Fits a generally altered, inflated, truncated and deflated Poisson regression by MLE. The GAITD combo model having 7 types of special values is implemented. This allows mixtures of Poissons on nested and/or partitioned support as well as a multinomial logit model for (nonparametric) altered, inflated and deflated values. Truncation may include the upper tail.
gaitdpoisson(a.mix = NULL, i.mix = NULL, d.mix = NULL,
a.mlm = NULL, i.mlm = NULL, d.mlm = NULL,
truncate = NULL, max.support = Inf,
zero = c("pobs", "pstr", "pdip"),
eq.ap = TRUE, eq.ip = TRUE, eq.dp = TRUE,
parallel.a = FALSE, parallel.i = FALSE, parallel.d = FALSE,
llambda.p = "loglink", llambda.a = llambda.p,
llambda.i = llambda.p, llambda.d = llambda.p,
type.fitted = c("mean", "lambdas", "pobs.mlm", "pstr.mlm",
"pdip.mlm", "pobs.mix", "pstr.mix", "pdip.mix",
"Pobs.mix", "Pstr.mix", "Pdip.mix", "nonspecial",
"Numer", "Denom.p", "sum.mlm.i", "sum.mix.i",
"sum.mlm.d", "sum.mix.d", "ptrunc.p",
"cdf.max.s"), gpstr.mix = ppoints(7) / 3,
gpstr.mlm = ppoints(7) / (3 + length(i.mlm)),
imethod = 1, mux.init = c(0.75, 0.5, 0.75),
ilambda.p = NULL, ilambda.a = ilambda.p,
ilambda.i = ilambda.p, ilambda.d = ilambda.p,
ipobs.mix = NULL, ipstr.mix = NULL, ipdip.mix = NULL,
ipobs.mlm = NULL, ipstr.mlm = NULL, ipdip.mlm = NULL,
byrow.aid = FALSE, ishrinkage = 0.95, probs.y = 0.35)
An object of class "vglmff"
(see vglmff-class
).
The object is used by modelling functions
such as vglm
,
rrvglm
and vgam
.
The fitted.values
slot of the fitted object,
which should be extracted by the
generic function fitted
,
returns the mean \(\mu\) by default.
See the information above on type.fitted
.
Vector of truncated values, i.e., nonnegative integers.
For the first seven arguments (for the special values)
a NULL
stands for an empty set, and
the seven sets must be mutually disjoint.
Argument max.support
enables RHS-truncation,
i.e., something equivalent to
truncate = (U+1):Inf
for some
upper support point U
specified by max.support
.
Vector of altered and inflated values corresponding to finite mixture models. These are described as parametric or structured.
The parameter lambda.p
is always estimated.
If length(a.mix)
is 1 or more then the parameter
pobs.mix
is estimated.
If length(i.mix)
is 1 or more then the parameter
pstr.mix
is estimated.
If length(d.mix)
is 1 or more then the parameter
pdip.mix
is estimated.
If length(a.mix)
is 2 or more then the parameter
lambda.a
is estimated.
If length(i.mix)
is 2 or more then the parameter
lambda.i
is estimated.
If length(d.mix)
is 2 or more then the parameter
lambda.d
is estimated.
If length(a.mix) == 1
, length(i.mix) == 1
or
length(d.mix) == 1
then lambda.a
,
lambda.i
and lambda.d
are unidentifiable and
therefore ignored. In such cases
it would be equivalent to moving a.mix
into
a.mlm
, etc.
Due to its great flexibility, it is easy to misuse this function
and ideally the values of the above arguments should be well
justified by the application on hand.
Adding inappropriate or
unnecessary values to these arguments willy-nilly
is a recipe for disaster, especially for
i.mix
and d.mix
.
Using a.mlm
effectively removes a subset of the data
from the main analysis, therefore may result in a substantial
loss of efficiency.
For seeped values, a.mix
, a.mlm
,
d.mix
and d.mlm
can be used only.
Heaped values can be handled by i.mlm
and i.mix
,
as well as a.mix
and a.mlm
.
Because of the NBP reason below, it sometimes may be necessary
to specify deflated values to altered values.
Vector of altered, inflated and deflated values corresponding
to the multinomial logit model (MLM) probabilities of
observing those values---see multinomial
.
These are described as nonparametric or unstructured.
Link functions for the parent,
altered, inflated and deflated distributions respectively.
