The Jouve-Cerebrals Test of Induction (JCTI) is a computer-adaptive nonverbal measure of inductive reasoning published by Cogn-IQ. Its evidence base now spans eight published studies, from the fixed-form psychometrics paper (Jouve, 2023) to a confirmatory factor analysis against the ACT (Jouve, 2026c), consolidated in the 2026 technical manual. This article summarizes the reliability evidence under both classical and adaptive (IRT/CAT) frameworks, the convergent and discriminant validity profile against major intelligence and academic criteria, the factor-structure findings, and how the score is anchored to the general population.
What the JCTI measures
The JCTI is designed as a focused index of inductive reasoning—pattern detection, rule extraction, and generalization from exemplars—delivered through nonverbal figural items that minimize verbal-knowledge contamination. Carroll’s (1993) Human Cognitive Abilities placed inductive reasoning (RG) within the broad fluid-reasoning factor (Gf) of the Cattell-Horn-Carroll three-stratum model, where it is one of the narrow abilities most central to the Gf construct. The JCTI’s nonverbal figural format aligns with that placement: by reducing reliance on accumulated verbal knowledge (Gc), the test targets the fluid component of cognitive ability more directly than mixed-content tests like the SAT or the Wechsler full-scale composites.
The item bank is the 52 figural items of the JCTI’s fixed form, each an incomplete visual pattern with six response options. Since 2025 the test has been delivered as a computer-adaptive test (CAT) that picks, one response at a time, the item most informative about the examinee at hand. The current adaptive configuration has run since 15 September 2026: sessions last between 30 and 52 items, are untimed, and produce a single score, the Inductive Reasoning Index (IRI; M = 100, SD = 15), reported on a range of 56 to 158.
Reliability: classical and adaptive evidence
The manual reports reliability under both frameworks and deliberately does not compare them, since a classical internal-consistency coefficient and an IRT empirical reliability index different quantities (Embretson & Reise, 2000).
Classical reliability (fixed form). In the published fixed-form study, the 52-item form reached an internal consistency of \(\alpha\) = .90 on 2,306 examinees (Jouve, 2023). On the 783 archived protocols in which all 52 items were answered, the full form’s KR-20 is .913. An earlier fixed-form sample of 1,020 examinees, analysed by age group, found \(\alpha\) between .92 and .96 in every age subgroup and .95 overall, with a mean standard error of measurement of 2.63 raw-score points (Jouve, 2010a): the fixed form’s precision does not depend on the examinee’s age.
Adaptive reliability (live design). The current configuration was evaluated by replaying the adaptive algorithm over those 783 complete fixed-form records, each examinee answering every item the algorithm chose exactly as they had on the fixed form. The replay gave an empirical reliability of .90, a median of 30 items per session (mean 36.3), and a correlation of .997 between the adaptive ability estimate and the estimate from all 52 items. About three sessions in four (74.6%) ended on the precision rule; the rest used the whole bank. These are simulations from real answers rather than results of live adaptive sittings, and the manual states plainly that no reliability study of live sittings under this configuration has yet been reported.
The earlier adaptive configuration. Before September 2026 the CAT ran under different stopping rules and a different calibration. Its published study on 1,003 examinees found sessions of 29.4 items on average, an empirical reliability of .87, and a mean conditional standard error of .42 on the ability metric (Jouve, 2026a). Those figures describe the retired configuration, not the live test.
| Form | Coefficient | Value | N |
|---|---|---|---|
| 52-item fixed form, age-group analysis (Jouve, 2010a) | Cronbach’s alpha | .95 | 1,020 |
| 52-item fixed form (Jouve, 2023) | Cronbach’s alpha | .90 | 2,306 |
| 52-item fixed form, complete archive records | KR-20 | .913 | 783 |
| Live adaptive design (replay) | Empirical reliability | .90 | 783 |
| Earlier adaptive configuration (Jouve, 2026a) | Empirical reliability | .87 | 1,003 |
Note. The 2010 age-group alphas ranged from .92 to .96. Classical and IRT coefficients index different quantities and are listed side by side, not compared. The earlier adaptive configuration was retired on 15 September 2026.
