Statscraft 0.15.0, release v0.15.0
PLUM
Ordinal regression: a cumulative link model of an ordered dependent variable.
Syntax
PLUM dependent [BY factors] [WITH covariates]
[/LINK=LOGIT | PROBIT | CLOGLOG | NLOGLOG | CAUCHIT] [/LOCATION=effects] [/SCALE=effects]
[/CRITERIA=[CIN(95)] [DELTA(0)] [BIAS(0)] [MXITER(100)] [MXSTEP(5)] [LCONVERGE(0)]
[PCONVERGE(1E-6)] [SINGULAR(1E-8)]] [/MISSING=EXCLUDE | INCLUDE]
[/PRINT=[FIT] [SUMMARY] [PARAMETER] [TPARALLEL] [CELLINFO] [COVB] [CORB] [HISTORY[(n)]] [KERNEL]]
[/SAVE=[ESTPROB[(root[:n])]] [PREDCAT[(name)]] [PCPROB[(name)]] [ACPROB[(name)]]]
[/TEST[(constants)]=['label'] effect values ...; ... | ALL values ...] ...
The values of the dependent, sorted ascending (numbers or strings), are the ordered
categories. Factors may be strings; covariates are numeric. A case missing on any variable is left out (INCLUDE
keeps user-missing values); the weight is rounded to a whole number and a case below 1 after
rounding is left out. The model puts the link of the probability of each category or a lower one
equal to its threshold minus the location effects, divided by exp of the scale effects. LOCATION
defaults to the covariates and then the main effect of each factor; effects are written as in
UNIANOVA (a*b, a BY b, a(b), a WITHIN b, x*x), and INTERCEPT alone leaves only the thresholds. A
factor has one parameter per level that occurs; the last is redundant and 0. The cases are pooled by
their factor levels and covariate values: DELTA goes into each empty cell of such a group, BIAS into
every cell. The fit is Newton-Raphson from thresholds at the cumulative shares of the categories and
the other parameters 0 (the location part first when SCALE is given), each step halved up to MXSTEP
times while the log likelihood would fall; it stops when no parameter moves PCONVERGE or more, the
log likelihood changes less than LCONVERGE (absolute or relative), or after MXITER iterations.
Tables: Case Processing Summary and Model Fitting Information (-2 log likelihood of the thresholds
alone and of the final model, their difference as a chi-square) always; Goodness-of-Fit (FIT:
Pearson and deviance over the pooled cells), Pseudo R-Square (SUMMARY: Cox and Snell, Nagelkerke,
McFadden) and Parameter Estimates (PARAMETER: estimate, standard error, Wald, df, significance and
CIN interval) without PRINT or when named; Test of Parallel Lines (TPARALLEL: a likelihood ratio
test against separate location parameters for each threshold), Cell Information (CELLINFO),
Parameter Covariance (COVB), Parameter Correlation (CORB) and Iteration History (HISTORY, every
n-th, the first and the last) when named. The -2 log likelihood includes the multinomial constant of
the pooled cells; KERNEL leaves it out. A warning counts the empty pooled cells.
SAVE adds EST1_m, EST2_m, ... (the first n categories, 25 by default), PRE_m, PCP_m and ACP_m with
the smallest free m, or root1, root2, ... and the names given: the probability of each category, the
most likely category, its probability and the probability of the observed category, for every case
with valid factors and covariates. TEST gives Wald tests of combinations of the parameters, in the
order of the Parameter Estimates (thresholds, location, scale, redundant ones included): INTERCEPT
takes one value per threshold, ALL one per parameter; one TEST uses either ALL in every row or
effect names, never both. Each row is tested against its constant (0 by default) on 1 df and the
rows together on the rank of their covariance; a coefficient on a redundant parameter multiplies
its 0.
Example
PLUM comfort BY seat_row WITH trip_hours /LINK=PROBIT /PRINT=FIT PARAMETER SUMMARY TPARALLEL /SAVE=ESTPROB PREDCAT.
statscraft help PLUM prints the same text in the terminal. All commands