Written and maintained by CASRAI Editorial Board
Published
On this page
A lexical decision task (LDT) is a foundational behavioral experimental paradigm in psycholinguistics and cognitive psychology in which a participant is presented with a visual or auditory letter or sound string and must decide, as rapidly and accurately as possible, whether the stimulus constitutes a real word in their language or an invented nonword foil. First introduced systematically by David E. Meyer and Roger W. Schvaneveldt in 1971, the lexical decision paradigm serves as a primary empirical window into the mental lexicon, visual word recognition, morphological parsing, and semantic memory organization.
Definition and Theoretical Foundations
The lexical decision task assesses the speed and accuracy with which the human cognitive system accesses stored lexical representations. In a classic visual lexical decision experiment, trials alternate between real words (targets such as doctor, library, or table) and nonwords (foils such as morth, blape, or zptkq). The participant indicates their binary judgment by pressing one of two dedicated response keys on a low-latency keyboard, response box, or touchscreen. The experimental software records two primary dependent variables on every trial: reaction time (RT) measured in milliseconds from stimulus onset to key press, and response accuracy (correct or incorrect).
Within cognitive linguistics and cognitive science, the task operates on the theoretical assumption that before an individual can judge a letter string as an authentic word, the orthographic input must activate a corresponding lexical entry within the mental lexicon. The latency of this verification reflects both the ease of lexical access and the cognitive architecture governing language processing. Across decades of empirical work in psycholinguistics within general linguistics, researchers have identified several robust behavioral phenomena using this paradigm:
- The Word Frequency Effect: High-frequency words (such as water or house) are recognized significantly faster and with fewer classification errors than low-frequency words (such as abacus or furlong). In massive empirical databases, this frequency advantage produces a reaction time reduction of 50 to 150 milliseconds.
- The Word Length Effect: Longer words typically require longer response latencies than shorter words, particularly when length exceeds five to six letters, reflecting serial sublexical decoding or eye-fixation constraints.
- Orthographic Neighborhood Effects: The number of real words that can be formed by altering a single letter of the target stimulus (Coltheart’s N) systematically modulates decision speed. High neighborhood density often facilitates word recognition in English but delays rejection of pseudowords due to lexical competition.
- Semantic and Associative Priming: Presenting a semantically related prime word immediately prior to a target word (such as nurse preceding doctor) significantly accelerates lexical decision response times relative to an unrelated prime, providing the empirical foundation for semantic priming experiments.
- Pseudoword Legality: Pseudowords that obey the phonotactic and orthographic rules of the language (such as flom or snark) produce substantially longer reaction times and higher false alarm rates than illegal consonant clusters (such as rkfp or btvlm).
Mathematical Formulas and Quantitative Metrics
Analyzing lexical decision performance requires statistical frameworks that handle both categorical accuracy (discriminating words from foils) and continuous latency distributions. Researchers rely on Signal Detection Theory (SDT), Ex-Gaussian distributional decomposition, and Linear Mixed-Effects Models (LMM).
1. Signal Detection Theory (SDT) Formulation
Rather than relying on uncorrected percentage correct, Signal Detection Theory isolates a participant’s perceptual and cognitive sensitivity to lexical status from their underlying response bias or cautiousness. Trials are cross-classified into a 2 × 2 contingency structure:
- Hit (H): Correctly identifying a real word as a word (“Yes” response to word stimulus).
- Miss (M): Incorrectly classifying a real word as a nonword (“No” response to word stimulus).
- False Alarm (FA): Incorrectly classifying a nonword foil as a real word (“Yes” response to nonword stimulus).
- Correct Rejection (CR): Correctly identifying a nonword foil as a nonword (“No” response to nonword stimulus).
The observable proportions are defined as the Hit Rate (H) and False Alarm Rate (FA):
H = Hits / (Hits + Misses) = Hits / Nwords
FA = False Alarms / (False Alarms + Correct Rejections) = False Alarms / Nnonwords
When hit rates reach 1.00 or false alarm rates reach 0.00, standard normal transformation produces infinite values. To resolve this boundary problem without distorting distributions, researchers apply the log-linear correction formulated by Hautus (1995) and recommended by Stanislaw and Todorov (1999):
Hcorrected = (Hits + 0.5) / (Nwords + 1)
FAcorrected = (False Alarms + 0.5) / (Nnonwords + 1)
Sensitivity (d’, pronounced d-prime) quantifies the distance between the signal (word) distribution and noise (nonword) distribution in standard deviation units:
d’ = Z(H) – Z(FA)
Where Z(p) represents the inverse cumulative distribution function (quantile function) of the standard normal distribution, Φ-1(p).
