Subject screening is the removal or down-weighting of raters whose votes do not reflect quality - viewers who misread the scale, lose attention, or rate at random - so their noise does not pollute the MOS. It happens in two stages. Pre-screening checks eyesight before the test (acuity on a Snellen chart, colour vision on an Ishihara plate). Post-screening, the subject of this term, runs after the data is in, with three families that trade simplicity for accuracy. The ITU-R BT.500-15 kurtosis rule is a hard test rejecting a subject frequently outside a 2-sigma band whose misses are balanced high and low - random voting - applied once, to small non-expert panels. The P.910 correlation method drops subjects whose ratings correlate weakly with the panel MOS, catching the consistently-wrong voter the kurtosis test misses. The newest, soft approach is the subject-behaviour model (standardized in P.913), which estimates each subject's bias and inconsistency and forms a bias-subtracted, consistency-weighted MOS instead of discarding data. Always fix the rule before seeing the data.