See Links
for more choices and information.
Single logical each.
Constrain the rate parameters to be equal?
See CommonVGAMffArguments
for information.
Having all three arguments TRUE
gives
greater stability in
the estimation because of fewer parameters and therefore
fewer initial values needed,
however if so then one should try relax some of
the arguments later.
For the GIT--Pois submodel,
after plotting the responses,
if the distribution of the spikes
above the nominal probabilities
has roughly the same shape
as the ordinary values then setting
eq.ip = TRUE
would be a good idea
so that lambda.i == lambda.p
.
And if i.mix
is of length 2 or a bit more, then
TRUE
should definitely be entertained.
Likewise, for heaped or seeped data, setting
eq.ap = TRUE
(so that lambda.p == lambda.p
)
would be a good idea for the
GAT--Pois if the shape of the altered probabilities
is roughly the same as the parent distribution.
Single logical each.
Constrain the MLM probabilities to be equal?
If so then this applies to all
length(a.mlm)
pobs.mlm
probabilities
or all
length(i.mlm)
pstr.mlm
probabilities
or all
length(d.mlm)
pdip.mlm
probabilities.
See CommonVGAMffArguments
for information.
The default means that the probabilities are generally
unconstrained and unstructured and will follow the shape
of the data.
See constraints
.
See CommonVGAMffArguments
and below for information.
The first value is the default, and this is usually the
unconditional mean.
Choosing an irrelevant value may result in
an NA
being returned and a warning, e.g.,
"pstr.mlm"
for a nonparametric GAT model.
The choice "lambdas"
returns a matrix with at least
one column and up to three others,
corresponding to all those estimated.
In order, their colnames
are
"lambda.p"
, "lambda.a"
, "lambda.i"
and "lambda.d"
.
For other distributions such as gaitdlog
type.fitted = "shapes"
is permitted and the
colnames
are
"shape.p"
, "shape.a"
, "shape.i"
and
"shape.d"
, etc.
Option "Pobs.mix"
provides more detail about
"pobs.mix"
by returning a matrix whose columns
correspond to each altered value; the row sums
(rowSums
)
of this matrix is "pobs.mix"
.
Likewise "Pstr.mix"
about "pstr.mix"
and "Pdip.mix"
about "pdip.mix"
.
The choice "cdf.max.s"
is the CDF evaluated
at max.support
using the parent distribution,
e.g., ppois(max.support, lambda.p)
for
gaitdpoisson
.
The value should be 1 if max.support = Inf
(the default).
The choice "nonspecial"
is the probability of a
nonspecial value.
The choices "Denom.p"
and "Numer"
are quantities
found in the GAITD combo PMF and are for convenience only.
The choice type.fitted = "pobs.mlm"
returns
a matrix whose columns are
the altered probabilities (Greek symbol \(\omega_s\)).
The choice "pstr.mlm"
returns
a matrix whose columns are
the inflated probabilities (Greek symbol \(\phi_s\)).
The choice "pdip.mlm"
returns
a matrix whose columns are
the deflated probabilities (Greek symbol \(\psi_s\)).
The choice "ptrunc.p"
returns the probability of having
a truncated value with respect to the parent distribution.
It includes any truncated values in the upper tail
beyond max.support
.
The probability of a value less than or equal to
max.support
with respect to the parent distribution
is "cdf.max.s"
.
The choice "sum.mlm.i"
adds two terms.
This gives the probability of an inflated value,
and the formula can be loosely written down
as something like
"pstr.mlm" + "Numer" * dpois(i.mlm, lambda.p) / "Denom.p"
.
The other three "sum.m*"
arguments are similar.
See CommonVGAMffArguments
for information.
Gridsearch values for the two parameters.
If failure occurs try a finer grid, especially closer to 0,
and/or experiment with mux.init
.
See CommonVGAMffArguments
for information.
Good initial values are difficult to compute because of
the great flexibility of GAITD regression, therefore
it is often necessary to use these arguments.
A careful examination of a spikeplot
of the data should lead to good choices.
See CommonVGAMffArguments
for information.
Numeric, of length 3. General downward multiplier for initial values for the sample proportions (MLEs actually). This is under development and more details are forthcoming. In general, 1 means unchanged and values should lie in (0, 1], and values about 0.5 are recommended. The elements apply in order to altered, inflated and deflated (no distinction between mix and MLM).
