我玩过这个。我不确定,但我认为AIC
包使用的公式biglm
是:
2 * (n.parameters + obs.added - 1) + deviance(a)
其中obs_added
是 中的观察chunk2
数加上 中的观察数chunk3
:
obs.added <- dim(chunk2)[1] + dim(chunk3)[1]
并且是由(其中是误差项)n.parameters
返回的估计系数的数量,并且是模型的偏差。summary(a) + 1
+1
deviance(a)
a
####################################################
data(trees)
ff <- log(Volume)~log(Girth)+log(Height)
n.parm <- 4
chunk1<-trees[1:10,]
chunk2<-trees[11:20,]
chunk3<-trees[21:31,]
obs.added <- dim(chunk2)[1] + dim(chunk3)[1]
library(biglm)
a <- biglm(ff,chunk1)
a <- update(a,chunk2)
a <- update(a,chunk3)
AIC(a)
summary(a)
deviance(a)
2 * (n.parm + obs.added - 1) + deviance(a)
round(AIC(a), 5) == round(2 * (n.parm + obs.added - 1) + deviance(a), 5)
# [1] TRUE
####################################################
由于我不能 100% 确定我的答案是正确的,因此您可以使用下面的代码,看看您是否可以找到 AIC 的建议公式不起作用的场景。如果我发现任何此类情况,我将尝试根据需要修改下面的代码和上面的公式。
#########################################################
# Generate some data
n <- 118 # number of observations
B0 <- 2 # intercept
B1 <- -1.5 # slope 1
B2 <- 0.4 # slope 2
B3 <- 2.0 # slope 3
B4 <- -0.8 # slope 4
sigma2 <- 5 # residual variance
x1 <- round(runif(n, -5 , 5), digits = 3) # covariate 1
x2 <- round(runif(n, 10 , 20), digits = 3) # covariate 2
x3 <- round(runif(n, 2 , 8), digits = 3) # covariate 3
x4 <- round(runif(n, 12 , 15), digits = 3) # covariate 4
eps <- rnorm(n, mean = 0, sd = sqrt(sigma2)) # error
y <- B0 + B1 * x1 + B2 * x2 + B3 * x3 + B4 * x4 + eps # dependent variable
my.data <- data.frame(y, x1, x2, x3, x4)
# analyze data with linear regression
model.1 <- lm(my.data$y ~ my.data$x1 + my.data$x2 + my.data$x3 + my.data$x4)
summary(model.1)
AIC(model.1)
n.parms <- length(model.1$coefficients) + 1
my.AIC <- 2 * n.parms - 2 * as.numeric(logLik(model.1))
my.AIC
#########################################################
ff0 <- y ~ 1
ff1 <- y ~ x1
ff2 <- y ~ x1 + x2
ff3 <- y ~ x1 + x2 + x3
ff4 <- y ~ x1 + x2 + x3 + x4
n.parm0 <- 2
n.parm1 <- 3
n.parm2 <- 4
n.parm3 <- 5
n.parm4 <- 6
n.chunks <- 5
chunk1<-my.data[ 1:round(((nrow(my.data)/n.chunks)*1)+0),]
chunk2<-my.data[round(((nrow(my.data)/n.chunks)*1)+1):round(((nrow(my.data)/n.chunks)*2)+0),]
chunk3<-my.data[round(((nrow(my.data)/n.chunks)*2)+1):round(((nrow(my.data)/n.chunks)*3)+0),]
chunk4<-my.data[round(((nrow(my.data)/n.chunks)*3)+1):round(((nrow(my.data)/n.chunks)*4)+0),]
chunk5<-my.data[round(((nrow(my.data)/n.chunks)*4)+1):nrow(my.data),]
obs.added <- dim(chunk2)[1] + dim(chunk3)[1] + dim(chunk4)[1] + dim(chunk5)[1]
# check division of data
foo <- list()
foo[[1]] <- chunk1
foo[[2]] <- chunk2
foo[[3]] <- chunk3
foo[[4]] <- chunk4
foo[[5]] <- chunk5
all.data.foo <- do.call(rbind, foo)
all.equal(my.data, all.data.foo)
####################################################
