see Camiz et al. 2006), is based on a
sequence of PCA-like eigenanalyses so that, at each step, a principal plane is created where we can
represent both the variables belonging to the group and the units as seen by these variables only. In
addition, the first principal component of each PCA is adopted as a representative variable of the
corresponding group, a role similar to the centroid in the clustering of units.
The following run performs cluster analysis:

------------------------------------------------------------------------------MAIN MENU
* T transform data
* O ordination
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): c
31
Ordination scores available: (1)principal coordinates analysis
Session IS saved.
Clustering method:
(1)simple linkage (nearest neighbor)
(2)complete linkage
(3)incremental sum of squares (Ward)
(4)average linkage (UPGMA)
(5)average linkage (WPGMA)
(6)hierarchical factorial (Denimal & Camiz)
(7)topologically constrained incremental sum of squares
Enter option: 3
//The following window is shown on screen (only for Macintosh versions):
//Under Mac OS X the graph can be captured by using the system utility Grab. To close the
window, click in the upper left box, after which the main menu will appear.
//Under Windows OS the graph will not be produced, but all possible partitions based on the
dendrogram are presented in Prinda.txt.
32
//The option that follows, for evaluating group partition sharpness using bootstrap resampling,
is only offered if the clustering is of sampling units. The maximum number of groups evaluated
is bounded by the number of sampling units minus one and sets the minimum initial sample size
in process sampling. Process sampling will take bootstrap samples with sizes varying from the
initial to the total number of sampling units. The number of steps in process sampling will be a
function of these choices. If the initial sample size is equal to the total number of sampling
units, there will be only one step in process sampling.
Evaluate group partition sharpness using bootstrap resampling? y/n y
Enter the maximum number of groups (group partition level) to be evaluated: 3
//A minimum of 1000 iterations in bootstrap resampling is recommended. At each step in

process sampling, this number of bootstrap samples will be taken with replacement and
subjected to the chosen clustering method. By specifying the same number for initialization the
analysis will reproduce identical results for the same data and options, but initialization of
random numbers should normally be automatic.
(1) automatic
(2) specify seed
Enter option: 1
Results saved on file Prinda.txt
//Since the option for bootstrapped resampling was set, in the Macintosh version the profiles
with the probabilities P(GG*) at increasing sample size for each group partition level are
shown in sequence on screen. Under Mac OS X you can use the utility Grab to capture and save
each profile. You have to close the profile window (click in the upper left box) for the next
graph, if any, to appear on screen.
Randomization tests: comparing variables

This option is offered when the available resemblance is between variables. The test
criterion is any pair-wise resemblance function already defined and pertinent to the type of
variables, e.g., correlation coefficient, chi-square, coherence coefficient, Jaccard, Euclidean
distance. If the Null Hypothesis (Ho) of no association (correlation) between the variables pairwise is true, the state observed in a variable in a given sampling unit is independent from the states
observed in the other variables. For a data set with t variables and n sampling units, the complete
reference set will contain (n!)t-1 permutations. See Manly (1991) for details. A random data set
member of the reference set is generated by shuffling the observations in each variable among the
sampling units. At each random iteration a random member of the reference set is generated, the
resemblance matrix is computed and each resemblance value for a variable pair (riko) is compared
to the corresponding one found in the observed data set (rik). After many iterations, if P(riko r) is
small (smaller than ), Ho is rejected and we conclude that there is significant association
(correlation) between variables i and k. It is recommended to use at least 5000 iterations when =
0.05. The results present probabilities for all variable pairs.
33
The following run implements a randomization test to evaluate the significance of the
coherence coefficient between qualitative variables:
Variable: 1 2 3 4
Type:
1 1 2 2
Session IS saved.
------------------------------------------------------------------------------HYPOTHESIS TESTING VIA RANDOMIZATION
Initialization of random numbers:
(1) automatically
(2) wish to specify a seed
Enter option: 1
Number of iterations for randomization test: 1000
Save random data and results of each iteration? y/n n
The results in file Prinda.txt are:

------------------------------------------------------------------------------RANDOMIZATION TEST
------------------------------------------------------------------------------Fri Feb 13 17:57:16 2004
Elapsed time: 1 seconds
Variable: 1 2 3 4
Type:
1 1 2 2
Session IS saved.
Number of iterations: 1000
Random number generation initializer: 1076695032
Test criterion (Lambda):
Results:
(Resemblance matrix below the main diagonal and
corresponding probabilities P(|Lambda random| >= |Lambda observed|) above the
main diagonal)
0
0.234
0.375
0.227
0.71909
0
0.342
0.575
0.54919
0.61006
0
1
0.71909
0.39375
0
0
//The matrix presents each resemblance value below the main diagonal in row i column k and
the corresponding probability above the main diagonal in row k column i. In this case, with a
34
4x4 matrix comparing the variables 1 and 3, the probability of finding a coherence coefficient
larger than 0.54919 if Ho were true is 0.375.
Randomization tests: comparing groups of sampling units

