gee, geese and glmer
On Tue, 18 Mar 2014, Ben Bolker wrote:
[forwarding a conversation about lme4/lme4.0 incompatibilities. This example looks pretty interesting, as it seems hard to prove that lme4 *isn't* giving the right answer/an answer that is numerically superior to lme4.0, yet the lme4.0 answer is biologically preferable/more similar to other estimation approaches. I don't know yet if we will eventually find out that (1) the data are weird in a way that explains the difference; (2) lme4 is actually misconverging, preliminary evidence to the contrary; (3) ??? Enlightening comments are welcome.]
Just to clarify further: All the cases are singletons(i.e. families of 1 member) while controls are from extended families (i.e. families of multiple members).
lme4 has got stuck. The setup as described is pretty pathological: a) sex is not usually correlated within families, and more significantly b) there are no families containing both cases and controls. So, the variance for famid should be zero, and we should get the same answer as a binomial regression. | David Duffy (MBBS PhD) | email: David.Duffy at qimrberghofer.edu.au ph: INT+61+7+3362-0217 fax: -0101 | Genetic Epidemiology, QIMR Berghofer Institute of Medical Research | 300 Herston Rd, Brisbane, Queensland 4006, Australia GPG 4D0B994A