Initial values for the rate parameters;
see CommonVGAMffArguments
for information.
See CommonVGAMffArguments
for information.
Details are at Gaitdpois
.
See CommonVGAMffArguments
for information.
By default, all the MLM probabilities are
modelled as simple as possible (intercept-only) to
help avoid numerical problems, especially when there
are many covariates.
The Poisson means are modelled by the covariates, and
the default zero
vector is pruned of any irrelevant values.
To model all the MLM probabilities with covariates
set zero = NULL
, however, the number of regression
coefficients could be excessive.
For the MLM probabilities,
to model pobs.mix
only with covariates
set zero = c('pstr', 'pobs.mlm', 'pdip')
.
Likewise,
to model pstr.mix
only with covariates
set zero = c('pobs', 'pstr.mlm', 'pdip')
.
It is noted that, amongst other things,
zipoisson
and zipoissonff
differ
with respect to zero
, and ditto for
zapoisson
and zapoissonff
.
Amateurs tend to be overzealous fitting
zero-inflated models when the
fitted mean is low---the warning of
ziP
should be heeded.
For GAITD regression the warning applies more
strongly and generally; here to all
i.mix
, i.mlm
, d.mix
and
d.mlm
values, not just 0. Even one
misspecified special value usually will cause
convergence problems.
Default values for this and similar family
functions may change in the future, e.g.,
eq.ap
and eq.ip
. Important
internal changes might occur too, such as the
ordering of the linear/additive predictors and
the quantities returned as the fitted values.
Using i.mlm
requires more caution
than a.mlm
because gross inflation
is ideally needed for it to work safely.
Ditto for i.mix
versus a.mix
.
Data exhibiting deflation or little to no
inflation will produce numerical problems,
hence set trace = TRUE
to monitor
convergence. More than c.10 IRLS iterations
should raise suspicion.
Ranking the four operators by difficulty, the easiest is truncation followed by alteration, then inflation and the most difficult is deflation. The latter needs good initial values and the current default will probably not work on some data sets. Studying the spikeplot is time very well spent. In general it is very easy to specify an overfitting model so it is a good idea to split the data into training and test sets.
This function is quite memory-hungry with
respect to length(c(a.mix, i.mix, d.mix,
a.mlm, i.mlm, d.mlm))
. On consuming something
different, because all values of the NBP vector
need to be positive it pays to be economical
with respect to d.mlm
especially so
that one does not consume up probabilities
unnecessarily so to speak.
It is often a good idea to set eq.ip =
TRUE
, especially when length(i.mix)
is not much more than 2 or the values of
i.mix
are not spread over the range
of the response. This way the estimation
can borrow strength from both the inflated
and non-inflated values. If the i.mix
values form a single small cluster then this
can easily create estimation difficulties---the
idea is somewhat similar to multicollinearity.
The same holds for d.mix
.
T. W. Yee
The full
GAITD--Pois combo model
may be fitted with this family function.
There are seven types of special values and all arguments for these
may be used in a single model.
Here, the MLM represents the nonparametric while the Pois
refers to the Poisson mixtures.
The defaults for this function correspond to an ordinary Poisson
regression so that poissonff
is called instead.
A MLM with only one probability to model is equivalent to
logistic regression
(binomialff
and logitlink
).
The order of the linear/additive predictors is best
explained by an example.
Suppose a combo model has
length(a.mix) > 2
and
length(i.mix) > 2
,
length(d.mix) > 2
,
a.mlm = 3:5
,
i.mlm = 6:9
and
d.mlm = 10:12
, say.
Then loglink(lambda.p)
is the first.
The second is multilogitlink(pobs.mix)
followed
by loglink(lambda.a)
because a.mix
is long enough.
The fourth is multilogitlink(pstr.mix)
followed
by loglink(lambda.i)
because i.mix
is long enough.
The sixth is multilogitlink(pdip.mix)
followed
by loglink(lambda.d)
because d.mix
is long enough.
Next are the probabilities for the a.mlm
values.
Then are the probabilities for the i.mlm
values.
Lastly are the probabilities for the d.mlm
values.
All the probabilities are estimated by one big MLM
and effectively
the "(Others)"
column of left over probabilities is
associated with the nonspecial values.
These might be called the
nonspecial baseline probabilities (NBP).
The dimension of the vector of linear/additive predictors here
is \(M=17\).
Two mixture submodels that may be fitted can be abbreviated
GAT--Pois or
GIT--Pois.
