File: extend_family.R

package info (click to toggle)
r-cran-projpred 2.0.2%2Bdfsg-1
  • links: PTS, VCS
  • area: main
  • in suites: bullseye
  • size: 740 kB
  • sloc: cpp: 355; sh: 14; makefile: 2
file content (249 lines) | stat: -rw-r--r-- 8,124 bytes parent folder | download
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
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
# Model-specific helper functions.
#
# \code{extend_family(family)} returns a family object augmented with auxiliary
# functions that
# are needed for computing KL-divergence, log predictive density, dispersion
# projection etc.
#
# Missing: Quasi-families are not implemented. If dis_gamma is the correct shape
# parameter for projected Gamma regression, everything should be OK for gamma.

#' Add extra fields to the family object.
#' @param family Family object.
#' @return Extended family object.
#' @export
extend_family <- function(family) {
  if (.has_family_extras(family)) {
    ## if the object already was created using this function, then return
    return(family)
  }
  extend_family_specific <- paste0("extend_family_", tolower(family$family))
  extend_family_specific <- get(extend_family_specific, mode = "function")
  extend_family_specific(family)
}

extend_family_binomial <- function(family) {
  kl_dev <- function(pref, data, psub) {
    if (NCOL(pref$mu) > 1) {
      w <- rep(data$weights, NCOL(pref$mu))
      colMeans(family$dev.resids(pref$mu, psub$mu, w)) / 2
    } else {
      mean(family$dev.resids(pref$mu, psub$mu, data$weights)) / 2
    }
  }
  dis_na <- function(pref, psub, wobs = 1) rep(0, ncol(pref$mu))
  predvar_na <- function(mu, dis, wsample = 1) {
    0
  }
  ll_binom <- function(mu, dis, y, weights = 1) {
    dbinom(y, weights, mu, log = TRUE)
  }
  dev_binom <- function(mu, y, weights = 1, dis = NULL) {
    if (NCOL(y) < NCOL(mu)) {
      y <- matrix(y, nrow = length(y), ncol = NCOL(mu))
    }
    -2 * weights * (y * log(mu) + (1 - y) * log(1 - mu))
  }
  ppd_binom <- function(mu, dis, weights = 1) rbinom(length(mu), weights, mu)
  initialize_binom <- expression({
    if (NCOL(y) == 1) {
      if (is.factor(y)) {
        y <- y != levels(y)[1L]
      }
      n <- rep.int(1, nobs)
      y[weights == 0] <- 0
      mustart <- (weights * y + 0.5) / (weights + 1)
      m <- weights * y
    }
    else if (NCOL(y) == 2) {
      n <- (y1 <- y[, 1L]) + y[, 2L]
      y <- y1 / n
      if (any(n0 <- n == 0)) {
        y[n0] <- 0
      }
      weights <- weights * n
      mustart <- (n * y + 0.5) / (n + 1)
    }
  })

  family$initialize <- initialize_binom
  family$kl <- kl_dev
  family$dis_fun <- dis_na
  family$predvar <- predvar_na
  family$ll_fun <- ll_binom
  family$deviance <- dev_binom
  family$ppd <- ppd_binom

  return(family)
}

extend_family_poisson <- function(family) {
  kl_dev <- function(pref, data, psub) {
    if (NCOL(pref$mu) > 1) {
      w <- rep(data$weights, NCOL(pref$mu))
      colMeans(family$dev.resids(pref$mu, psub$mu, w)) / 2
    } else {
      mean(family$dev.resids(pref$mu, psub$mu, data$weights)) / 2
    }
  }
  dis_na <- function(pref, psub, wobs = 1) rep(0, ncol(pref$mu))
  predvar_na <- function(mu, dis, wsample = 1) {
    0
  }
  ll_poiss <- function(mu, dis, y, weights = 1)
    weights * dpois(y, mu, log = TRUE)
  dev_poiss <- function(mu, y, weights = 1, dis = NULL) {
    if (NCOL(y) < NCOL(mu)) {
      y <- matrix(y, nrow = length(y), ncol = NCOL(mu))
    }
    -2 * weights * (y * log(mu) - mu)
  }
  ppd_poiss <- function(mu, dis, weights = 1) rpois(length(mu), mu)

  family$kl <- kl_dev
  family$dis_fun <- dis_na
  family$predvar <- predvar_na
  family$ll_fun <- ll_poiss
  family$deviance <- dev_poiss
  family$ppd <- ppd_poiss

  return(family)
}

extend_family_gaussian <- function(family) {
  kl_gauss <- function(pref, data, psub) {
    colSums(data$weights * (psub$mu - pref$mu)^2)
  } # not the actual KL but reasonable surrogate..
  dis_gauss <- function(pref, psub, wobs = 1) {
    sqrt(colSums(wobs / sum(wobs) * (pref$var + (pref$mu - psub$mu)^2)))
  }
  predvar_gauss <- function(mu, dis, wsample = 1) {
    wsample <- wsample / sum(wsample)
    mu_mean <- mu %*% wsample
    mu_var <- mu^2 %*% wsample - mu_mean^2
    as.vector(sum(wsample * dis^2) + mu_var)
  }
  ll_gauss <- function(mu, dis, y, weights = 1) {
    dis <- matrix(rep(dis, each = length(y)), ncol = NCOL(mu))
    weights * dnorm(y, mu, dis, log = TRUE)
  }
  dev_gauss <- function(mu, y, weights = 1, dis = NULL) {
    if (is.null(dis)) {
      dis <- 1
    } else {
      dis <- matrix(rep(dis, each = length(y)), ncol = NCOL(mu))
    }
    if (NCOL(y) < NCOL(mu)) {
      y <- matrix(y, nrow = length(y), ncol = NCOL(mu))
    }
    -2 * weights * (-0.5 / dis * (y - mu)^2 - log(dis))
  }
  ppd_gauss <- function(mu, dis, weights = 1) rnorm(length(mu), mu, dis)

