What is Item Discrimination?
Item discrimination is a statistic that shows how well a single question (item) separates respondents who score high on the overall questionnaire from those who score low. A high-discrimination item helps the survey measure the intended trait or opinion reliably.
Item discrimination describes the extent to which responses to one question are related to respondents' overall scores on the scale or survey. If an item discriminates well, people who generally have high scores on the construct (for example, high satisfaction or high ability) tend to answer that item differently than people with low overall scores. Psychometricians measure discrimination in different ways — common methods include the discrimination index (difference in endorsement between top and bottom groups) and the corrected item-total correlation or point-biserial correlation. In IRT (item response theory), a discrimination parameter (often called a
) captures how steeply the probability of endorsing an item rises with the latent trait. Values near zero mean the item doesn't distinguish between respondents; negative values suggest the item may be miskeyed or confusing.
Usage example
You run a parent engagement survey. For the overall parent-engagement score you identify the top 27% and bottom 27% of respondents. On the question School communications are clear,
the top group averages 4.4/5 and the bottom group averages 2.0/5 — this large difference indicates good discrimination. If another question shows nearly identical averages in both groups, it has poor discrimination and may need to be rewritten or removed.
Practical application
Why it matters: good item discrimination helps you build shorter, more reliable surveys and trust the conclusions you draw. For organisations using Hearo, discrimination is also a useful check on multilingual quality: if an item discriminates strongly in one language but not in another, that can flag a translation or cultural interpretation problem. Monitoring discrimination helps you refine wording, detect ambiguous or biased items, and ensure the survey measures the same thing across communities — improving response quality, comparability and inclusion.
FAQ
What values count as good discrimination?
Rules of thumb vary by method. For corrected item-total correlations, values above about 0.30 are often considered acceptable; 0.40+ is good. For discrimination indices (top-minus-bottom group), values above 0.30–0.40 are usually strong. Negative values are a red flag and suggest the item may be miskeyed or confusing.
How is item discrimination calculated?
Common approaches: (1) Discrimination index — compare the average item score or proportion endorsing the item in a high-scoring group versus a low-scoring group (often the top and bottom 27%). (2) Corrected item-total correlation or point-biserial correlation — correlate the item score with the total score excluding that item. Statistical software or spreadsheets can compute these quickly.
Can translation or language affect discrimination?
Yes. A question that discriminates well in one language may perform poorly in another if wording, nuance or cultural relevance changes. Differences in discrimination across languages are useful diagnostics: they point you to items that need rewording, review, or participant feedback to improve translation quality.
Is discrimination the same as difficulty?
No. Difficulty (or item mean) describes how easy or commonly endorsed an item is (for example, the average score or percent correct). Discrimination describes how well the item distinguishes between higher- and lower-scoring respondents. An item can be easy but still discriminate well, or be of medium difficulty and discriminate poorly.