A single reliability also hides how precision varies across the scale. Under the live design, the standard error of the IRI is lowest around the average range (about 4 points with the full bank at IRI 90–100) and largest at the extremes. Above about IRI 140 the reporting metric is compressed and the ability estimate itself is least precise, so the manual advises reading scores above 140 as a range rather than a point.
Convergent and discriminant validity
The criterion evidence comes from the convergent-validity study (Jouve, 2026b) and the fixed-form study (Jouve, 2023). Because the JCTI’s online examinees are self-selected and able, observed correlations are depressed by restriction of range; the 2026 study therefore reports each coefficient both as observed and corrected with Thorndike’s (1949) Case II formula, anchored to the criterion’s published normative standard deviation.
Intelligence tests. Convergence is highest exactly where a fluid-reasoning measure should converge: Raven’s Advanced Progressive Matrices at r = .80 observed and .93 corrected (n = 52), and WAIS Matrix Reasoning at .77 observed and .88 corrected (n = 62). Omnibus IQ composites follow in the .60s—WAIS Full Scale IQ .63 observed and .65 corrected (n = 163)—while verbal composites trail, with WAIS Verbal IQ at .41 observed and .43 corrected (n = 116).
| Criterion | n | Observed r | Corrected r |
|---|---|---|---|
| Raven’s Advanced Progressive Matrices | 52 | .80 | .93 |
| WAIS Matrix Reasoning | 62 | .77 | .88 |
| Culture Fair Intelligence Test | 61 | .70 | .67 |
| WAIS Full Scale IQ | 163 | .63 | .65 |
| WAIS Performance IQ | 113 | .65 | .62 |
| WAIS Verbal IQ | 116 | .41 | .43 |
| Armed Forces Qualification Test | 149 | .45 | .67 |
| ACT composite | 157 | .58 | .66 |
| Wonderlic Personnel Test | 64 | .54 | .57 |
| GRE Quantitative | 77 | .35 | .56 |
| GRE Verbal | 78 | .07 | .09 |
Note. Corrected = Thorndike Case II, anchored to each criterion’s normative standard deviation. Selected rows; the manual also reports GAMA, WISC, and Cerebrals CAT criteria. WAIS = Wechsler Adult Intelligence Scale; GRE = Graduate Record Examinations.
The discriminant pattern. The gradient repeats inside single instruments, where test method is held constant: WAIS Performance IQ above Verbal IQ, GRE Quantitative (.56 corrected) far above GRE Verbal (.09). The fixed-form study adds the cleanest within-test contrast. On 34 examinees who took the Reynolds Intellectual Screening Test (RIST), the JCTI correlated .86 with its nonverbal Odd Item Out subtest and only .35 with its verbal Guess What subtest; against self-reported SAT scores (n = 63) it correlated .84 with Mathematics and .38 with Verbal (Jouve, 2023).
| Criterion | n | r |
|---|---|---|
| SAT Mathematics | 63 | .84 |
| SAT Composite | 63 | .79 |
| SAT Verbal | 63 | .38 |
| RIST Odd Item Out (nonverbal) | 34 | .86 |
| RIST Index | 34 | .70 |
| RIST Guess What (verbal) | 34 | .35 |
Note. SAT scores are self-reported. The gap between the nonverbal and verbal subtests of the same battery is the discriminant-validity evidence: the JCTI tracks figural reasoning, not verbal knowledge.
Factor-structure evidence
Three published studies place the JCTI inside joint structures with external instruments, and they agree on where it sits: with the reasoning-and-quantitative side of whatever battery it joins, away from the verbal side.
