Gaussian-Gaussian model

68 Views Asked by At

Can we rewrite N Gaussian measurements ~iid, in one measurement with a Gaussian distribution? However, a book talking about these really confused me.

Suppose we have data from multiple related group. For example $x_{ij}$ can be the test score for student i in school j, for $j = 1:D$ and $i = 1:N_j$. We want to estimate the mean score for each school $\theta_j $and we assume that $\theta_j$ come from a common distribution $N(\mu,r^2)$ The join distribtuion has the form: $$p(\theta,X |\mu,r,\sigma) = \prod_{j=1}^D N(\theta_j|\mu,r^2) \prod_{i=1}^{N_j} N(x_{ij} | \theta_j,\sigma^2)$$
Once we have estimated $(\mu,r)$, we can compute the posteriors over the $\theta_j$'s. To do that, it simplifies matters to rewrite the joint distribution in the follow form, exploiting the fact that $N_j$ Gaussian measurements with values $x_{ij}$ and variance $\sigma^2$ are equivalent to one measurement of value $y_j := (1/N_j)\sum_{j=1}^{N_j} x_{ij}$ with variance $\sigma_j^2 := \sigma^2/N_j$
This yields $$p(\theta,X |\hat \mu,\hat r,\sigma) = \prod_{j=1}^D N(\theta_j|\hat\mu,\hat r^2) N(y_j | \theta_j,\sigma_j^2)$$

But a particular event A in which $X_{1j}= x_{1j},..., X_{nj} = x_{nj}$ is a subset of event B = {y:$y_j := (1/N_j)\sum_{j=1}^{N_j} x_{ij}$}

Can someone explain?

1

There are 1 best solutions below

1
On BEST ANSWER

The real question here is why we can substitute the whole data with sufficient statistics in finding the posterior probability.

This is best to summarized in an equation

$p(\theta|t(x1,...,xn)) = (p(t|\theta)*p(\theta))/p(t)$

In regard of the getting the posteria, we can map to a statistic t= T(x1,...,xn) upon getting the observation(x1,...,xn) and because it simplified the calcuation of the likelihood $p(t(x)|\theta)$, we should focus the pdf of the statistic instead of the whole data