TOPIC: Selection of Initial Values
[ The following is a reply to an email message of a reader of the book]
Q: By the way, could you tell readers how to set the initial
values in your book?
REPLY
There is a very small section in page 107 (section 3.4.6.7) for initial values.
Some rough rules are the following
- avoid values that cause over-flow or underflows.
- in normal regression usual starting points are intercept = mean of y, all
other coefficients =zero, precision =1. These usually do not cause any
problems. Similar approach can be used in other models.
- If you want to have good starting points use MLEs when the prior is not
informative or the posterior mode if the prior is informative. This assumes
that you will have to estimate them in another package.