This option is offered when the available resemblance is between sampling units. The
procedures are identical for univariate or multivariate comparisons. Groups of sampling units must
also be specified (see at the end of this section). Different test statistics can be chosen. When the
sum of squares between groups or the pseudo F-ratio is selected as test statistic, the results are
interpreted similarly to the ones in an analysis of variance table; the method is described in Pillar
& Orlci (1996), with improvements after Pillar (2004). MULTIV also implements MRPP
(Mielke & Berry 2001). Pillar (2004) evaluated by data simulation the accuracy and power of the
different methods described in the sequel.
The methods are based on two sets of information. The first is the matrix of p variables
describing n units and the second is one or more factors with discrete states defining k groups of
units. An n-by-n dissimilarity matrix obtained from the first matrix is needed for computing the
chosen test statistic according to the k groups. MULTIV can perform the analysis with any type of
resemblance matrix, irrespective of bearing or not Euclidean metric properties.
The test criterion may be the sum of squares between groups ( Qb ) computed according to
Pillar and Orlci (1996):
Qb = Qt Qw
where
n1
Qt =
1
n
d
2
hi
h=1 i= h +1
is the total sum of squares of n(n-1)/2 pair-wise squared dissimilarities between n sampling units
and
k
Qw = Qwc
c=1
is the sum of squares within k groups, such that

n c 1 n c
Qwc =
1
nc
d
2
hi|c
h=1 i= h +1
2
where d hi|c
is comparing units belonging to group c, which contains nc sampling units.
In one-factor analysis, the k groups are given by the states of the factor. In two-factor
analysis, as explained later, Qb will be partitioned and the groups will be the ones defined by each
factor in isolation, for the main effects, or their joint states for the interaction.
This partitioning of the total sum of squares based on the distance matrix is not constrained
by the type of variables describing the units (e.g., means are not defined for qualitative variables),
provided an appropriate dissimilarity measure is found that meets the requirements for the
partitioning, allowing the use of randomization testing with qualitative and mixed data types.
The test criterion may as well be the pseudo F-ratio
35
F = Qb / Qw
In randomization testing, Qb / Qw is equivalent to the pseudo F-ratio used by Edgington (1987) and
Anderson & ter Braak (2003), i.e. for the same permutations both statistics will give identical
probabilities. There is no need to divide the sum of squares by the degrees of freedom in the
numerator and denominator, since the degrees of freedom are constant over all permutation
iterations. Furthermore, the F-ratio is equivalent to Qb in randomization testing for one-factor
multivariate analysis of variance; in two-factor designs they are also equivalent when the test is
exact (this will be further explained in the next sections).
Furthermore, the chosen test criterion may be the average dissimilarity within groups, as
described for the multiresponse permutation procedure (MRPP, Mielke & Berry 2001). In this the
test criterion is
k
= n1c c
c=1
is a weighted average dissimilarity within k groups, such that

n c 1 n c
c =
1
n c (n c 1)/ 2
d
hi|c
h=1 i= h +1
The only option implemented for MRPP in MULTIV is this MRPP statistic using the non-squared
dissimilarities (v = 1 in Mielke & Berry 2001). Another choice would be using the squared
distances (v = 2, Mielke & Berry 2001) and Cc = (nc1)/(nk) as group weight for each group c
with nc sampling units, where k is the number of groups, but this statistic is equivalent to Qb, that
is, will give identical probabilities in a randomization test with the same set of permutations.
Random permutations in one-factor designs
If the null hypothesis (Ho) is true, the observation vector in a given unit is independent
from the group to which the unit belongs. The observed data set is seen as one of the possible
permutations of observation vectors among the units and their groups in the data matrix. An
example is given in the following table:
Factor groups
Observation vector identities
1 2 3 4 1 2 3 4
1 2 3 4 5 6 7 8
One permutation:
Factor groups
Observation vector identities
1 2 3 4 1 2 3 4
6 8 1 4 5 7 2 3
The observation vectors contain p variables and are permuted intact, preserving the
correlation structure in the data. In some cases with more than one factor, random permutations of
the vectors may be restricted within the levels of one factor or units belonging to the same factor
combination may be permuted as blocks (more details will be given later). Over permutations, Qt
remains constant; it is only redistributed among Qb and Qw . Therefore, the test is exact (Anderson
& ter Braak 2003). The same is valid for the MRPP.
36
Thus, the basis of randomization testing is to randomly permute the data according to Ho
and for each of these permutations compute the test criterion (e.g. Qbo ), comparing it to the value
of Qb found in the observed data. The probability P( Qbo Qb ) will be given by the proportion of
permutations in which Qbo Qb . For the MRPP, since measures within group dispersion, the
probability P(0 ) is the relevant one, which will be given by the proportion of permutations in
which the test statistic under H0 is smaller than or equal to the observed test statistic (0 ).
The collection of all possible permutations, the reference set, can be generated
systematically, but the number of permutations may be too many to be all considered in
computations. A random but still large sample of this reference set is sufficient for generating the
probabilities (Hope 1968, Edgington 1987). A minimum of 1000 and 5000 iterations for tests at
the 5% and 1% significance levels is recommended to obtain reliable results, that is, P-values close
to the exact ones that would be obtained in complete systematic data permutation (Edgington
1987, Manly 1991). By convention, the observed data set is included as one of these permutations,
thus determining a minimum probability of 1/B, where B is the number of random permutations.
Ho will be rejected if P( Qbo Qb ) , where is the significance level chosen for the test.
If Ho is rejected, we conclude that the groups differ.
Contrasts
When there are more than two groups, contrasts are used to find which groups are different
from each other. Contrasts are as well evaluated by randomization testing (Pillar & Orlci 1996).
The computation of sum of squares for evaluating contrasts is done similarly as in the previous
explanations. Contrast coefficients help indicating which groups are involved in the contrast. For
instance, with three groups, these contrasts could be defined as:
cd: 0 1 -1 -> compares 2 vs. 3
ce: 2 -1 -1 -> compares 1 vs. 2 and 3 taken with equal weight
For any contrast i there will be a total sum of squares involving the dissimilarities of units
belonging to the groups in the contrast ( Qti ), a sum of squares within groups with positive contrast
coefficients ( Qwi+ ), and a sum of squares within groups with negative coefficients ( Qwi ). The
between groups sum of squares for the contrast i is
Qbi = Qti (Qwi+ + Qwi )
For u = k-1 independent contrasts,

