I agree that the linked blog explains things well, but isn't it unorthodox to distinguish "sum coding" and "deviation coding"? I thought the two terms were usually synonyms.
And if they are to be distinguished, it struck me as odd that +0.5/-0.5 was being labelled "deviation coding", when as the blog shows, it's actually under +1/-1 coding that the coefficients represent the deviations of each group from the grand mean. Dan On 11 Apr 2015, at 23:22, Dan McCloy <[email protected]> wrote: Hi Sunfa, regarding your question #1: There is a good explanation of the difference between -1/+1 and -0.5/+0.5 coding on Dale Barr's blog: http://talklab.psy.gla.ac.uk/tvw/catpred/ regarding your question #3: As far as I know there is no built-in function for the -0.5/+0.5 coding scheme. You can achieve it either by specifying the contrast matrix yourself, or by setting contr.sum() and then dividing by 2: contrasts(foo) <- contr.sum contrasts(foo) <- contrasts(foo) / 2 regarding your question #2, what is preferred is likely to vary by researcher, and probably depends on what your research question is, although the aforementioned blog post offers some reasons to prefer the -0.5/+0.5 scheme. -- dan Daniel McCloy http://dan.mccloy.info/ Postdoctoral Research Fellow Institute for Learning and Brain Sciences University of Washington On Sun, Apr 12, 2015 at 10:08 AM, Sunfa Kim <[email protected]> wrote: > Dear R users, > > I would like to ask questions regarding the process of simplifying random > effects structures in mixed logit models, and the deviation coding. > > I have an independent variable with four levels and a participant variable > with numerical values as my two predictor variables. My data are > categorical data and so I used mixed logit models (Jaeger, 2008). > > In my experiment, I examined if structural priming would interact with > bilinguals' L2 proficiency. The participant variable (i.e., bilinguals' L2 > proficiency) was centered to its mean before it was included in the model. > > Prime Type Active Passive Unaccusative intransitive Unergative > intransitive Active target responses Passive target responses > > I followed the "Keep it maximal" standard proposed by Barr, Levy, > Scheepers, and Tily (2013). But, the model with the maximal random effects > structure failed to converge. Barr et al. suggested two ways to simplify > the maximal random effects structure when it fails to converge. > > (1) remove correlation parameters > > TargetResponse~PrimeType*L2Proficiency+(1 | Subject)+(0+PrimeType | > Subject)+(1 | Item)+(0+PrimeType*L2Proficiency | Item) > > > (2) Remove random intercepts for any within-unit > > TargetResponse~PrimeType*L2Proficiency+(0+PrimeType | > Subject)+(0+PrimeType*L2Proficiency | Item) > > Now, Barr et al. warned that, with these simplified models, the treatment > coding vs. the deviation coding will make difference, with the deviation > coding performing generally better. > > Throughout the analyses, I have been using the treatment coding of (0 and > 1), and I would like to re-code the "Prime Type" with the deviation coding. > In Barr et al.'s article, the deviation coding is introduced as (-.5 and > .5). > > <***Question 1***> Will the deviation coding of (-.5 and .5) be different > from the deviation coding of (-1 and 1)? In other words, when the "-.5 and > .5" coding is used, will it yield different results from the "-1 and 1" > coding? If so, why? > > <***Question 2***> Which deviation coding is preferred for data analyses? > > <***Question 3***> If the (-.5 and .5) coding is preferred over the (-1 > and 1) coding, is there a R built-in function for the "-.5 and .5" coding?: > > contr.sum(4) > > The above R built-in function provides the (-1 and 1) coding. I am > wondering if there is a R built-in function for the (-.5 and .5) coding? > > Thank you very much for any advice you may have on the above questions. > > Warmest regards, > Sunfa Kim >
