Army General Classification Test (AGCT)

Background

The Army General Classification Test (AGCT) was a World War II intelligence test used by the U.S. Army to classify recruits according to their "ability to learn quickly the duties of a soldier.¹" It remains one of the largest normed cognitive ability tests ever administered and shows strong g-loading and predictive validity.

After many concerns during World War II over the misassignment of soldiers into unsuitable roles and the underutilization of more capable soldiers, the US Army spent lots of resources on commissioning an intelligence and aptitude test, resulting in the early forms of the AGCT. After the end of World War II, the AGCT continued to undergo constant improvements and revisions to ensure its accuracy. Amassing an enormous sample of more than 12 million soldiers, this transcends the samples of modern professional tests by over five thousand times.²

Because the drafted soldiers spanned a wide range of ages, the test was tailored to provide accurate scores for teenagers to middle-aged adults. Accordingly, it has been found that performance on the AGCT is not related to age, as shown by a correlation of .02 for 4,330 inductees¹. Furthermore, with drafted soldiers of all classes and lifestyles being the intended testees, the test was designed with questions that minimized prior knowledge from education and culture.

A test of g

In order to rehabilitate this test for modern use, a few things had to be done:

1. The original score distribution had to be re-normalized by correcting for skew.
2. Norm obsolescence, if any, had to be ascertained and accounted for.
3. The g-loading has to be estimated.

Original distribution

The original distribution is highly left-skewed¹³. This is because those charged with the norming underestimated the number of easy questions on the test. This resulted in a test that discriminates well in the low range (you don’t want to draft morons), but not as effectively in the higher range.

In order to correct for this flaw, the test had to be re-normalized. With percentile rank-equating, it is possible to generate new aligned norms.

This is the original distribution:

This is the fixed distribution:

Overall, most of the changes happened in the low range; however, this step was necessary for psychometric rigor.

Norm obsolescence

It is normal to wonder if a test from 1941, 82 years ago, is still valid today.

Consider this:

In 1980, during the renorming of the ASVAB, the AGCT was pitted against it⁴. It was found that the percentiles matched nicely at all ranges. 39 years later, where Flynn effects would have predicted a systematic inflation of nearly 12 pts, what was found was a simple fluctuation of the sign of the difference between the tests throughout the range. This can be easily attributed to either sampling or error of measurement. There are absolutely no Flynn effects for this test. This makes sense, given the absence of the particular test elements that are responsible for the majority of the Flynn effect.

Before it was released on the subreddit, it was given to dozens of people within the community with known scores from professional tests. More often than not, AGCT ended up being one of their lower rather than higher scores. This provides strong confidence to declare that the AGCT is not an obsolete test.

Construct Validity

The ‘g-loading’ is the degree to which a test correlates with the ‘g factor’ or general intelligence. A higher g-loading means a test is better, and figures above 0.8 are generally considered to be great. These correlations are often derived through factor analysis. As item data for this test is impossible to get by, we can first estimate this test’s accuracy by its proxy g-loading from its successors, the ASVAB and AFOQT. Factor analyzing these two batteries and deriving composites from subtests that most closely resemble the AGCT in content was the only way to assess its construct validity.

From the ASVAB, the pseudo-AGCT composite yielded a g-loading of .92, whereas the AFOQT pseudo-AGCT composite had a g-loading of .90. Averaging the two gives an estimate of ~.91. 

Furthermore, using data from the automated AGCT form at CognitiveMetrics, the g-loading for the AGCT can be calculated. With a sample size of 1734 and M 121.7 SD 12.95, we can calculate the reliability at 0.941 and after being corrected for range, 0.956. 

The g-loading of this sample is 0.816, and after being corrected for range restriction and SLODR, the g-loading has been calculated at 0.925, further aligning with our estimations above. The g-loading unadjusted for V is 0.535, Q is 0.733, and S is 0.597. It isn’t possible to correct for SLODR due to a lack of individual norms, but after correcting for range restriction, the g-loadings are 0.659 for V, 0.733 for Q, and 0.646 for S.

A g-loading of 0.925 is highly impressive for an 82-year-old test. Factorial validity is manifest.

