Many of these questions are rather poor. In this case, it's a binary choice, so guessing at random gives a 50% chance of getting it right. This severely limits the g-loading. Multiple choice questions should have many alternative options, not below 4.
RE realised fertility, since the data is from a dating site it's not going to include married parents. I assume % single parents differs by race and IQ, which might explain why results for some races appeared to be reversed from intentions, and some of the IQ difference. Doesn't affect the main point about preferred number of kids, though.
"The sex interaction wasn't seen -- that female dysgenics is stronger -- but probably just due to low power (standard error is 0.046)."
I think there's a good chance that the measure of intelligence can be improved to get more power.
You should try to see if the OK cupid dataset has a decent measure of job status. I forget whether this works for income too (maybe that could be used instead), but since g is the source of all predictive validity in predicting job status, you could just run a linear regression predicting job status from the 14 items, and this would be the weighting of the items which yields the best estimate of g. It should be even better even than a weighting derived from factor analyzing the items themselves, as this still suffers from the problem of restricted diversity of test content, but the predictive validity (and hence g-loading) of a content domain isn't affected by what content domains you've chosen to measure.
I wouldn't try this with education though. It may not work for this purpose due to the non-g gains it causes; g is only the source of predictive validity when pre-dicting the amount of education students will obtain which they don't actually have yet, but using it retro-dictively won't work the same way.
I laughed at the "mile or kilometer" question. An European person would be like WTF is this?
Many of these questions are rather poor. In this case, it's a binary choice, so guessing at random gives a 50% chance of getting it right. This severely limits the g-loading. Multiple choice questions should have many alternative options, not below 4.
How many % failed 'earth or the sun'?
RE realised fertility, since the data is from a dating site it's not going to include married parents. I assume % single parents differs by race and IQ, which might explain why results for some races appeared to be reversed from intentions, and some of the IQ difference. Doesn't affect the main point about preferred number of kids, though.
"The sex interaction wasn't seen -- that female dysgenics is stronger -- but probably just due to low power (standard error is 0.046)."
I think there's a good chance that the measure of intelligence can be improved to get more power.
You should try to see if the OK cupid dataset has a decent measure of job status. I forget whether this works for income too (maybe that could be used instead), but since g is the source of all predictive validity in predicting job status, you could just run a linear regression predicting job status from the 14 items, and this would be the weighting of the items which yields the best estimate of g. It should be even better even than a weighting derived from factor analyzing the items themselves, as this still suffers from the problem of restricted diversity of test content, but the predictive validity (and hence g-loading) of a content domain isn't affected by what content domains you've chosen to measure.
I wouldn't try this with education though. It may not work for this purpose due to the non-g gains it causes; g is only the source of predictive validity when pre-dicting the amount of education students will obtain which they don't actually have yet, but using it retro-dictively won't work the same way.
No good data on job status or education, and many are too young anyway.