Response criterion (c) measures the location of the participant’s internal decision threshold relative to the neutral crossover point:
c = -0.5 × [Z(H) + Z(FA)]
When c = 0, the participant exhibits no bias. When c < 0, the participant adopts a liberal criterion (predisposing them to respond "word"). When c > 0, the participant adopts a conservative criterion (requiring substantial evidence before judging a stimulus to be a word).
The relative decision criterion or likelihood ratio (β) provides an alternative index of response bias:
β = exp(d’ × c) = exp[-0.5 × (Z(H)2 – Z(FA)2)]
A value of β = 1.0 reflects neutral decision behavior; β 1.0 indicates a conservative bias.
2. Ex-Gaussian Response Time Decomposition
Raw reaction times in lexical decision tasks exhibit pronounced positive skew, rendering standard normal assumptions invalid for raw latency averages. The ex-Gaussian distribution models reaction time distributions as the convolution of a Gaussian (normal) component and an exponential component:
- μ (Mu): The mean of the Gaussian component, reflecting the leading edge of the response time distribution, primarily capturing basic sensory encoding and motor execution speed.
- σ (Sigma): The standard deviation of the Gaussian component, measuring variability in early sensory-motor processes.
- τ (Tau): The mean and standard deviation of the exponential tail, capturing the frequency of prolonged retrieval operations and decision difficulty.
The total mean and variance of the distribution relate directly to these parameters: Mean RT = μ + τ, and Variance(RT) = σ2 + τ2.
3. Linear Mixed-Effects Model (LMM) Specification
Because lexical decision experiments cross human participants with specific linguistic stimuli (words and nonwords), traditional repeated-measures ANOVA suffers from the language-as-fixed-effect fallacy identified by Clark (1973). As demonstrated by Baayen, Davidson, and Bates (2008), researchers must fit linear mixed-effects models with crossed random intercepts and slopes for both subjects and items:
log(RTij) = β0 + β1 × Frequencyj + β2 × Lengthj + β3 × Neighborhoodj + u0i + w0j + εij
Where u0i ~ N(0, σu2) is the random intercept for participant i, w0j ~ N(0, σw2) is the random intercept for item j, and εij ~ N(0, σε2) is residual trial error.
Worked Example With Real Trial Counts and Calculations
To demonstrate the exact mathematical workflow, consider an experimental evaluation in which 40 participants complete a visual lexical decision task comprising 400 total trials: 200 real English words (split equally into 100 high-frequency and 100 low-frequency items) and 200 orthographically legal pseudowords (such as trand and stumpel). We examine both trial-level Signal Detection Theory metrics and latency distribution parameters for a representative participant.
Step 1: Compile the Contingency Matrix
For the representative participant, trial counts across conditions are recorded as follows:
| Stimulus Condition | Target Trials | “Word” Responses | “Nonword” Responses | Empirical Rate |
|---|---|---|---|---|
| High-Frequency Words (Signal) | 100 | 96 (Hits) | 4 (Misses) | Hit Rate = 0.960 |
| Low-Frequency Words (Signal) | 100 | 80 (Hits) | 20 (Misses) | Hit Rate = 0.800 |
| Combined Words (Total Signal) | 200 | 176 (Hits) | 24 (Misses) | Hit Rate = 0.880 |
| Pseudoword Foils (Noise) | 200 | 24 (False Alarms) | 176 (Correct Rejections) | FA Rate = 0.120 |
Step 2: Calculate Combined Sensitivity and Criterion
Across all 200 word trials and 200 pseudoword trials:
H = 176 / 200 = 0.8800
FA = 24 / 200 = 0.1200
Applying the standard normal quantile transformation Φ-1(p):
Z(0.8800) = +1.1750
Z(0.1200) = -1.1750
Computing discrimination sensitivity (d’):
d’ = Z(H) – Z(FA) = 1.1750 – (-1.1750) = 1.1750 + 1.1750 = 2.3500
Computing response criterion (c):
c = -0.5 × [Z(H) + Z(FA)] = -0.5 × [1.1750 + (-1.1750)] = -0.5 × 0.0000 = 0.0000
Computing likelihood ratio bias (β):
β = exp(d’ × c) = exp(2.3500 × 0.0000) = exp(0) = 1.0000
The participant demonstrates robust discrimination between words and nonwords (d’ = 2.35) while maintaining a perfectly neutral decision criterion (c = 0.00, β = 1.00).
