Statscraft 0.7.0, release v0.7.0

TREE

A classification tree grown by CHAID or exhaustive CHAID.

Syntax

TREE dependent [level] BY predictor [level] ... [FORCE = var [level]]
     [/TREE DISPLAY=TOPDOWN | LEFTTORIGHT | RIGHTTOLEFT | NONE  NODES=STATISTICS | CHART | BOTH
            BRANCHSTATISTICS=YES | NO  NODEDEFS=YES | NO  SCALE=AUTO | n]
     [/DEPCATEGORIES USEVALUES=[VALID MISSING value ...] TARGET=[value ...]]
     [/PRINT=MODELSUMMARY CLASSIFICATION RISK CPS TREETABLE | NONE]
     [/GAIN CATEGORYTABLE=YES | NO  TYPE=[NODE]  SORT=DESCENDING | ASCENDING  CUMULATIVE=YES | NO]
     [/SAVE NODEID[(name)] PREDVAL[(name)] PREDPROB[(root)]]
     [/METHOD TYPE=CHAID | EXHAUSTIVECHAID]  [/VALIDATION TYPE=NONE]
     [/GROWTHLIMIT MAXDEPTH=AUTO | n  MINPARENTSIZE=100  MINCHILDSIZE=50]
     [/CHAID ALPHASPLIT=.05 ALPHAMERGE=.05 SPLITMERGED=NO CHISQUARE=PEARSON | LR CONVERGE=.001
             MAXITERATIONS=100 ADJUST=BONFERRONI | NONE INTERVALS=10 | varlist(n) ...]
     [/COSTS EQUAL] [/SCORES EQUALINCREMENTS | CUSTOM=value [score] ...]
     [/MISSING NOMINALMISSING=MISSING | VALID].
Levels: [S] scale, [O] ordinal, [N] or [C] nominal; without one the dictionary's level holds,
and an unknown level comes from the data as in a data pass. The dependent is nominal
(Pearson or likelihood-ratio chi-square) or ordinal (likelihood ratio against Goodman's row
effects model, scored 1, 2, ... or by SCORES). Scale predictors are cut into INTERVALS groups of
about equal size. Predictor categories that do not differ at ALPHAMERGE are merged (ordinal ones
only when adjacent); a missing value is a category of its own that may merge with any. A category
smaller than MINCHILDSIZE is merged with its most similar neighbour. The predictor with the
smallest Bonferroni-adjusted significance splits the node when that is at most ALPHASPLIT. A node
is not split when it is pure, smaller than MINPARENTSIZE or at MAXDEPTH (AUTO is 3). Exhaustive
CHAID merges down to two categories and keeps the most significant set. The tree is shown as a
table with one row per node, indented by depth; charts are not drawn yet. TREETABLE lists the nodes
in order with their parents. Gain tables (one per TARGET category) show terminal nodes by index.
SAVE adds the terminal node, the predicted category and one probability per category for the
cases in the analysis (default names NodeID, PredictedValue, PredictedProbability_1, ...).
A weight counts each case as often as its rounded value. SPLIT FILE is ignored.
PLOT, PRIORS, CRT, QUEST, the settings of CRT and QUEST and PRINT IMPORTANCE, SURROGATES and
CATEGORYSPECS are accepted and ignored with a warning. Not supported
yet: METHOD TYPE=CRT and QUEST, a scale dependent, VALIDATION SPLITSAMPLE and CROSSVALIDATION,
SPLITMERGED=YES, COSTS CUSTOM, GAIN TYPE PTILE, RULES, PROFITS, INFLUENCE, OUTFILE, TARGETRESPONSE.

Example

TREE churn [n] BY region [n] age_group [o] income [s] /GROWTHLIMIT MAXDEPTH=2 /SAVE NODEID PREDVAL.

statscraft help TREE prints the same text in the terminal. All commands