I'm playing with some Monte Carlo simulations to get an idea of the properties of some linear and non-linear models. The linear OLS model in my case is specified as:

$Y_t = \beta_0 + \beta_1x+ \epsilon,$

with $\beta_0=\beta_1=0.5$

My point of interest here is to figure out if the estimates $\hat{\beta_0}$ and $\hat{\beta_1}$ are unbiased and roughly equal to the true parameters $\beta_0$ and $\beta_1$.

Below, you can see my R code for doing a Monte Carlo simulation with both the linear OLS model and a non-linear quasibinomial probit GLM version of it.

#Monte Carlo simulation

N <- 100 #number of observations
mean <- 0 #mean
sd <- 1 #standard deviation
x <- rnorm(N,mean = mean,sd=sd) #simulate random exogeneous variables

n_sim <- 1000 #number of simulations

b0 <- 0.5 #true parameter beta 0
b1 <- 0.5 #true parameter beta 1

b0_lin <- numeric(n_sim) #stores the estimates of b0
b1_lin <- numeric(n_sim) #stores the estimates of b1
b0_probit <- numeric(n_sim) #stores the estimates of b0
b1_probit <- numeric(n_sim) #stores the estimates of b1

#apply the Monte Carlo simulation
for (i in 1:n_sim){
  #simulate dependent variable
  y_lin <- b0+b1*x+rnorm(N, mean=mean, sd=sd) 
  #linear model
  mod <- lm(y_lin~x)
  #save vector of estimated parameters
  b0_lin[i] <- coef(mod)[[1]]
  b1_lin[i] <- coef(mod)[[2]]
  #transform dependend variable with inverse probit link funktion, thus it lies between zero and 1
  y_prob <- pnorm(y_lin)
  #quasibinomial probit GLM model
  mod <- glm(y_prob~x, family = quasibinomial("probit"))
  #save vector of estimated parameters
  b0_probit[i] <- coef(mod)[[1]]
  b1_probit[i] <- coef(mod)[[2]]

#Save means of the parameter vectors
b0_lin_mean <- mean(b0_lin)
b1_lin_mean <- mean(b1_lin)

b0_probit_mean <- mean(b0_probit)
b1_probit_mean <- mean(b1_probit)

Montecarlo_comparison <- data.frame("b0_mean"=c("True parameters"=b0,

As you can see below, the estimator from a linear OLS model is not biased. However, the quasibinomial probit GLM estimator appears to be severely biased.

enter image description here

My questions: Why is the quasibinomial probit GLM estimator biased in my Monte Carlo simulation? Is the process of my Monte Carlo simulation technically correct or did I maybe make a mistake in the model specifications? If I did it correct, what is the explanation for the fact that the quasibinomial probit GLM estimator is biased here?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.