I discovered the mlogit-package for multinomial logit models in search of estimating a multinomial mixed logit model. After reading the excellent vignette I discovered that I could not apply my data on any of the described examples.
我发现了多项logit模型的mlogit-package，用于估计多项混合的logit模型。在阅读了优秀的vignette之后，我发现我无法将我的数据应用到任何描述的示例中。
I now write in hope of help with my problem and created a minimal example to illustrate my situation.
现在我希望能帮助解决我的问题，并创建了一个最小的例子来说明我的情况。
The Problem is as follows: There are words with the consonant 'Q' somewhere. Now an experiment was conducted with people who were tasked to listen to these words and say if they heard a Q, an U or some OTHER consonant. This has to modeled in dependence of some factors like syllable position or real/non-real-word.
问题是这样的:在某个地方有辅音“Q”的单词。现在有一项实验是对那些被要求听这些单词的人进行的，他们会说，如果他们听到了一个Q，一个U或者其他的辅音。这必须依赖于一些因素，如音节位置或真实/非实词。
In the minimal example I created 4 people and their answers with the syllable position.
在最小的例子中，我用音节的位置创造了4个人和他们的答案。
library(mlogit)
library(nnet)
set.seed(1234)
data <- data.frame(personID = as.factor(sample(1:4, 40, replace=TRUE)),
decision = as.factor(sample(c("Q","U", "other"), 40, replace=TRUE)),
syllable = as.factor(sample(1:4, 40, replace=TRUE)))
summary(data)
personID decision syllable
1:11 other:10 1:18
2:10 Q :18 2: 9
3:10 U :12 3: 5
4: 9 4: 8
As far as I know nnet's multinomfunction does not cover mixed models.
据我所知，nnet的多项功能不包括混合模型。
modNnet1 <- multinom(decision ~ syllable, data=data)
First I used the mlogit.data-function to reshape the file. After discussion with a colleague we came to the conclusion that there is no alternative.specific.variable.
首先，我使用了mlogit。数据功能来重塑文件。在和同事讨论之后，我们得出结论:没有其他选择。
dataMod <- mlogit.data(data, shape="wide", choice="decision", id.var="personID")
mod1 <- mlogit(formula = decision ~ 0|syllable,
data = dataMod,
reflevel="Q", rpar=c(personID="n"), panel=TRUE)
Error in names(sup.coef) <- names.sup.coef :
'names' attribute [1] must be the same length as the vector [0]
mod2 <- mlogit(formula = decision ~ personID|syllable,
data = dataMod,
reflevel="Q", rpar=c(personID="n"), panel=TRUE)
Error in solve.default(H, g[!fixed]) :
Lapack routine dgesv: system is exactly singular: U[3,3] = 0
No I do not know what to do, so I ask for help in here. But I believe this kind of problem can be solved with mlogit and I just don't see it yet ;)
不，我不知道该怎么做，所以我在这里寻求帮助。但是我相信这类问题可以用mlogit解决，而我只是还没有看到。
1 个解决方案
#1
3
The rpar argument accepts only alternative-specific variables. There is no need to specify the person-specific id in the model formula -- this is handled by including id.var = something in the mlogit.data command. For example, if you had an alternative specific covariate acov, you could allow random slopes for acov across a panel:
rpar参数只接受可替换的特定变量。没有必要在模型公式中指定特定于人的id——这是通过包含id.var =在mlogit中的某个东西来处理的。数据的命令。例如，如果您有一个可选的特定covariate acov，您可以在一个面板上允许acov的随机斜坡:
N = 200
dat <- data.frame(personID = as.factor(sample(1:4, N, replace=TRUE)),
decision = as.factor(sample(c("Q","U", "other"), N, replace=TRUE)),
syllable = as.factor(sample(1:4, N, replace=TRUE)),
acov.Q = rnorm(N), acov.U = rnorm(N), acov.other = rnorm(N))
dataMod <- mlogit.data(dat, shape="wide", choice="decision", id.var="personID", varying = 4:6)
mlogit(formula = decision ~ acov|syllable, rpar = c(acov = "n"), panel = T, data = dataMod)
It seems you are trying to fit a model with a random, person-specific intercept for each alternative (not random slopes). Unfortunately, I don't think you can do so in mlogit (but see this post).
