To calculate the posterior distribution of the latent variables.

To maximize the expected complete-data log-likelihood with respect to $\theta$.

To compute the observed Fisher information.

To generate new bootstrap samples.