R help - Nov 2009 - optim or nlminb for minimization, which to believe?

I have constructed the function mml2 (below) based on the likelihood function
described in the minimal latex I have pasted below for anyone who wants to look
at it. This function finds parameter estimates for a basic Rasch (IRT) model.
Using the function without the gradient, using either nlminb or optim returns
the correct parameter estimates and, in the case of optim, the correct standard
errors.

By correct, I mean they match another software program as well as the rasch()
function in the ltm package.

Now, when I pass the gradient to optim, I get a message of successful
convergence, but the parameter estimates are not correct, but they are *very*
close to being correct. But, when I use nlminb with the gradient, I get a
message of false convergence and again, the estimates are off, but again very
close to being correct.

This is best illustrated via the examples:

### Sample data set
set.seed(1234)
tmp <- data.frame(item1 = sample(c(0,1), 20, replace=TRUE), item2 =
sample(c(0,1), 20, replace=TRUE), item3 = sample(c(0,1), 20, replace=TRUE),item4
= sample(c(0,1), 20, replace=TRUE),item5 = sample(c(0,1), 20, replace=TRUE))

## Use function mml2 (below) with optim  with use of
gradient
> mml2(tmp,Q=10)
$par
[1] -0.2438733  0.4889333 -0.2438733  0.4889333  0.7464162

$value
[1] 63.86376

$counts
function gradient
      45        6

$convergence
[1] 0

$message
NULL

$hessian
         [,1]     [,2]     [,3]     [,4]     [,5]
[1,] 4.095479 0.000000 0.000000 0.000000 0.000000
[2,] 0.000000 3.986293 0.000000 0.000000 0.000000
[3,] 0.000000 0.000000 4.095479 0.000000 0.000000
[4,] 0.000000 0.000000 0.000000 3.986293 0.000000
[5,] 0.000000 0.000000 0.000000 0.000000 3.800898

## Use same function but use nlminb with use of gradient
> mml2(tmp,Q=10)
$par
[1] -0.2456398  0.4948889 -0.2456398  0.4948889  0.7516308

$objective
[1] 63.86364

$convergence
[1] 1

$message
[1] "false convergence (8)"

$iterations
[1] 4

$evaluations
function gradient
      41        4


### use nlminb but turn off use of gradient
> mml2(tmp,Q=10)
$par
[1] -0.2517961  0.4898682 -0.2517961  0.4898682  0.7500994

$objective
[1] 63.8635

$convergence
[1] 0

$message
[1] "relative convergence (4)"

$iterations
[1] 8

$evaluations
function gradient
      11       64

### Use optim and turn off gradient
> mml2(tmp,Q=10)
$par
[1] -0.2517990  0.4898676 -0.2517990  0.4898676  0.7500906

$value
[1] 63.8635

$counts
function gradient
      22        7

$convergence
[1] 0

$message
NULL

$hessian
           [,1]       [,2]       [,3]       [,4]       [,5]
[1,]  3.6311153 -0.3992959 -0.4224747 -0.3992959 -0.3764526
[2,] -0.3992959  3.5338195 -0.3992959 -0.3960956 -0.3798141
[3,] -0.4224747 -0.3992959  3.6311153 -0.3992959 -0.3764526
[4,] -0.3992959 -0.3960956 -0.3992959  3.5338195 -0.3798141
[5,] -0.3764526 -0.3798141 -0.3764526 -0.3798141  3.3784816

The parameter estimates with and without the use of the gradient are so close
that I am inclined to believe that the gradient is correct and maybe the problem
is elsewhere.

It seems odd that optim seems to converge but nlminb does not with the use of
the gradient. But, with the use of the gradient in either case, the parameter
estimates differ from what I think are the correct values. So, at this point I
am unclear if the problem is somewhere in the way the functions are used or how
I am passing the gradient or if the problem lies in the way I have constructed
the gradient itself.

Below is the function and also some latex for those interested in looking at the
likelihood function.

Thanks for any reactions
Harold
> sessionInfo()
R version 2.10.0 (2009-10-26)
i386-pc-mingw32

locale:
[1] LC_COLLATE=English_United States.1252  LC_CTYPE=English_United States.1252
[3] LC_MONETARY=English_United States.1252 LC_NUMERIC=C
[5] LC_TIME=English_United States.1252

