Questions: Item Response Functions and Item Characteristic Curves

Question 1

On an item characteristic curve (ICC), what does the difficulty parameter (b) indicate?

• A. The maximum probability of a correct response for any person
• B. The ability level at which a person has a 50% probability of answering correctly
• C. The slope of the ICC at its steepest point
• D. The probability that a very low-ability person guesses correctly

Answer: (B)

In a 1PL or 2PL IRT model, the difficulty parameter b is the ability level (θ) at which the probability of a correct response equals 0.50. An item with b = 1.5 is harder than average because only people with θ ≥ 1.5 have better-than-even odds of answering it correctly. This is a direct read-off from the ICC's horizontal axis.

Question 2

In item response theory, two items with the same difficulty parameter (b) will typically produce identical item characteristic curves.

• T. True
• F. False

Answer: False

Two items can share the same difficulty (b) but differ in their discrimination (a) parameter. The discrimination parameter controls how steeply the ICC rises around b. A high-discrimination item has a steep S-curve that sharply differentiates people just above and below the difficulty threshold, while a low-discrimination item has a flatter curve and is less useful for distinguishing ability levels near b. Matching only on difficulty does not make two ICCs identical.

Question 3

Classical test theory assigns a single difficulty index (p-value) to each item. What is the key limitation of this approach compared to the IRT item characteristic curve?

Model answer: Classical difficulty (p-value) depends on the sample — an item looks easy in a high-ability group and hard in a low-ability group. The IRT ICC describes difficulty as a point on the ability scale that is independent of the sample tested, making it a more stable and generalizable property of the item.

This is the central advantage of IRT over CTT for item analysis. CTT item statistics are group-dependent: the same item administered to Harvard students versus a general population will have very different p-values, even though the item itself has not changed. The IRT b parameter, estimated using the logistic model, places difficulty on the latent ability scale where it remains constant (assuming the model fits), enabling fair comparisons across test forms and populations.