Step 3: Calculate Subscale Sensitivity by Word Frequency
Next, we evaluate whether lexical sensitivity changes between high-frequency and low-frequency vocabulary items against the shared pseudoword baseline (FA = 0.1200, Z(FA) = -1.1750):
High-Frequency Condition:
HHF = 96 / 100 = 0.9600; Z(0.9600) = +1.7507
d’HF = 1.7507 – (-1.1750) = 2.9257
cHF = -0.5 × [1.7507 + (-1.1750)] = -0.5 × 0.5757 = -0.2878
βHF = exp(2.9257 × -0.2878) = exp(-0.8420) = 0.4308
Low-Frequency Condition:
HLF = 80 / 100 = 0.8000; Z(0.8000) = +0.8416
d’LF = 0.8416 – (-1.1750) = 2.0166
cLF = -0.5 × [0.8416 + (-1.1750)] = -0.5 × (-0.3334) = +0.1667
βLF = exp(2.0166 × 0.1667) = exp(0.3362) = 1.3996
The arithmetic reveals that high-frequency words elicit higher discrimination sensitivity (d’ = 2.93 vs. 2.02) and a more liberal response criterion (c = -0.29 vs. +0.17), indicating that participants are more willing to endorse familiar letter sequences rapidly as words.
Step 4: Compute Ex-Gaussian Latency Metrics
Correct reaction times were fitted using maximum likelihood estimation to the ex-Gaussian distribution:
- High-Frequency Words: μ = 480.0 ms, σ = 45.0 ms, τ = 95.0 ms. Mean RT = 480.0 + 95.0 = 575.0 ms. Variance = 45.02 + 95.02 = 2025.0 + 9025.0 = 11050.0 ms2 (SD = 105.1 ms).
- Low-Frequency Words: μ = 515.0 ms, σ = 50.0 ms, τ = 155.0 ms. Mean RT = 515.0 + 155.0 = 670.0 ms. Variance = 50.02 + 155.02 = 2500.0 + 24025.0 = 26525.0 ms2 (SD = 162.9 ms).
- Pseudowords: μ = 560.0 ms, σ = 65.0 ms, τ = 190.0 ms. Mean RT = 560.0 + 190.0 = 750.0 ms. Variance = 65.02 + 190.02 = 4225.0 + 36100.0 = 40325.0 ms2 (SD = 200.8 ms).
The mean difference in reaction time between high- and low-frequency words is 95.0 ms (670.0 – 575.0 ms). Crucially, the ex-Gaussian decomposition demonstrates that 60.0 ms (63.2%) of this effect is driven by the exponential tail parameter τ (155.0 – 95.0 ms), whereas only 35.0 ms (36.8%) is driven by the Gaussian shift μ (515.0 – 480.0 ms). This confirms that low-frequency words produce intermittent prolonged lexical search rather than a uniform shift across all processing stages.
Interpretation and Benchmark Values
Interpreting lexical decision metrics requires comparing observed discrimination values and response latencies against established psycholinguistic benchmarks. In general cognitive research, d’ scores above 2.0 indicate strong lexical discrimination, whereas values approaching 0 reflect chance guessing or inattention.
