Upper-Lower Thirds Discrimination Procedure

• The upper-and-lower-thirds discrimination procedure is a psychometric method that ranks participants by estimated ability and measures the performance gap between high and low ability groups.
• It utilizes a normalized discrimination index, computed from binary responses and supported by IRT and Kullback–Leibler divergence formulations, to guide efficient item selection.
• Empirical studies show that this method improves short-form test construction by ensuring robust predictive validity and effective discrimination in adaptive testing.

The upper-and-lower-thirds discrimination procedure is a classical psychometric method for quantifying and exploiting item-level discriminative power specifically for fixed-threshold, binary response settings. Its principal application is efficient ability group discrimination, including in adaptive testing and cognitive assessment frameworks. The procedure has been extensively employed in both theoretical investigations of optimal sequential testing and practical construction of short-form diagnostic tools, where it enables highly interpretable item selection and robust generalization to new populations (Bassamboo et al., 2020, Xu et al., 31 Jan 2026).

1. Formal Definition and Mathematical Formulation

In the upper-and-lower-thirds discrimination framework, participants are assessed on a set of binary items (questions or tasks). Response data are modeled as $Y_{ij} \in \{0,1\}$, the outcome for participant $j$ on item $i$. To quantify the discriminative value of each item, individuals are first rank-ordered by estimated latent ability $\alpha_j$. The sample is then split into three groups of equal size $N$: bottom third (low ability), middle third, and top third (high ability). For item $i$, let

$$U_i = \sum_{j \in \text{top third}} Y_{ij}, \quad L_i = \sum_{j \in \text{bottom third}} Y_{ij}$$

be the counts of correct responses in the high- and low-ability groups, respectively. The discrimination index is defined as

$D_i = \frac{U_i - L_i}{N}.$

This index summarizes the normalized performance gap between high- and low-ability participants for each item.

2. Theoretical Foundations in Adaptive and Sequential Testing

The discrimination concept arises prominently in the context of sequential adaptive questioning, particularly for the problem of distinguishing between upper and lower ability segments relative to pre-defined thresholds. Formally, let the latent ability $\theta \in \mathbb{R}$ be classified into "low" ($\theta \leq \theta_L$) or "high" ($j$0) categories, where thresholds are determined by a reference distribution $j$1: $j$2, $j$3. Respondent performance is modeled via a psychometric function $j$4, monotonic in ability and difficulty $j$5.

In the fixed-confidence (or $j$6-correct) framework, the aim is to minimize expected sample size $j$7 subject to stringent error probability constraints for misclassification at the ability boundaries:

$j$8

Information-theoretic lower bounds, derived via change-of-measure arguments, establish that discriminating between $j$9 and $i$0 at level $i$1 requires at least

$i$2

where

$i$3

represents the maximal Kullback–Leibler divergence between binary item response models at the two thresholds (Bassamboo et al., 2020).

3. Algorithmic Implementation and Item Selection

Practical application of the upper-and-lower-thirds discrimination procedure entails the following sequence: calibrate a two-parameter logistic Item Response Theory (IRT) model with participant ability $i$4, item difficulty $i$5, and item discrimination $i$6. Participants are divided into thirds by estimated $i$7. For each item $i$8, calculate $i$9 as previously defined. The primary use-case is item selection: items are rank-ordered by $\alpha_j$0, and those with the highest values are retained for test construction.

In settings such as handwriting assessment, the procedure yields a test form consisting of items on which high-ability and low-ability participants differ most strongly—directly optimizing for maximal observable discrimination between target groups (Xu et al., 31 Jan 2026).

4. Applications in Test Construction and Assessment

The upper-and-lower-thirds discrimination index has been implemented for constructing short, diagnostic assessments. In the study of Chinese character amnesia, a 30-item short form was constructed by ranking 440 calibrated character-writing items by $\alpha_j$1 and selecting the top 30, without further adjustment for difficulty or balance. The result is a compact test that matches the full-length battery's ability to preserve individual differences, achieving within-sample correlation $\alpha_j$2 and cross-validated correlation $\alpha_j$3 (Xu et al., 31 Jan 2026).

A summary of item selection schemes and their empirical predictive performance in this context is given below:

Scheme	Mean $\alpha_j$4	95% CI
Upper-and-Lower-Thirds	0.74	[0.69, 0.80]
Maximum Discrimination ($\alpha_j$5)	0.68	[0.61, 0.75]
Diverse Difficulty	0.35	—
Random	0.53	—

Empirical superiority of the upper-and-lower-thirds method is observed in both in-sample and out-of-sample predictive settings, indicating its robustness to variations in participant ability estimation and its practical advantage for efficient, high-fidelity assessment.

5. Relation to Information-Theoretic and Decision-Theoretic Analysis

The procedure connects to optimal sequential hypothesis testing and active learning under statistical efficiency criteria. The index $\alpha_j$6 operationalizes the gap in observable response distributions between the ability extremes, akin to maximizing Kullback–Leibler divergence between the conditional models at $\alpha_j$7 and $\alpha_j$8. Recent theoretical analysis formalizes the minimax optimality, showing that, under mild regularity conditions, the best performance is realized by adaptively focusing on the single most discriminative item or difficulty level as indexed by this quantity (Bassamboo et al., 2020). No forced exploration is needed: sampling at the optimal level retains endogenous adaptivity.

6. Parameters, Indices, and Interpretation

Key parameters and variables in the upper-and-lower-thirds discrimination context are summarized below:

Symbol	Description	Typical Range/Type
$\alpha_j$9	Binary response for participant $N$0 on item $N$1	$N$2
$N$3	Model-predicted correct response probability	$N$4
$N$5	Latent participant ability	$N$6
$N$7	Item difficulty	$N$8
$N$9	Item discrimination parameter	$i$0
$i$1	Count correct in top/bottom thirds	$i$2
$i$3	Number of participants per tercile	$i$4
$i$5	Upper-and-lower-thirds discrimination score	$i$6 (typically $i$7)

A high $i$8 indicates that an item successfully distinguishes between ability groups and is a strong candidate for inclusion in discriminative short forms.

7. Empirical Performance and Practical Considerations

Assessment via the upper-and-lower-thirds discrimination procedure leads to parsimonious instruments that retain measurement precision while greatly reducing length. Empirical comparisons demonstrate superior out-of-sample predictive validity compared to alternative selection strategies—including simple reliance on the highest IRT discrimination parameters or diversity by item difficulty. A plausible implication is that raw ability-based discrimination captures critical item-level variation not reflected in parameter-based selection alone. The approach is directly extensible to other settings involving binary outcomes, latent trait estimation, and diagnostic screening within fixed-confidence or $i$9-correct frameworks (Xu et al., 31 Jan 2026).