When it comes to the effect of education on intelligence, mostly I feel confused, and this post doesn't fully solve my confusion. Here's my understanding:
I'll start with the basics; what is cognition and the g factor, and why is it important? As a simplified model, we might say that people face various tasks in their life; for instance maybe they need to go buy some food, or they need to solve some sort of issue at work, or whatever. Some of these tasks are cognitive tasks, in the sense that performing them is significantly dependent on intelligence; e.g. "abstraction, logic, understanding, self-awareness, learning, emotional knowledge, reasoning, planning, creativity, critical thinking, and problem-solving".
For any given cognitive task, people vary in their ability to solve it. This can happen for a lot of reasons, e.g. task familiarity or domain-specific knowledge, or more generally applicable skills and biological (especially brain) factors that are useful for the task. But there are a lot of cognitive tasks, and so that naturally raises the question of, what are the usual patterns in individual differences in task solving ability?
And the overwhelmingly important and dominant finding here is then the g factor: There is variation in a cognitive process or set of processes which is useful for just about any task you can think of. This variation appears to be mostly biological in origin. Why is this important? Because the modern world is so heavily dependent on cognitive abilities, and because the g factor is so broad, it becomes one of the biggest influences on outcomes of interest, in particular on "broad" outcomes. Whenever there is some outcome we want to study, there are likely some cognitive challenges that people need to face in order to achieve it, and those cognitive challenges will always depend fairly strongly on the g factor, so g automatically becomes important for the outcome.
I believe the above is pretty standard, and pretty compatible with your post? Now the parts that confuse me or which I think are less compatible with your post:
According to the above argument, g is primarily important because it improves your cognitive performance in a variety of real-world scenarios. But this seems like it implies that if there is any other factor, which also improves your cognitive performance in a similar variety of real-world scenarios, then it should also similarly improve your outcomes. It doesn't have to have the exact same pattern of associations as g has, or have the exact same mechanisms as g has, presumably the main thing that matters is that it improves your performance sufficiently broadly and sufficiently much.
Now suppose we take the first study you looked at in this post. It claims that education improves scores across all of the test types they tested, and that it does so in a persistent way. If it is right, then I think whether it is "on g" shouldn't matter for whether it has knock-on effects on things like economic productivity, because per the above argument even non-g factors which have broad effects should still be useful in practice. (That doesn't mean that it doesn't matter at all whether it is "on g". After all, whether the factor is "on g" can be useful for understanding other topics, such as the nature of race differences in intelligence; if education is off g but race differences are on g, then differences in education cannot be the cause of differences in intelligence.)
After analyzing this study, in the rest of your post, you argue that education doesn't improve real-world outcomes. This would be weird according to my theory, except you also argue that education doesn't improve test scores on all tests, which makes it make sense. Maybe education improves some abilities that are not useful, while there are other abilities that are useful that don't get improved by education. I'm not sure I buy the idea though, since the claims that there are such tests seem kind of sketch in some ways (e.g. reliant on multiple comparisons). But supposing we grant it, it seems like that should make it a huge priority to gather a battery of such tests, since they can presumably be more valid, and causally cleaner.
> "I think whether it is "on g" shouldn't matter for whether it has knock-on effects on things like economic productivity, because per the above argument even non-g factors which have broad effects should still be useful in practice."
Non-g factors aren't well correlated with measures of life outcome. This is very well known that nearly all the correlation between life outcome and IQ that are usually talked about are nearly entirely driven by g. If you take the least g-loaded items of an IQ tests, you get less correlation with life outcome.
Emil probably wrote this post assuming it was a known fact.
It is correct that the degree to which a cognitive indicator is associated with outcomes is very tightly correlated to its g-loading. However, the flaw in your argument is in assuming that this means non-g factors don't matter. In actuality, the model I described in my comment *also* predicts a very strong correlation between g-loadings and predictor-outcome correlations, even though my model doesn't count the IQ points you gain from education any lower than it counts the IQ points you get due to g.
