I want to fit a model to a number of observations, each of them being a k-dimensional binary vector $(x_1, x_2, ..., x_k)$ where $x_i \in \{0,1\}$.
Naturally I would like to fit a multivariate bernoulli distribution $\mathbf X = (X_1, X_2, ... , X_k)$ where each $X_i$ is a bernoulli variable. I realize that there are $2^k - 1$ parameters for such a multivariate bernoulli distribution but for simplicity's sake I would like to simplify it by parameterizing the distribution by mean and second order interaction (higher order interaction are ignored):
- Marginal expectation $E[\mathbf X] = (p_1, p_2, ..., p_k)$
- Pairwise correlation $\rho_{ij}, 1 \le i,j \le k $
I have read a number of papers on this distribution and found that sometimes it is called Ising Model. But I still need some major pointers/directions due my limited statistical background:
- How do I find an MLE of the parameters given my observations?
- If I have a partial observation $(x_1, x_2, ..., x_j), j < k$, how do I estimate the remaining values? In other words, is there a simple formula for the conditional expectation? $$ E[X_{j+1}, X_{j+2},..., X_k | X_1=x_1, X_2=x_2, ..., X_j=x_j], 1 < j < k $$