| Metric | Range / Value | Cognitive Interpretation | Methodological Implications |
|---|---|---|---|
| Sensitivity (d’) | Below 1.00 | Poor lexical discrimination; near-chance performance | Examine participant engagement, language proficiency, or excessive foil difficulty. |
| Sensitivity (d’) | 1.00 to 1.99 | Moderate discrimination ability | Typical for low-frequency technical terms, second-language readers, or clinical populations. |
| Sensitivity (d’) | 2.00 to 3.00 | Normal adult lexical discrimination | Standard baseline for skilled adult native speakers responding to matched stimuli. |
| Sensitivity (d’) | Above 3.00 | Exceptional discrimination; potential ceiling effect | Foils may be too obvious (e.g., illegal consonant strings instead of pseudowords). |
| Criterion (c) | Below -0.25 | Liberal response bias (“Word” bias) | Participant guesses “word” under uncertainty; elevates hit rate at the expense of high false alarms. |
| Criterion (c) | -0.25 to +0.25 | Neutral decision threshold | Balanced response strategy; optimal for unbiased latency and accuracy comparisons. |
| Criterion (c) | Above +0.25 | Conservative response bias (“Nonword” bias) | Participant requires extensive lexical verification; suppresses false alarms but increases misses. |
| Ex-Gaussian τ | Above 200 ms | Substantial right-tail latency skewness | Indicates severe lexical retrieval bottlenecks, attentional lapses, or competition from dense neighborhoods. |
Large-scale behavioral megastudies provide normative benchmarks for visual word recognition. The dataset table below compiles cross-linguistic benchmarks across six major psycholinguistic lexical decision megastudies, including the English Lexicon Project (Balota et al., 2007), British Lexicon Project (Keuleers et al., 2012), Dutch Lexicon Project (Keuleers et al., 2010), French Lexicon Project (Ferrand et al., 2010), Spanish Lexicon Project (Aguirre et al., 2018), and Chinese Lexicon Project (Tse et al., 2017). Across these diverse linguistic databases, skilled adult native speakers consistently average response latencies between 540 ms and 690 ms for monosyllabic and disyllabic words, with word recognition accuracy exceeding 88%.
Methodological Assumptions and Common Mistakes
Designing, analyzing, and interpreting a lexical decision task involves subtle methodological pitfalls that can compromise both internal validity and statistical inference.
1. Pseudoword Foil Design and Legal Orthography
The single most consequential design factor in an LDT is the nature of the nonword foils. If pseudowords are orthographically or phonotactically illegal (such as xgtrb or qzfm), participants do not need to access their mental lexicon; they simply perform visual pattern matching or bigram legality screening. To force participants to consult lexical representations, foils must be pronounceable pseudowords created by altering one or two letters of real words (e.g., generating flirp from flip) while matching bigram frequency, length, and syllable structure using automated generators like Wuggy.
2. The Nonword Ratio and Decision Strategies
Standard experimental designs maintain an exact 1:1 ratio between word and nonword trials. Skewing this proportion (for example, presenting 70% words and 30% nonwords) induces an artificial liberal response bias, shifting criterion c negative as participants adapt to the base rate. Furthermore, repeating stimuli within an experiment introduces repetition priming and episodic memory confounds, artificially compressing response times.
3. Reaction Time Trimming and Outlier Exclusion
Raw response times invariably contain anticipatory errors (responses under 200 ms) and attentional lapses (responses exceeding 2,500 ms). Researchers frequently make the error of applying arbitrary absolute cutoffs or asymmetric exclusion rules that discard valid tail data. The recommended practice is to exclude trials faster than 200 to 250 ms as physiological anticipation, apply logarithmic transformations to normalize positive skewness, or model raw latencies directly using generalized linear mixed models with a Gamma error distribution and log link.
4. Ignoring Item Variance (Clark’s Fallacy)
A persistent flaw in older literature is calculating participant means per condition and running a standard repeated-measures ANOVA (F1), or calculating item means and running an item analysis (F2). As demonstrated by Clark (1973) and generalized by Baayen et al. (2008), this procedure dramatically inflates Type I error rates because words are not interchangeable replicates. Linear mixed-effects models with crossed random effects for subjects and items are required to generalize findings simultaneously across participant and word populations.
5. Speed-Accuracy Trade-Offs
Comparing reaction times across experimental conditions without inspecting error rates introduces interpretative error. If a condition yields faster latencies but higher error rates, participants may simply have lowered their decision threshold rather than experiencing facilitated lexical access. Calculating Signal Detection Theory metrics (d’ and c) or fitting Ratcliff’s (2004) drift diffusion model ensures that latency differences reflect evidence accumulation rate (drift rate v) rather than boundary collapse (boundary separation a).
How to Report Lexical Decision Results in APA 7th Style
When reporting lexical decision findings in scientific manuscripts, provide comprehensive descriptive statistics (mean RTs, standard deviations, and error percentages), specify the data trimming protocol, report Signal Detection Theory parameters, and present inferential mixed-effects model coefficients with test statistics, standard errors, and confidence intervals. Below are copy-ready templates formatted according to APA 7th edition guidelines.
