1
$\begingroup$

Assume the following bivariate regression model:

$y_i = \beta x_i + u_i$ where $u_i$ is i.i.d $N(0, \sigma^2 = 9)$ for $i = 1, 2, ..., n$.

Assume a noninformative prior of the form:

$p(\beta) \propto constant$, then it can be shown that the posterior pdf for $\beta$ is:

$p(\beta|\mathbf{y}) = (18\pi)^{-\frac{1}{2}}\left(\sum_{i=1}^n x_i^2\right)^{\frac{1}{2}} \exp\left[-\frac{1}{18}\sum_{i=1}^n x_i^2 (\beta - \hat{\beta})^2\right]$

where $\displaystyle{\hat{\beta} = \frac{\sum_{i=1}^n y_ix_i}{\sum_{i=1}^n x_i^2}}$

Now consider the value of $y$ with a given future value of $x$, $x_{n+1}$:

$y_{n+1} = \beta x_{n+1} + u_{n+1}$ where u$_{n+1}$ is i.i.d $N(0, \sigma^2 = 9)$ , show that:

$p(y_{n+1}|x_{n+1},\mathbf{y}) = \int_{\beta} p(y_{n+1}|x_{n+1}, \beta, \mathbf{y}) p(\beta|\mathbf{y})d\beta$ is a normal density with:

$E[y_{n+1}|x_{n+1},\mathbf{y}] = \hat{\beta}x_{n+1}$

and

$\displaystyle{var[y_{n+1}|x_{n+1},\mathbf{y}] = \frac{9[x_{n+1}^2 + \sum_{i=1}^n x_i^2]}{\sum_{i=1}^n x_i^2}}$

Now my approach to this question is like this:

We can derive $p(y_{n+1}|x_{n+1}, \beta, \mathbf{y}) = (18\pi)^{-\frac{1}{2}} \exp\left[-\frac{1}{18}(y_{n+1} - \beta x_{n+1})^2\right]$

So $p(y_{n+1}|x_{n+1}, \beta, \mathbf{y}) \propto \exp\left[-\frac{1}{18}(y_{n+1} - \beta x_{n+1})^2\right]$

and $p(\beta|\mathbf{y}) \propto \exp\left[-\frac{1}{18}\sum_{i=1}^n x_i^2 (\beta - \hat{\beta})^2\right]$

Hence: $p(y_{n+1}|x_{n+1},\mathbf{y}) \propto \int_{\beta} \exp\left[-\frac{1}{18}(y_{n+1} - \beta x_{n+1})^2\right] \exp\left[-\frac{1}{18}\sum_{i=1}^n x_i^2 (\beta - \hat{\beta})^2\right] d\beta$

Now I am stuck on how I should manipulate the integrand to a known distribution so I can recover the integrating constants and hence find an expression for the integral.

Any help would be appreciated.

$\endgroup$
2
  • $\begingroup$ Add the two exponents together and rewrite them such that you collect all the terms including $\beta$ and then see what you're left with. $\endgroup$ – hejseb Aug 8 '13 at 12:58
  • $\begingroup$ Okay, so the integrand becomes: $\exp\left[-\frac{1}{18}\left(y_{n+1}^2-2y_{n+1}\beta x_{n+1} + \beta^2 x_{n+1}^2 + (\beta^2 - 2\beta \hat{\beta} + \hat{\beta}^2)\sum_{i=1}^n x_i^2\right)\right]$ and further becomes... $\exp\left[-\frac{1}{18}\left(y_{n+1}^2-2y_{n+1}\beta x_{n+1} + \beta^2 x_{n+1}^2 + \beta^2 \sum_{i=1}^n x_i^2 - 2\beta \hat{\beta} \sum_{i=1}^n x_i^2 + \hat{\beta}^2\sum_{i=1}^n x_i^2\right) \right]$. Then, I've tried factoring out the $\beta$ but nothing seems to strike out to me. $\endgroup$ – Trts Aug 8 '13 at 13:06
1
$\begingroup$