u
Qb = Qbi
i=1
(To be noted is that two contrasts are independent when the product of the coefficient vectors is
zero). Therefore, for each contrast i there is a test criterion Qti for which a probability will be
found in randomization.
The permutations of observation vectors are done similarly as for the evaluation of the
general Ho, but only for the vectors and units belonging to the groups in the contrast. In versions
of MULTIV previous to version 2.3 permutations involved all vectors irrespective of belonging or
not to the contrast and therefore some contrast probabilities, depending on the data, may have been
inaccurate or overestimated. For instance, power may have been lower in contrasts having smaller
differences when other contrasts in the same set had larger differences.
37
Random permutations in two-factor designs

In two-factor designs the k groups are defined by the joint states of the two factors, i.e., k =
l A x lB , where each factor is individually defining l groups (levels). The sum of squares between
these k groups can be partitioned as
Qb = Qb| A + Qb|B + Qb| AB
where Qb| A is the sum of squares between l A groups according to factor A disregarding factor B,
Qb|B is the sum of squares between lB groups according to factor B disregarding factor A, and
Qb| AB is the sum of squares of the interaction AB, obtained by difference.
In two-factor designs, there is no general agreement on how permutations should be done
(Edgington 1987, Manly 1991, Anderson & ter Braak 2003). Permutations may be unrestricted,
similarly as for one-factor designs, but with the groups defined by the joint states of the two
factors. For each permutation, the test criteria for each factor and for the interaction are computed
and compared to the observed ones. In case the permutations are unrestricted, the denominator
used in the F-ratio computation should be Qw . An example is given in the following table:
Groups factor A
Groups factor B
Observation vector
identities
11112222
11221122
12345678
One permutation:
Groups factor A
Groups factor B
Observation vector
identities
11112222
11221122
68145723
Unrestricted permutations will lead the test not being exact, for Qt will be redistributed
between more than two terms, i.e., Qb| A , Qb|B and Qw (Anderson & ter Braak 2003). However,
depending on the data, the probabilities when Ho is true (type I error) may be acceptably close to
the nominal significance level. No general solution has been found for an exact test of the
interaction (Anderson & ter Braak 2003).
In two-factor crossed (not nested) designs, for testing one factor, permutations may be
restricted to occur within the levels of the other factor (Edgington 1987). In this case, permutations
will have to be done separately for finding the probabilities for the main effect of each factor. In
the absence of interaction, such a restriction will give an exact test, for when testing for factor A,
Qb|B will be kept constant over permutations and hence Qt will be redistributed among Qb| A and
Qw , and vice-versa (Anderson & ter Braak 2003). The F-ratio is calculated using Qw in the
denominator. The following table gives an example of restricted permutation within the levels of
factor A:
38
Groups factor A
Groups factor B
Observation vector
identities
One permutation:
Groups factor A
Groups factor B
Observation vector
identities
11112222
11221122
12345678
11112222
11221122
24315786
In two-factor nested designs, for testing the nested factor, permutations should be restricted
to occur within the levels of the nesting factor (Edgington 1987). This will give an exact test. For
testing the nesting factor, however, observation vectors belonging to the same combination of the
two factors would be permuted as blocks, which as well will give an exact test.
Block designs are nested designs in that the reference set is defined by restricting random
allocations to within the blocks. The analysis is similar to a design with two factors (blocks,
treatments), but in this case the sum of squares for the interaction term is not defined and the
probability for the block factor is not computed.
Use of residuals
The use of observation vectors in which the effects of one or of both factors were removed
has been suggested to overcome the impossibility of exact tests (e.g., for interactions) or when the
available exact tests are too restrictive regarding power (McArdle & Anderson 2001, Anderson &
ter Braak 2003). For this, residuals are computed in the data before obtaining the dissimilarity
matrix. For testing the interaction in two-factor, crossed designs with multivariate data. the
residuals are computed as
zhijk = yhijk yhi.. yh. j. + yh...
In this, yhijk is the observation of variable h in unit k, belonging to group i in factor A and to group j
in factor B; yhi .. is the mean for variable h in factor A group i; y h. j. is the mean for variable h in
factor B group j; and y h... is the overall mean for variable h in the data set. For testing factor A, the
residuals were given by zhijk = yhijk yh. j. and for testing factor B they were zhijk = yhijk yhi.. .
General guidelines
In one-factor designs sum of squares and F-ratio will give identical probabilities (for the
same random number generation initializer). The simulation results in Pillar (2004) indicated that
for evaluating interactions in randomization testing in two-factor, crossed designs, the use of the
F-ratio ( F = Qb / Qw ), with residuals, can improve accuracy and power. For testing factor main