To prove it, here is a simulation, which runs my model and still finds the classical pattern of education being negatively g-loaded and outcomes being positively g-loaded, despite education improving outcomes through improved cognition.
There is something I don't understand. Why do you have
edu = g * G_TO_EDU_EFFECT + np.random.normal(0, np.sqrt(1-G_TO_EDU_EFFECT**2), SAMPLE_SIZE)?
The design of the study is that some people randomly (at least the hope is it was random enough) get more year of education. You should instead have something like:
edu = np.random.normal(0, 1, SAMPLE_SIZE)
Because the added year of education aren't from "g" but from an external froce: the government.
The "g * G_TO_EDU_EFFECT" represents that smarter (higher g) people get more education. How much education you get isn't purely random, it's also dependent on how smart you are.
(I don't think this has any particular effect on the result, though; you could set G_TO_EDU_EFFECT to 0, and I think that would still yield the basic observations that even though education improves the outcome, 1) education's impact on IQ is perfectly negatively g-loaded, and 2) the outcome-ability correlations are perfectly positively g-loaded.
But we're talking about a study and situation where education would not be correlation to g! We're talking about the effects of increasing people's Education alone. By having G causally related to Education, you're just making a model where the outcome-Edu correlation isn't necessarily causal, but just a correlation
Setting G_TO_EDU_EFFECT = 0 actually kills most of the correlation between outcome and education though.
And with the remaining correlation being weak enough that the simplicity and linearity of the assumtions may be the cause. For example, the simplicity of the model made it so the least g-loaded still have a very big 0.2 correlation with outcome. I don't know whether this is true in real life. (Don't have the time right now to check tbh).
For example, if at line 23 I write "factors[:,0] = 0" it will make it so the first ability is just a random normally distributed thing, right? like something with 0 G-loading? Yet it is still correlated to outcome. So if we add a random item into an IQ test (like the lenghth of your nostril hairs) it'll still have a 0.08 correlation with outcome? This looks like what your model says.
I think you are misunderstanding my view and my claims.
I absolutely agree that the correlation between education and outcomes is not just due to education improving outcomes, but is also confounded by g (and that this confounding shouldn't be considered a benefit of education); that is precisely *why* I put in the "g * G_TO_EDU_EFFECT" term, to represent this confounding. As such, I also agree that the correlation between education and outcomes in my model is not just due to education improving outcomes.
"Setting G_TO_EDU_EFFECT = 0 actually kills most of the correlation between outcome and education though.
And with the remaining correlation being weak enough that the simplicity and linearity of the assumtions may be the cause."
It is correct that setting G_TO_EDU_EFFECT = 0 reduces the correlation between the outcome and education due to removing the confounded fragment.
However, it is incorrect to think that the remaining correlation between education and outcome is due to confounding. You can read it off in the code; on line 20, education influences the ability factors, and on line 22, the ability factors influence overall ability, and on line 23, overall ability influences the outcomes. So you've got a causal chain: education -> ability factors -> overall ability ("iq") -> outcome. This causal chain implies that education must improve the outcome.
"For example, if at line 23 I write "factors[:,0] = 0" it will make it so the first ability is just a random normally distributed thing, right? like something with 0 G-loading?"
By line 23, the factors have already asserted their effect, so that is a bad place to put it. Instead I recommend between line 14 and 15 to set G_LOADINGS[0] = 0, and between line 15 and 16 to set EDU_LOADINGS[0] = 0. This would make the first ability a random normally distributed thing.
"So if we add a random item into an IQ test (like the lenghth of your nostril hairs) it'll still have a 0.08 correlation with outcome? This looks like what your model says."
You seem to be misunderstanding my model. Partly this is my fault because it's a simplified model, so while I believe different abilities have different degrees of effect on outcomes, these differences have not been included in the model. So if you want a modified model that also describes variation in the relationship between different abilities and outcomes of interest, then feel free to request that and I will make it.
