1
$\begingroup$

I am reading the answer to this post (Prediction based on bayesian model) about new predictions, and I have a question. Could I find beta0 and beta1 distributions in regular way:

 beta0 ~ dnorm(0, 10)
 beta1 ~ dnorm(0, 10)
 sigma ~ dunif(0, 50)
 for (i in 1:N) {
    y[i] ~ dnorm(beta0 + beta1 * x[i], sigma)
 }

And then in R do something like this (beta0 and beta1 would be taken from the posterior distribution (regression) and x_new (1:40) would be an array where I need extrapolation):

beta0 beta1 x_new y_new
1.09  0.01  1     1.1
0.98  0.015 1     0.995
1.08  0.012 1     1.092
...
...
...
1.09  0.01  2     1.1
0.98  0.015 2     1.01
1.08  0.012 2     1.11
...
...
...
1.09  0.01  40     1.49
0.98  0.015 40     1.59
1.08  0.012 40     1.56
...
...
...

Can I use this approach to create Bayes inference? Is it wrong? Why do I need "sigma" in the answer from the link for y_new?

Thank you very much for explanation

$\endgroup$

1 Answer 1

2
$\begingroup$

Writing your code move verbosely:

for (i in 1:N) {
    mu[i] <- beta0 + beta1 * x[i]
    y[i] ~ dnorm(mu[i], sigma)
}

Your model is a Bayesian linear regression, so as standard linear regression, it predicts the mean $E[y_i|x_i] = \mu_i = \beta_0 + \beta_1 x_i$. If you are interested in the mean $\mu_i$, sure, take the predictions for the mean. However if you want to look at the samples from the conditional distribution of $y_i|x_i$ itself, then you sample from the posterior predictive distribution. In many cases, saving both the mean and samples from the posterior predictive distribution would be useful.

$\endgroup$
3
  • $\begingroup$ Hi Tim, thank you very much for your answer. I see that my question was not clear enough; I edited it. So, instead of using the mean values for beta0 and beta 1, I was thinking to use beta0 and beta1 samples from the posterior distributions. Can you please see my edits (table) in the question? Thank you again. $\endgroup$
    – eod
    Oct 28, 2021 at 13:19
  • 1
    $\begingroup$ @eod it was clear. mu would be samples from $E[y|x]$ while y would be samples from $y|x$ itself. $\endgroup$
    – Tim
    Oct 28, 2021 at 13:23
  • $\begingroup$ Thank you again for clarification! $\endgroup$
    – eod
    Oct 28, 2021 at 13:28

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.