Statscraft 0.15.0, release v0.15.0
PROBIT
Probit and logit models of response counts out of subject counts, with effective levels.
Syntax
PROBIT responses OF subjects WITH covariates [BY factor(min,max)]
[/MODEL=PROBIT | LOGIT | BOTH] [/LOG=[base | NONE]]
[/CRITERIA=[OPTOL(n) | CONVERGE(n)] [ITERATE(n)] [P(0.15)] [STEPLIMIT(0.1)]]
[/NATRES[=rate]] [/PRINT=[FREQ] [CI] [RMP] [PARALL] [DEFAULT] [ALL] [NONE]]
[/MISSING=EXCLUDE | INCLUDE].
Each case is a group of subjects: how many there were and how many responded. The
chance to respond is F(a + b'x), where F is the standard normal distribution (PROBIT, the
default) or the logistic one (LOGIT); BOTH fits the logit model first, then the probit
one. With a factor (BY, whole-number range; cases outside it are left out) each level has
its own intercept and all share the slopes. Covariates are replaced by their base 10
logarithm unless LOG says otherwise: a bare LOG means base e, NONE no transformation. A
case with a covariate of 0 is a control group; under a logarithm it stays out of the fit.
NATRES adds a natural response rate c, so the chance becomes c + (1 - c) F(...): alone it
is estimated (starting from the control groups' rate, or the lowest observed rate), with
a value it is fixed. Estimates are maximum likelihood by Newton steps from zero slopes,
with standard errors from the observed information. More responses than subjects is an
error; a case weight multiplies both counts.
Tables: Data Information, Convergence Information, Parameter Estimates (with z tests and
95% intervals), Natural Response Rate, Covariances and Correlations of Parameter
Estimates and the Pearson goodness-of-fit test. PRINT left out or bare shows FREQ (observed
and expected responses per case), CI (the covariate level at which each share from 0.01
to 0.99 respond, with 95% fiducial limits, on both scales under a logarithm) and RMP (the
relative median potency of every pair of factor levels); PARALL tests whether the levels
share the slope. CI, RMP and PARALL need a single covariate. When the goodness-of-fit
test's significance is below P the limits widen by the heterogeneity factor chi-square/df
and use t instead of the normal; P(0) never does that, P(1) always. With one covariate a
scatter plot shows the transformed observed rates. MISSING=INCLUDE keeps user-missing
values.
Example
PROBIT hatched OF eggs BY strain(1,3) WITH dose /MODEL=BOTH /PRINT=ALL.
statscraft help PROBIT prints the same text in the terminal. All commands