Concept map / IRT: item curves
Share
Item characteristic curves
Compare how ability, difficulty, discrimination, and the lower asymptote affect the probability of a correct response.
1A student and a question on one scale
Item response theory (IRT) puts students and questions on one scale. A student’s ability θ and a question’s difficulty b are in the same units, so it makes sense to say a student sits one unit above a question. How likely is a correct answer then, and what else decides it?
An item characteristic curve (ICC) gives that chance at every ability. It is a logistic regression of the item score on ability. The 1PL curve has only b. The 2PL adds a discrimination a, which sets how steeply the curve rises, and the 3PL adds a lower asymptote c, the chance a student far below the item still has, often from guessing. Drag the student and the question, then switch models.
Drag the student and the question along the ability scale.
- Chance at θ = b
- 50.0%
- Steepest slope
- 0.25
73.1% chance of a correct answer. Ability is 1.0 above difficulty.
A student at θ = 1.0 meets a 2PL item with difficulty b = 0.0 and discrimination a = 1.00. Ability is 1.0 above difficulty, so the chance of a correct answer is 73.1%. Far below the item, the chance falls toward zero.
2Predict what a steeper item does
A student’s ability equals the item’s difficulty, θ = b, and there is no guessing (c = 0). The chance of a correct answer is 50%. What happens to that chance if the item’s discrimination a rises, so the curve gets steeper?
3Where a, b and c enter the formula
Every item curve in this chapter is this logistic function, written without the 1.7 scaling constant some textbooks use. Point at a symbol and its part of the chart lights up.
- θ is the student’s ability, the green circle.
- b is the difficulty, the orange diamond. The curve is halfway between c and 1 there.
- a is the discrimination. The blue tangent marks the slope at θ = b, a(1 − c)/4 = 0.25.
- c is the lower asymptote, the gold dashed floor. At θ = b the chance is (1 + c)/2 = 50.0%.
Show where this item measures best
Item information peaks at 0.25 at ability 0.00, exactly at b. Raise a and the peak grows taller and narrower. The item information page adds items together.
4Check yourself
Each task ticks itself when the lab meets its condition.
- With c at zero, build an item that gives a student at θ = 0 a 70% chance of a correct answer (rounded to the nearest percent). You will need to move both a and b.Now: θ = 1.0, c = 0.00, a = 1.00, b = 0.0, chance 73.1%
- Make the steepest slope at least 0.50, so a step of 0.1 in ability adds about 5 percentage points there, and put the student exactly at that steepest point.Now: Steepest slope 0.25 at θ = b = 0.0; student at θ = 1.0
- Choose the 3PL and set c to at least 0.20. Then give a student at θ = −3 a chance within one percentage point of c.Now: 2PL, c = 0.00, student at θ = 1.0 has 73.1%, 73.1 points above c
5Compute the curve and fit it with mirt
The first lines compute the curve by hand for the page’s starting item, with and without c = 0.20. The rest simulates 2,000 students answering five 2PL items and fits them with mirt, which reports an intercept d = −ab. coef(fit, IRTpars = TRUE) turns each item back into a and b. The first item was generated with a = 1 and b = 0 and comes back as a = 0.97 and b = −0.06.
Compute an item curve and fit a 2PL item with mirtShow the code
library(mirt)
# The page's starting item: a = 1, b = 0, student at theta = 1
icc <- function(theta, a, b, c = 0) {
c + (1 - c) / (1 + exp(-a * (theta - b)))
}
round(icc(1, a = 1, b = 0), 3)
round(icc(1, a = 1, b = 0, c = 0.2), 3)
# Steepest slope, reached at theta = b: a * (1 - c) / 4
1 * (1 - 0.2) / 4
# Simulate 2000 students answering five 2PL items
set.seed(1)
a <- c(1, 0.6, 1.4, 2, 0.8)
b <- c(0, -1, 0.5, 1, -0.5)
# mirt writes the 2PL as a * theta + d, with d = -a * b
resp <- simdata(matrix(a), matrix(-a * b), N = 2000,
itemtype = "2PL")
fit <- mirt(resp, 1, itemtype = "2PL", verbose = FALSE)
# Item 1 in the page's a and b form (mirt calls c "g")
round(coef(fit, IRTpars = TRUE)$Item_1, 2)Uses mirt. Install once with install.packages(c("mirt")).
What the three parameters do
- Ability θ and difficulty b share units, so the gap θ − b sets the chance of a correct answer, not either number alone.
- Raising b slides the curve to the right without changing its shape. With c = 0, a student whose ability equals b has a 50% chance at any slope.
- The curve is steepest at θ = b, where its slope is a(1 − c)/4. In the 3PL the chance never drops below c, and at θ = b it is (1 + c)/2.
- This site writes the curve without the 1.7 scaling constant, as
mirtdoes withIRTpars = TRUE. A textbook that multiplies a by 1.7 reports a smaller a for the same curve. - A steep curve gives the most item information near its difficulty. Adding curves gives the test characteristic curve, and multiplying them gives the likelihood used to estimate ability.
Recommended reading
- The Basics of Item Response Theory Using RFrank B. Baker and Seock-Ho KimChapters 2–3, item characteristic curves and their models