Hi, It seems that I'm quite lost in this wide and powerful R's universe, so I permit myself to ask your help about issues with which I'm struggling. Thank you, I would like to know if the answer’s accuracy (correct = 1; incorrect = 0) varies depending on 2 categorical variables which are the group (A and B) and the condition (a, b and c) knowing that I’ve got n subjects and 12 trials by conditions for each subject (i.e. 12 repetitions). To do that, I’m focusing on logistic regression analysis. I’ve got no problem with this kind of analysis until now (logistic regression with numeric predictor variables and/or categorical predictor with 2 levels only) but, in this new context, I think I have to focus more specifically on logistic regression including *nested (or random?) factors* in a*repeated measures design* (because of the variables “Subject” and “Trial”) with a categorical predictor variable with *more than 2 levels* (the variable “Condition”) and I never did such a thing…yet. mydata mydata$Subject: Factor w/38 levels: "i01", "i02", "i03", "i04" mydata$Group: Factor w/2 levels: "A", "B" mydata$Condition: Factor w/3 levels: "a", "b", "c" mydata$Trial: Factor w/12 levels: "t01", "t02", ..."t12" mydata$Accuracy: Factor w/2 levels: "0", "1" Subject Group Trial Condition Accuracy i01 A t01 a 0 i01 A t02 a 1 ... i01 A t12 a 1 i01 A t01 b 1 i01 A t02 b 1 ... i01 A t12 b 0 i01 A t01 c 0 i01 A t02 c 1 ... i01 A t12 c 1 i02 B t01 a 1 ... First, I’m wondering if I have to calculate a % of accuracy for each subject and each condition and thus “remove” the variable “Trial” but “lose” data (power?) in the same time… or to take into account this variable in the analysis and in this case, how to do that? Second, I don’t know which function I’ve to choose (lmer, glm, glmer…)? Third, I’m not sure I proceed correctly to specify in this analysis that the responses all come from the same subject: within-subject design …+(1|Subject) as I can do for a repeated measures ANOVA to analyze the effect of my different variables on a numeric one such as the response time: test=aov(Int~Group*Condition+*Error(Subject/(Group*Condition)*),data=mydata) and here again how can I add the variable "Trial" if I don't work on an average reaction time for each subject in the different conditions? Below, examples of models I can write with glmer(), fit.1=glmer(Accuracy~Group* Condition +(1|Subject),data=mydata,family=binomial) fit.2=glmer(Accuracy~Group* Condition +(1|Subject)-1,data=mydata,family=binomial) (“without intercept”) fit.3=glmer(Accuracy~Group* Condition +(1|Subject)+(1|Trial)...?? I believed the analysis I've to conduct will be in the range of my qualifications then I realize it could be more complicated than that of course (ex GLMMs), I can hear "do it as we do usually" (=repeated measures ANOVA on a percentage of correct answers for each subject ??) as if there's only one way to follow but I think there's a lot, which one's revelant for my data, that's I want to find. Hope you can put me on the track, Best Suzon [[alternative HTML version deleted]]
Mitchell Maltenfort
2014-Jul-01 15:13 UTC
[R] logistic regression for data with repeated measures
http://stats.stackexchange.com/questions/62225/conditional-logistic-regression-vs-glmm-in-r might be a good start ____________________________ Ersatzistician and Chutzpahthologist I can answer any question. "I don't know" is an answer. "I don't know yet" is a better answer. On Tue, Jul 1, 2014 at 10:24 AM, Suzon Sepp <suzon.sepp at gmail.com> wrote:> Hi, > > It seems that I'm quite lost in this wide and powerful R's universe, so I > permit myself to ask your help about issues with which I'm struggling. > Thank you, > > I would like to know if the answer?s accuracy (correct = 1; incorrect = 0) > varies depending on 2 categorical variables which are the group (A and B) > and the condition (a, b and c) knowing that I?ve got n subjects and 12 > trials by conditions for each subject (i.e. 12 repetitions). > > To do that, I?m focusing on logistic regression analysis. I?ve got no > problem with this kind of analysis until now (logistic regression with > numeric predictor variables and/or categorical predictor with 2 levels > only) but, in this new context, I think I have to focus more specifically > on logistic regression including *nested (or random?) factors* in a*repeated > measures design* (because of the variables ?Subject? and ?Trial?) with a > categorical predictor variable with *more than 2 levels* (the variable > ?Condition?) and I never did such a thing?yet. > > mydata > mydata$Subject: Factor w/38 levels: "i01", "i02", "i03", "i04" > mydata$Group: Factor w/2 levels: "A", "B" > mydata$Condition: Factor w/3 levels: "a", "b", "c" > mydata$Trial: Factor w/12 levels: "t01", "t02", ..."t12" > mydata$Accuracy: Factor w/2 levels: "0", "1" > > Subject Group Trial Condition Accuracy > i01 A t01 a 0 > i01 A t02 a 1 > ... > i01 A t12 a 1 > i01 A t01 b 1 > i01 A t02 b 1 > ... > i01 A t12 b 0 > i01 A t01 c 0 > i01 A t02 c 1 > ... > i01 A t12 c 1 > i02 B t01 a 1 > ... > > First, I?m wondering if I have to calculate a % of accuracy for each > subject and each condition and thus ?remove? the variable ?Trial? but > ?lose? data (power?) in the same time? or to take into account this > variable in the analysis and in this case, how to do that? > > Second, I don?t know which function I?ve to choose (lmer, glm, glmer?)? > > Third, I?m not sure I proceed correctly to specify in this analysis that > the responses all come from the same subject: within-subject design > ?+(1|Subject) as I can do for a repeated measures ANOVA to analyze the > effect of my different variables on a numeric one such as the response > time: test=aov(Int~Group*Condition+*Error(Subject/(Group*Condition)*),data=mydata) > and here again how can I add the variable "Trial" if I don't work on an > average reaction time for each subject in the different conditions? > > Below, examples of models I can write with glmer(), > > fit.1=glmer(Accuracy~Group* Condition > +(1|Subject),data=mydata,family=binomial) > > fit.2=glmer(Accuracy~Group* Condition > +(1|Subject)-1,data=mydata,family=binomial) (?without intercept?) > > fit.3=glmer(Accuracy~Group* Condition +(1|Subject)+(1|Trial)...?? > > > I believed the analysis I've to conduct will be in the range of my > qualifications then I realize it could be more complicated than that of > course (ex GLMMs), I can hear "do it as we do usually" (=repeated measures > ANOVA on a percentage of correct answers for each subject ??) as if there's > only one way to follow but I think there's a lot, which one's revelant for > my data, that's I want to find. > > Hope you can put me on the track, > > Best > > Suzon > > [[alternative HTML version deleted]] > > > ______________________________________________ > R-help at r-project.org mailing list > https://stat.ethz.ch/mailman/listinfo/r-help > PLEASE do read the posting guide http://www.R-project.org/posting-guide.html > and provide commented, minimal, self-contained, reproducible code. >