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base

other attached packages:
[1] ltm_0.9-2      polycor_0.7-7  sfsmisc_1.0-9  mvtnorm_0.9-8  msm_0.9.4     
MASS_7.3-3     MiscPsycho_1.5
[8] statmod_1.4.1

loaded via a namespace (and not attached):
[1] splines_2.10.0  survival_2.35-7 tools_2.10.0


mml2 <- function(data, Q, startVal = NULL, gr = TRUE, ...){
                if(!is.null(startVal) && length(startVal) != ncol(data)
){
                                stop("Length of argument startVal not equal
to the number of parameters estimated")
                }
                if(!is.null(startVal)){
                                startVal <- startVal
                                } else {
                                p <- colMeans(data)
                                startVal <- as.vector(log((1 - p)/p))
                }
                qq <- gauss.quad.prob(Q, dist = 'normal')
                rr1 <- matrix(0, nrow = Q, ncol = nrow(data))
                data <- as.matrix(data)
                L <- nrow(data)
                C <- ncol(data)
                fn <- function(b){
                                for(j in 1:Q){
                                                for(i in 1:nrow(data)){
                                                                rr1[j,i] <-
exp(sum(dbinom(data[i,], 1, plogis(qq$nodes[j]-b), log = TRUE))) *
                                                                qq$weights[j]
                                                }
                                }
                -sum(log(colSums(rr1)))
                }
                gradient <- function(b){
                                p <- outer(qq$nodes, b, plogis) * L
                                x <- t(matrix(colSums(data), nrow= C, Q))
                                w <- matrix(qq$weights, Q, C)
                                rr2 <- (-x + p) * w
                                -colSums(rr2)
                }
                opt <- optim(startVal, fn, gr = gradient, hessian = TRUE,
method = "BFGS")
                #opt <- nlminb(startVal, fn, gradient=gradient)
                #list("coefficients" = opt$par, "LogLik" =
-opt$value, "Std.Error" = 1/sqrt(diag(opt$hessian)))
                opt
}



### latex describing the likelihood function
\documentclass{article}
\usepackage{latexsym}
\usepackage{amssymb,amsmath,bm}
\newcommand{\Prb}{\operatorname{Prob}}
\title{Marginal Maximum Likelihood for IRT Models}
\author{Harold C. Doran}
\begin{document}
\maketitle
The likelihood function is written as:
\begin{equation}
\label{eqn:mml}
f(x) = \prod^N_{i=1}\int_{\Theta}\prod^K_{j=1}p(x|\theta,\beta)f(\theta)d\theta
\end{equation}
\noindent where $N$ indexes the total number of individuals, $K$ indexes the
total number of items, $p(x|\theta,\beta)$ is the data likelihood and
$f(\theta)$ is a population distribution. For the rasch model, the data
likelihood is:
\begin{equation}
p(x|\theta,\beta) = \prod^N_{i=1}\prod^K_{j=1} \Pr(x_{ij} = 1 | \theta_i,
\beta_j)^{x_{ij}} \left[1 - \Pr(X_{ij} = 1 | \theta_i,
\beta_j)\right]^{(1-x_{ij})}
\end{equation}
\begin{equation}
\label{rasch}
\Pr(x_{ij} = 1 | \theta_i, \beta_j) = \frac{1}{1 + e^{-(\theta_i-\beta_j)}}
\quad i = (1, \ldots, K); j = (1, \ldots, N)
\end{equation}
\noindent where $\theta_i$ is the ability estimate of person $i$ and $\beta_j$
is the location parameter for item $j$. The integral in Equation~(\ref{eqn:mml})
has no closed form expression, so it is approximated using Gauss-Hermite
quadrature as:
\begin{equation}
\label{eqn:mml:approx}
f(x) \approx \prod^N_{i=1}\sum_{q=1}^{Q}\prod^K_{j=1}p(x|\theta_q,\beta)w_q
\end{equation}
\noindent where $q$ indexes the node at quadrature point $q$ and $w$ is the
weight at quadrature point $q$. With Equation~(\ref{eqn:mml:approx}), the
remaining challenge is to find $\underset{x}{\operatorname{arg\,max}} \, f(x)$.
\end{document}

	[[alternative HTML version deleted]]

Your function named 'gradient' is not the correct gradient. Take as an
example the following point x0, very near to the true minimum,

    x0 <- c(-0.2517964, 0.4898680, -0.2517962, 0.4898681, 0.7500995)
 
then you get

    > gradient(x0)
    [1] -0.0372110470  0.0001816991 -0.0372102284  0.0001820976 
0.0144292657

but the numerical gradient is different:

    > library(numDeriv)
    > grad(fn, x0)
    [1] -6.151645e-07 -5.507219e-07  1.969143e-07 -1.563892e-07
-4.955502e-08

that is the derivative is close to 0 in any direction -- as it should be for
an optimum.

No wonder, optim et al. get confused when applying this 'gradient'.

Regards
Hans Werner


Doran, Harold wrote:
> 
> I have constructed the function mml2 (below) based on the likelihood
> function described in the minimal latex I have pasted below for anyone who
> wants to look at it. This function finds parameter estimates for a basic
> Rasch (IRT) model. Using the function without the gradient, using either
> nlminb or optim returns the correct parameter estimates and, in the case
> of optim, the correct standard errors.
> 
> By correct, I mean they match another software program as well as the
> rasch() function in the ltm package.
> 
> Now, when I pass the gradient to optim, I get a message of successful
> convergence, but the parameter estimates are not correct, but they are
> *very* close to being correct. But, when I use nlminb with the gradient, I
> get a message of false convergence and again, the estimates are off, but
> again very close to being correct.
> 
> This is best illustrated via the examples:
> 
> ### Sample data set
> set.seed(1234)
> tmp <- data.frame(item1 = sample(c(0,1), 20, replace=TRUE), item2 >
sample(c(0,1), 20, replace=TRUE), item3 = sample(c(0,1), 20,
> replace=TRUE),item4 = sample(c(0,1), 20, replace=TRUE),item5 >
sample(c(0,1), 20, replace=TRUE))
> 
> ## Use function mml2 (below) with optim  with use of gradient
>> mml2(tmp,Q=10)
> $par
> [1] -0.2438733  0.4889333 -0.2438733  0.4889333  0.7464162
> 
> $value
> [1] 63.86376
> 
> $counts
> function gradient
>       45        6
> 
> $convergence
> [1] 0
> 
> $message
> NULL
> 
> $hessian
>          [,1]     [,2]     [,3]     [,4]     [,5]
> [1,] 4.095479 0.000000 0.000000 0.000000 0.000000
> [2,] 0.000000 3.986293 0.000000 0.000000 0.000000
> [3,] 0.000000 0.000000 4.095479 0.000000 0.000000
> [4,] 0.000000 0.000000 0.000000 3.986293 0.000000
> [5,] 0.000000 0.000000 0.000000 0.000000 3.800898
> 
> ## Use same function but use nlminb with use of gradient
>> mml2(tmp,Q=10)
> $par
> [1] -0.2456398  0.4948889 -0.2456398  0.4948889  0.7516308
> 
> $objective
> [1] 63.86364
> 
> $convergence
> [1] 1
> 
> $message
> [1] "false convergence (8)"
> 
> $iterations
> [1] 4
> 
> $evaluations
> function gradient
>       41        4
> 
> 
> ### use nlminb but turn off use of gradient
>> mml2(tmp,Q=10)
> $par
> [1] -0.2517961  0.4898682 -0.2517961  0.4898682  0.7500994
> 
> $objective
> [1] 63.8635
> 
> $convergence
> [1] 0
> 
> $message
> [1] "relative convergence (4)"
> 
> $iterations
> [1] 8
> 
> $evaluations
> function gradient
>       11       64
> 
> ### Use optim and turn off gradient
> 
>> mml2(tmp,Q=10)
> $par
> [1] -0.2517990  0.4898676 -0.2517990  0.4898676  0.7500906
> 
> $value
> [1] 63.8635
> 
> $counts
> function gradient
>       22        7
> 
> $convergence
> [1] 0
> 
> $message
> NULL
> 
> $hessian
>            [,1]       [,2]       [,3]       [,4]       [,5]
> [1,]  3.6311153 -0.3992959 -0.4224747 -0.3992959 -0.3764526
> [2,] -0.3992959  3.5338195 -0.3992959 -0.3960956 -0.3798141
> [3,] -0.4224747 -0.3992959  3.6311153 -0.3992959 -0.3764526
> [4,] -0.3992959 -0.3960956 -0.3992959  3.5338195 -0.3798141
> [5,] -0.3764526 -0.3798141 -0.3764526 -0.3798141  3.3784816
> 
> The parameter estimates with and without the use of the gradient are so
> close that I am inclined to believe that the gradient is correct and maybe
> the problem is elsewhere.
> 
> It seems odd that optim seems to converge but nlminb does not with the use
> of the gradient. But, with the use of the gradient in either case, the
> parameter estimates differ from what I think are the correct values. So,
> at this point I am unclear if the problem is somewhere in the way the
> functions are used or how I am passing the gradient or if the problem lies
> in the way I have constructed the gradient itself.
> 
> Below is the function and also some latex for those interested in looking
> at the likelihood function.
> 
> Thanks for any reactions
> Harold
> 
>> sessionInfo()
> R version 2.10.0 (2009-10-26)
> i386-pc-mingw32
> 
> locale:
> [1] LC_COLLATE=English_United States.1252  LC_CTYPE=English_United
> States.1252
> [3] LC_MONETARY=English_United States.1252 LC_NUMERIC=C
> [5] LC_TIME=English_United States.1252
> 
> attached base packages:
> [1] stats     graphics  grDevices utils     datasets  methods   base
> 
> other attached packages:
> [1] ltm_0.9-2      polycor_0.7-7  sfsmisc_1.0-9  mvtnorm_0.9-8  msm_0.9.4
> MASS_7.3-3     MiscPsycho_1.5
> [8] statmod_1.4.1
> 
> loaded via a namespace (and not attached):
> [1] splines_2.10.0  survival_2.35-7 tools_2.10.0
> 
> 
> mml2 <- function(data, Q, startVal = NULL, gr = TRUE, ...){
>                 if(!is.null(startVal) && length(startVal) !=
ncol(data) ){
>                                 stop("Length of argument startVal not
> equal to the number of parameters estimated")
>                 }
>                 if(!is.null(startVal)){
>                                 startVal <- startVal
>                                 } else {
>                                 p <- colMeans(data)
>                                 startVal <- as.vector(log((1 - p)/p))
>                 }
>                 qq <- gauss.quad.prob(Q, dist = 'normal')
>                 rr1 <- matrix(0, nrow = Q, ncol = nrow(data))
>                 data <- as.matrix(data)
>                 L <- nrow(data)
>                 C <- ncol(data)
>                 fn <- function(b){
>                                 for(j in 1:Q){
>                                                 for(i in 1:nrow(data)){
>                                                                 rr1[j,i]
> <- exp(sum(dbinom(data[i,], 1, plogis(qq$nodes[j]-b), log = TRUE))) *
>                                                                
> qq$weights[j]
>                                                 }
>                                 }
>                 -sum(log(colSums(rr1)))
>                 }
>                 gradient <- function(b){
>                                 p <- outer(qq$nodes, b, plogis) * L
>                                 x <- t(matrix(colSums(data), nrow= C,
Q))
>                                 w <- matrix(qq$weights, Q, C)
>                                 rr2 <- (-x + p) * w
>                                 -colSums(rr2)
>                 }
>                 opt <- optim(startVal, fn, gr = gradient, hessian =
TRUE,
> method = "BFGS")
>                 #opt <- nlminb(startVal, fn, gradient=gradient)
>                 #list("coefficients" = opt$par,
"LogLik" = -opt$value,
> "Std.Error" = 1/sqrt(diag(opt$hessian)))
>                 opt
> }
> 
> 
> 
> ### latex describing the likelihood function
> \documentclass{article}
> \usepackage{latexsym}
> \usepackage{amssymb,amsmath,bm}
> \newcommand{\Prb}{\operatorname{Prob}}
> \title{Marginal Maximum Likelihood for IRT Models}
> \author{Harold C. Doran}
> \begin{document}
> \maketitle
> The likelihood function is written as:
> \begin{equation}
> \label{eqn:mml}
> f(x) >
\prod^N_{i=1}\int_{\Theta}\prod^K_{j=1}p(x|\theta,\beta)f(\theta)d\theta
> \end{equation}
> \noindent where $N$ indexes the total number of individuals, $K$ indexes
> the total number of items, $p(x|\theta,\beta)$ is the data likelihood and
> $f(\theta)$ is a population distribution. For the rasch model, the data
> likelihood is:
> \begin{equation}
> p(x|\theta,\beta) = \prod^N_{i=1}\prod^K_{j=1} \Pr(x_{ij} = 1 | \theta_i,
> \beta_j)^{x_{ij}} \left[1 - \Pr(X_{ij} = 1 | \theta_i,
> \beta_j)\right]^{(1-x_{ij})}
> \end{equation}
> \begin{equation}
> \label{rasch}
> \Pr(x_{ij} = 1 | \theta_i, \beta_j) = \frac{1}{1 +
> e^{-(\theta_i-\beta_j)}} \quad i = (1, \ldots, K); j = (1, \ldots, N)
> \end{equation}
> \noindent where $\theta_i$ is the ability estimate of person $i$ and
> $\beta_j$ is the location parameter for item $j$. The integral in
> Equation~(\ref{eqn:mml}) has no closed form expression, so it is
> approximated using Gauss-Hermite quadrature as:
> \begin{equation}
> \label{eqn:mml:approx}
> f(x) \approx
> \prod^N_{i=1}\sum_{q=1}^{Q}\prod^K_{j=1}p(x|\theta_q,\beta)w_q
> \end{equation}
> \noindent where $q$ indexes the node at quadrature point $q$ and $w$ is
> the weight at quadrature point $q$. With Equation~(\ref{eqn:mml:approx}),
> the remaining challenge is to find $\underset{x}{\operatorname{arg\,max}}
> \, f(x)$.
> \end{document}
> 
> 	[[alternative HTML version deleted]]
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
> 
> 
-- 
View this message in context:
http://n4.nabble.com/optim-or-nlminb-for-minimization-which-to-believe-tp930841p930942.html
Sent from the R help mailing list archive at Nabble.com.