Statistics of The Sargasso Test

Norms as of 29 April 2022

• Quality of norms: √N × r² = 11.28
• Weighted mean correlation with tests used for norming: 0.653 (N = 700)
• Threshold correlation to select tests used for norming: 0.53
• Norming method: Rank equation
• Sample-independent hardness: 0.47

Norms are obtained by anchoring the object test to those of the other tests that equal or exceed a threshold correlation with the object test. The threshold is set to maximize the quality of norms. All of the scores on thus selected tests are used, each score forming a pair with a raw score on the object test, so that there are more pairs than initial raw scores. Where needed, the scores from the other tests are converted to protonorms using a relevant conversion formula or table. The raw scores on the object test, which form pairs with the used scores, are equated to the used scores with the above indicated method, and where needed there is interpolation or extrapolation. Non-extrapolated floor scores are normed at the median of their range after equating but reported as Below the measured range of the test; in case of a single outlying non-extrapolated floor score, it is normed low in its range after equating. Possible non-extrapolated ceiling scores are normed at the median of their range after equating when rare, but are normed at the bottom of their range after equating when abundant (ceiling effect); in case of a single outlying non-extrapolated ceiling score, it is normed high in its range after equating.

Scores on The Sargasso Test as of 30 April 2022

Contents type: Verbal, numerical, spatial, logical.   Period: 2007-present

• n: 64
• Median: 33.0
• Quartile deviation: 4.0
• Range: 40
• Maximum possible: 65
• Male median: 33.0 (n = 61)
• Female median: 30.0 (n = 3)
• Resolution: 0.41 (0.58 if half points are frequently given)

Correlation of The Sargasso Test with other tests by I.Q. Tests for the High Range

Weighted average of correlations: 0.538 (N = 1100, weighted sum = 591.84)

Conservatively estimated minimum g loading: 0.73

Ranking in above table is based on the unrounded correlations. All available data is present in this table, no tests are left out except for those with less than 4 score pairs. All known pairs are used, including possible floor/ceiling scores or outliers.

Correlation of The Sargasso Test with tests by others

Weighted average of correlations: 0.318 (N = 108, weighted sum = 34.31)

Ranking in above table is based on the unrounded correlations. All available data is present in this table, no tests are left out except for those with less than 4 score pairs. All known pairs are used, including possible floor/ceiling scores or outliers.

Please be aware that correlations with these external tests are in most cases affected (depressed, typically) by one or more of the following: (1) Little overlap with the object test because of the much lower ceilings and inherent ceiling effects of the tests used in regular psychology; (2) Candidates reporting scores selectively, for instance only the higher ones while withholding lower ones; (3) Candidates reporting, or having been reported by psychometricians, incorrect scores.

Estimated loadings of The Sargasso Test on particular item types

These are estimated g factor loadings, but against homogeneous tests (containing only particular item types) as opposed to non-compound heterogeneous tests. Although tending to surprise the lay person, it is not uncommon for tests to have high loadings on item types they do not actually contain themselves. Such loadings reflect the empirical fact that most tests for mental abilities measure primarily g, regardless of their contents; that the major part of test score variance is caused by g, and only a minor part by factors germane to particular item types. It is of key importance to understand that this is a fact of nature, a natural phenomenon, and not something that was built into the tests by the test constructors.

N = 859

Compound tests have been left out of this table to avoid overlap.

Balanced g loading = 0.69

National medians for The Sargasso Test

For reasons of privacy, only countries with 2 or more candidates are included in this table. Ranking is based on the medians, and then alphabetic.

Correlation with national I.Q.'s of The Sargasso Test

Correlation of this test with national average I.Q.'s published by Lynn and Vanhanen:

• r = -0.01 (n = 63)

Correlation of The Sargasso Test with personal details

Estimated g factor loadings upward and downward of particular scores

In parentheses the number of score pairs on which that estimated g factor loading is based. The goal of this is to verify the hypothesis that g becomes less important, accounts for a smaller proportion of the variance, at higher I.Q. levels. The mere fact of restricting the range like this also depresses the g loading compared to computing it over the test's full range, so it would be normal for both values to be lower than the test's full-range g loading.

Reliability

• Split-half (odd-even) = 0.86
• Split-half (other division) = 0.82
• Cronbach's alpha = 0.84

Remark: These reliability coefficients are low for a stand-alone I.Q. test (.9 or higher is normal) but logical if one considers that the test consists exclusively of "bad" items that were previously removed from other tests.

Error

• Standard error = 2.7 raw score points

Scores by age

N = 63

Scores by year taken

r_{year taken × median score} = -0.00 (N = 64)

Robustness and overall test quality

• Robustness by chronological rank = 0.74
• Robustness by month = 0.74 (r_{raw scores × months} = -0.10)
• Quality = 0.704

Item analysis

Item statistics are not published as that would help future candidates. To detect bad items, answers and comments from candidates are studied, as well as, for each problem, the correlation with total score and the proportion of candidates getting it wrong (hardness of the item). Possible bad items are removed or revised, resulting in a revised version of the test; however, in the present test, bad items are left in as the test was purposely constructed out of bad items previously removed from other tests.

Nevertheless, about 60 of the 65 items display normal statistical behaviour in this test environment.

[More statistical reports]