1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
|
## simple wrapper function to specify fitter and return class
glmtree <- function(formula, data, subset, na.action, weights, offset, cluster,
family = gaussian, epsilon = 1e-8, maxit = 25, method = "glm.fit", ...)
{
## use dots for setting up mob_control
control <- mob_control(...)
## keep call
cl <- match.call(expand.dots = TRUE)
## extend formula if necessary
f <- Formula::Formula(formula)
if(length(f)[2L] == 1L) {
attr(f, "rhs") <- c(list(1), attr(f, "rhs"))
formula[[3L]] <- formula(f)[[3L]]
} else {
f <- NULL
}
## process family
if(inherits(family, "family")) {
fam <- TRUE
} else {
fam <- FALSE
if(is.character(family)) family <- get(family)
if(is.function(family)) family <- family()
}
## distinguish whether glm should be fixed for case weights or not
glmfit0 <- function(y, x, start = NULL, weights = NULL, offset = NULL, cluster = NULL, ...,
estfun = FALSE, object = FALSE, caseweights = TRUE)
{
glmfit(y = y, x = x, start = start, weights = weights, offset = offset, cluster = cluster, ...,
estfun = estfun, object = object, caseweights = control$caseweights)
}
## call mob
m <- match.call(expand.dots = FALSE)
if(!is.null(f)) m$formula <- formula
m$fit <- glmfit0
m$control <- control
m$epsilon <- epsilon
m$maxit <- maxit
m$method <- method
if("..." %in% names(m)) m[["..."]] <- NULL
if(!fam) m$family <- family
m[[1L]] <- as.call(quote(partykit::mob))
rval <- eval(m, parent.frame())
## extend class and keep original call
rval$info$call <- cl
rval$info$family <- family$family
class(rval) <- c("glmtree", class(rval))
return(rval)
}
## actual fitting function for mob()
glmfit <- function(y, x, start = NULL, weights = NULL, offset = NULL, cluster = NULL, ...,
estfun = FALSE, object = FALSE, caseweights = TRUE)
{
## catch control arguments
args <- list(...)
ctrl <- list()
for(n in c("epsilon", "maxit")) {
if(n %in% names(args)) {
ctrl[[n]] <- args[[n]]
args[[n]] <- NULL
}
}
args$control <- do.call("glm.control", ctrl)
## add intercept-only regressor matrix (if missing)
## NOTE: does not have terms/formula
if(is.null(x)) x <- matrix(1, nrow = NROW(y), ncol = 1L,
dimnames = list(NULL, "(Intercept)"))
## call glm fitting function (defaulting to glm.fit)
glm.method <- if("method" %in% names(args)) args[["method"]] else "glm.fit"
args[["method"]] <- NULL
args <- c(list(x = x, y = y, start = start, weights = weights, offset = offset), args)
z <- do.call(glm.method, args)
## degrees of freedom
df <- z$rank
if(z$family$family %in% c("gaussian", "Gamma", "inverse.gaussian")) df <- df + 1
if(substr(z$family$family, 1L, 5L) != "quasi") objfun <- z$aic/2 - df else objfun <- z$deviance
## list structure
rval <- list(
coefficients = z$coefficients,
objfun = objfun,
estfun = NULL,
object = NULL
)
## add estimating functions (if desired)
if(estfun) {
wres <- as.vector(z$residuals) * z$weights
dispersion <- if(substr(z$family$family, 1L, 17L) %in% c("poisson", "binomial", "Negative Binomial")) {
1
} else {
## for case weights: fix dispersion estimate
if(!is.null(weights) && caseweights) {
sum(wres^2/weights, na.rm = TRUE)/sum(z$weights, na.rm = TRUE)
} else {
sum(wres^2, na.rm = TRUE)/sum(z$weights, na.rm = TRUE)
}
}
rval$estfun <- wres * x[, !is.na(z$coefficients), drop = FALSE]/dispersion
}
## add model (if desired)
if(object) {
class(z) <- c("glm", "lm")
z$offset <- if(is.null(offset)) 0 else offset
z$contrasts <- attr(x, "contrasts")
z$xlevels <- attr(x, "xlevels")
cl <- as.call(expression(glm))
cl$formula <- attr(x, "formula")
if(!is.null(offset)) cl$offset <- attr(x, "offset")
z$call <- cl
z$terms <- attr(x, "terms")
## for case weights: change degrees of freedom
if(!is.null(weights) && caseweights) {
z$df.null <- z$df.null - sum(weights > 0) + sum(weights)
z$df.residual <- z$df.residual - sum(weights > 0) + sum(weights)
}
rval$object <- z
}
return(rval)
}
## methods
print.glmtree <- function(x,
title = NULL, objfun = NULL, ...)
{
if(is.null(title)) title <- sprintf("Generalized linear model tree (family: %s)", x$info$family)
if(is.null(objfun)) objfun <- if(substr(x$info$family, 1L, 5L) != "quasi") "negative log-likelihood" else "deviance"
print.modelparty(x, title = title, objfun = objfun, ...)
}
predict.glmtree <- function(object, newdata = NULL, type = "response", ...)
{
## FIXME: possible to get default?
if(is.null(newdata) & !identical(type, "node")) stop("newdata has to be provided")
predict.modelparty(object, newdata = newdata, type = type, ...)
}
plot.glmtree <- function(x, terminal_panel = node_bivplot,
tp_args = list(), tnex = NULL, drop_terminal = NULL, ...)
{
nreg <- if(is.null(tp_args$which)) x$info$nreg else length(tp_args$which)
if(nreg < 1L & missing(terminal_panel)) {
plot.constparty(as.constparty(x),
tp_args = tp_args, tnex = tnex, drop_terminal = drop_terminal, ...)
} else {
if(is.null(tnex)) tnex <- if(is.null(terminal_panel)) 1L else 2L * nreg
if(is.null(drop_terminal)) drop_terminal <- !is.null(terminal_panel)
plot.modelparty(x, terminal_panel = terminal_panel,
tp_args = tp_args, tnex = tnex, drop_terminal = drop_terminal, ...)
}
}
|
