Module 5 — The 3PL Model — Guessing

Core concept · hands-on · about 25 minutes.

The 2PL model is powerful, but it has one blind spot: it assumes that a student with zero knowledge of the subject — ability infinitely low — has essentially a zero probability of answering correctly. In a short-answer or constructed-response format that might be reasonable. But in a multiple-choice test, even a student who knows nothing can still guess. On a four-option item, random guessing gives a 25% floor. The 2PL ignores this completely, and the consequence is that it systematically overestimates the difficulty of items (or underestimates the ability of low-scoring students) when guessing is a real option. The three-parameter logistic model (3PL) fixes this by adding a third item parameter: the lower asymptote \( c \).

The full 3PL formula

The 3PL is the model QuantegyAI uses in its adaptive assessments. It has three parameters for each item: discrimination \( a \), difficulty \( b \), and lower asymptote \( c \). The probability of a correct response is:

\[ P(\theta) = c + (1 - c)\,\frac{1}{1 + e^{-a(\theta - b)}} \]

Read this carefully. The right side is not just the logistic function — it is a scaled and shifted version of it. The logistic part still ranges from 0 to 1, but multiplying it by \( (1 - c) \) compresses it to the range \( [0, 1-c] \), and then adding \( c \) shifts the whole thing up so it ranges from \( c \) to 1.

As \( \theta \to -\infty \) (ability far below difficulty), the logistic term goes to 0 and \( P \to c \). The ICC approaches \( c \) from above — \( c \) is the lower asymptote, the floor probability.
As \( \theta \to +\infty \) (ability far above difficulty), the logistic term goes to 1 and \( P \to c + (1-c) \cdot 1 = 1 \). The upper asymptote is still 1.
When \( c = 0 \), the formula collapses exactly to the 2PL.

What c represents: the pseudo-guessing parameter

The parameter \( c \) is often called the pseudo-guessing parameter or the lower asymptote. It represents the probability that even a very low-ability student answers the item correctly — not because they know anything, but because they can guess. For a four-option multiple-choice item, a purely random guesser has a 25% chance, so \( c \approx 0.25 \) is a natural starting point. In practice, items are often written so that one or two distractors are obviously wrong, raising the effective guessing rate slightly above 0.25 for students who can eliminate options.

\( c \) is called "pseudo-guessing" and not simply "guessing" because it represents the lower asymptote of the modeled response curve — not the probability of guessing for any particular student. Low-ability students who do not blindly guess at random but instead apply partial knowledge may still have a floor probability near \( c \). Empirically estimated \( c \) values fold in guessing plus any other sources of floor probability.

How c lifts the left tail off zero

The most visible effect of \( c \) is on the left (low-ability) end of the ICC. In the 2PL, the curve approaches zero as \( \theta \to -\infty \). With \( c = 0.25 \), the curve instead approaches 0.25 — the entire left tail is lifted up by \( c \). The curve no longer starts from the x-axis; it starts from a horizontal floor at height \( c \).

The right tail is unaffected: both models approach 1 for very high ability. The only difference is at low ability, where guessing raises the floor above zero. This lifting effect is easy to see in the interactive activity below — slide \( c \) up from zero and watch the left tail rise.

How c changes the meaning of b

In the 2PL, \( b \) was the ability where \( P(\theta) = 0.5 \) exactly. Adding a guessing floor changes this. With the 3PL, the inflection point of the curve (the ability where the curve is steepest) is still near \( \theta = b \), but the probability at \( \theta = b \) is now:

\[ P(b) = c + (1-c)\,\frac{1}{2} = \frac{1+c}{2} \]

For \( c = 0.25 \), this gives \( P(b) = 0.625 \). So the difficulty parameter \( b \) in the 3PL is not the ability at which \( P = 0.5 \) — it is the ability at which \( P = (1+c)/2 \), which is always above 0.5. Intuitively, because even very-low-ability students can guess correctly 25% of the time, a student at the difficulty point needs more than just a coin-flip probability to stand out — the floor has been raised for everyone.

Where guessing modeling matters — in high-stakes exams The GRE Subject Tests removed their guessing penalty in 2011, explicitly encouraging test-takers to attempt every item. When an IRT model accounts for the guessing floor at the scoring stage, a separate penalty formula becomes redundant — and can disadvantage risk-averse students. More broadly, high-stakes multiple-choice exams advise candidates to eliminate options and answer every item rather than leave blanks, because modern scoring models already anticipate a floor probability of success on multiple-choice items. Without modeling that floor, a test systematically misestimates the ability of low-scoring examinees.

Typical values and estimation challenges

For a four-option item, \( c \) is often fixed or constrained to be near 0.25 rather than freely estimated. The reason is statistical: \( c \) is hard to estimate accurately because it governs behavior at the far left tail of the ICC, where very few students in any typical sample actually fall. Estimating a probability in a region with sparse data yields noisy results. Many practical IRT implementations therefore use a Bayesian prior that pulls \( c \) estimates toward a sensible value (like 0.2–0.25) to prevent them from drifting to extreme values. QuantegyAI uses item-type-specific priors on \( c \) during calibration.

This activity needs JavaScript. Adjust the a, b, and c sliders to see how the 3PL ICC changes, with readouts showing the floor probability, probability at b, and probability at θ = 2.

Sort: guessing concepts

For each statement, decide whether it is True or False about the 3PL lower asymptote \( c \).

This activity needs JavaScript.

Why this matters next You now have the complete 3-parameter model — the exact model QuantegyAI uses. Parameters \( a \), \( b \), \( c \) describe each item; ability \( \theta \) describes each student. The remaining question is: given a set of responses, how do we figure out where on the \( \theta \) scale a student sits? Module 6 answers this with maximum likelihood ability estimation — the engine that turns a response pattern into a score.

One-sentence summary: the 3PL model adds a lower asymptote \( c \) (the pseudo-guessing parameter) that lifts the left tail of the ICC off zero, so that even very-low-ability students have a floor probability \( c \) of a correct answer — for a four-option multiple-choice item, \( c \approx 0.25 \), and the difficulty \( b \) is now the ability at which \( P = (1+c)/2 \), not 0.5.

Next: Estimating Ability →