For the GAT model
the distribution being fitted is a (spliced) mixture
of two Poissons with differing (partitioned) support.
Likewise, for the GIT model
the distribution being fitted is a mixture
of two Poissons with nested support.
The two rate parameters may be constrained to be equal using
eq.ap
and eq.ip
.
A good first step is to apply spikeplot
for selecting
candidate values for altering, inflating and deflating.
Deciding between
parametrically or nonparametrically can also be determined from
examining the spike plot. Misspecified
a.mix/a.mlm/i.mix/i.mlm/d.mix/d.mlm
will result in convergence problems
(setting trace = TRUE
is a very good idea.)
This function currently does not handle multiple responses.
Further details are at Gaitdpois
.
A well-conditioned data--model combination should pose no
difficulties for the automatic starting value selection
being successful.
Failure to obtain initial values from this self-starting
family function indicates the degree of inflation/deflation
may be marginal and/or a misspecified model.
If this problem is worth surmounting
the arguments to focus on especially are
mux.init
,
gpstr.mix
, gpstr.mlm
,
ipdip.mix
and ipdip.mlm
.
See below for the stepping-stone trick.
Apart from the order of the linear/additive predictors,
the following are (or should be) equivalent:
gaitdpoisson()
and poissonff()
,
gaitdpoisson(a.mix = 0)
and zapoisson(zero = "pobs0")
,
gaitdpoisson(i.mix = 0)
and zipoisson(zero = "pstr0")
,
gaitdpoisson(truncate = 0)
and pospoisson()
.
Likewise, if
a.mix
and i.mix
are assigned a scalar then
it effectively moves that scalar to a.mlm
and i.mlm
because there is no lambda.a
or lambda.i
being estimated.
Thus
gaitdpoisson(a.mix = 0)
and gaitdpoisson(a.mlm = 0)
are the effectively same, and ditto for
gaitdpoisson(i.mix = 0)
and gaitdpoisson(i.mlm = 0)
.
Yee, T. W. and Ma, C. (2023) Generally altered, inflated, truncated and deflated regression. In preparation.
Gaitdpois
,
multinomial
,
rootogram4
,
specials
,
plotdgaitd
,
spikeplot
,
meangaitd
,
KLD
,
goffset
,
Trunc
,
gaitdnbinomial
,
gaitdlog
,
gaitdzeta
,
multilogitlink
,
multinomial
,
residualsvglm
,
poissonff
,
zapoisson
,
zipoisson
,
pospoisson
,
CommonVGAMffArguments
,
simulate.vlm
.
i.mix <- c(5, 10) # Inflate these values parametrically
i.mlm <- c(14, 15) # Inflate these values
a.mix <- c(1, 13) # Alter these values
tvec <- c(3, 11) # Truncate these values
pstr.mlm <- 0.1 # So parallel.i = TRUE
pobs.mix <- pstr.mix <- 0.1
max.support <- 20; set.seed(1)
gdata <- data.frame(x2 = runif(nn <- 1000))
gdata <- transform(gdata, lambda.p = exp(2 + 0.0 * x2))
gdata <- transform(gdata,
y1 = rgaitdpois(nn, lambda.p, a.mix = a.mix, i.mix = i.mix,
pobs.mix = pobs.mix, pstr.mix = pstr.mix,
i.mlm = i.mlm, pstr.mlm = pstr.mlm,
truncate = tvec, max.support = max.support))
gaitdpoisson(a.mix = a.mix, i.mix = i.mix, i.mlm = i.mlm)
with(gdata, table(y1))
fit1 <- vglm(y1 ~ 1, crit = "coef", trace = TRUE, data = gdata,
gaitdpoisson(a.mix = a.mix, i.mix = i.mix,
i.mlm = i.mlm, parallel.i = TRUE,
eq.ap = TRUE, eq.ip = TRUE, truncate =
tvec, max.support = max.support))
head(fitted(fit1, type.fitted = "Pstr.mix"))
head(predict(fit1))
t(coef(fit1, matrix = TRUE)) # Easier to see with t()
summary(fit1) # No HDE test by default but HDEtest = TRUE is ideal
if (FALSE) spikeplot(with(gdata, y1), lwd = 2)
plotdgaitd(fit1, new.plot = FALSE, offset.x = 0.2, all.lwd = 2)
Run the code above in your browser using DataLab