I’ve been using a two-step approach, where in the first step I use all data (including fillers, but not practice items) of an experiment to fit a model of log-transformed raw reading times with:
- word length (Wlen)
- position of word in stimulus (Wpos)
- position of stimulus in list (Lpos)
- experiment ID: different stimulus types may have an effect. This is really just a bad, but easy way to account for the possibility that what matters is not the linear position in a stimulus, but somehow the position relative to syntactic and prosodic phrasing. By including a variable that capture the construction type, some the variance associated with the word’s position relative to phrasing can be accounted for.
- subject differences
library(languageR) l <- lmer(logRT ~ EXPTsimple + Wlen + log(Lpos) + rcs(Wpos) + (1 | SUBJ), d) d$logRTresidual <- residuals(l)
library(languageR) l <- lmer(logRTresidual ~ CONDITION + SPILLOVER_1 + SPILLOVER_2 + SPILLOVER_3 + (1 | SUBJ) + (1 | ITEM), target , subset= EXPT != "ProdComp3" ) ml <- mcmcsamp(l, n=10000)