library(biglm)
####################################################
a0 <- biglm(ff0, chunk1)
a0 <- update(a0, chunk2)
a0 <- update(a0, chunk3)
a0 <- update(a0, chunk4)
a0 <- update(a0, chunk5)
AIC(a0)
summary(a0)
deviance(a0)
print(a0)
2 * (n.parm0 + obs.added - 1) + deviance(a0)
round(AIC(a0), 5) == round(2 * (n.parm0 + obs.added - 1) + deviance(a0), 5)
####################################################
a1 <- biglm(ff1, chunk1)
a1 <- update(a1, chunk2)
a1 <- update(a1, chunk3)
a1 <- update(a1, chunk4)
a1 <- update(a1, chunk5)
AIC(a1)
summary(a1)
deviance(a1)
print(a1)
2 * (n.parm1 + obs.added - 1) + deviance(a1)
round(AIC(a1), 5) == round(2 * (n.parm1 + obs.added - 1) + deviance(a1), 5)
####################################################
a2 <- biglm(ff2, chunk1)
a2 <- update(a2, chunk2)
a2 <- update(a2, chunk3)
a2 <- update(a2, chunk4)
a2 <- update(a2, chunk5)
AIC(a2)
summary(a2)
deviance(a2)
print(a2)
2 * (n.parm2 + obs.added - 1) + deviance(a2)
round(AIC(a2), 5) == round(2 * (n.parm2 + obs.added - 1) + deviance(a2), 5)
####################################################
a3 <- biglm(ff3, chunk1)
a3 <- update(a3, chunk2)
a3 <- update(a3, chunk3)
a3 <- update(a3, chunk4)
a3 <- update(a3, chunk5)
AIC(a3)
summary(a3)
deviance(a3)
print(a3)
2 * (n.parm3 + obs.added - 1) + deviance(a3)
round(AIC(a3), 5) == round(2 * (n.parm3 + obs.added - 1) + deviance(a3), 5)
####################################################
a4 <- biglm(ff4, chunk1)
a4 <- update(a4, chunk2)
a4 <- update(a4, chunk3)
a4 <- update(a4, chunk4)
a4 <- update(a4, chunk5)
AIC(a4)
summary(a4)
deviance(a4)
print(a4)
2 * (n.parm4 + obs.added - 1) + deviance(a4)
round(AIC(a4), 5) == round(2 * (n.parm4 + obs.added - 1) + deviance(a4), 5)
####################################################
编辑
我建议biglm
使用以下等式AIC
:
2 * (n.parameters + obs.added - 1) + deviance(a)
Ben Bolker 指出,用于的等式biglm
是AIC
:
deviance(object) + k * (object$n - object$df.resid)
Ben 还确定biglm
没有更新残差 df 的第一个值。
鉴于这些新信息,我现在看到这两个方程是等价的。
首先,将两个方程限制为以下,这是它们唯一不同的地方:
(n.parameters + obs.added - 1) # mine
(object$n - object$df.resid) # Ben's
重新排列我的,以说明我在参数数量上加 1,然后减 1:
((n.parameters-1) + obs.added) = ((4-1) + obs.added) = (3 + 21) = 24
现在把我的方程变成 Ben 的方程:
我3
的是一样的:
(number of observations in chunk1 - object$df.resid) = (10 - 7) = 3
给予:
((number of obs in chunk1 - object$df.resid) + obs.added) = ((10-7) + 21)
或者:
(3 + 21) = 24
改编:
((number of obs in chunk1 + obs.added) - object$df.resid) = ((10 + 21) - 7)
或者:
(31 - 7) = 24
和:
((number of observations in chunk1 + obs.added) - object$df.resid)
是相同的:
(total number of observations - object$df.resid)
这与以下内容相同:
(object$n - object$df.resid) = (31 - 7) = 24
看来我提出的方程确实是biglm
用于的方程AIC
,只是以不同的形式表示。
当然,我之所以能够意识到这一点,是因为 Ben 提供了关键代码和错误的关键解释。