如何使用非线性模型测试分组变量的效果?

机器算法验证 r 混合模式 nls
2022-01-18 14:17:22

我有一个关于在非线性模型中使用分组变量的问题。由于 nls() 函数不允许因子变量,我一直在努力弄清楚是否可以测试一个因子对模型拟合的影响。我在下面提供了一个示例,我希望将“季节性 von Bertalanffy”生长模型拟合到不同的生长处理(最常应用于鱼类生长)。我想测试鱼生长的湖泊以及提供的食物的效果(只是一个人为的例子)。我熟悉解决此问题的方法 - 应用 F 检验比较模型拟合合并数据与 Chen 等人概述的单独拟合。(1992 年)(ARSS -“残差平方和分析”)。换句话说,对于下面的示例,

在此处输入图像描述

我想在 R 中有一种更简单的方法可以使用 nlme() 执行此操作,但我遇到了问题。首先,通过使用分组变量,自由度高于我通过拟合单独模型获得的自由度。其次,我无法嵌套分组变量——我看不出我的问题出在哪里。非常感谢使用 nlme 或其他方法的任何帮助。下面是我的人工示例的代码:

###seasonalized von Bertalanffy growth model
soVBGF <- function(S.inf, k, age, age.0, age.s, c){
    S.inf * (1-exp(-k*((age-age.0)+(c*sin(2*pi*(age-age.s))/2*pi)-(c*sin(2*pi*(age.0-age.s))/2*pi))))
}

###Make artificial data
food <- c("corn", "corn", "wheat", "wheat")
lake <- c("king", "queen", "king", "queen")

#cornking, cornqueen, wheatking, wheatqueen
S.inf <- c(140, 140, 130, 130)
k <- c(0.5, 0.6, 0.8, 0.9)
age.0 <- c(-0.1, -0.05, -0.12, -0.052)
age.s <- c(0.5, 0.5, 0.5, 0.5)
cs <- c(0.05, 0.1, 0.05, 0.1)

PARS <- data.frame(food=food, lake=lake, S.inf=S.inf, k=k, age.0=age.0, age.s=age.s, c=cs)

#make data
set.seed(3)
db <- c()
PCH <- NaN*seq(4)
COL <- NaN*seq(4)
for(i in seq(4)){
    age <- runif(min=0.2, max=5, 100)
    age <- age[order(age)]
    size <- soVBGF(PARS$S.inf[i], PARS$k[i], age, PARS$age.0[i], PARS$age.s[i], PARS$c[i]) + rnorm(length(age), sd=3)
	PCH[i] <- c(1,2)[which(levels(PARS$food) == PARS$food[i])]
	COL[i] <- c(2,3)[which(levels(PARS$lake) == PARS$lake[i])]
	db <- rbind(db, data.frame(age=age, size=size, food=PARS$food[i], lake=PARS$lake[i], pch=PCH[i], col=COL[i]))
}

#visualize data
plot(db$size ~ db$age, col=db$col, pch=db$pch)
legend("bottomright", legend=paste(PARS$food, PARS$lake), col=COL, pch=PCH)


###fit growth model
library(nlme)

starting.values <- c(S.inf=140, k=0.5, c=0.1, age.0=0, age.s=0)

#fit to pooled data ("small model")
fit0 <- nls(size ~ soVBGF(S.inf, k, age, age.0, age.s, c), 
  data=db,
  start=starting.values
)
summary(fit0)

#fit to each lake separatly ("large model")
fit.king <- nls(size ~ soVBGF(S.inf, k, age, age.0, age.s, c), 
  data=db,
  start=starting.values,
  subset=db$lake=="king"
)
summary(fit.king)

fit.queen <- nls(size ~ soVBGF(S.inf, k, age, age.0, age.s, c), 
  data=db,
  start=starting.values,
  subset=db$lake=="queen"
)
summary(fit.queen)


#analysis of residual sum of squares (F-test)
resid.small <- resid(fit0)
resid.big <- c(resid(fit.king),resid(fit.queen))
df.small <- summary(fit0)$df
df.big <- summary(fit.king)$df+summary(fit.queen)$df

F.value <- ((sum(resid.small^2)-sum(resid.big^2))/(df.big[1]-df.small[1])) / (sum(resid.big^2)/(df.big[2]))
P.value <- pf(F.value , (df.big[1]-df.small[1]), df.big[2], lower.tail = FALSE)
F.value; P.value


