Concept map / Person–item map
Share
Person–item maps and test targeting
Compare examinee abilities with item difficulties on a common scale and inspect where measurement is most precise.
1Is this test aimed at these students?
An eight-item reading test was written for students in the middle of the scale. A teacher wants to give it to beginning readers. Will it measure them well? Item information says precision depends on how close the items sit to each student, so the question is where the items sit relative to the group.
A person–item map, or Wright map, shows that at a glance. Under the Rasch model, ability θ and difficulty b share one scale in logits (log-odds units), so students and items go on the same vertical ruler. Switch groups, shift the items, and move θ to see the standard error, the margin of error of an ability estimate, change.
The beginning readers average −1.60 and the items average 0.84, so the items sit 2.44 logits above the group. At θ = −1.6, 0 of 8 items lie within 0.75 logits and the standard error is 1.21. Averaged over the group it is 1.30, against 0.81 for the advanced readers.
2What “above an item” means
On the map, a student sits exactly 1 logit above an item. Under the Rasch model, what is that student’s chance of answering the item correctly?
3Why the gap sets precision
Under the Rasch model every item has discrimination 1, so only the gap θ − b matters. Each item adds P(1 − P) to the information at θ: 0.25 when the item sits exactly at θ, close to zero when it sits far above or below. Point at a symbol to light up its part of the map.
Items far from θ = −1.6 add almost nothing to the sum, so a test aimed at another group measures this student poorly.
4Check yourself
Each task ticks itself when the lab meets its condition.
- With the beginning readers, shift the items until the mean item difficulty is within 0.10 of the group’s mean ability.Now: Beginning readers, group mean −1.60, item mean 0.84, gap 2.44
- Find one shift that keeps the average standard error below 1.00 for both groups at once.Now: shift 0.0: beginning 1.30, advanced 0.81
- Set the shift back to zero, then move the inspected ability to where the standard error is smallest (to three decimals). Compare that ability with the item mean.Now: shift 0.0, standard error 1.208 at θ = −1.6; smallest 0.760 at θ = 0.8
5Person and item locations from mirt
The code simulates 500 beginning readers (mean ability −1.6, as in the lab) answering the eight items, fits a Rasch model with mirt, and prints both sides of the map. mirt fixes the group’s mean ability at 0, so every item comes out about 2.4 logits higher than on the page. The mean item difficulty, 2.43, is the gap between items and group; the lab shows 2.44. Most students answer none or one item correctly, so the person estimates take only a few values: −0.75 for zero correct and 0.56 for one.
Fit a Rasch model and print person and item locationsShow the code
library(mirt)
# Eight Rasch items and 500 beginning readers, as on the page
set.seed(3)
b <- c(-0.4, 0, 0.3, 0.7, 1, 1.3, 1.7, 2.1)
theta <- rnorm(500, mean = -1.6, sd = 0.6)
p <- plogis(outer(theta, b, "-"))
resp <- (matrix(runif(length(p)), nrow(p)) < p) * 1
colnames(resp) <- paste0("Item_", 1:8)
fit <- mirt(resp, 1, itemtype = "Rasch", verbose = FALSE)
# Item locations (difficulty b)
pars <- coef(fit, IRTpars = TRUE, simplify = TRUE)
items <- pars$items[, "b"]
round(items, 2)
# Person locations: weighted likelihood estimates, which
# stay finite for students with no correct answers
persons <- fscores(fit, method = "WLE")[, 1]
round(summary(persons), 2)
# mirt fixes the group's mean ability at 0, so the mean item
# difficulty is how far the items sit above the group
round(mean(items), 2)Uses mirt. Install once with install.packages(c("mirt")).
What the map tells you about targeting
- Under the Rasch model, ability and difficulty are in the same logit units, so one ruler holds both. Software picks the zero point, often the mean ability or the mean item difficulty, so only differences between locations carry meaning.
- A student 1 logit above an item has a 73.1% chance of answering it. Three logits above, the student still misses about one time in 21.
- Each item adds P(1 − P) to the information at θ, and adds most when it sits at θ. The starting items sit 2.44 logits above the beginning readers, whose average standard error is 1.30, against 0.81 for the advanced readers.
- Shifting the items toward one group lowers that group’s standard errors and raises the other’s. One form can serve both only by spreading its items.
- A map shows where new items are needed. Adaptive testing goes further and picks each item to sit near the current ability estimate.
Recommended reading
- Applying the Rasch ModelTrevor G. Bond, Zi Yan, and Moritz Heene4th ed., person–item maps