  family$kl <- kl_gauss
  family$dis_fun <- dis_gauss
  family$predvar <- predvar_gauss
  family$ll_fun <- ll_gauss
  family$deviance <- dev_gauss
  family$ppd <- ppd_gauss

  return(family)
}

extend_family_gamma <- function(family) {
  kl_gamma <- function(pref, data, psub) {
    stop("KL-divergence for gamma not implemented yet.")
    ## mean(data$weights*(
    ##   p_sub$dis*(log(pref$dis)-log(p_sub$dis)+log(psub$mu)-log(pref$mu)) +
    ##     digamma(pref$dis)*(pref$dis - p_sub$dis) - lgamma(pref$dis) +
    ##     lgamma(p_sub$dis) + pref$mu*p_sub$dis/p_sub$mu - pref$dis))
  }
  dis_gamma <- function(pref, psub, wobs = 1) {
    ## TODO, IMPLEMENT THIS
    stop("Projection of dispersion parameter not yet implemented for family",
         " Gamma.")
    ## mean(data$weights*((pref$mu - p_sub$mu)/
    ##                      family$mu.eta(family$linkfun(p_sub$mu))^2))
  }
  predvar_gamma <- function(mu, dis, wsample = 1) {
    stop("Family Gamma not implemented yet.")
  }
  ll_gamma <- function(mu, dis, y, weights = 1) {
    dis <- matrix(rep(dis, each = length(y)), ncol = NCOL(mu))
    weights * dgamma(y, dis, dis / matrix(mu), log = TRUE)
  }
  dev_gamma <- function(mu, dis, y, weights = 1) {
    stop("Loss function not implemented for Gamma-family yet.")
    ## dis <- matrix(rep(dis, each=length(y)), ncol=NCOL(mu))
    ## weights*dgamma(y, dis, dis/matrix(mu), log= TRUE)
  }
  ppd_gamma <- function(mu, dis, weights = 1) rgamma(length(mu), dis, dis / mu)

  family$kl <- kl_gamma
  family$dis_fun <- dis_gamma
  family$predvar <- predvar_gamma
  family$ll_fun <- ll_gamma
  family$deviance <- dev_gamma
  family$ppd <- ppd_gamma

  return(family)
}

extend_family_student_t <- function(family) {
  kl_student_t <- function(pref, data, psub) {
    log(psub$dis)
  } #- 0.5*log(pref$var) # FIX THIS, NOT CORRECT
  dis_student_t <- function(pref, psub, wobs = 1) {
    s2 <- colSums(psub$w / sum(wobs) *
                  (pref$var + (pref$mu - psub$mu)^2)) # CHECK THIS
    sqrt(s2)
    ## stop('Projection of dispersion not yet implemented for student-t')
  }
  predvar_student_t <- function(mu, dis, wsample = 1) {
    wsample <- wsample / sum(wsample)
    mu_mean <- mu %*% wsample
    mu_var <- mu^2 %*% wsample - mu_mean^2
    as.vector(family$nu / (family$nu - 2) * sum(wsample * dis^2) + mu_var)
  }
  ll_student_t <- function(mu, dis, y, weights = 1) {
    dis <- matrix(rep(dis, each = length(y)), ncol = NCOL(mu))
    if (NCOL(y) < NCOL(mu)) {
      y <- matrix(y, nrow = length(y), ncol = NCOL(mu))
    }
    weights * (dt((y - mu) / dis, family$nu, log = TRUE) - log(dis))
  }
  dev_student_t <- function(mu, y, weights = 1, dis = NULL) {
    if (is.null(dis)) {
      dis <- 1
    } else {
      dis <- matrix(rep(dis, each = length(y)), ncol = NCOL(mu))
    }
    if (NCOL(y) < NCOL(mu)) {
      y <- matrix(y, nrow = length(y), ncol = NCOL(mu))
    }
    (-2 * weights * (-0.5 * (family$nu + 1)
      * log(1 + 1 / family$nu * ((y - mu) / dis)^2) - log(dis)))
  }
  ppd_student_t <- function(mu, dis, weights = 1)
    rt(length(mu), family$nu) * dis + mu

  family$kl <- kl_student_t
  family$dis_fun <- dis_student_t
  family$predvar <- predvar_student_t
  family$ll_fun <- ll_student_t
  family$deviance <- dev_student_t
  family$ppd <- ppd_student_t

  return(family)
}

.has_dispersion <- function(family) {
  # a function for checking whether the family has a dispersion parameter
  family$family %in% c("gaussian", "Student_t", "Gamma")
}

.has_family_extras <- function(family) {
  # check whether the family object has the extra functions, that is, whether it
  # was created by extend_family
  !is.null(family$deviance)
}