R 使用optimx软件包的最大似然估计误差

我试图使用R中使用optimx包的收入插补的最大似然法估计参数。我定义了函数，并提供了包含23个元素的初始参数向量。在23个向量中，第21个向量的值被约束为1

经过几次迭代后，它说收敛已经达到，当我试图找出参数值时，它给了我所有参数的“NA”。我的代码是

cat("\014")
data = read.csv("DRCOG.csv")
drcog = data    # please make a column in the dataset indicating inc grp. Named as data$grp
param = c(1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1)  #initial value of the parameter  


imput = function(param) {              #defining funtion   
  rn = nrow(drcog)
  nc = ncol(drcog)
  inc_thres = matrix(c(-100, 1.5, 2.0, 3.0, 4.0, 5.0, 6.0, 7.50, 10.0, 13.5, 15.0, 100.0 ),ncol = 12) 
  inc_grp = ncol(inc_thres) - 1
  tinc_thres = t(inc_thres)

  #creating matrix for reportage and group variable
  # equation for reporting income

  reportage = as.matrix(cbind(drcog$uno,drcog$DUGRAD, drcog$AGE5P, drcog$sero, drcog$sero, drcog$sero, drcog$sero, drcog$sero, drcog$sero, drcog$sero))
  nparmr = param[1:10] #number of input variables, change in initial parameter value and equation accordingly

  grp = as.matrix(cbind(drcog$sero, drcog$sero, drcog$sero, drcog$uno, drcog$NVEH, drcog$MALEP, drcog$RENTED, drcog$EMPLP, drcog$SINGUNI, drcog$NWHITE)) 
  nparmg = param[11:21] #number of input variables in grp variable, change in initial parameter value and equation accordingly
  treportage = t(reportage)
  tgrp = t(grp)
  btheq = cbind(treportage, tgrp)
  cncol = ncol(reportage)


  gammar = param[1:10]             #parameter vector of regression
  gammar = as.matrix(gammar)
  gammal = param[11:20]            #parameter vector of group
  gammal = as.matrix(gammal)

  #choleskey matrix as input parameter, variance-covariance matrix in the formula 
  omega_cvar = param[21:23]
  omegaL = matrix(data = 0, nrow = 2, ncol =2)
  omegaL[1, 1] = omega_cvar[1]
  omegaL[2, 1] = omega_cvar[2]
  omegaL[2, 2] = omega_cvar[3]
  omega = omegaL%*%t(omegaL)

  x <- matrix(data=0,nrow =rn,ncol=14) #first col is reporting/missing, 2-12 is probability for each individual and each group 
  library(pbivnorm)        
  x[, 1] = matrix(drcog$grp)

  # threshold matrix calculation 

  th <- matrix(data=0,nrow =rn,ncol=2) 
  for (i in 1:rn){
    curr_cat = x[i, 1]
    if (curr_cat>0) {
      th[i, 1] = inc_thres[1, curr_cat] 
      th[i, 2] = inc_thres[1, curr_cat+1]
    }
  }

  y11 =  reportage%*%gammar  
  y22 =  (grp%*%gammal)
  cor1 = -1* cov2cor(omega)
  LL = matrix(data = 0, nrow = rn, ncol = 1)

  upper_lim = 1e-5
  for (i in 1:rn) {
    curr_cat = x[i, 1]
    if (curr_cat>0) {
      th_low = (th[i, 1] - y22[i, ])/omega[2,2]
      th_up = (th[i, 2] - y22[i, ])/omega[2,2]

      nnn = (pbivnorm(th_up, y11[i, 1], rho = cor1[1, 2]) - pbivnorm(th_low, y11[i, 1], rho = cor1[1, 2]))
      if (nnn > upper_lim) {
        LL[i, 1] = log(nnn)
      }
      else {
        LL[i, 1] = log(upper_lim)
      }

    }
    else {
      nnn = (1 - pnorm(y11[i, 1]))
      if (nnn > upper_lim){
        LL[i, 1] = log(nnn)
      }
      else {
        LL[i, 1] = log(upper_lim)
      }
    }
  }
  return(sum(LL))
}

grad.norm <- function(imput) {
  require(numDeriv)
  return(sum(grad(imput, param)^2))
}


#mle estimation
#install package stats4
library(stats4)
set.seed(123)


#fit1 = optim(par = param, fn = imput, gr = NULL, lower = c(-Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, 1, -Inf, -Inf), upper = c(Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, 1.01, Inf, Inf), method = "L-BFGS-B", control = list(trace = 1, REPORT =1))

library(optimx)
fit1 = optimx(param, imput, gr = NULL, lower = c(-Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, -Inf, 1, -Inf, -Inf), upper = c(Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, Inf, 1.01, Inf, Inf), method = "L-BFGS-B", control = list(trace = 1, REPORT =1))

---

我尝试过其他优化包，比如maxLik、maxBFGS、optim。出于某种原因，那里也没有趋同。如果您需要用于编译的数据集，请告诉我

不知道这是如何在8个月内没有得到回答的。很明显，你的目标函数中有一个错误，它给出了“if（nnn>upper_lim）{：需要真/假的地方缺少值”。你修正了吗

Analytic gradient not made available.
Analytic Hessian not made available.
Scale check -- log parameter ratio= 0   log bounds ratio= 0 
Method:  L-BFGS-B 
iter    1 value -63967.987853
iter    2 value -66454.130893
iter    3 value -66870.500976
iter    4 value -68290.490185
iter    5 value -71406.527209
iter    6 value -71569.074462
Error in if (nnn > upper_lim) { : missing value where TRUE/FALSE needed
optim function evaluation failure
Post processing for method  L-BFGS-B 
Save results from method  L-BFGS-B 
$fevals
[1] NA

$convcode
[1] 9999

$value
[1] 8.988466e+307

$par
[1] NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA NA