APA 7th Template: Linear Mixed-Effects Model
Trimming excluded trials with reaction times faster than 250 ms or slower than 2,000 ms, accounting for 2.1% of trials. Analyses were conducted on log-transformed reaction times for correct word trials using linear mixed-effects models with crossed random intercepts for participants and items, fitted via restricted maximum likelihood in R (lme4 package). High-frequency words were identified significantly faster (M = 575 ms, SD = 105 ms) and with higher accuracy (M = 88.0%, SD = 6.2%) than low-frequency words (M = 670 ms, SD = 163 ms; M = 77.0%, SD = 8.4%). The fixed effect of log word frequency was statistically significant, b = -0.048, SE = 0.009, t(198.4) = -5.33, p < .001, 95% CI [-0.066, -0.030], demonstrating that each unit increase in log lexical frequency yielded an approximate 4.8% reduction in recognition latency.
APA 7th Template: Signal Detection Analysis
Signal detection metrics were calculated using the Hautus (1995) log-linear correction to account for extreme response proportions. Participants exhibited robust lexical discrimination sensitivity (overall d’ = 2.35, 95% CI [2.12, 2.58]), with significantly higher sensitivity for high-frequency words (d’ = 2.93, SD = 0.41) than for low-frequency words (d’ = 2.02, SD = 0.48), t(39) = 8.74, p < .001, Cohen's d = 1.38. Response criterion did not differ significantly from zero in the aggregate condition (c = 0.00, SD = 0.18), t(39) = 0.12, p = .905, confirming that participants maintained an unbiased decision threshold between real words and pseudowords.
When reporting effect size metrics alongside sample planning, consult established protocols for power analysis and sample size estimation to ensure that both participant and item sampling levels achieve at least 80% statistical power.
How to Calculate Lexical Decision Metrics in Excel, R, and Python
Below are verified, copy-ready code snippets and formulas for calculating Signal Detection Theory metrics and fitting linear mixed-effects models in Microsoft Excel, R, and Python.
1. Microsoft Excel Formulas
Set up your trial counts in a worksheet: cell B2 = Hits, B3 = Total Word Trials (Nwords), C2 = False Alarms, C3 = Total Nonword Trials (Nnonwords).
- Corrected Hit Rate (Hautus log-linear rule):
=(B2 + 0.5) / (B3 + 1)(Enter in cellB4) - Corrected False Alarm Rate:
=(C2 + 0.5) / (C3 + 1)(Enter in cellC4) - Sensitivity (d’):
=NORM.S.INV(B4) - NORM.S.INV(C4)(Enter in cellD4) - Response Criterion (c):
=-0.5 * (NORM.S.INV(B4) + NORM.S.INV(C4))(Enter in cellE4) - Likelihood Ratio Bias (β):
=EXP(D4 * E4)(Enter in cellF4)
2. R Implementation (SDT and Linear Mixed-Effects Model)
This script computes SDT parameters and fits a crossed random-effects model using the lme4 package:
# 1. Compute Signal Detection Metrics
compute_ldt_sdt = function(hits, n_words, fa, n_nonwords) {
# Apply Hautus (1995) log-linear correction
h_corr = (hits + 0.5) / (n_words + 1)
fa_corr = (fa + 0.5) / (n_nonwords + 1)
z_h = qnorm(h_corr)
z_fa = qnorm(fa_corr)
d_prime = z_h - z_fa
c_crit = -0.5 * (z_h + z_fa)
beta_val = exp(d_prime * c_crit)
data.frame(
HitRate = round(h_corr, 4),
FARate = round(fa_corr, 4),
dPrime = round(d_prime, 4),
Criterion_c = round(c_crit, 4),
Beta = round(beta_val, 4)
)
}
# Example execution
sdt_results = compute_ldt_sdt(hits = 176, n_words = 200, fa = 24, n_nonwords = 200)
print(sdt_results)
# 2. Linear Mixed-Effects Model for LDT Response Latencies
library(lme4)
library(lmerTest)
# Fit crossed random intercepts for participants and items (Baayen et al., 2008)
# ldt_data requires columns: log_rt, log_freq, length, subject_id, item_id
lmm_model = lmer(
log_rt ~ log_freq + length + (1 | subject_id) + (1 | item_id),
data = ldt_data
)
summary(lmm_model)
3. Python Implementation (Standard Library and SciPy)
This script calculates exact SDT parameters using Python’s built-in statistics.NormalDist, requiring no external dependencies:
import math
from statistics import NormalDist
def calculate_ldt_metrics(hits: int, n_words: int, fa: int, n_nonwords: int):
