Differential Item Functioning in Educational Testing

Impact is the unconditional difference in item performance between groups — it can simply reflect a true difference in ability. DIF is the difference that remains after matching examinees on ability. An item can show large impact (because one group is more proficient overall) yet no DIF (because, at equal ability, both groups perform the same). Only DIF speaks to fairness; impact alone does not imply bias.

Uniform DIF means one group is favored by a roughly constant amount across the entire ability range — the item characteristic curves are separated but parallel. Nonuniform DIF means the direction or size of the advantage changes with ability, so the curves cross; an item might favor one group among low scorers and the other among high scorers. Logistic regression and IRT methods detect both, whereas the basic Mantel-Haenszel test is insensitive to symmetric nonuniform DIF.

Not automatically. DIF is a statistical signal that an item functions differently for matched examinees; whether that constitutes bias is a substantive judgment. A content review panel examines flagged items for construct-irrelevant features that could disadvantage a group. If the differential difficulty is judged construct-relevant (genuinely part of what the test should measure), the item may be retained; if it reflects irrelevant interference, it is revised or dropped.