But let's just have a natural-language explanation of what I would make:
Different abilities differ in their effect on outcomes based on what real-world tasks people encounter, how much people's real-world tasks depend on the abilities, and how much people vary in their abilities.
So for instance, people who know how to program are likely to work as programmers, and their various programming skills like debugging ability etc. vary a lot and influence what sorts of tasks they can solve and how quickly they can solve them, and in effect influences what jobs they get and how much money they make. This is _heavily_ influenced by g, but it's one of the _mechanisms_ by which g acts; if there was a non-g way for a programmer to improve their skills, then that would also improve their outcomes.
As another example, consider ability to play the 1992 video game "Night Trap". It's not useful for anything; there aren't even progaming contests you could play it in to win prizes AFAIK, so "Night Trap" ability won't improve your outcomes, even if it is g-loaded (as it likely is, just like any other cognitive ability). However, since it is g-loaded, it will likely still correlate with your outcomes, because g improves your outcomes.
However, consider something like hard work. Conscientiousness doesn't correlate with g, but conscientiousness still contributes to positive outcomes, because it makes you make safer and more productive decisions, even if they aren't smarter per se. So conscientiousness would be an example of a non-g-loaded variable which improves your outcomes.
On the other hand, yes, there are things like the length of your nostril hairs, which is neither g-loaded nor improves your outcomes. Probably most variables are like this. I didn't include such variables in my simulation for efficiency reasons, but if you would like me to, I can add them in another simulation.
Possible way out: Education may improve real world outcomes that are not associated with g or intelligence, but are often associated with non-cognitive skills, for example socialization and empathy, which can impact physical health and longevity.
I think that is the opposite of a way out? Emil's post semi-successfully argues that education does not do much to improve real-world outcomes. However, I'm not sure I buy his argument that education does not improve intelligence. So the paradox is that education does improve intelligence but does not improve outcomes, whereas your proposal gives a mechanism by which education could improve outcomes without improving intelligence.
I do not know if he is referring to fluid intelligence or crystalline intelligence, where former is g-strong while the latter is education-strong (or grit-strong, in reference to Freddie DeBoer).
Physics nitpicking here: having an absolute zero doesn't imply a meaningful or well-defined multiplication operation. In fact, the example of the Kelvin scale is good for illustrating this: as it is a little-known fact that negative Kelvin degrees are not only theoretically possible but have also been obtainable in the lab for several decades now - and negative Kelvin degrees are actually hotter than any finite positive-Kelvin temperature. So clearly -10K is not -1 hotter than 10K. As it happens, temperature just isn't a concept with a natural preferred scale - e.g., there is no objective advantage of using Kelvin over Rankine.
Just the blanket term education implies a "good" education which is a joke. How many bad teachers have people had who were pretty much worthless or even destructive to teaching anything worthwhile? On the other hand I believe that constantly reading on one's own, especially good books actually does increase IQ. A person who reads constantly will undoubtedly be quicker on the draw mentally to his peers who do not. I mean think about it. No brainer. Reading is to the mind what exercise is to the body. Case and point is the life story of Dr. Ben Carson the famous brain surgeon turned Department Of Housing Secretary. Carson hailed from the poorest of slums of Detroit with no father. How did he succeed? His mother made him take out any two books from the library he wanted every week and do a book report on them. Engendering a life long love of reading early is the hugest factor in increasing mental acuity.
Yep. The Flynn Effect is essentially this, scaled up to the level of entire populations.
When it comes to the effect of education on intelligence, mostly I feel confused, and this post doesn't fully solve my confusion. Here's my understanding:
I'll start with the basics; what is cognition and the g factor, and why is it important? As a simplified model, we might say that people face various tasks in their life; for instance maybe they need to go buy some food, or they need to solve some sort of issue at work, or whatever. Some of these tasks are cognitive tasks, in the sense that performing them is significantly dependent on intelligence; e.g. "abstraction, logic, understanding, self-awareness, learning, emotional knowledge, reasoning, planning, creativity, critical thinking, and problem-solving".