effects, however, the exact test, involving restricted permutations, with Qb or F-ratio, is simpler
and gave equivalent accuracy and power compared to using residuals and the F-ratio.
39
The accuracy and power in testing for factor effects is influenced by whether interaction is
present or not. Therefore, in judging the tests for factor effects, it is better to interpret the
probabilities for factors only when the interaction is detected as not significant. In fact, if the
interaction is significant, it is recommended splitting the data according to the levels of one factor
and then testing for effects of the other factor in each data set separately.
MULTIV offers the choice of using default options, which automatically will select for
each case the combination of procedures that had been found by Pillar (2004) the most accurate
and with higher power in data simulation. Users more acquainted with the intricacies of
randomization testing and analysis of variance may decide to follow the advanced options.
If the groups are defined by one factor only, the default option will use the sum of squares
as test statistics. If there are two factors, the first one explicitly defining blocks, the default option
will as well use the sum of squares as test criterion, performing restricted permutations within
blocks. For the probabilities in these cases it is irrelevant whether Qb or F-ratio is used. For
contrasts as well, the test criterion is irrelevant.
If the groups are defined by more than one factor, the default option will use different
procedures and separate randomization tests for interactions and for main effects (the results,
however, will be presented altogether in one table). For double interactions, residuals removing
both factors involved in the interaction and the F-ratio as test criterion will be used for generating
the probabilities (the observed sum of squares and F-ratio presented in the results, though, will be
based on the raw data). For each main factor, restricted permutations (raw data, not residuals)
within the groups defined by the combination of the other factors will be used. For comparing the
groups defined by the joint states of all factors, unrestricted permutations with raw data will be
used. In all cases, except for interactions as explained above, Qb will be the test criterion.
Guidelines for testing factors and interactions under other multi-factor designs are found in
Anderson & ter Braak (2003). In multifactor designs with more than 2 factors, MULTIV will be
restricted to evaluate main effects and double interactions.
Specifying groups of sampling units

Groups must be specified before randomization comparing groups of sampling units. The
user must inform the number and names of factors defining groups. If the design is a block design,
the first factor always is blocks. The following run exemplifies the procedure (data from Orlci et
al. 1987, p.106):
Analysis status:
------------------------------------------------------------------------------MAIN MENU
* T transform data
40
*
*
*
*
*
D scatter diagrams
O ordination
C cluster analysis
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): g
//Data set okop106.txt (from Orlci et al. 1997, p.106) contains one variable describing 12
experimental units (the fact the data set is univariate will not change the procedure):
14 16
18
17
19
22
12
16
17
15
16
18
------------------------------------------------------------------------------SPECIFY GROUPS
File: okop106.txt
Enter the number of factors that are defining groups: 2
Block design? y/n y
Enter the name of each factor:
Factor 1: Blocks
Factor 2: Diet
Read group memberships from file (f) or from the keyboard (k)? f/k k
Enter the group of each sampling unit:
Sampling unit:
1 2 3 4 5 6 7 8 9 10 11 12
Factor Blocks:
Groups:
1 1 1 2 2 2 3 3 3 4 4 4
Factor Diet:
Groups:
1 2 3 1 2 3 1 2 3 1 2 3
//If you choose option f the program will ask the name of the file containing the group
memberships in the same order as if they were entered via keyboard.
//If option G is selected again, this partition will be shown on screen and the user may choose to
specify new factors and partitions or leave as is.
The following run implements a randomization testing of treatment effect in data obtained in a
block design, using default options:
Analysis status:
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
* T transform data
41
* O ordination
* P randomization testing comparing groups of sampling units
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): p
//Irrespective of having already specified the groups, the program will present the current
groups partition for confirmation.
------------------------------------------------------------------------------SPECIFYING GROUPS
File: p106.txt
Current group partition:
Sampling units:
1
2
3
4
5
6
7
8
9
10 11 12
Factor Blocks:
Groups:
1
1
1
2
2
2
3
3
3
4
4
4
Factor Diet:
Groups:
1
2
3
1
2
3
1
2
3
1
2
3
Change? y/n n
Analysis status:
Session IS saved.
(1) automatically
Enter option: 1
Use default (d) or wish to specify advanced (a) options? d/a d
Contrasts of groups:
(0)no contrasts
(1)perform all pairwise group contrasts
(2)specify contrasts
Enter option: 2
Enter the number of contrasts for factor Blocks (max. 3): 0
Enter the number of contrasts for factor Diet (max. 2): 2
Specify the coefficients in each contrast:
Groups:
1 2 3
1 -1 0
Contrast 1:
1 1 -2
Contrast 2:
Save random data and results at each iteration? y/n n
The results presented on file Prinda.txt were:

-------------------------------------------------------------------------------
42
RANDOMIZATION TEST
------------------------------------------------------------------------------Fri Feb 13 19:59:32 2004
Elapsed time: 1 seconds
Number of iterations (random permutations): 1000
Group partition of sampling units:
Sampling units:
1 2 3 4 5 6 7 8 9 10 11 12
Factor Blocks:
Groups:
1 1 1 2 2 2 3 3 3 4 4 4
Order of groups in contrasts: 1 2 3 4
Factor Diet:
Groups:
1 2 3 1 2 3 1 2 3 1 2 3
Order of groups in contrasts: 1 2 3
Permuted data were vectors of raw data.
Randomization restrictions:
Random permutation within nesting groups:
1 1 1 2 2 2 3 3 3 4 4 4
Source of variation
Sum of squares(Q)
P(QbNULL>=Qb)
-----------------------------------------------------------------------------Blocks:
Between groups
31.333
-----------------------------------------------------------------------------Diet:
Between groups
36.167
0.007
Contrasts:
1 -1 0
10.125
0.123
1 1 -2
26.042
0.029
Within groups
3.1667
-----------------------------------------------------------------------------Total
70.667
Mean vectors of each group:
Factor Blocks:
Group 1 (n=3):
16
Group 2 (n=3):
19.333
Group 3 (n=3):
15
Group 4 (n=3):
16.333
Factor Diet:
Group 1 (n=4):
14.5
Group 2 (n=4):
16.75
Group 3 (n=4):
18.75
Interaction
Group 1 x 1
Group 1 x 2
Group 1 x 3
Group 2 x 1
Group 2 x 2
Group 2 x 3
Group 3 x 1
Group 3 x 2
Group 3 x 3
Group 4 x 1
Group 4 x 2
Group 4 x 3
factors Blocks x Diet:

(n=1):
14
(n=1):
16
(n=1):
18
(n=1):
17
(n=1):
19
(n=1):
22
(n=1):
12
(n=1):
16
(n=1):
17
(n=1):
15
(n=1):
16
(n=1):
18
//The probabilities indicate the effect of treatment was significant (P = 0.007) and that
treatments 1 and 2 differed significantly (P = 0.029) from treatment 3.
43
The next run illustrates the analysis of a two-factor design (data from Orlci et al. 1987, p.109),
using default options:
Analysis status:
Session IS saved.
------------------------------------------------------------------------------MAIN MENU
* T transform data
* O ordination
* P randomization testing comparing groups of sampling units
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): p
//In the data matrix in file okop109.txt 1 variable is describing 24 experimental units (the fact
the data set is univariate will not change the procedure):
40 32 32
35
38
44
36
34
31
33
40
42
38
35
33
32
42
44
38
35
36
36
40
42
------------------------------------------------------------------------------SPECIFYING GROUPS
File: okop109.txt
Current group partition:
Sampling units:
1
2
3
4
5
6
7
8
9
10 11 12 13 14 15
16 17 18 19 20 21 22 23 24
Factor Lime:
Groups:
1
1
1
2
2
2
1
1
1
2
2
2
1
1
1
2
2
2
1
1
1
2
2
2
Factor Nitrogen:
Groups:
3
1
2
3
1
Change? y/n n
1
3
2
1
3
2
Analysis status:
Session IS saved.
44
(1) automatically
Enter option: 1
Use default (d) or wish to specify advanced (a) options? d/a d
(0)no contrasts
Enter option: 1
The results were:

------------------------------------------------------------------------------Tue Nov 28 17:37:11 2006
Elapsed time: 0.561544 seconds
Number of random permutations plus observed data set: 1000
Sampling units:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
24
Factor Lime:
Groups:
1 1 1 2 2 2 1 1 1 2 2 2 1 1 1 2 2 2 1 1 1 2 2 2
Order of groups in contrasts: 1 2
Factor Nitrogen:
Groups:
1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3
Vectors of raw data were used to generate probabilities for the main effects
while vectors of residuals were used for double interactions (both factor
effects removed).
For each main effect permutations were restricted within the groups defined by
the combination of the other factors.
For the all-factor combination effect permutations were unrestricted.
(*) Probabilities P generated for sum of squares (Qb), except for interactions,
where F=Qb/Qw was used as test criterion.
Source of variation
Sum of squares(Q)
P(QbNULL>=Qb)*
-----------------------------------------------------------------------------Factor Lime:
Between groups
96
0.017
Contrasts:
1 -1
96
0.02
-----------------------------------------------------------------------------Factor Nitrogen:
Between groups
16
0.51
Contrasts:
1 -1 0
4
0.608
1 0 -1
16
0.362
0 1 -1
4
0.387
-----------------------------------------------------------------------------Lime
x Nitrogen
208
0.001
------------------------------------------------------------------------------
45
Between groups
320
0.001
Within groups
50
-----------------------------------------------------------------------------Total
370
Factor Lime:
Group 1 (n=12):
35
Group 2 (n=12):
39
Factor Nitrogen:
Group 1 (n=8):
36
Group 2 (n=8):
37
Group 3 (n=8):
38
Interaction
Group 1 x 1
Group 1 x 2
Group 1 x 3
Group 2 x 1
Group 2 x 2
Group 2 x 3
factors Lime x Nitrogen:

(n=4):
38
(n=4):
34
(n=4):
33
(n=4):
34
(n=4):
40
(n=4):
43
//Since the interaction between the two factors was significant, for examining main factor
effects the data set should be splitted so that, e.g, the effect of nitrogen is tested within each
level of lime, or vice versa.
In the next run, advanced options were set for the analysis with the same data set okop109.txt and
groups defined by two factors. The analysis used permutation of residuals removing the effects of
both factors:
Analysis status:
Session IS saved.
(1) automatically
Enter option: 1
Use default (d) or wish to specify advanced (a) options? d/a a
Choose the test statistic:
(1) Sum of squares between groups (Qb, Pillar & Orloci 1996)
(2) Pseudo F-ratio (Anderson 2001)
(3) Average dissimilarity within groups (MRPP for v=1, Mielke & Berry 2001)
Enter option: 2
Use residuals in the analysis?
(0) No
(1) Remove the effect of one factor
(2) Remove the effect of two factors
Enter option: 2
Specify the factors whose effects are to be removed ( 1.Lime 2.Nitrogen ):
1 2
46
Nesting (like a block design or a nested design):
(0) None
(1) Permutations restricted within groups of one factor already specified
(2) Permutations restricted within groups to be specified
Enter option: 0
Additional randomization restrictions:
Permutation among unit bundles (e.g., main plots in a nested design or when
subsampling is involved):
(0) None
(1) Bundles according to one factor already specified
(2) Bundles to be specified
Enter option: 0
(0)no contrasts
Enter option: 0
The results were:

------------------------------------------------------------------------------Tue Nov 28 17:47:28 2006
Sampling units:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
24
Factor Lime:
Groups:
1 1 1 2 2 2 1 1 1 2 2 2 1 1 1 2 2 2 1 1 1 2 2 2
Factor Nitrogen:
Groups:
1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3
Permuted data were vectors of residuals (effects of factors Lime and Nitrogen
were removed).
Randomization restrictions: none
Source of variation
Sum of squares(Q)
F=Qb/Qw
P(FNULL>=F)
-----------------------------------------------------------------------------Factor Lime:
Between groups
0
0
1
-----------------------------------------------------------------------------Factor Nitrogen:
Between groups
0
0
1
-----------------------------------------------------------------------------Lime
x Nitrogen
208
4.16
0.0001
-----------------------------------------------------------------------------Between groups
208
4.16
0.0001
Within groups
50
47
-----------------------------------------------------------------------------Total
258
Factor Lime:
Group 1 (n=12):
35
Group 2 (n=12):
39
Factor Nitrogen:
Group 1 (n=8):
36
Group 2 (n=8):
37
Group 3 (n=8):
38
Interaction
Group 1 x 1
Group 1 x 2
Group 1 x 3
Group 2 x 1
Group 2 x 2
Group 2 x 3

(n=4):
38
(n=4):
34
(n=4):
33
(n=4):
34
(n=4):
40
(n=4):
43
Analysis status:
Session IS saved.
//The sum of squares for the factors Lime and Nitrogen are zero since the effects of both factors
were removed in the observed data.
In the next analysis advanced options were set to perform the test with random permutations
restricted within the levels of the factor Lime:
Analysis status:
Session IS saved.
(1) automatically
Enter option: 1
Use default (d) or wish to specify advanced (a) options? d/a a
Use as statistics sum of squares between groups (s) or an F-ratio (f)? s/f s
Use residuals in the analysis?
(0) No
(1) Remove the effect of one factor
(2) Remove the effect of two factors
Enter option: 0
Randomization restrictions (for testing factor effects and interactions):
48
Nesting:
(0) None
(1) Permutations restricted within levels of one of the factors
(2) Permutations restricted within levels to be specified
Enter option: 1
Specify which is the nesting factor ( 1.Lime 2.Nitrogen ):
Blocking (sets of units permuted as blocks):

