For $X = {(Z_{i}, Y_{i}) : i = 1, ... ,n}$, consider the model:

$Y_{i} = \beta_{1} + \beta_{2}Z_{i} + \epsilon_{i}$ where $\epsilon_{1}, ... ,\epsilon_{n}$ are i.i.d $N(0,\sigma^2)$, $Z_{i},...Z_{i}$ are i.i.d $N(\mu_{1},\sigma_{1}^2)$ and independent of $\epsilon_{1}, ... ,\epsilon_{n}$.

For $1\le i\le m$, we observe both $Z_i$ and $Y_i$, and for $m + 1 \le i\le n$, we observe only $Y_i$.

I try to complete the E- and M-steps of the EM algorithm for estimating $\theta=(\mu_{1},\beta_{1},\sigma_{1}^2,\sigma^2,\beta_{2})$

E-step: $Q(\theta; \theta_{old}) = E [\ell(θ; X , Y) | X , \theta_{old}]$

M-step: $\theta_{new} := \max\limits_{\theta} Q(\theta; \theta_{old})$.

still stuggle how to apply these step to solve the problem

  • $\begingroup$ Have you lost interest in the question? $\endgroup$
    – Xi'an
    Mar 23, 2020 at 15:17

1 Answer 1


You could express the complete log-likelihood as $$\ell(\theta) = -n\log\sigma_1-\frac{1}{2}\sum_{i=1}^n \sigma_1^{-2}(z_i-\mu_1)^2 -n\log\sigma^2-\frac{1}{2}\sum_{i=1}^n \sigma^{-2}(y_i-\beta_1-\beta_2 z_i)^2$$ to derive $$\mathbb E_{\theta^\text{old}}[\ell(\theta)|X^\text{obs}]$$as \begin{align}&-n\log\sigma_1-\frac{1}{2}\sum_{i=1}^m \sigma_1^{-2}(z_i-\mu_1)^2\\&-\frac{\sigma_1^{-2}}{2}\sum_{i=m+1}^n \mathbb E_{\theta^\text{old}}[(Z_i-\mu_1)^2|y_i]\\ &-n\log\sigma^2-\frac{1}{2}\sum_{i=1}^m \sigma^{-2}(y_i-\beta_1-\beta_2 z_i)^2\\ &-\frac{\sigma^{-2}}{2}\sum_{i=m+1}^n \mathbb E_{\theta^\text{old}}[(y_i-\beta_1-\beta_2 Z_i)^2|y_i] \end{align} and then use the conditional distribution of $Z_i$ given $Y_i=y_i$ and $\theta^\text{old}$: $$f(z_i|y_i,\theta)\propto\exp\{-\sigma^{-2}(y_i-\beta_1-\beta_2z_i)^2/2\}\times\exp\{-\sigma_1^{-2}(z_i-\mu_1)^2/2\}$$ which means $$Z_i|Y_i,\theta\sim\mathcal N\left(\{\sigma^-2\beta_2^2+\sigma_1^{-1}\}^{-1}\{\sigma^-2\beta_2[y_i-\beta_1]\},\{\sigma^-2\beta_2^2+\sigma_1^{-1}\}^{-1}\right)$$ from which $$\mathbb E_{\theta^\text{old}}[(Z_i-\mu_1)^2|y_i]\qquad\text{and}\qquad\mathbb E_{\theta^\text{old}}[(y_i-\beta_1-\beta_2 Z_i)^2|y_i]$$ are easily derived.


Your Answer

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

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