Confidence Intervals for Parameters of VGLMs
Computes confidence intervals (CIs)
for one or more parameters in a fitted model.
Currently the object must be a
"vglm" object.
confintvglm(object, parm, level = 0.95, method = c("wald", "profile"),
trace = NULL, ...)object |
A fitted model object. |
parm, level, ... |
Same as |
method |
Character.
The default is the first method.
Abbreviations are allowed.
Currently |
trace |
Logical. If |
The default for
this methods function is based on confint.default
and assumes
asymptotic normality. In particular,
the coef and
vcov methods functions are used for
vglm-class objects.
When method = "profile" the function
profilevglm
is called to do the profiling. The code is very heavily
based on profile.glm
which was originally written by
D. M. Bates and W. N. Venables (For S in 1996)
and subsequently corrected by B. D. Ripley.
Sometimes the profiling method can give problems, for
example, cumulative requires the M
linear predictors not to intersect in the data cloud.
Such numerical problems are less common when
method = "wald", however, it is well-known
that inference based on profile likelihoods is generally
more accurate than Wald, especially when the sample size
is small.
The deviance (deviance(object)) is used if possible,
else the difference
2 * (logLik(object) - ell) is computed,
where ell are the values of the loglikelihood on a grid.
For
Wald CIs and
rrvglm-class
objects, currently an error message is produced because
I haven't gotten around to write the methods function;
it's not too hard, but am too busy!
An interim measure is to
coerce the object into a "vglm" object,
but then the confidence intervals will tend to be too narrow because
the estimated constraint matrices are treated as known.
For
Wald CIs and
vgam-class
objects, currently an error message is produced because
the theory is undeveloped.
Same as confint.
The order of the values of argument method may change
in the future without notice.
The functions
plot.profile.glm
and
pairs.profile.glm
from MASS
appear to work with output from this function.
Thomas Yee adapted confint.lm
to handle "vglm" objects, for Wald-type
confidence intervals.
Also, profile.glm
was originally written by
D. M. Bates and W. N. Venables (For S in 1996)
and subsequently corrected by B. D. Ripley.
This function effectively calls confint.profile.glm()
in MASS.
vcovvlm,
summaryvglm,
confint,
profile.glm,
lrt.stat.vlm,
wald.stat,
plot.profile.glm,
pairs.profile.glm.
# Example 1: this is based on a glm example
counts <- c(18,17,15,20,10,20,25,13,12)
outcome <- gl(3, 1, 9); treatment <- gl(3, 3)
glm.D93 <- glm(counts ~ outcome + treatment, family = poisson())
vglm.D93 <- vglm(counts ~ outcome + treatment, family = poissonff)
confint(glm.D93) # needs MASS to be present on the system
confint.default(glm.D93) # based on asymptotic normality
confint(vglm.D93)
confint(vglm.D93) - confint(glm.D93) # Should be all 0s
confint(vglm.D93) - confint.default(glm.D93) # based on asympt. normality
# Example 2: simulated negative binomial data with multiple responses
ndata <- data.frame(x2 = runif(nn <- 100))
ndata <- transform(ndata, y1 = rnbinom(nn, mu = exp(3+x2), size = exp(1)),
y2 = rnbinom(nn, mu = exp(2-x2), size = exp(0)))
fit1 <- vglm(cbind(y1, y2) ~ x2, negbinomial, data = ndata, trace = TRUE)
coef(fit1)
coef(fit1, matrix = TRUE)
confint(fit1)
confint(fit1, "x2:1") # This might be improved to "x2" some day...
## Not run:
confint(fit1, method = "profile") # Computationally expensive
confint(fit1, "x2:1", method = "profile", trace = FALSE)
## End(Not run)
fit2 <- rrvglm(y1 ~ x2, negbinomial(zero = NULL), data = ndata)
confint(as(fit2, "vglm")) # Too narrow (SEs are biased downwards)Please choose more modern alternatives, such as Google Chrome or Mozilla Firefox.