Statscraft 0.15.0, release v0.15.0
ALSCAL
Multidimensional scaling of symmetric proximity matrices by alternating least squares on squared distances.
Syntax
ALSCAL VARIABLES=varlist | /MATRIX=IN('file' | dataset | *)
[/INPUT=ROWS(ALL | n)] [/SHAPE=SYMMETRIC]
[/LEVEL=ORDINAL[([UNTIE] [SIMILAR])] | INTERVAL[(1..4)] | RATIO[(1..4)] | NOMINAL]
[/CONDITION=MATRIX | UNCONDITIONAL] [/MODEL=EUCLID | INDSCAL]
[/CRITERIA=[NEGATIVE] [CUTOFF(0)] [CONVERGE(.001)] [ITER(30)] [STRESSMIN(.005)] [DIMENS(2[,max])]]
[/PRINT=[DATA] [HEADER]] [/PLOT[=DEFAULT | ALL]] [/OUTFILE='file' | dataset].
Each case is a row of a square matrix of dissimilarities and each variable a column; only the
triangle below the diagonal is read, and more cases than variables hold one matrix per subject.
MATRIX=IN reads a matrix file of PROXIMITIES or CLUSTER instead, VARIABLES picking its items; a
similarity file is scaled as SIMILAR, or turned into dissimilarities from its largest value for
INTERVAL and RATIO. A missing value or one below CUTOFF is left out; the weight variable is not used. The
model fits squared distances to squared disparities and reports S-stress, the root of the mean over
the matrices of their squared misfit relative to the fourth powers of the disparities. ORDINAL
keeps the order of the data (UNTIE lets ties part, SIMILAR turns the order), NOMINAL only which values
are equal; INTERVAL(n) and RATIO(n) make the squared disparities a polynomial of degree n of the
squared data, with and without an intercept. CONDITION=MATRIX fits each matrix on its own,
UNCONDITIONAL all at once. MODEL=INDSCAL weights the dimensions per matrix (no negative weights
unless NEGATIVE). The start is classical scaling of the mean scalar products of all matrices, after
each got the smallest constant that makes it a metric. Iterations stop when S-stress improves by less
than CONVERGE, falls to STRESSMIN or after ITER; DIMENS(min,max) solves from max down to min (1 only
for EUCLID), each smaller solution from the first dimensions of the one before.
Tables per dimensionality: Iteration History (S-stress and its improvement), Stress and Squared
Correlation (RSQ) in Distances per matrix (Kruskal's stress formula 1 and the squared correlation of
disparities and distances, with their average), Stimulus Coordinates and for INDSCAL Subject Weights;
PRINT=DATA adds the data and the optimally scaled data, HEADER the options. The coordinates are
centered; EUCLID turns them to principal axes with a mean square of 1, INDSCAL scales each dimension
to a sum of squares equal to the number of stimuli (the weights the other way) and orders the
dimensions by their weights; every dimension is signed so that the first stimulus is positive. PLOT
gives the configuration, the subject weights, the linear fit (disparities against distances), the
nonlinear fit (data against distances) for ORDINAL and NOMINAL and the transformation (data against
disparities), with ALL one per matrix. OUTFILE writes TYPE_ (CON for coordinates, SUB for subject
weights), DIMENS, MATNUM (0 for coordinates) and DIM1, DIM2, ... for every dimensionality. Not
supported yet: SHAPE=ASYMMETRIC and RECTANGULAR, CONDITION=ROW, the models ASCAL, AINDS and GEMSCAL,
FILE.
Example
ALSCAL VARIABLES=s1 TO s9 /LEVEL=ORDINAL /CRITERIA=DIMENS(2,3) /PLOT.
statscraft help ALSCAL prints the same text in the terminal. All commands