What is Item Characteristic Curve (ICC)?
An Item Characteristic Curve (ICC) is a graph that shows how likely people with different levels of an underlying trait (like ability, confidence or agreement) are to give a particular response to one survey question. ICCs help you see where an item is most informative and how well it distinguishes between respondents.
An ICC comes from item response theory (IRT). The horizontal axis represents the latent trait being measured (often called theta) — for example, how confident a parent feels about supporting their child’s learning. The vertical axis shows the probability that a respondent with a given trait level will choose a specific response (often a correct answer, a ‘yes’, or a high-agreement option). For simple yes/no or right/wrong items the ICC is typically an S-shaped curve: people with low trait levels have a low probability of endorsing the item, and that probability rises as trait level increases. Key features of an ICC are:
- Location (difficulty): where on the trait scale the probability crosses midline — which group the item targets.
- Slope (discrimination): how sharply the probability changes — how well the item separates nearby trait levels.
- Lower/upper limits (guessing or ceiling effects): the asymptotic probabilities at extreme trait values.
ICCs can also be drawn for each response category of a multi-point question (category characteristic curves). They are estimated from response data using IRT models and are used to evaluate and compare items beyond simple summary statistics.
Usage example
A school runs a short survey asking parents whether they feel confident helping with homework (agree/disagree). The ICC for that question shows a low probability of 'agree' for parents with low confidence, but the curve rises steeply around the average confidence level — indicating the item is good at distinguishing between less and more confident parents. If a translated version of the question shows a much flatter ICC for one language group, that may signal a translation or cultural issue.
Practical application
ICCs matter because they tell you how each question behaves across the full range of the thing you are trying to measure. Practical benefits include:
- Selecting items that target the parts of the population you care about (e.g., items that pick out people with low service access).
- Choosing questions that discriminate well so scores are more reliable with fewer items.
- Diagnosing problems with translated items by comparing curves across language groups (to spot bias or loss of meaning).
- Informing adaptive surveys that present questions matched to a respondent’s likely trait level.
For teams running multilingual, inclusive surveys, ICCs provide an evidence-based way to check that questions work equally well for different communities and to improve items rather than relying only on intuition.
FAQ
How do you get an ICC from survey data?
You estimate ICCs by fitting an item response model (for example a one- or two-parameter logistic model for dichotomous items, or graded-response models for ordered responses) to a dataset of responses. The model produces parameter estimates that are used to plot the curve. This requires a reasonably sized sample and software that supports IRT modelling.
Can ICCs tell me if a translated question is working the same way?
Yes. By comparing ICCs (or item parameters) across language groups you can detect differential item functioning (DIF) — situations where people with the same underlying trait have different probabilities of endorsing an item depending on language or group. That flags items that may need rewording or cultural adaptation.
Do I need to be a statistician to use ICCs?
You don’t have to be a specialist to use the concept, but producing and interpreting ICCs requires some statistical tools and judgement. Many user-friendly packages and services can estimate ICCs (e.g., R packages like mirt or ltm, or commercial IRT software). For practical survey design, working with a simple IRT summary or a consultant can be enough to apply the insights.