The text below is based on an ongoing simulation study, as of 2024-10-22.
Usage
RIbootRestscore(
dat,
iterations = 200,
samplesize = 600,
cpu = 4,
output = "table",
cutoff = 5
)Arguments
- iterations
How many bootstrap samples to run
- samplesize
How large sample to use in each bootstrap
- cpu
How many CPU's to use
- output
Optional "dataframe", or "quarto" for
knitr::kable()output- cutoff
Filter to include only rows with % of results above this value
- data
Dataframe with only response data, with 0 as lowest response
Details
Item-restscore will often indicate false positives (item misfit when it is not misfitting) if the sample size is above 400 and there is one truly misfitting item in the data. If there is more than one misfitting item, false positives can occur at sample sizes of n = 150-200.
Conversely, when sample size is below n = 800, the detection rate of truly misfitting items is below 90%, particularly if misfitting items have location > 1.5 logits from the sample mean.
Thus, if one has a large dataset it may be useful to be able to use non-parametric bootstrapping with replacement to get a more nuanced view of the probability of items actually being misfit.