Suppose for fixed $t>0$ we have two discrete random variables $X_t$ and $X_{t+1}$ where $X_t$ takes on $11$ values $$ X_t\in\{0,1,2,\cdots, 10\}. $$ Also, we have $$ P(X_{t+1}=X_t)=1-\lambda,\quad P(X_{t+1}=\min\{X_t+1,10\})=\lambda. $$ What is the ML estimator for $\lambda$?

The question is motivated by a model about bus engine replacement in microeconometrics. I have searched along this Wikipedia article and this lecture note. But I don't see how to do it since two random variables are involved. Could anyone come up with a theoretical reference for such an issue (MLE method for a statistical model with two random variables)? (Or this could be trivially reduced to the one-variable case?)

  • 1
    $\begingroup$ Do you just have two random variables or several consisting of the complete histories of more than one bus engine with the values of $X_1, X_2,..., X_t$ ? That would also mean that we would assume that $t$ also takes on the discrete values $1, 2, 3, ..., t$ ? And does $X_1=0$ with probability 1 ? $\endgroup$
    – JimB
    Nov 29 '15 at 5:01
  • $\begingroup$ I'm now focusing on the case that $t$ is fixed. Thus there are only two random variables. $\endgroup$
    – Jack
    Nov 29 '15 at 14:57

This is more of an extended comment than an answer.

I wonder if you meant to give conditional probabilities such that $P(X_{n+1}=x_n | X_n = x_n)=1-\lambda$ and $P(X_{n+1}=\min{(x_t+1,10)}|X_n=x_n)=\lambda$ with $P(X_1=0)=1$.

That way the likelihood of an observed history with $t=6$ of 0 0 1 1 2 3 would be $(1-\lambda)^2 \lambda^3$ with the maximum likelihood estimator being 3/5.

In general with an arbitrary value of $t$ and $m$ occurrences where there was a change from $X_i$ to $X_{i+1}$ the likelihood would be $(1-\lambda)^{t-m-1} \lambda^m$, the maximum likelihood estimator of $\lambda$ would be $m/(t-1)$ (except when $X_t=10$ as then $X_{t+1}$ must also equal 10 resulting in no information about $\lambda$).

But if you only know say two values $X_5$ and $X_6$, then you only have a sample size of 1 and the maximum likelihood estimator of $\lambda$ is 1 if $x_6 > x_5$ and 0 otherwise (unless $x_5=10$ when means that $x_6$ also equals 10 with probability 1 and for that case there is no information about $\lambda$).

Alternatively, if you have $n$ multiple sets of $X_t$ and $X_{t+1}$ (where $X_t < 10$ and if you can assume independence among the sets),then you simply have a binomial distribution and the maximum likelihood estimator is simply the number of observations of increases (m) divided by n.

I still think you need to give more specifics.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.