###plot models
plot(db$size ~ db$age, col=db$col, pch=db$pch)
legend("bottomright", legend=paste(PARS$food, PARS$lake), col=COL, pch=PCH)
legend("topleft", legend=c("soVGBF pooled", "soVGBF king", "soVGBF queen"), col=c(1,2,3), lwd=2)

#plot "small" model (pooled data)
tmp <- data.frame(age=seq(min(db$age), max(db$age),,100))
pred <- predict(fit0, tmp)
lines(tmp$age, pred, col=1, lwd=2)

#plot "large" model (seperate fits)
tmp <- data.frame(age=seq(min(db$age), max(db$age),,100), lake="king")
pred <- predict(fit.king, tmp)
lines(tmp$age, pred, col=2, lwd=2)
tmp <- data.frame(age=seq(min(db$age), max(db$age),,100), lake="queen")
pred <- predict(fit.queen, tmp)
lines(tmp$age, pred, col=3, lwd=2)



###Can this be done in one step using a grouping variable?
#with "lake" as grouping variable
starting.values <- c(S.inf=140, k=0.5, c=0.1, age.0=0, age.s=0)
fit1 <- nlme(model = size ~ soVBGF(S.inf, k, age, age.0, age.s, c), 
  data=db,
  fixed = S.inf + k + c + age.0 + age.s ~ 1,
  group = ~ lake,
  start=starting.values
)
summary(fit1)

#similar residuals to the seperatly fitted models
sum(resid(fit.king)^2+resid(fit.queen)^2)
sum(resid(fit1)^2)

#but different degrees of freedom? (10 vs. 21?)
summary(fit.king)$df+summary(fit.queen)$df
AIC(fit1, fit0)


###I would also like to nest my grouping factors. This doesn't work...
#with "lake" and "food" as grouping variables
starting.values <- c(S.inf=140, k=0.5, c=0.1, age.0=0, age.s=0)
fit2 <- nlme(model = size ~ soVBGF(S.inf, k, age, age.0, age.s, c), 
  data=db,
  fixed = S.inf + k + c + age.0 + age.s ~ 1,
  group = ~ lake/food,
  start=starting.values
)

参考文献:Chen, Y.、Jackson, DA 和 Harvey, HH, 1992。von Bertalanffy 和多项式函数在鱼类生长数据建模中的比较。49, 6: 1228-1235。

2个回答

您可以按分类预测变量的值进行分层并比较拟合。例如,假设您有连续的预测变量X1,...,Xp和因变量Y. 我相信 nls() 给出了最大似然估计f这样

Y=f(X1,...,Xp)+ε

在哪里εN(0,σ2)f以某种非线性方式参数化(请参阅 nls 帮助文件)。假设你有一个分类预测器Bm水平并根据值对数据进行分层B并在每个层内拟合模型。由于这些是数据的不相交子集,因此分层模型的对数似然,L1是每个层内的似然之和,可以使用 logLik 函数从 nls 模型中提取(您也可以从未分层模型中获得对数似然,L0, 使用 logLik)。

未分层模型显然是分层模型的子模型,因此似然比检验适合查看较大的模型是否值得增加复杂性 - 检验统计量为

λ=2(L1L0)

如果分类预测器真的没有效果,λ有一个χ2自由度等于的分布mpp=p(m1), 在哪里p是您的非线性回归函数的参数数量。您可以使用 pchisq() 来计算χ2p 值。

我发现可以使用 nls() 对分类变量进行编码,只需将真/假向量乘以方程即可。例子:

# null model (no difference between groups; all have the same coefficients)
nls.null <- nls(formula = percent_on_cells ~ vmax*(Time/(Time+km)),
            data = mehg,
            start = list(vmax = 0.6, km = 10))

# alternative model (each group has different coefficients)
nls.alt <- nls(formula = percent_on_cells ~ 
              as.numeric(DOC==0)*(vmax1)*(Time/(Time+(km1))) 
            + as.numeric(DOC==1)*(vmax2)*(Time/(Time+(km2)))
            + as.numeric(DOC==10)*(vmax3)*(Time/(Time+(km3)))
            + as.numeric(DOC==100)*(vmax4)*(Time/(Time+(km4))),
            data = mehg, 
            start = list(vmax1=0.63, km1=3.6, 
                         vmax2=0.64, km2=3.6, 
                         vmax3=0.50, km3=3.2,
                         vmax4= 0.40, km4=9.7))