#The problem with analyzing all the factors together is that in the "GRM Function" assumes that all of the items on the #measure are assessing the same underlying construct - and that this construct can be #estimated by averaging across all of #the items within the questionnaire. Thus, you are using “ sexual #anxiety” items to estimate “sexual responsibility”, #“sexual responsibility” items to estimate “sexual #competence”, and so on. A better practice might be to evaluate the items #associated with each of the four #scales, and check to see how well they predict the scale with which they are expected to #be associated.
###Sexual Anxiety IRT
discovery_n = round(nrow(data)*.9) # get the 90% N rounded to the nearest whole number discovery = data[1:discovery_n,] # defines a new dataset (discovery) as containing rows 1 through n_discovery from the full dataset (msscq)
#Subset just the items for the Sexual Anxiety factor short form
myvars<-c("Q70","Q81","Q41","Q37") AnxietySF<-discovery[myvars]
#Fit a graded response model since it handles ordinal polytomous data #Constrained models assume that each item have the same level of discrimination. #Perform ANOVA to determine which model is best, the results suggest an unconstrained model
fit.1<-grm(AnxietySF, constrained = TRUE) fit.2<-grm(AnxietySF, constrained = FALSE) anova(fit.1,fit.2)
#Plot each items' response categorical curve #An item is better at discriminating between individuals when the curves are peaked and dispersed across all levels of the #latent trait. For example, an item with high discrimination would have 6 peaks dispersed from low levels of the latent #trait to high levels of the latent trait.
fit.AnxietySF<- grm(AnxietySF, constrained = FALSE) summary(fit.AnxietySF) plot(fit.AnxietySF)