Conditional logistic regression vs lmer
On 3 Mar 2010, at 22:23, David Duffy wrote:
Was the study a matched case-control design, or is "match" the family or some natural clustering variable? How many cases and controls for each level of match? (maybe this was in your earlier email.) If the strata are small (eg 1:1), then I would be nervous about (tIgE.Resp|match) giving a meaningful estimate, cf c.c ~ tIgE.Resp + (1|match)+ (1|Run) + (1|Box).
The study is a matched case/control, with matched 3 controls for each case. On top of that I have 4 independent cohorts, which I would see as random terms --though my example above does not include cohort because I had data for only one at that point.
You might read over the intro to the paper by Liang, Qaqish and Zeger on the GEE for binary data in JRSSB (?1990 or so), where they contrast the different approaches to clustered data. Basically, the (usual) GLMM fits a parametric model for the intercepts for each stratum, while the CLR discards this information, and for a categorical covariate is "just" a permutation test.
I'll look into it. Thanks! F -- Federico C. F. Calboli Department of Epidemiology and Biostatistics Imperial College, St. Mary's Campus Norfolk Place, London W2 1PG Tel +44 (0)20 75941602 Fax +44 (0)20 75943193 f.calboli [.a.t] imperial.ac.uk f.calboli [.a.t] gmail.com