# Calculate Signal Detection Theory metrics for a Lexical Decision Task.
# Applies the Hautus (1995) log-linear correction.
h = (hits + 0.5) / (n_words + 1.0)
f = (fa + 0.5) / (n_nonwords + 1.0)
# Standard normal quantile function
dist = NormalDist()
z_h = dist.inv_cdf(h)
z_fa = dist.inv_cdf(f)
# SDT metrics
d_prime = z_h - z_fa
criterion_c = -0.5 * (z_h + z_fa)
beta = math.exp(d_prime * criterion_c)
return {
"hit_rate_corrected": round(h, 4),
"false_alarm_rate_corrected": round(f, 4),
"z_hit": round(z_h, 4),
"z_false_alarm": round(z_fa, 4),
"d_prime": round(d_prime, 4),
"criterion_c": round(criterion_c, 4),
"beta": round(beta, 4)
}
# Execute calculation for representative participant data
participant_metrics = calculate_ldt_metrics(
hits=176, n_words=200,
fa=24, n_nonwords=200
)
for metric, value in participant_metrics.items():
print(f"{metric}: {value}")
Frequently asked questions
What is the difference between a visual and auditory lexical decision task?
In a visual lexical decision task, orthographic letter strings are displayed simultaneously on a screen, allowing participants to process the stimulus at their own reading pace. In an auditory lexical decision task, spoken acoustic stimuli unfold over time, meaning lexical access depends on phonological uniqueness points and continuous acoustic-phonetic decoding rather than instantaneous visual apprehension.
How are pseudowords generated to avoid confounding lexical decision trials?
High-quality pseudowords are constructed using specialized algorithms such as Wuggy, which generate pronounceable nonwords by matching the phonotactic transition probabilities, syllable length, and subsyllabic structure of the real words in the experiment. This ensures participants must consult lexical memory rather than relying on superficial orthographic illegality.
Why is reaction time trimmed or log-transformed in lexical decision research?
Human reaction times naturally exhibit severe right skew, with a long tail of delayed responses reflecting occasional lapses in attention or lexical competition. Trimming extreme outliers (such as responses faster than 200 ms or slower than 2,000 ms) and applying logarithmic transformations normalizes the residual error variance, satisfying the distributional assumptions of linear models.
What does a negative criterion c value indicate in a lexical decision task?
A negative criterion (c < 0) signifies a liberal response bias. It indicates that the participant adopts a low evidentiary threshold and is predisposed to respond "word" whenever there is slight familiarity, resulting in higher hit rates but also elevated false alarm rates on pseudoword trials.
Why are linear mixed-effects models preferred over repeated-measures ANOVA?
Repeated-measures ANOVA requires averaging over items to analyze participants (F1) or averaging over participants to analyze items (F2), which creates Clark’s language-as-fixed-effect fallacy and inflates false-positive rates. Linear mixed-effects models simultaneously account for random variance across both human participants and linguistic items within a single unified model.
How does word frequency influence the ex-Gaussian parameters mu and tau?
Empirical studies demonstrate that increasing word frequency primarily reduces the exponential tail parameter τ (tau) rather than shifting the normal distribution parameter μ (mu). This reveals that high word frequency alleviates the cognitive burden of prolonged lexical search rather than accelerating initial sensory-motor processing.