Convergent Validity

Of the AGCT's convergent validity, the Examiner Manual for the AGCT states¹:

The relationship of the AGCT to other well-known tests of general learning ability is high (7)³. Correlations of the ACCT with three intelligence tests were: .90 with the Army Alpha, .83 with the Otis Higher Mental Ability Examination, and .79 with the American Council on Education Psychological Examination. (p. 2)

Additionally, Jensen states that the AGCT is, "as highly correlated (r≈.80) with various IQ tests as the IQ tests are correlated with each other." (p. 376)⁵. We tested whether the convergent validity of the AGCT holds up today.

In 2023, the AGCT was proctored to 58 individuals with verified scores on professional tests, and the relationship depicted below was calculated. The correlation was calculated at r = 0.7219, with the average estimated g-loading of the professional test composite being 0.924.

Here are the correlations for some of the most prominent tests directly:

Test	n	r	rRR	MeanTest	SDTest	MeanAGCT*	SDAGCT*
Composite	58	0.7219	0.8621	132.06	9.98	128.48	9.20
Old SAT	29	0.6964	0.8477	133.21	10.33	130.72	9.11
Old GRE	20	0.7623	0.8755	134.25	10.10	131.55	9.75
CAIT	37	0.6957	0.8378	135.49	11.76	129.49	9.47
WAIS-IV	14	0.6484	0.7515	132.07	12.47	130.64	11.22
SB-V	9	0.7435	0.8644	127.33	9.11	127.11	9.70

Note. r_RR: Since these correlations were calculated on a sample from a portion of the general population, they were corrected for indirect range restriction so that they better approximate the correlations that would be observed in an unrestricted population. Since not every participant took every test, the table shows the corresponding AGCT scores for those who took the specific test.

Ultimately, such similar and strong convergent validity eighty years later using modern batteries shows how robust the AGCT is to the Flynn Effect and how it remains a great test of intelligence.

Note. The reliability of the AGCT under good administration is not less than .95³. The test-retest reliability for the ACCT was .82³. The actual gain in the ACCT score upon retesting was only 1.3 points¹.

Predictive Validity

The Examiner Manual¹ for the AGCT states that, even with substantial restriction of range, the predictive validity of the AGCT is high. For 3,000 clerical trainees, a validity of .40 was reported between AGCT score and success in clerical school. The validity of .40 means that the chances that a subject would do average or better in the course were only 1 in 100 if his AGCT score corresponded to 74 IQ (SD = 15); 5 in 100 for 85 IQ; 20 in 100 for ~100 IQ; 47 in 100 for 113 IQ; and 76 for 125 IQ⁶. Moreover, in officer candidate schools for the various arms and services, AGCT usually correlated with academic grades around .40 as well, despite the rigid selection¹.

The tables below are from Army studies conducted since WW2¹⁶. The AGCT scores are in the original standard deviation of 20.

Mean AGCT Scores per Individual Occupation

AGCT Scores per Individual Occupation	Mean
Accountant	121.1
Lawyer	120.7
Public Relations Man	119.5
Auditor	119.4
Chemist	118.6
Reporter	118.4
Chief Clerk	118.2
Teacher	117.1
Draftsman	116.5
Stenographer	115.8
Pharmacist	115.4
Tabulating Machine Operator	115.1
Bookkeeper	115.0
Manager, Sales	114.3
Purchasing Agent	114.0
Production Manager	113.6
Photographer	113.2
Clerk, General	113.1
Clerk, Typist	112.6
Installer, Telephone and Telegraph	111.9
Cashier	111.9
Instrument Repairman	111.6
Radio Repairman	111.5
Artist	111.2
Manager, Retail Store	110.5
Laboratory Assistant	110.1
Tool Maker	109.4
Stock Clerk	108.9
Musician	108.2
Machinist	107.6
Watchmaker	107.4
Airplane Mechanic	107.0
Sales Clerk	106.9
Electrician	106.8
Lathe Operator	106.4
Receiving and Shipping Checker	105.7
Sheet Metal Worker	105.6
Lineman, Power and Tel. & Tel.	105.3
Auto Service Man	103.2
Riveter	103.1
Cabinetmaker	102.6
Upholsterer	102.5
Butcher	102.2
Plumber	102.0
Bartender	101.7
Carpenter, Construction	101.6
Pipe Fitter	101.4
Welder	101.4
Auto Mechanic	101.0
Molder	100.8
Chauffeur	100.6
Tractor Driver	99.6
Painter, General	98.7
Crane Hoist Operator	98.4
Weaver	97.8
Barber	96.5
Farmer	94.5
Farmhand	93.6
Miner	92.9
Teamster	90.8