(0) None
(1) Levels of one factor forming blocks
(2) Blocks to be specified
Enter option: 0
(0)no contrasts
Enter option: 1
The results were:

------------------------------------------------------------------------------Tue Nov 28 17:51:26 2006
Sampling units:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
24
Factor Lime:
Groups:
1 1 1 2 2 2 1 1 1 2 2 2 1 1 1 2 2 2 1 1 1 2 2 2
Order of groups in contrasts: 1 2
Factor Nitrogen:
Groups:
1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3
Permuted data were vectors of raw data.
Random permutation within nesting groups:
1 1 1 2 2 2
1 1 1 2 2 2 1 1 1 2 2 2 1 1 1 2 2 2
Source of variation
Sum of squares(Q)
P(QbNULL>=Qb)
-----------------------------------------------------------------------------Factor Lime:
Between groups
96
1
Contrasts:
1 -1
96
1
-----------------------------------------------------------------------------Factor Nitrogen:
Between groups
16
0.5385
Contrasts:
1 -1 0
4
0.5915
1 0 -1
16
0.3875
0 1 -1
4
0.3734
------------------------------------------------------------------------------
49
Lime
x Nitrogen
208
0.0001
-----------------------------------------------------------------------------Between groups
320
0.0001
Within groups
50
-----------------------------------------------------------------------------Total
370
Factor Lime:
Group 1 (n=12):
35
Group 2 (n=12):
39
Factor Nitrogen:
Group 1 (n=8):
36
Group 2 (n=8):
37
Group 3 (n=8):
38
Interaction
Group 1 x 1
Group 1 x 2
Group 1 x 3
Group 2 x 1
Group 2 x 2
Group 2 x 3

(n=4):
38
(n=4):
34
(n=4):
33
(n=4):
34
(n=4):
40
(n=4):
43
//In this case random permutations were restricted within factor Lime and thus the
corresponding probabilities for factor Lime are all equal to one.
Setting preferences
The user can change the default settings for the language of the interface and results
(English or Portuguese), the sizes of dendrograms and scatter diagrams, and the number of
significant digits in outputs. The default settings are used until anything different is selected in the
Preferences menu, after which a file named MULTIV.prf is placed in the same folder with the
program. The new settings are then used as long as this file is kept with the program. When no
MULTIV.prf file is found, the default settings are used.
------------------------------------------------------------------------------MAIN MENU
* T transform data
* O ordination
* P randomization tests
* E preferences
* S save session
* X exit
------------------------------------------------------------------------------Enter option (valid options with *): e
PREFERENCES
Set preferences of:
L language (change to Portuguese/mudar para portugus)
S scatter diagram
50
D dendrograms
F output
C return to main menu
Type option: s
------------------------------------------------------------------------------Dimensions of the scattergram in mm (zero for default
height 88 x width 88
mm):
height (max. 183): 100
Scattergram height and width are the same.
PREFERENCES
Set preferences of:
S scatter diagram
D dendrograms
F output
Type option: d
------------------------------------------------------------------------------Dimensions of the dendrograms in mm (zero for default
height 73 x width 73
mm):
height (max. 187): 100
width (max. 240): 100
PREFERENCES
Set preferences of:
S scatter diagram
D dendrograms
F output
Type option: f
------------------------------------------------------------------------------Number of significant digits on printouts (enter zero for default):
10
PREFERENCES
Set preferences of:
S scatter diagram
D dendrograms
F output
Type option: c
References
Anand, M. & L. Orlci. 1996. Complexity in plant communities: the notion and quantification.
Journal of Theoretical Biology 179: 179-186.
Anderson, M.J. & C. ter Braak. 2003. Permutation tests for multi-factorial analysis of variance.
Journal of Statistical Computations and Simulations 73:85-113.
Camiz, S., J.-J. Denimal, & V.D. Pillar. 2006. Hierarchical factor classification of variables in
ecology. Community Ecology 7: 165-179.
Denimal, J.J. 2001. Hierarchical Factorial Analysis, Actes du 10th International Symposium on
Applied Stochastic Models and Data Analysis. Compigne, 12-15 Juin 2001.
Edgington, E. S. 1987. Randomization Tests. Marcel Dekker, New York.
Efron, B., and R. Tibshirani. 1993. An Introduction to the Bootstrap. Chapman & Hall, London.
51
Digby, P. G. N. & R. A. Kempton. 1987. Multivariate Analysis of Ecological Communities.

