Predictive distribution with divergent integral

246 Views Asked by At

I am having some trouble understanding the basic definition of a posterior predictive distribution and how to apply it to simple examples.

I have that if $$y \sim g(y\mid \theta, x)$$

then the predictive posterior distribution is

$$g(y\mid x)=\int g(y\mid\theta,x)\pi(\theta\mid x) \,d\theta$$

So I am trying to work through an example of this where

$x \sim \operatorname{Normal}(\theta, \sigma^2)$ and $y \sim \operatorname{Normal}(ex,\sigma^2)$ and $\pi(\theta,\sigma^2)=\frac{1}{\sigma^2}$

And I am not sure I understand all the notations correctly. I don't have any examples either of actually solving for this, but I would be interested in seeing any.

So I know that $$\pi(\theta\mid x) \propto \pi(x\mid\theta) \pi(\theta)$$

so $$\pi(\theta,\sigma^2\mid x) \propto \exp\left(\frac{-1}{2\sigma^2}{(x-\theta)^2}\right) \frac{1}{\sigma^2}$$

But then I am not sure how to proceed.

I assume I am looking for $\pi(Y\mid X=x)$

Well $P(Y\mid X=x,\theta , \sigma^2) = \exp(\frac{-1}{2\sigma^2}(y-ex)^2)$

so,

$$P(Y\mid X=x,\theta , \sigma^2) \pi(\theta,\sigma^2 \mid x)= \frac{1}{\sigma^2} \exp\left(\frac{-1}{2\sigma^2}[(y-ex)^2+(x-\theta)^2]\right)$$

Note that if we try to double integral to normalise over $\pi(x\mid\theta , \sigma^{2})\pi(\theta,\sigma^2)$ we get a divergent integral. So we cannot make equality.

1

There are 1 best solutions below

3
On

To sum up, you should know the prior distribution of $\sigma^2$. And, yes, you need double integral to integrate out both $\sigma^2, \theta$, unless $\sigma^2$ has been assumed to be a given constant.

Let's clarify what you are looking for. Your are looking for the posterior predictive, i.e. the distribution of $Y | X=x$. Your model is as follows. Given constant $e$,

$$ \begin{aligned} X \mid \theta, \sigma^2 &\sim N(\theta,\sigma^2) \\[7pt] Y \mid x, \theta, \sigma^2 &\sim N(ex,\sigma^2) \\[7pt] π(θ \mid \sigma^2) &= 1/\sigma^2 I_\Theta \\[7pt] \sigma^2 &\sim f_{\sigma^2} \end{aligned} $$

So you have

$$ \begin{aligned} \pi(\theta, \sigma^2 \mid x ) &\propto \pi(x \mid \theta, \sigma^2) \pi(\theta \mid \sigma^2) \pi(\sigma^2)\\[7pt] &\propto \exp(-\frac{1}{2\sigma^2} (x-\theta)^2) 1/\sigma^2 \cdot f_{\sigma^2} \end{aligned} $$

Thus

$$ \begin{aligned} P(Y \mid X=x) &= \int_{\sigma^2} \int_\theta P( Y, \theta, \sigma^2 \mid X=x) \\[7pt] &= \int_{\sigma^2} \int_\theta \pi(Y \mid x, \theta, \sigma^2) \pi (\theta \mid \sigma^2) f_{\sigma^2} \\[7pt] &= \int_{\sigma^2} \int_\theta N(ex,\sigma^2) 1/\sigma^2 f_{\sigma^2} \\[7pt] \end{aligned} $$