Screening, trimming, and aggregating response time data before fitting an evidence accumulation model.
The package was written for a tutorial manuscript on response time preprocessing. That manuscript and the simulation studies behind it are held back until the paper is published; this repository carries the package itself.
The problem
Every response time analysis makes an exclusion decision, and most inherit it by convention: 200 ms and 3 s because that is what the last paper did, or ±2.5 SD because that is what the field does. The consequences are invisible in the data the researcher can see, and they are not small: contaminant trials bias evidence accumulation model parameters, and which contaminants a method can reach depends on how far they sit from the RT distribution the decision process itself produces.
The methods to deal with this exist and are scattered. What does not exist is a way to compare them, because every implementation returns a different kind of object: a trimmed data frame, a vector of per-trial probabilities, a set of summary statistics. rtprep gives them one interface, so you can ask what a different cutoff would have removed, and how far two defensible choices disagree on your own data.
What it provides
- Screening under one return type: absolute cutoffs, standard deviation and median absolute deviation criteria, recursive moving criteria, and model-based mixture flagging
- Aggregation into EZ-diffusion summary statistics, with mixture down-weighting and accuracy adjustment
- Diagnostics reporting what each rule removed, per cell, and where rules disagree
- Generators for response time data with contaminants of known type, so a chosen pipeline can be tested against ground truth on data shaped like yours
That last one matters most. Contamination is invisible in your own data; simulating it is the only way to check whether your pipeline did what you thought it did.
Installation
# the development version, from GitHub
# install.packages("remotes")
remotes::install_github("GidonFrischkorn/rtprep")
# from CRAN, once the first release is accepted
# install.packages("rtprep")A first look
Every rule is a constructor, one engine applies it, and the result has the same four columns whichever rule went in:
library(rtprep)
set.seed(1)
dat <- r_contaminated(400, process = "mixed", rate = 0.1)
scr <- rt_screen(dat$rt, rule_sd(2.5))
head(scr)
#> <rtprep screen> 6 trials, sd(2.5, mean, sd)
#> kept 6 (100.0%), dropped 0 (0.0%)
#> policy: keep where .prob > 0.5
#> per-group diagnostics: dropped by subsetting; screen_fits() to refit
#>
#> .keep .prob .rule .reason
#> 1 TRUE 1 sd(2.5, mean, sd) <NA>
#> 2 TRUE 1 sd(2.5, mean, sd) <NA>
#> 3 TRUE 1 sd(2.5, mean, sd) <NA>
#> 4 TRUE 1 sd(2.5, mean, sd) <NA>
#> 5 TRUE 1 sd(2.5, mean, sd) <NA>
#> 6 TRUE 1 sd(2.5, mean, sd) <NA>
# how much each rule would have removed, and how much the removed sets overlap
cmp <- screen_compare(
dat$rt,
list(cutoff = rule_cutoff(0.18, 3), sd = rule_sd(2.5), mad = rule_mad(2.5))
)
cmp
#> <rtprep comparison> 400 trials, 3 rules
#>
#> cutoff dropped 0.0%
#> sd dropped 3.8%
#> mad dropped 10.8%
#>
#> least agreement: cutoff vs mad, 89.2% of decisions (Jaccard 0.00)Because r_contaminated() carries the ground truth, the screen can be scored rather than trusted:
table(flagged = !scr$.keep, contaminant = dat$contaminant)
#> contaminant
#> flagged FALSE TRUE
#> FALSE 357 28
#> TRUE 2 13Relationship to other packages
rtprep implements its screening rules itself and carries no runtime dependency on other preprocessing packages. Where reference implementations exist, namely trimr for the trimming families and bmm for the mixture EM and EZ aggregation, the test suite checks equivalence against them.