Roger, Corey, On 30 Sep 2007, at 00:47, Roger Levy wrote: > So with that explanation, the recommendation I'd give you for what you > want to do is: rerun lmer four different ways, each time resetting the > contrasts for place and voice such that you cover the four logical > possibilities, and report the CI for each cell on the basis of the CIs > (either classical or HPD) obtained when that cell is the default.
Do you see a problem with explicitly getting the model to estimate the constrasts for all the factor combinations without parameters for the main effects an the intercept, using model <- glmmPQL(data=d, fixed=c ~ (a*b) - a - b - 1, random=..., family=...) (This is with the MASS / nlme libraries, syntax from the top of my head.) and then using `intervals.lme', intervals(model, level=0.95) Would this produce the same results? > This means that if > you wanted to construct a simultaneous confidence "interval" on > multiple > parameter of the model, it would actually be a confidence *region* in > the form of an ellipsoid. The symmetric confidence intervals we > look at > around the point estimate of a single parameter of the model have > simply > marginalized out all the other model parameters. In the above approach, we would basically only estimate parameters for a single factor - which is simply the product of factors a and b, e.g. all combinations of their respective levels. That way, there shouldn't be interactions between multiple parameters. - David -- David Reitter ICCS/HCRC, Informatics, University of Edinburgh http://www.david-reitter.com _______________________________________________ R-lang mailing list R-lang@ling.ucsd.edu http://pidgin.ucsd.edu/mailman/listinfo/r-lang