per Major Occupational Group

AGCT Scores per Major Occupational Group	Mean
Professional	117.2
Managerial	114.1
Semiprofessional	113.2
Sales	109.1
Clerical	103.3
Skilled	101.3
Semiskilled	99.7
Personal Service	99.0
Agricultural	94.0

per Type of Work

AGCT Scores per Type of Work	Mean
Literary Work	118.9
Technical Work	117.3
Public Service	117.1
Managerial Work	112.8
Artistic Work	112.2
Recording Work	111.8
Public Contact Work	109.1
Musical Work	108.2
Manipulative Work	104.5
Crafts	103.8
Machine Trades	102.6
Observational Work	100.2
Personal Service Work	99.0
Farming	92.9

per Field of Specialization & Respective Degree Level

AGCT Scores per Field of Specialization	Degree Level	10th	25th	50th	75th	90th
Natural Sciences	AB	111	116	121	126	132
	Graduate students	114	119	125	130	135
	PhD	117	123	129	136	144
Chemistry	AB	112	117	123	128	134
	Graduate students	114	120	126	132	136
	PhD	119	124	130	136	143
Physical Sciences, other	AB	112	117	124	129	137
	Graduate students	117	122	127	132	136
	PhD	117	126	132	141	146
Earth Sciences	AB	111	115	120	126	129
	Graduate students	111	116	122	128	133
	PhD	120	125	129	137	145
Biological Sciences	AB	109	114	120	125	130
	Graduate students	113	117	123	129	134
	PhD	115	120	126	132	138
Psychology	AB	110	114	121	126	132
	Graduate students	117	123	128	132	137
	PhD	119	125	132	141	147
Social Sciences	AB	108	113	120	124	129
	Graduate students	111	116	122	129	134
Economics	AB	111	115	120	126	132
	Graduate students	111	116	123	129	134
History	AB	108	114	119	124	129
	Graduate students	111	116	122	127	133
Other Social Sciences	AB	106	111	117	123	128
	Graduate students	111	116	122	129	134
Humanities and Arts	AB	110	115	120	126	131
	Graduate students	111	117	123	129	135
English	AB	111	116	121	127	132
	Graduate students	115	120	126	131	135
Languages	AB	111	116	121	126	132
	Graduate students	111	117	123	130	136
Philosophy and other Humanities	AB	107	114	117	125	129
	Graduate students	113	120	126	132	136
Fine Arts	AB	109	114	120	124	130
	Graduate students	109	114	120	126	132
Engineering	AB	111	117	122	128	134
	Graduate students	114	117	123	129	134
	PhD	116	123	129	137	140
Applied Biology	AB	105	111	116	120	126
	Graduate students	113	117	129	126	131
Agriculture	AB	111	114	118	123	128
	Graduate students	116	120	124	129	133
	PhD	110	116	123	128	133
Home Economics	AB	100	108	114	118	123
	Graduate students	108	112	116	120	123
Health Fields	Graduate students	112	117	123	128	133
Medicine	Medical school students	114	119	124	129	134
Dentistry	Dental school students	109	114	120	126	132
Nursing	AB	110	114	119	126	132
Other	Graduate students	112	117	123	129	134
Business and Commerce	AB	108	113	118	123	128
	Graduate students	109	114	120	125	130
Education	AB	104	111	117	122	126
	Graduate students	109	114	120	125	129
Education, general	AB	105	112	117	123	127
	Graduate students	110	114	120	126	129
Physical Education	AB	99	108	113	118	126
	Graduate students	106	111	115	119	122
Other Fields
Law	Law school graduates	113	115	122	125	130
Social Work	Graduate students	109	114	120	124	129
All Fields Combined (weighted averages)	AB	109	114	120	125	130
	Graduate students	111	116	122	128	133

References

Related Reading. See "Stephen Jay Gould’s Analysis of the Army Beta Test in The Mismeasure of Man: Distortions and Misconceptions Regarding a Pioneering Mental Test" (2019).