Expanding on your comment, you can do it like this:

$$ \text{exp}\left[-\frac{1}{18}\left(y^2_{n+1}-2y_{n+1}\beta x_{n+1}+\beta^2x^2_{n+1}+\beta^2\sum x^2_i-2\beta\hat{\beta}\sum x^2_i+\hat{\beta}^2\sum x_i^2\right)\right]\\ \propto \text{exp}\left[-\frac{1}{18}\left(\beta^2(x^2_{n+1}+\sum x^2_i)-2\beta(\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1})\right)\right]\\ =\text{exp}\left[-\frac{1}{18}\left(\left(\beta\sqrt{(x^2_{n+1}+\sum x^2_i)}\right)^2-2\left(\beta\sqrt{(x^2_{n+1}+\sum x^2_i)}\right)\left(\frac{\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}}{\sqrt{(x^2_{n+1}+\sum x^2_i)}}\right)\right)\right]\\ =\text{exp}\left[-\frac{1}{18}\left(\left(\beta\sqrt{(x^2_{n+1}+\sum x^2_i)}\right)^2-2\left(\beta\sqrt{(x^2_{n+1}+\sum x^2_i)}\right)\left(\frac{\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}}{\sqrt{(x^2_{n+1}+\sum x^2_i)}}\right)+\left(\frac{\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}}{\sqrt{(x^2_{n+1}+\sum x^2_i)}}\right)^2-\left(\frac{\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}}{\sqrt{(x^2_{n+1}+\sum x^2_i)}}\right)^2\right)\right]\\ \propto\text{exp}\left[-\frac{1}{18}\left(\left(\beta\sqrt{(x^2_{n+1}+\sum x^2_i)}\right)^2-2\left(\beta\sqrt{(x^2_{n+1}+\sum x^2_i)}\right)\left(\frac{\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}}{\sqrt{(x^2_{n+1}+\sum x^2_i)}}\right)+\left(\frac{\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}}{\sqrt{(x^2_{n+1}+\sum x^2_i)}}\right)^2\right)\right]\\ $$ Note that the proportionality here is used in terms of the integral. You need to keep these terms in the density function (as it’s the density for $y$), but not in the integrand.

Let’s call the terms next to the $\beta$ $a$ and the other term $\mu$. The variance is $\sigma^2$. Then the expression is: $$ \text{exp}[-\frac{1}{2\sigma^2}(\beta a - \mu)]= \text{exp}[-\frac{a^2}{2\sigma^2}(\beta -\frac{ \mu}{a})]=)]= \text{exp}[-\frac{1}{2(\sigma a^{-1})^2}(\beta -\frac{ \mu}{a})] $$ Thus, you can rewrite the integrand as the density of a normal distribution (where $\beta$ is the random variable) with mean $\frac{\mu}{a}$ and variance $\sigma^2a^{-2}$.

$\endgroup$
3
  • $\begingroup$ Thanks, just wondering, in your third line of working, where did the extra $\sqrt{(x_{n+1}^2 + \sum x_i^2)}$ in $2\left(\beta\sqrt{(x^2_{n+1}+\sum x^2_i)}\right)(\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1})$ come from? $\endgroup$ – Trts Aug 8 '13 at 14:47
  • 1
    $\begingroup$ @TrueTears Oops. Nice catch, I forgot to add the inverse of the term as well. The trick is to rewrite $2\left(\beta(\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}\right)$ as $2\left(\beta\left[\sqrt{(x^2_{n+1}+\sum x_i^2)}\right]\left[\sqrt{(x^2_{n+1}+\sum x_i^2)}\right]^{-1}(\hat{\beta}\sum x^2_i+y_{n+1}x_{n+1}\right)$. I'm on a way too slow computer to be able to edit the post right away, but I hope you get the idea how to solve it anyway. $\endgroup$ – hejseb Aug 8 '13 at 14:53
  • $\begingroup$ Certainly, I can finish it off now :) $\endgroup$ – Trts Aug 8 '13 at 14:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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