Question

我想在一些数据集上拟合曲线，按治疗分组。这对nlslist非常有效，但现在我想介绍我的参数的上限。

当我使用nls单独使用每个组时，介绍边界确实非常好，但显然不是当我想用nlslist加速我的工作（我有更多的数据集）时。

有人可以帮我解决这个问题吗？

我的数据集示例：

DF1<-data.frame(treatment = rep(c("mineral","residues"),4),
            N_level = c(0,0,100,100,200,200,300,300),
            yield = c(8,8.5,10,10.5,11,9.8,9.5,9.7))

DF1

    treatment N_level yield
1   mineral       0   8.0
2  residues       0   8.5
3   mineral     100  10.0
4  residues     100  10.5
5   mineral     200  11.0
6  residues     200   9.8
7   mineral     300   9.5
8  residues     300   9.7

尝试仅使用nls适合此数据集效果很好：

fit_mineral <- nls(formula = yield ~ a + b*0.99^N_level +c*N_level, 
                data=subset(DF1, subset = treatment == "mineral"), 
                algorithm = "port", start = list(a = 12, b = -8, c= -0.01), 
                upper = list(a=1000, b=-0.000001, c=-0.000001))

fit_mineral

Nonlinear regression model
model: yield ~ a + b * 0.99^N_level + c * N_level
data: subset(DF1, subset = treatment == "mineral")
    a       b       c 
13.7882 -5.8685 -0.0126 
residual sum-of-squares: 0.4679

但是一旦我尝试在nlslist中组合内容，它就不起作用了：

fit_mineral_and_residues <- nlsList(model = yield ~ a + b*0.99^N_level +c*N_level 
                      | treatment, data=DF1, 
                      algorithm = "port", start = list(a = 12, b = -8, c= -0.01), 
                      upper = list(a=1000, b=-0.000001, c=-0.000001))

错误消息：

Error in nlsList(model = yield ~ a + b * 0.99^N_level + c * N_level |  : 
unused arguments (algorithm = "port", upper = list(a = 1000, b = -1e-06, c = -1e-06))

Answer 1

我刚遇到同样的问题 - 我认为这个问题确实应该在源代码级别修复！

作为一种解决方法，您或许可以尝试自己构建nlsList对象，这与

相似。

library(nlme)
DF1=data.frame(treatment = rep(c("mineral","residues"),4),
               N_level = c(0,0,100,100,200,200,300,300),
               yield = c(8,8.5,10,10.5,11,9.8,9.5,9.7))
nlslist=lapply(unique(DF1$treatment),function(i) {datasubs=DF1[DF1$treatment==i,];
                                                             nls(yield ~ a + b*0.99^N_level +c*N_level, 
                                                                 data=datasubs, 
                                                                 start = list(a = 12, b = -8, c= -0.01), 
                                                                 upper = list(a=1000, b=-0.000001, c=-0.000001), 
                                                                 algorithm="port", 
                                                                 control=list(maxit=100000,tol=1e-10,warnOnly=T,minFactor=1e-10) )
                                                  })
names(nlslist)=unique(DF1$treatment)
attr(nlslist, "dims")=list(N = nrow(DF1), M = length(nlslist))
attr(nlslist, "call")=NA # this line is not correct - should be fixed
attr(nlslist,"groups")=names(nlslist)
attr(nlslist, "origOrder")=1:length(unique(DF1$treatment))
attr(nlslist, "pool")=TRUE
attr(nlslist, "groupsForm")=formula(~treatment)
class(nlslist)=c("nlsList", "lmList")

这几乎让我在那里，除了我不知道如何正确填写"call"广告位（在nlsList中使用match.call()构建 - 任何知道如何什么机会这样做？

如果您想检查正确的结构，可以通过查看例如在

test=nlsList(uptake ~ SSasympOff(conc, Asym, lrc, c0),
               data = CO2, start = c(Asym = 30, lrc = -4.5, c0 = 52))
class(test)=NULL
test

Answer 2

调整Tom建议的解决方法，我现在使用以下代码：

DF2 <- lapply(unique(DF1$treatment),function(i) {datasubs=DF1[DF1$treatment==i,];
                                              coef(nls(yield ~ a + b*0.99^N_level +c*N_level, 
                                                  data=datasubs, 
                                                  start = list(a = 12, b = -8, c= -0.01), 
                                                  upper = list(a=1000, b=-0.000001, c=-0.000001), 
                                                  algorithm="port", 
                                                  control=list(maxit=100000,tol=1e-10,warnOnly=T,minFactor=1e-10)))
})

DF2 <- data.frame(DF2)
names(DF2) <- levels(DF1$treatment)
DF2 <- t(DF2)
DF2

这给出了拟合系数：

                a         b           c
mineral  13.78825 -5.868506 -0.01260393
residues 12.76157 -4.201832 -0.01006236

目前对我来说已经足够好了，对于更大的数据集也是如此。

R如何在nlslist中使用边界和“端口”算法？

2 个答案: