Revisions to Using generalized method of moments (GMM) to calculate logistic regression parameter

Bounty Ended with 100 reputation awarded by user20650

occurred Jul 26, 2016 at 8:41

Adjust relative tolerance termination criteria so both procedures finish with convergence code 0

Source Link

edited Jul 22, 2016 at 4:17

4.4k
29
44

init.coef <- lm(IsVersicolor~., data=as.data.frame(dat[,-2]))$coefficients
library(gmm)
fitted <- gmm(moments, x = dat, t0 = init.coef, type = "iterative", crit = 1e-2519,
              wmatrix = "optimal", method = "Nelder-Mead",
              control = list(reltol = 1e-2519, maxit = 20000))
fitted
#  (Intercept)  Sepal.Length   Sepal.Width  Petal.Length   Petal.Width  
#      7.37849      -0.24536      -2.79657       1.31364      -2.77834  
# 
# Convergence code =  0

gmm.objective <- function(theta, x, momentFun) {
  avg.moment <- colMeans(momentFun(theta, x))
  sum(avg.moment^2)
}
optim(init.coef, gmm.objective, x=dat, momentFun=moments,
      control = list(reltol = 1e-2519, maxit = 20000))$par
#  (Intercept) Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
#    7.3784866   -0.2453567   -2.7965681    1.3136433   -2.7783439

init.coef <- lm(IsVersicolor~., data=as.data.frame(dat[,-2]))$coefficients
library(gmm)
fitted <- gmm(moments, x = dat, t0 = init.coef, type = "iterative", crit = 1e-25,
              wmatrix = "optimal", method = "Nelder-Mead",
              control = list(reltol = 1e-25, maxit = 20000))
fitted
#  (Intercept)  Sepal.Length   Sepal.Width  Petal.Length   Petal.Width  
#      7.37849      -0.24536      -2.79657       1.31364      -2.77834  
# 
# Convergence code =  0

gmm.objective <- function(theta, x, momentFun) {
  avg.moment <- colMeans(momentFun(theta, x))
  sum(avg.moment^2)
}
optim(init.coef, gmm.objective, x=dat, momentFun=moments,
      control = list(reltol = 1e-25, maxit = 20000))$par
#  (Intercept) Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
#    7.3784866   -0.2453567   -2.7965681    1.3136433   -2.7783439

init.coef <- lm(IsVersicolor~., data=as.data.frame(dat[,-2]))$coefficients
library(gmm)
fitted <- gmm(moments, x = dat, t0 = init.coef, type = "iterative", crit = 1e-19,
              wmatrix = "optimal", method = "Nelder-Mead",
              control = list(reltol = 1e-19, maxit = 20000))
fitted
#  (Intercept)  Sepal.Length   Sepal.Width  Petal.Length   Petal.Width  
#      7.37849      -0.24536      -2.79657       1.31364      -2.77834  
# 
# Convergence code =  0

gmm.objective <- function(theta, x, momentFun) {
  avg.moment <- colMeans(momentFun(theta, x))
  sum(avg.moment^2)
}
optim(init.coef, gmm.objective, x=dat, momentFun=moments,
      control = list(reltol = 1e-19, maxit = 20000))$par
#  (Intercept) Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
#    7.3784866   -0.2453567   -2.7965681    1.3136433   -2.7783439

added 973 characters in body

Source Link

edited Jul 20, 2016 at 19:19

josliber

4.4k
29
44

$$ \frac{1}{N}\sum_{i=1}^N \bigg[\bigg(Y_i-\frac{A}{1+e^{-X_i'b}}\bigg)X_i\bigg] = 0 $$$$ m = \frac{1}{N}\sum_{i=1}^N \bigg[\bigg(Y_i-\frac{A}{1+e^{-X_i'b}}\bigg)X_i\bigg] = 0 $$

This is practically solved by minimizing $m'm$ across all possible coefficient values $b$ (below we will use the Nelder-Mead simplex to perform this optimization).

The convergence code of 0 indicates the procedure converged, and the parameters are identical to those returned by logistic regression.

A quick look at the gmm package source (functions momentEstim.baseGmm.iterative and gmm:::.obj1 for the parameters provided) shows that the gmm package is minimizing $m'm$ as indicated above. The following equivalent code calls the R optim function directly, performing the same optimization we achieved above with the call to gmm:

gmm.objective <- function(theta, x, momentFun) {
  avg.moment <- colMeans(momentFun(theta, x))
  sum(avg.moment^2)
}
optim(init.coef, gmm.objective, x=dat, momentFun=moments,
      control = list(reltol = 1e-25, maxit = 20000))$par
#  (Intercept) Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
#    7.3784866   -0.2453567   -2.7965681    1.3136433   -2.7783439

$$ \frac{1}{N}\sum_{i=1}^N \bigg[\bigg(Y_i-\frac{A}{1+e^{-X_i'b}}\bigg)X_i\bigg] = 0 $$

The convergence code of 0 indicates the procedure converged, and the parameters are identical to those returned by logistic regression.

$$ m = \frac{1}{N}\sum_{i=1}^N \bigg[\bigg(Y_i-\frac{A}{1+e^{-X_i'b}}\bigg)X_i\bigg] = 0 $$

This is practically solved by minimizing $m'm$ across all possible coefficient values $b$ (below we will use the Nelder-Mead simplex to perform this optimization).

The convergence code of 0 indicates the procedure converged, and the parameters are identical to those returned by logistic regression.

A quick look at the gmm package source (functions momentEstim.baseGmm.iterative and gmm:::.obj1 for the parameters provided) shows that the gmm package is minimizing $m'm$ as indicated above. The following equivalent code calls the R optim function directly, performing the same optimization we achieved above with the call to gmm:

gmm.objective <- function(theta, x, momentFun) {
  avg.moment <- colMeans(momentFun(theta, x))
  sum(avg.moment^2)
}
optim(init.coef, gmm.objective, x=dat, momentFun=moments,
      control = list(reltol = 1e-25, maxit = 20000))$par
#  (Intercept) Sepal.Length  Sepal.Width Petal.Length  Petal.Width 
#    7.3784866   -0.2453567   -2.7965681    1.3136433   -2.7783439

added 132 characters in body

Source Link

edited Jul 20, 2016 at 17:57

josliber

4.4k
29
44

dat <- as.matrix(cbind(data.frame(IsVersicolor = as.numeric(iris$Species == "versicolor"), Intercept=1), iris[,1:4]))
head(dat)
#      IsVersicolor Intercept Sepal.Length Sepal.Width Petal.Length Petal.Width
# [1,]            0         1          5.1         3.5          1.4         0.2
# [2,]            0         1          4.9         3.0          1.4         0.2
# [3,]            0         1          4.7         3.2          1.3         0.2
# [4,]            0         1          4.6         3.1          1.5         0.2
# [5,]            0         1          5.0         3.6          1.4         0.2
# [6,]            0         1          5.4         3.9          1.7         0.4

dat <- as.matrix(cbind(data.frame(IsVersicolor = as.numeric(iris$Species == "versicolor"), Intercept=1), iris[,1:4]))
head(dat)
#      IsVersicolor Intercept Sepal.Length Sepal.Width Petal.Length Petal.Width
# [1,]            0         1          5.1         3.5          1.4         0.2
# [2,]            0         1          4.9         3.0          1.4         0.2
# [3,]            0         1          4.7         3.2          1.3         0.2
# [4,]            0         1          4.6         3.1          1.5         0.2
# [5,]            0         1          5.0         3.6          1.4         0.2
# [6,]            0         1          5.4         3.9          1.7         0.4

summary(glm(IsVersicolor~., data=as.data.frame(dat[,-2]), family="binomial"))
# Coefficients:
#              Estimate Std. Error z value Pr(>|z|)    
# (Intercept)    7.3785     2.4993   2.952 0.003155 ** 
# Sepal.Length  -0.2454     0.6496  -0.378 0.705634    
# Sepal.Width   -2.7966     0.7835  -3.569 0.000358 ***
# Petal.Length   1.3136     0.6838   1.921 0.054713 .  
# Petal.Width   -2.7783     1.1731  -2.368 0.017868 *

summary(glm(IsVersicolor~., data=as.data.frame(dat[,-2]), family="binomial"))
# Coefficients:
#              Estimate Std. Error z value Pr(>|z|)    
# (Intercept)    7.3785     2.4993   2.952 0.003155 ** 
# Sepal.Length  -0.2454     0.6496  -0.378 0.705634    
# Sepal.Width   -2.7966     0.7835  -3.569 0.000358 ***
# Petal.Length   1.3136     0.6838   1.921 0.054713 .  
# Petal.Width   -2.7783     1.1731  -2.368 0.017868 *

moments <- function(b, X) {
  A <- 1
  as.vector(X[,1] - A / (1 + exp(-(X[,-1] %*% cbind(b))))) * X[,-1]
}

moments <- function(b, X) {
  A <- 1
  as.vector(X[,1] - A / (1 + exp(-(X[,-1] %*% cbind(b))))) * X[,-1]
}

init.coef <- lm(IsVersicolor~., data=as.data.frame(dat[,-2]))$coefficients
library(gmm)
fitted <- gmm(moments, x = dat, t0 = init.coef,
              type = "iterative", crit = 1e-25,wmatrix = "optimal", method = "Nelder-Mead",
              control = list(reltol = 1e-25, maxit = 20000))
fitted
#  (Intercept)  Sepal.Length   Sepal.Width  Petal.Length   Petal.Width  
#      7.37849      -0.24536      -2.79657       1.31364      -2.77834  
# 
# Convergence code =  0