London, Chapman & Hall. 206 p.
Feoli, E., M. Lagonegro & L. Orlci. 1984. Information Analysis of Vegetation Data. Junk, The
Hague.
Gower, J. C. 1971. A general coefficient of similarity and some of its properties. Biometrics 27:
857-874
Grimm, E.C. 1987. CONISS: A FORTRAN 77 program for stratigraphically constrained cluster
analysis by the method of incremental sum of squares. Computers and Geosciences 13: 13-35.
Kendall, M. & J. D. Gibbons. 1990. Rank Correlation Methods 5th ed. Edward Arnold, London.
Legendre, P. & L. Legendre. 1998. Numerical Ecology 2nd English Edition. Elsevier, Amsterdam.
853 p.
McArdle, B.H. & M.J. Anderson. 2001. Fitting multivariate models to community data: a
comment on distance-based redundancy analysis. Ecology 82:290-297.
Manly, B. F. J. 1991. Randomization and Monte Carlo Methods in Biology. London, Chapman &
Hall. 281 p.
Mielke, P. W., and K. J. Berry. 2001. Permutation Methods: A Distance Approach. SpringerVerlag, New York, USA.
Orlci, L. 1967. An agglomerative method for classification of plant communities. Journal of
Ecology 55: 193-205.
Orlci, L. 1978. Multivariate Analysis in Vegetation Research. 2.ed. The Hague, W. Junk. 451 p.
Orlci, L. & V. D. Pillar. 1991. On sample optimality in ecosystem survey. In: Feoli, E. & L.
Orlci (eds.) Computer Assisted Vegetation Analysis. p.41-46. Kluwer, Dordrecht. 498p.
Orlci, L. & V. D. Pillar. 1989. On sample size optimality in ecosystem survey. BiometriePraximetrie 29: 173-184.
Orlci, L. 1991. Entropy and Information. SPB Academic Publishing, The Hague.
Orlci, L., N. C. Kenkel & M. Orlci. 1987. Data Analysis in Population and Community
Ecology. University of Hawaii, Honolulu / New Mexico State University, Las Cruces. 211p.
Pielou, E. C. 1984. The Interpretation of Ecological Data; a Primer on Classification and
Ordination. New York, J. Wiley. 263 p.
Pillar, V. D. & L. Orlci. 1993. Character-Based Community Analysis; the Theory and an
Application Program. SPB Academic Publishing, The Hague.
Pillar, V. D. & L. Orlci. 1996. On randomization testing in vegetation science: multifactor
comparisons of relev groups. Journal of Vegetation Science 7: 585-592.
Pillar, V. D. 1998. Sampling sufficiency in ecological surveys. Abstracta Botanica 22: 37-48.
Pillar, V. D. 1999a. How sharp are classifications? Ecology 80: 2508-2516.
Pillar, V. D. 1999b. The bootstrapped ordination reexamined. Journal of Vegetation Science 10:
895-902.
Pillar, V.D. 2006. How accurate and powerful are randomization tests in multivariate analysis of
variance with ecological data? Ecology (submitted).
Podani, J. 2000. Introduction to the Exploration of Multivariate Biological Data. Leiden, Backuys
Publishers. 407 p.
Ward, J.H. 1963. Hierarchical grouping to optimize and objective function. J. Amer. Stat. Assoc.
58: 236-244.

Multiv: Users Guide v. 2.4

Uploaded by

Copyright:

Available Formats

Multiv: Users Guide v. 2.4

Uploaded by

Document Information

Original Title

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

Multiv: Users Guide v. 2.4

Uploaded by

Copyright:

Available Formats

MULTIV

Multivariate Exploratory Analysis, Randomization Testing and Bootstrap

Users Guide v. 2.4

Copyright 2006 by Valrio DePatta Pillar

means, entropy, complexity

between variables or sampling units

Descriptive attributes and bootstrap resampling

Enter option (valid options with *): n

Quantitative and /or binary data

Chord distance. This distance between sampling units is equivalent to computing

Qualitative and/or binary data

interval 0 except = 1. When approaches 1 ( = 1 specified by the user), the function

varying from zero (no association) to 1 (strong association), defined by ik =

dik is the Rajski's metric (see below).

2Iik = 2f..Hik = 2 [fhj ln (fhj/fohj)]

, being the terms as defined above.

Variable i (sampling unit ) 1

Principal coordinates analysis.

Principal components analysis

uhk where p is the number of variables (sampling units).

The following run performs a principal coordinates analysis:

//A minimum of 1000 iterations in bootstrap resampling is recommended. At each step in

Cluster analysis and dendrograms

Single linkage (nearest neighbour)

Incremental sum of squares

Hierarchical factorial clustering of variables

The following run performs cluster analysis:

//The following window is shown on screen (only for Macintosh versions):

//A minimum of 1000 iterations in bootstrap resampling is recommended. At each step in

Randomization tests: comparing variables

The results in file Prinda.txt are:

Randomization tests: comparing groups of sampling units

is the sum of squares within k groups, such that

is a weighted average dissimilarity within k groups, such that

For u = k-1 independent contrasts,

Random permutations in two-factor designs

Specifying groups of sampling units

The results presented on file Prinda.txt were:

factors Blocks x Diet:

The results were:

factors Lime x Nitrogen:

The results were:

factors Lime x Nitrogen:

Blocking (sets of units permuted as blocks):

The results were:

factors Lime x Nitrogen:

Digby, P. G. N. & R. A. Kempton. 1987. Multivariate Analysis of Ecological Communities.

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.