With the SAT (N = 106). A single component carrying 67.4% of the variance organized the JCTI, SAT Mathematics, and SAT Verbal. The JCTI’s association with SAT Mathematics survived partialing (r = .79 controlling Verbal), while its association with SAT Verbal vanished once Mathematics was controlled (r = −.10) (Jouve, 2010b).
With the GAMA (N = 118). Joined to the four subtests of the General Ability Measure for Adults, the JCTI participated in a single common factor: parallel analysis retained one dimension, which carried 58.2% of the variance, with subtest loadings from .48 (Matching) to .90 (Construction), the JCTI loading .85, and a composite reliability of \(\omega\) = .87 (Jouve, 2018).
With the ACT (n = 95). The most stringent test is confirmatory. Three prespecified models were compared: one general factor; the ACT’s own two-factor structure with the JCTI assigned to its STEM factor (Mathematics, Science); and the same structure with the JCTI assigned to the ELA factor (English, Reading). The one-factor model failed, the ELA placement was rejected, and the STEM placement fit well on every index, with the JCTI loading .62 on STEM and the two factors correlating .83 (Jouve, 2026c).
| Model | χ² | df | RMSEA | CFI | SRMR |
|---|---|---|---|---|---|
| One general factor | 34.60 | 5 | .251 | .908 | .048 |
| ELA / STEM, JCTI on STEM | 4.96 | 4 | .050 | .997 | .016 |
| ELA / STEM, JCTI on ELA | 14.10 | 4 | .164 | .968 | .050 |
Note. Maximum-likelihood estimation, n = 95. Lower RMSEA and SRMR and higher CFI indicate better fit; only the STEM placement fits well.
Scoring and the population anchor
The IRI is computed on the server from the recorded responses and converted from the ability estimate through a fixed table built by equipercentile linking on the 783 complete archive records. Because the people who take the JCTI online are an able group—the calibration cohort averages 124.9 on the IRI metric (SD 16.1)—the score is anchored to the general population through an external measure rather than ranked within that cohort.
The anchor comes from an internal linking study: fixed-form raw scores were linked to a composite of three GAMA subtests on 49 examinees with both scores, corrected to the population metric, and corroborated on a separate subsample of 34 examinees with a WAIS Performance IQ, which gave the same slope and an intercept about five points lower. The manual is explicit about the limits: 49 is a small linking sample, and the five-point gap between the GAMA and WAIS lines bounds how exactly the level of the metric is known. The earlier norm-group conversion used before September 2026 is retired and no longer used for scoring.
Where the evidence is strongest and where it is thinnest
Strengths. Convergence with the reference nonverbal reasoning tests is very high after range correction (.93 with Raven’s APM, .88 with WAIS Matrix Reasoning), and the discriminant pattern repeats across four batteries (WAIS, GRE, RIST, SAT). The structural placement replicates across three partner instruments and three methods, ending in a confirmatory test that a competing placement failed. Reliability is .90 on both the fixed form and the replayed adaptive design.
Limitations. Every criterion sample is self-selected, several are small (the RIST sample is 34), and the corrected coefficients rest on the assumptions of the range-restriction correction. The live adaptive design is documented by a replay on real answers, not by live sittings. The population anchor rests on 49 examinees. Differential item functioning on the live bank has not yet been studied.
Practical implications
The evidence supports the JCTI where the assessment goal is a focused, low-cost index of fluid, inductive reasoning in examinees aged 16 and over. Sessions run 30 to 52 items, untimed, typically in 30 to 45 minutes. The score is a measure of one construct: it is not a full-scale IQ and says nothing about verbal, memory, or speed abilities. Because the test is taken online without supervision, high-stakes uses should establish the conditions of the sitting independently and weigh the IRI alongside other evidence.
For research, the JCTI’s documentation is unusually open: the manual publishes the adaptive configuration, the reliability replay, the score-linking samples, and the full criterion tables, information commercial publishers often gate behind purchase agreements.
Frequently asked questions
What does the JCTI measure?
