~0.83 g-loading for RC is kind of laughable. That's clearly false. It's well-known that on modern batteries, tests of fluid reasoning tend to be the most highly g-loaded. Just take a look at the factor structures of the WAIS-5, WJ-V, WISC-V, KABC-II, DAS-II, CORE, etc. etc. etc. It's also very uncommon for individual subtests in a cognitive battery to produce g-loadings above 0.80.
Came here to say the "number-series" quiz is bad and just does the same thing over and over. I understood all 50, but presumably made a typo in one of them because I got 49 out of 50. That put me at 88th percentile among test takers (so 12% of test takers got all 50 correct). Around problem 37 I got bored of doing mental arithmetic and wrote a simple 1-line Python program that handles about 90% of the quiz problems. The patterns need to be more interesting and more varied if they're going to actually test pattern discovery/recognition.
I agree it's bad, but how would you make very difficult number series to satisfy online autists/Mensans? The kinds of tests suitable for them are not so suitable for normal people. The test as it is is too long, so I have made an adaptive version to speed it up. It still lacks high range, so I added some more items, but we have to wait to gather some data on their item parameters until they can be utilized in the adaptive scoring.
ICAR60's "true g-loading" would be 0.9+? How did you conjure such an absurd figure? That's on the order of tests like WAIS and CORE. The authors themselves reported an ωₕ of 0.61, so the g-loading they found is about 0.78. Kind of meh for such a long test.
He's estimating 0.9+ based purely on the fact that most long tests tend to have high g-loadings, but also deliberately contradicting the reported figure, instead of acknowledging that ICAR is crap.
It's based on 4 subtests, and batteries have high g-loadings.
The ICAR validation papers show the test works fine compared to other established tests. A low correlation on the site more likely reflects that a lot of the data are bad because people didn't have patience to sit there for 50 minutes rotating cubes.
when I made tests like this I’d infer language from the user’s brower, which exposes a list of locales, and I’d use the first that wasn’t English (because a user with two language options tends to be ESL, since native English speakers rarely have a second language option)
The questionnaire has a question about native language(s) now, but since it wasn't there originally, we only have this information going forward, about 15k users right now. https://taketest.xyz/demographics
Final results look pretty consistent with the aggregate ranking I made
https://www.technotheoria.org/p/what-ability-best-measures-intelligence
~0.83 g-loading for RC is kind of laughable. That's clearly false. It's well-known that on modern batteries, tests of fluid reasoning tend to be the most highly g-loaded. Just take a look at the factor structures of the WAIS-5, WJ-V, WISC-V, KABC-II, DAS-II, CORE, etc. etc. etc. It's also very uncommon for individual subtests in a cognitive battery to produce g-loadings above 0.80.
That's because those loadings are not corrected for unreliability.
Came here to say the "number-series" quiz is bad and just does the same thing over and over. I understood all 50, but presumably made a typo in one of them because I got 49 out of 50. That put me at 88th percentile among test takers (so 12% of test takers got all 50 correct). Around problem 37 I got bored of doing mental arithmetic and wrote a simple 1-line Python program that handles about 90% of the quiz problems. The patterns need to be more interesting and more varied if they're going to actually test pattern discovery/recognition.
I agree it's bad, but how would you make very difficult number series to satisfy online autists/Mensans? The kinds of tests suitable for them are not so suitable for normal people. The test as it is is too long, so I have made an adaptive version to speed it up. It still lacks high range, so I added some more items, but we have to wait to gather some data on their item parameters until they can be utilized in the adaptive scoring.
Yeah most of the IQ tests on this site are mediocre or straight up junk/AI-generated. Kind of disappointing from such an esteemed researcher.
ICAR60's "true g-loading" would be 0.9+? How did you conjure such an absurd figure? That's on the order of tests like WAIS and CORE. The authors themselves reported an ωₕ of 0.61, so the g-loading they found is about 0.78. Kind of meh for such a long test.
Yes, it should be slightly worse than those. Similar to AFQT.
He's estimating 0.9+ based purely on the fact that most long tests tend to have high g-loadings, but also deliberately contradicting the reported figure, instead of acknowledging that ICAR is crap.
It's based on 4 subtests, and batteries have high g-loadings.
The ICAR validation papers show the test works fine compared to other established tests. A low correlation on the site more likely reflects that a lot of the data are bad because people didn't have patience to sit there for 50 minutes rotating cubes.
when I made tests like this I’d infer language from the user’s brower, which exposes a list of locales, and I’d use the first that wasn’t English (because a user with two language options tends to be ESL, since native English speakers rarely have a second language option)
The questionnaire has a question about native language(s) now, but since it wasn't there originally, we only have this information going forward, about 15k users right now. https://taketest.xyz/demographics