Find the maximum likelihood estimate of probabilities of ordered pairs of random variables

36 Views Asked by At

$P((X_k,Y_k)=(i,j))=\pi_{ij}$ where $\sum \pi_{ij}=1$ and the frequency of $(i,j)$ is $n_{ij}$.

I'm now asked to find the MLE of $\pi_{ij}$ if $X$ and $Y$ are independent and again if they are not.

I found that the likelihood funciton is: $L(\pi;x)= \prod_{ij} \pi_{ij}^{n_{ij}}$

If $X$ and $Y$ are independent then we can say $\pi_{ij}=\alpha_i \beta_j$ with constraints $\sum \alpha_i = 1$ and $\sum \beta_j =1$ and $\sum \sum \pi_{ij}=1$

Using Lagrange multipliers I find the MLE of $\alpha_i$ and $\beta_j$ are $\frac {2}{\sum_j n_{ij}}$ and $\frac {2}{\sum_i n_{ij}}$

Then the MLE of $\pi_{ij}$ is $\frac {4}{\sum_j n_{ij} \sum_i n_{ij}}$ which doesn't make logical sense to me... Is there some explanation as to why this answer makes sense?

Many thanks for your help!

1

There are 1 best solutions below

0
On

No, it doesn't make sense. That could even be greater than 1.

Let $N=\sum\sum n_{i,j}$. If they are independent, the mle of $\alpha_i$ is $\frac{\sum_{j}n_{i,j}}{N}$ and the mle of $\beta_j$ is $\frac{\sum_{i}n_{i,j}}{N}$.
Just re-write the likelihood in terms of the $\alpha_i$ and $\beta_j$ to derive those.
$$\prod_{i,j}(\alpha_i \beta_j)^{n_{i,j}}$$ Try to maximize this wrt $\alpha_1$ for example.
It's easier to see if you maximize the log-likelihood.