4
$\begingroup$

Suppose we wanted to do inference on the mean $\mu$ of some population for which the variance is known.

The conventional frequentist approach is as follows. Given a random sample $X_1, \dots, X_n$ with $\text{E}(X_i) = \mu$ and $\text{Var}(X_i) = \sigma^2$, the CLT tells us that

$$\frac{\bar{X} - \mu}{\sigma / \sqrt{n}}$$

converges in distribution to a standard normal. We can then use this fact to construct confidence intervals or compute p-values using actual data $x_1, \dots, x_n$ collected from the population, provided $n$ is large enough.

Here's my question: Is there a Bayesian reformulation of this problem that allows us to essentially treat the above sampling distribution of $\bar{X}$ as a posterior distribution on $\mu$? In other words, is there a case in which

$$ \mu \sim N(\bar{x}, \sigma^2/n) $$

is actually a valid conclusion?

$\endgroup$
  • 1
    $\begingroup$ The question is: why would you like to have a Bayesian estimate to obtasin frequentist confidence intervals and p-values? $\endgroup$ – Tim Oct 24 '17 at 7:58
  • 1
    $\begingroup$ I think my question was motivated by the fact that non-statisticians are often tempted to use frequentist methods to derive a sampling distribution and then use that distribution to make probabilistic statements that treat the parameter of interest as a random variable, e.g. computing $P(\mu > 1)$. $\endgroup$ – tddevlin Oct 25 '17 at 6:48
4
$\begingroup$

Yes. If $$X_1,\ldots,X_n \stackrel{ind}{\sim} N(\mu,\sigma^2)$$ with $\sigma^2$ known and your prior on $\mu$ is an improper prior over the whole real line, i.e. $p(\mu)\propto 1$, then $$\mu|x_1,\ldots,x_n \sim N(\bar{x},\sigma^2/n).$$

| cite | improve this answer | |
$\endgroup$

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.