Questions: Three-Parameter Logistic IRT Model (3PL)

Question 1

A test developer creates multiple-choice items with four options and carefully writes distractors that reliably attract low-ability examinees — each wrong answer reflects a specific, common misconception. When fitting IRT models, the empirical lower asymptote on these items is near zero. Which model is most appropriate?

• A. 3PL — all multiple-choice data require modeling the guessing parameter
• B. 2PL — because effective distractors draw low-ability examinees toward wrong answers systematically, the floor stays near zero and c adds no practical value
• C. 3PL — the guessing parameter is necessary whenever items have more than two options
• D. 1PL — guessing corrections are only needed when items discriminate poorly

Answer: (B)

The 3PL's c parameter earns its complexity when low-ability examinees perform at chance levels above what the 2PL predicts. When distractors are effective — attracting low-ability examinees to specific wrong options — the empirical lower asymptote can be near zero even on multiple-choice items. In that case, 2PL fits adequately and provides cleaner, more stable estimates. The misconception is that multiple-choice format automatically requires 3PL; the format creates the possibility of guessing, but item design determines whether guessing actually occurs.

Question 2

In the 3PL model, what does the c parameter represent?

• A. The probability of a correct response for a randomly selected examinee, equal to 1/k where k is the number of options
• B. The lower asymptote of the item characteristic curve — the probability of a correct response as ability (θ) approaches negative infinity
• C. The slope of the ICC at the point of maximum discrimination
• D. The difficulty value at which 50% of examinees with high ability answer correctly

Answer: (B)

The c parameter is the lower asymptote of the ICC — P(θ) = c + (1-c)×logistic(a(θ-b)) — representing the floor probability of a correct response no matter how low ability is. It is often near 1/k (chance for k options) but is not defined as equal to it; it is estimated from the data or constrained during calibration. The b parameter sets the 50% point on the logistic portion of the curve (adjusted for c), and a governs the slope.

Question 3

In the 3PL model, the pseudo-guessing parameter c is generally equal to 1/k, where k is the number of response options.

• T. True
• F. False

Answer: False

c is estimated from the data and represents the empirical lower asymptote of the ICC — the floor probability of a correct response for the lowest-ability examinees. While it is often near 1/k (random guessing probability), it need not equal it. Items with obvious distractors may yield c > 1/k (examinees can eliminate options), while items where low-ability examinees are drawn to a specific wrong answer may yield c near zero. In practice, c is often fixed at 1/k or constrained precisely because it is so difficult to estimate precisely from data.

Question 4

The 3PL model is preferable to the 2PL for any multiple-choice test because it more realistically models the possibility of guessing.

• T. True
• F. False

Answer: False

Added realism comes at a real cost. The c parameter is notoriously difficult to estimate precisely — its likelihood surface is correlated with the difficulty parameter b, creating flat regions where many (c, b) combinations fit equally well. Stable estimation typically requires 1,000+ examinees. When items have effective distractors and the empirical lower asymptote is near zero, the 2PL fits adequately without these complications. The 3PL is warranted when items are susceptible to systematic chance-level performance (obvious distractors, speeded tests, item-writing flaws) — not as a universal default.

Question 5

Why is the c parameter in the 3PL model particularly difficult to estimate from data, and how do practitioners address this problem?

Model answer: c governs only the behavior of the lowest-ability examinees — a small, imprecisely measured group at the tail of the ability distribution. Its likelihood surface is correlated with the difficulty parameter b, producing flat regions where many (c, b) combinations fit nearly equally well. This instability is amplified by small sample sizes. Practitioners address it by fixing c at a theoretical value (e.g., 1/k), constraining it within a plausible range during estimation, or requiring samples of 1,000+ to get stable estimates.

The parameter identification problem for c is structural: to distinguish a truly difficult item (high b) from an easy-to-guess item (high c), you need many observations from very low-ability examinees — which are rare by definition. This is why c estimates are often unreliable even with moderate sample sizes, and why many testing programs constrain or fix c rather than freely estimating it from data.