Sources
- Balota, D. A., Yap, M. J., Cortese, M. J., Hutchison, K. A., Kessler, B., Loftis, B., Neely, J. H., Nelson, D. L., Simpson, G. B., and Treiman, R. (2007). The English Lexicon Project. Behavior Research Methods, 39(3), 445–459. https://doi.org/10.3758/BF03193014 (Checked 2026-11-14)
- Keuleers, E., Diependaele, K., and Brysbaert, M. (2010). Practice effects in large-scale visual word recognition studies: A lexical decision study on 14,000 Dutch mono- and disyllabic words and nonwords. Frontiers in Psychology, 1, 174. https://doi.org/10.3389/fpsyg.2010.00174 (Checked 2026-11-14)
- Keuleers, E., Lacey, P., Rastle, K., and Brysbaert, M. (2012). The British Lexicon Project: Lexical decision data for 28,730 monosyllabic and disyllabic English words. Behavior Research Methods, 44(1), 287–304. https://doi.org/10.3758/s13428-011-0118-4 (Checked 2026-11-14)
- Ferrand, L., New, B., Brysbaert, M., Keuleers, E., Bonin, P., Méot, A., Augustinova, M., and Pallier, C. (2010). The French Lexicon Project: Lexical decision data for 38,840 French words and 38,840 pseudowords. Behavior Research Methods, 42(2), 488–496. https://doi.org/10.3758/BRM.42.2.488 (Checked 2026-11-14)
- Aguirre, A., Glez-Nosti, M., Gómez, P., and Duchon, A. (2018). SPALEX: A Spanish lexical decision database from a massive online crowdsourcing experiment. Behavior Research Methods, 50(2), 528–540. https://doi.org/10.3758/s13428-017-0885-0 (Checked 2026-11-14)
- Tse, C.-S., Yap, M. J., Chan, Y.-L., Sze, W. P., Shaoul, C., and Lin, D. (2017). The Chinese Lexicon Project: A megastudy of lexical decision on 25,000 two-character Chinese words. Behavior Research Methods, 49(4), 1503–1519. https://doi.org/10.3758/s13428-016-0810-5 (Checked 2026-11-14)
- Stanislaw, H., and Todorov, N. (1999). Calculation of signal detection theory measures. Behavior Research Methods, Instruments, and Computers, 31(1), 137–149. https://doi.org/10.3758/BF03207704 (Checked 2026-11-14)
- Ratcliff, R., Gomez, P., and McKoon, G. (2004). A diffusion model account of the lexical decision task. Psychological Review, 111(1), 159–182. https://doi.org/10.1037/0033-295X.111.1.159 (Checked 2026-11-14)
- Baayen, R. H., Davidson, D. J., and Bates, D. M. (2008). Mixed-effects modeling with crossed random effects for subjects and items. Journal of Memory and Language, 59(4), 390–412. https://doi.org/10.1016/j.jml.2007.12.005 (Checked 2026-11-14)
- Meyer, D. E., and Schvaneveldt, R. W. (1971). Facilitation in recognizing pairs of words: Evidence of a dependence between retrieval operations. Journal of Experimental Psychology, 90(2), 227–234. https://doi.org/10.1037/h0031564 (Checked 2026-11-14)
Data: Cross-Linguistic Lexical Decision Megastudy Benchmarks
Comparative behavioral and structural benchmarks across six major psycholinguistic lexical decision megastudies, compiling item counts, participant sample sizes, response times, accuracy rates, and pseudoword generation methods from primary research publications.
Period covered: 2007/2018. Coverage: Global.
| Megastudy | Language | Word Count (items) | Pseudoword Count (items) | Participants (individuals) | Mean Word RT (ms) | Mean Word Accuracy (%) | Pseudoword Generation Method |
|---|---|---|---|---|---|---|---|
| English Lexicon Project (ELP) | American English | 40481 | 40481 | 816 | 686 | 88.5 | Phonotactically legal letter strings matched on length and bigram structure |
| Dutch Lexicon Project (DLP1) | Dutch | 14034 | 14034 | 39 | 541 | 94.7 | Wuggy pseudoword generator matching transition frequencies and length |
| British Lexicon Project (BLP) | British English | 28730 | 28730 | 78 | 588 | 93.2 | Wuggy algorithm matched for length and subsyllabic structure |
| French Lexicon Project (FLP) | French | 38840 | 38840 | 975 | 673 | 90.1 | Automated generator with legal phonotactic bigram transitions |
| Spanish Lexicon Project (SPALEX) | Spanish | 45389 | 45389 | 2826 | 642 | 92.4 | Syllable recombination and phonotactically legal substitution algorithms |
| Chinese Lexicon Project (CLP) | Mandarin Chinese | 25000 | 25000 | 640 | 624 | 91.8 | Plausible two-character compound combinations of real characters |
Sources
- The English Lexicon Project (Balota et al., 2007) (accessed 2026-11-14)
- Dutch Lexicon Project (Keuleers et al., 2010) (accessed 2026-11-14)
- The British Lexicon Project (Keuleers et al., 2012) (accessed 2026-11-14)
- The French Lexicon Project (Ferrand et al., 2010) (accessed 2026-11-14)
- SPALEX: Spanish Lexical Decision Database (Aguirre et al., 2018) (accessed 2026-11-14)
- The Chinese Lexicon Project (Tse et al., 2017) (accessed 2026-11-14)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