看起来你是在尝试用一个随机的、人为的截距来匹配一个模型(不是随机的斜率)。不幸的是，我不认为您可以在mlogit中这样做(但是请参阅本文)。
One option that would work to fit random intercepts in the absence of alternative-specific covariates is MCMCglmm.
MCMCglmm是一种选择，可以在没有替代特定的协变量的情况下进行随机截取。
library(MCMCglmm)
priors = list(R = list(fix = 1, V = 0.5 * diag(2), n = 2),
G = list(G1 = list(V = diag(2), n = 2)))
m <- MCMCglmm(decision ~ -1 + trait + syllable,
random = ~ idh(trait):personID,
rcov = ~ us(trait):units,
prior = priors,
nitt = 30000, thin = 20, burnin = 10000,
family = "categorical",
data = dat)
Relevant issues are prior selection, convergence of Markov chains, etc. Florian Jaeger's lab's blog has a short tutorial on multinomial models via MCMCglmm that you might find helpful, in addition to the MCMCglmm documentation.
相关的问题是之前的选择，Markov链的聚合，等等。Florian Jaeger的实验室的博客上有一个关于MCMCglmm的多项式模型的简短教程，除了MCMCglmm文档之外，你可能会发现它是有用的。
#1
3
The rpar argument accepts only alternative-specific variables. There is no need to specify the person-specific id in the model formula -- this is handled by including id.var = something in the mlogit.data command. For example, if you had an alternative specific covariate acov, you could allow random slopes for acov across a panel:
rpar参数只接受可替换的特定变量。没有必要在模型公式中指定特定于人的id——这是通过包含id.var =在mlogit中的某个东西来处理的。数据的命令。例如，如果您有一个可选的特定covariate acov，您可以在一个面板上允许acov的随机斜坡:
N = 200
dat <- data.frame(personID = as.factor(sample(1:4, N, replace=TRUE)),
decision = as.factor(sample(c("Q","U", "other"), N, replace=TRUE)),
syllable = as.factor(sample(1:4, N, replace=TRUE)),
acov.Q = rnorm(N), acov.U = rnorm(N), acov.other = rnorm(N))
dataMod <- mlogit.data(dat, shape="wide", choice="decision", id.var="personID", varying = 4:6)
mlogit(formula = decision ~ acov|syllable, rpar = c(acov = "n"), panel = T, data = dataMod)
It seems you are trying to fit a model with a random, person-specific intercept for each alternative (not random slopes). Unfortunately, I don't think you can do so in mlogit (but see this post).
看起来你是在尝试用一个随机的、人为的截距来匹配一个模型(不是随机的斜率)。不幸的是，我不认为您可以在mlogit中这样做(但是请参阅本文)。
One option that would work to fit random intercepts in the absence of alternative-specific covariates is MCMCglmm.
MCMCglmm是一种选择，可以在没有替代特定的协变量的情况下进行随机截取。
library(MCMCglmm)
priors = list(R = list(fix = 1, V = 0.5 * diag(2), n = 2),
G = list(G1 = list(V = diag(2), n = 2)))
m <- MCMCglmm(decision ~ -1 + trait + syllable,
random = ~ idh(trait):personID,
rcov = ~ us(trait):units,
prior = priors,
nitt = 30000, thin = 20, burnin = 10000,
family = "categorical",
data = dat)
Relevant issues are prior selection, convergence of Markov chains, etc. Florian Jaeger's lab's blog has a short tutorial on multinomial models via MCMCglmm that you might find helpful, in addition to the MCMCglmm documentation.
相关的问题是之前的选择，Markov链的聚合，等等。Florian Jaeger的实验室的博客上有一个关于MCMCglmm的多项式模型的简短教程，除了MCMCglmm文档之外，你可能会发现它是有用的。