Keyboard shortcuts

Press ← or → to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

Statscraft 0.15.0, release v0.15.0

NOMREG

Multinomial logistic regression: a nominal dependent variable on factors and covariates.

Syntax

NOMREG dependent [(BASE=FIRST | LAST | value  ORDER=ASCENDING | DATA | DESCENDING)]
   [BY factors] [WITH covariates]
   [/MODEL=[effects] [| BACKWARD | FORWARD | BSTEP | FSTEP = effects] | /FULLFACTORIAL]
   [/STEPWISE=[RULE(SINGLE | SFACTOR | CONTAINMENT | NONE)] [MINEFFECT(0)] [MAXEFFECT(n)]
              [PIN(0.05)] [POUT(0.1)] [ENTRYMETHOD(LR | SCORE)] [REMOVALMETHOD(LR | WALD)]]
   [/INTERCEPT=INCLUDE | EXCLUDE]
   [/CRITERIA=[CIN(95)] [DELTA(0)] [MXITER(100)] [MXSTEP(5)] [LCONVERGE(0)] [PCONVERGE(1E-6)]
              [SINGULAR(1E-8)] [BIAS(0)] [CHKSEP(20)]]
   [/PRINT=[CPS] [MFI] [PARAMETER] [LRT] [FIT] [SUMMARY] [CLASSTABLE] [IC] [KERNEL] [COVB] [CORB]
          [STEP] [NONE]]
   [/SAVE=[ESTPROB[(root[:n])]] [PREDCAT[(name)]] [PCPROB[(name)]] [ACPROB[(name)]]]
   [/SCALE=1 | n | DEVIANCE | PEARSON] [/SUBPOP=varlist] [/MISSING=EXCLUDE | INCLUDE].
effects: a factor or covariate, a*b or a BY b, a(b) or a WITHIN b, x*x. The dependent and the
factors may be numeric or string; covariates are numeric. The categories of the dependent are
sorted by ORDER (DATA keeps the order in which they first occur); one of them, the last unless
BASE says otherwise, is the reference and has no parameters. Without MODEL the model is the
intercept, the covariates and the main effects of the factors; FULLFACTORIAL adds every
interaction of the factors. A factor gets one parameter per level and an interaction one per
level combination that occurs; a parameter that repeats others (SINGULAR) is redundant and
set to 0, so the last level of a factor is the reference. A case missing on any variable is
left out, as is one with a weight of 0 or less; weights are used as they are. The cases are
grouped into subpopulations by the factors and covariates, or by the SUBPOP variables, which
must include every variable of the model. The fit is Newton-Raphson with step-halving (MXSTEP),
starting from the intercept-only fit; it stops when the log likelihood changes by less than
LCONVERGE, no parameter changes by PCONVERGE or more, or after MXITER iterations. DELTA is added
to empty cells of a subpopulation and BIAS to every cell. Tables: Case Processing Summary (CPS),
Model Fitting Information against the intercept-only model, or the model of equal
probabilities without intercept (MFI, with AIC and BIC under IC), Pseudo R-Square with Cox and
Snell, Nagelkerke and McFadden (SUMMARY), Goodness-of-Fit with Pearson and deviance
chi-square (FIT), Likelihood Ratio Tests of each effect against the model without it (LRT),
Parameter Estimates with Wald tests and the CIN interval of Exp(B) (PARAMETER), the
Classification table (CLASSTABLE) and the covariance or correlation matrix of the estimates
(COVB, CORB). Without PRINT: CPS, MFI and STEP. -2 log likelihoods include the multinomial constant
of the subpopulations; KERNEL leaves it out. SCALE multiplies the covariance matrix by the
value, or by the Pearson or deviance chi-square over its df. SAVE adds the probability of each
response category (EST1_1, EST2_1, ... or root1, root2, ...; the first n, 25 by default), the
predicted category (PRE_1), its probability (PCP_1) and the probability of the observed
category (ACP_1), for every case with valid factors and covariates.
Stepwise: the intercept and the effects before | are always in; the method picks among the
effects after it. FORWARD adds, BACKWARD starts with all of them and drops, FSTEP adds one and
then drops while it can, BSTEP starts with all, drops while it can and then adds one. An effect
is added when its p (likelihood ratio, or the score test with ENTRYMETHOD(SCORE)) is the smallest
and below PIN, and dropped when its p (likelihood ratio, or Wald with REMOVALMETHOD(WALD)) is the
largest and above POUT; on a tie the effect listed first goes (entry p values too small to
hold are compared by their logarithms). The steps stop when nothing qualifies or a model would
repeat. RULE(SINGLE) holds an effect back until the listed effects made of a part of its
variables are in, SFACTOR does that only for effects of factors alone, CONTAINMENT waits for the
effects it contains (same covariates, fewer factors), NONE never waits; an effect that another
one in the model needs is not dropped. MAXEFFECT limits adding under FORWARD and FSTEP,
MINEFFECT dropping under BACKWARD and BSTEP, counting effects without the intercept. The Step
Summary (STEP, shown by default) lists each step with the -2 log likelihood after it and the
test that decided it; the other tables describe the final model, which without the intercept
can be the null model of equal probabilities. Not supported yet: TEST, OUTFILE and PRINT
CELLPROB, HISTORY and ASSOCIATION.

Example

NOMREG commute (BASE=FIRST) BY district WITH distance /PRINT=PARAMETER LRT FIT SUMMARY CLASSTABLE /SAVE=ESTPROB PREDCAT.

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