init.coef <- lm(IsVersicolor~., data=as.data.frame(dat[,-2]))$coefficients
library(gmm)
fitted <- gmm(moments, x = dat, t0 = init.coef, type = "iterative", crit = 1e-25,
              wmatrix = "optimal", method = "Nelder-Mead",
              control = list(reltol = 1e-25, maxit = 20000))
fitted
#  (Intercept)  Sepal.Length   Sepal.Width  Petal.Length   Petal.Width  
#      7.37849      -0.24536      -2.79657       1.31364      -2.77834  
# 
# Convergence code =  0

dat <- as.matrix(cbind(data.frame(IsVersicolor = as.numeric(iris$Species == "versicolor"), Intercept=1), iris[,1:4]))
head(dat)
#      IsVersicolor Intercept Sepal.Length Sepal.Width Petal.Length Petal.Width
# [1,]            0         1          5.1         3.5          1.4         0.2
# [2,]            0         1          4.9         3.0          1.4         0.2
# [3,]            0         1          4.7         3.2          1.3         0.2
# [4,]            0         1          4.6         3.1          1.5         0.2
# [5,]            0         1          5.0         3.6          1.4         0.2
# [6,]            0         1          5.4         3.9          1.7         0.4

summary(glm(IsVersicolor~., data=as.data.frame(dat[,-2]), family="binomial"))
# Coefficients:
#              Estimate Std. Error z value Pr(>|z|)    
# (Intercept)    7.3785     2.4993   2.952 0.003155 ** 
# Sepal.Length  -0.2454     0.6496  -0.378 0.705634    
# Sepal.Width   -2.7966     0.7835  -3.569 0.000358 ***
# Petal.Length   1.3136     0.6838   1.921 0.054713 .  
# Petal.Width   -2.7783     1.1731  -2.368 0.017868 *

moments <- function(b, X) {
  A <- 1
  as.vector(X[,1] - A / (1 + exp(-(X[,-1] %*% cbind(b))))) * X[,-1]
}

init.coef <- lm(IsVersicolor~., data=as.data.frame(dat[,-2]))$coefficients
library(gmm)
fitted <- gmm(moments, x = dat, t0 = init.coef,
              type = "iterative", crit = 1e-25,wmatrix = "optimal", method = "Nelder-Mead",
              control = list(reltol = 1e-25, maxit = 20000))
fitted
#  (Intercept)  Sepal.Length   Sepal.Width  Petal.Length   Petal.Width  
#      7.37849      -0.24536      -2.79657       1.31364      -2.77834  
# 
# Convergence code =  0

dat <- as.matrix(cbind(data.frame(IsVersicolor = as.numeric(iris$Species == "versicolor"), Intercept=1), iris[,1:4]))
head(dat)
#      IsVersicolor Intercept Sepal.Length Sepal.Width Petal.Length Petal.Width
# [1,]            0         1          5.1         3.5          1.4         0.2
# [2,]            0         1          4.9         3.0          1.4         0.2
# [3,]            0         1          4.7         3.2          1.3         0.2
# [4,]            0         1          4.6         3.1          1.5         0.2
# [5,]            0         1          5.0         3.6          1.4         0.2
# [6,]            0         1          5.4         3.9          1.7         0.4

summary(glm(IsVersicolor~., data=as.data.frame(dat[,-2]), family="binomial"))
# Coefficients:
#              Estimate Std. Error z value Pr(>|z|)    
# (Intercept)    7.3785     2.4993   2.952 0.003155 ** 
# Sepal.Length  -0.2454     0.6496  -0.378 0.705634    
# Sepal.Width   -2.7966     0.7835  -3.569 0.000358 ***
# Petal.Length   1.3136     0.6838   1.921 0.054713 .  
# Petal.Width   -2.7783     1.1731  -2.368 0.017868 *

moments <- function(b, X) {
  A <- 1
  as.vector(X[,1] - A / (1 + exp(-(X[,-1] %*% cbind(b))))) * X[,-1]
}

init.coef <- lm(IsVersicolor~., data=as.data.frame(dat[,-2]))$coefficients
library(gmm)
fitted <- gmm(moments, x = dat, t0 = init.coef, type = "iterative", crit = 1e-25,
              wmatrix = "optimal", method = "Nelder-Mead",
              control = list(reltol = 1e-25, maxit = 20000))
fitted
#  (Intercept)  Sepal.Length   Sepal.Width  Petal.Length   Petal.Width  
#      7.37849      -0.24536      -2.79657       1.31364      -2.77834  
# 
# Convergence code =  0

Source Link

created Jul 20, 2016 at 17:47

josliber

4.4k
29
44

Loading

Stack Exchange Network

Return to Answer