Hi,
I am afraid I'm looking for an answer to a very basic question (& am embarrassed at admitting my ignorance):
For years I've hidden from the implications of George Box's "all models are wrong" by arm-wavy appeals to the reasonable robustness of approaches like regression, but I'm not really convinced. I still worry that the confidence/credibility intervals around predictions given by "best" models (or sets of models, whether frequentist or Bayesian) are rooted in assumptions about the correctness of representations.
Can anyone point me towards accessible documents to guide decisions on when particular models "are useful", and whether/when/how to deal with minor model misspecifications?
(By useful I mean provide reliable estimates of effect sizes and the uncertainty around them. I suspect that requires more than the carrying out of tests for significant patterns in residuals, and would be pleasantly surprised to find the topic covered by an introductory textbook.)
Thanks,
Mike.
The University of Dundee is a registered Scottish Charity, No: SC015096
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|