I have a dataset collected using the Method of Limits (https://en.wikipedia.org/wiki/Psychophysics#Method_of_limits) and am looking for any approaches to outlier detection that are accepted practice.
e.g., for vision, the dataset consists of thresholds of perception beyond which the person is unable to see a light as flickering. For each person, there are 8 data points in ascending direction (starting at 1Hz and increasing till person no longer sees a flicker) and 8 points in descending direction (starting at 100Hz and decreasing till person sees a flicker), yielding 16 data points b/w 1Hz and 100Hz.
One approach I can think of is looking at the standard deviation of the set and setting a threshold of acceptability. That is, if std_dev > 5 Hz for the set, remove the point farthest away (in either direction) from it. However, this seems a bit weird, since you can be in a situation where you end up tossing 5 data points from the descending runs and 0 from ascending. Am I overthinking this or is there a better approach to cleaning such data?