The Jouve-Cerebrals Test of Induction (JCTI) measures inductive reasoning—pattern detection, rule extraction, and generalization from exemplars—using nonverbal figural items. In Cattell-Horn-Carroll terms, inductive reasoning is one of the narrow abilities most central to fluid reasoning (Gf).
How reliable is the JCTI?
The 52-item fixed form has an internal consistency of \(\alpha\) = .90 (N = 2,306; Jouve, 2023). The current adaptive design, replayed on 783 complete records, has an empirical reliability of .90 with a median of 30 items per session. No reliability study of live adaptive sittings under the current configuration has yet been published.
How does the JCTI correlate with the WAIS?
Corrected for restriction of range, the JCTI correlates .88 with WAIS Matrix Reasoning, .65 with WAIS Full Scale IQ, .62 with Performance IQ, and .43 with Verbal IQ (Jouve, 2026b). The higher correlations with the nonverbal measures are the signature of a fluid-reasoning instrument.
How does the JCTI compare with Raven’s Progressive Matrices?
The JCTI correlates .80 observed and .93 corrected with Raven’s Advanced Progressive Matrices (n = 52), the highest convergent coefficient in its validity panel.
How many items does the adaptive JCTI give?
Between 30 and 52. A session ends once the standard error of the ability estimate reaches 0.30 after at least 30 items, or when the 52-item bank is exhausted; in the replay, about three sessions in four ended on the precision rule, and the median session was 30 items.
How is the JCTI scored?
The JCTI reports one score, the Inductive Reasoning Index (M = 100, SD = 15, range 56–158). The level of the scale is anchored to the general population through a linking study with the GAMA (n = 49), corroborated by a WAIS Performance IQ subsample (n = 34).
References
- Carroll, J. B. (1993). Human Cognitive Abilities: A Survey of Factor-Analytic Studies. Cambridge University Press. https://doi.org/10.1017/CBO9780511571312
- Embretson, S. E., & Reise, S. P. (2000). Item Response Theory for Psychologists. Lawrence Erlbaum Associates.
- Jouve, X. (2010a). Age-based reliability analysis of the Jouve Cerebrals Test of Induction (Version 2) [Research report]. PsychoLogic. https://doi.org/10.5281/zenodo.23033428
- Jouve, X. (2010b). Inductive reasoning and SAT performance: The JCTI relates strongly to SAT-Mathematics but shows no detectable independent association with SAT-Verbal. Cogn-IQ Research Papers. https://doi.org/10.5281/zenodo.22363358
- Jouve, X. (2018). Exploring underlying factors in cognitive tests: Evidence for a unidimensional structure. Cogn-IQ Research Papers. https://doi.org/10.5281/zenodo.22373708
- Jouve, X. (2023). Reliability and concurrent validity of the Jouve Cerebrals Test of Induction: A correlational study with SAT and RIST. Cogn-IQ Research Papers. https://doi.org/10.5281/zenodo.22363302
- Jouve, X. (2026a). Adaptive precision of the Jouve-Cerebrals Test of Induction: Empirical reliability and conditional standard error of the computerized adaptive form. Cogn-IQ Research Papers. https://doi.org/10.5281/zenodo.22363383
- Jouve, X. (2026b). Convergent and criterion validity of the Jouve-Cerebrals Test of Induction: A fluid-reasoning signature across intelligence and academic tests. Cogn-IQ Research Papers. https://doi.org/10.5281/zenodo.22363338
- Jouve, X. (2026c). Inductive reasoning in the structure of the ACT: The JCTI loads on the STEM factor, not the ELA factor. Cogn-IQ Research Papers. https://doi.org/10.5281/zenodo.22363368
- Jouve, X. (2026d). JCTI Technical and Interpretive Manual (Version 2026.3). Cogn-IQ. https://www.cogn-iq.org/methods/jcti-manual/
- Thorndike, R. L. (1949). Personnel Selection: Test and Measurement Techniques. John Wiley & Sons.