Statscraft 0.15.0, release v0.15.0
HILOGLINEAR
Hierarchical loglinear models of a multiway table of counts, with backward elimination.
Syntax
HILOGLINEAR varlist (min,max) [varlist (min,max) ...]
[/METHOD[=BACKWARD]] [/MAXORDER=k]
[/CRITERIA=[CONVERGE(n)] [ITERATE(20)] [P(0.05)] [DELTA(0.5)] [MAXSTEPS(10)] [DEFAULT]]
[/CWEIGHT=variable | (w w n*w ...)]
[/PRINT=[FREQ] [RESID] [ESTIM] [ASSOCIATION] | DEFAULT | ALL | NONE]
[/PLOT=[RESID] [NORMPROB] | DEFAULT | ALL | NONE] [/MISSING=EXCLUDE | INCLUDE]
[/DESIGN=term [term ...]] [/DESIGN ...].
terms: a variable or variables joined by * (or BY). At least two and at most ten numeric
variables, each group followed by its whole-number range; every integer of the range is a
category, values are cut to integers, and a case with a value outside a range is counted as
out of range and left out. A case missing on a variable is left out unless INCLUDE makes
user-missing values valid; case weights are used as they are.
A DESIGN lists the generating class: the highest terms, each of which brings in all the terms
made of part of its variables. Every DESIGN fits its own model with the CRITERIA, PRINT,
PLOT, MAXORDER and CWEIGHT given earlier in the command (they carry over); METHOD applies to
the next DESIGN only. A DESIGN with no terms, or the one added (with a warning) when the
command does not end with DESIGN, is the saturated model; MAXORDER k cuts every term down to
its k-variable parts. The model is fitted by iterative proportional fitting until no cell
moves by CONVERGE (by default the larger of 0.25 and a thousandth of the largest weighted
count) or ITERATE cycles have run. CWEIGHT gives a weight per cell, from a variable (its mean
in the cell, 1 for an empty cell) or as a list in cell order with the last variable changing
fastest; a weight of 0 makes a structural zero and weights are cut to the range 0 to 1.
Tables: Data Information, Convergence Information, Cell Counts and Residuals (FREQ: observed
and expected counts; RESID: raw and standardized residuals) and the likelihood-ratio and
Pearson Goodness-of-Fit Tests, whose df leaves out the cells fitted as 0. A saturated model
is fitted to the counts plus DELTA and adds the tests that all effects of order k, or of
order k and higher, are zero; ESTIM gives its effect-coded parameters with standard errors
(when no cell is 0) and ASSOCIATION the partial chi-square of each lower-order term. By
default an unsaturated model shows FREQ and RESID, a saturated one all four. METHOD=BACKWARD
removes, one per step and at most MAXSTEPS times, the term of the generating class whose
removal has the largest likelihood-ratio p, while that p is above P, and lists the steps.
PLOT (unsaturated models only) draws the standardized residuals against the observed and
expected counts (RESID) and a normal Q-Q plot of the adjusted residuals (NORMPROB, also
NORMPLOT); PLOT alone means both.
Example
HILOGLINEAR route(1,3) season(1,4) delay(0,1) /METHOD=BACKWARD /CRITERIA=P(0.01) /DESIGN.
statscraft help HILOGLINEAR prints the same text in the terminal. All commands