Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
| From | Nick Cox <n.j.cox@durham.ac.uk> |
| To | "'statalist@hsphsun2.harvard.edu'" <statalist@hsphsun2.harvard.edu> |
| Subject | RE: st: too good to be true : lr test in mlogit? |
| Date | Fri, 13 May 2011 12:50:28 +0100 |
An equivalent argument is that exactly or to a good approximation most kinds of sampling error scale inversely with the square root of sample size. At some large sample size, therefore, biases arising from selection or non-response or other measurement issues, which typically do not depend on sample size, will swamp sampling error and become what you should most worry about. In fact in many fields large surveys may have bigger measurement problems than tightly controlled small surveys, so the difference is intensified. A broader issue is that many statisticians, and not just Bayesians, manage without doing significance tests. See e.g. Nelder, John A. 1999.