For any given cognitive task, people vary in their ability to solve it. This can happen for a lot of reasons, e.g. task familiarity or domain-specific knowledge, or more generally applicable skills and biological (especially brain) factors that are useful for the task. But there are a lot of cognitive tasks, and so that naturally raises the question of, what are the usual patterns in individual differences in task solving ability?
And the overwhelmingly important and dominant finding here is then the g factor: There is variation in a cognitive process or set of processes which is useful for just about any task you can think of. This variation appears to be mostly biological in origin. Why is this important? Because the modern world is so heavily dependent on cognitive abilities, and because the g factor is so broad, it becomes one of the biggest influences on outcomes of interest, in particular on "broad" outcomes. Whenever there is some outcome we want to study, there are likely some cognitive challenges that people need to face in order to achieve it, and those cognitive challenges will always depend fairly strongly on the g factor, so g automatically becomes important for the outcome.
I believe the above is pretty standard, and pretty compatible with your post? Now the parts that confuse me or which I think are less compatible with your post:
According to the above argument, g is primarily important because it improves your cognitive performance in a variety of real-world scenarios. But this seems like it implies that if there is any other factor, which also improves your cognitive performance in a similar variety of real-world scenarios, then it should also similarly improve your outcomes. It doesn't have to have the exact same pattern of associations as g has, or have the exact same mechanisms as g has, presumably the main thing that matters is that it improves your performance sufficiently broadly and sufficiently much.
Now suppose we take the first study you looked at in this post. It claims that education improves scores across all of the test types they tested, and that it does so in a persistent way. If it is right, then I think whether it is "on g" shouldn't matter for whether it has knock-on effects on things like economic productivity, because per the above argument even non-g factors which have broad effects should still be useful in practice. (That doesn't mean that it doesn't matter at all whether it is "on g". After all, whether the factor is "on g" can be useful for understanding other topics, such as the nature of race differences in intelligence; if education is off g but race differences are on g, then differences in education cannot be the cause of differences in intelligence.)
After analyzing this study, in the rest of your post, you argue that education doesn't improve real-world outcomes. This would be weird according to my theory, except you also argue that education doesn't improve test scores on all tests, which makes it make sense. Maybe education improves some abilities that are not useful, while there are other abilities that are useful that don't get improved by education. I'm not sure I buy the idea though, since the claims that there are such tests seem kind of sketch in some ways (e.g. reliant on multiple comparisons). But supposing we grant it, it seems like that should make it a huge priority to gather a battery of such tests, since they can presumably be more valid, and causally cleaner.
> "I think whether it is "on g" shouldn't matter for whether it has knock-on effects on things like economic productivity, because per the above argument even non-g factors which have broad effects should still be useful in practice."
Non-g factors aren't well correlated with measures of life outcome. This is very well known that nearly all the correlation between life outcome and IQ that are usually talked about are nearly entirely driven by g. If you take the least g-loaded items of an IQ tests, you get less correlation with life outcome.
Emil probably wrote this post assuming it was a known fact.
It is correct that the degree to which a cognitive indicator is associated with outcomes is very tightly correlated to its g-loading. However, the flaw in your argument is in assuming that this means non-g factors don't matter. In actuality, the model I described in my comment *also* predicts a very strong correlation between g-loadings and predictor-outcome correlations, even though my model doesn't count the IQ points you gain from education any lower than it counts the IQ points you get due to g.
To prove it, here is a simulation, which runs my model and still finds the classical pattern of education being negatively g-loaded and outcomes being positively g-loaded, despite education improving outcomes through improved cognition.
https://pastebin.com/aHYh6nH7
There is something I don't understand. Why do you have
edu = g * G_TO_EDU_EFFECT + np.random.normal(0, np.sqrt(1-G_TO_EDU_EFFECT**2), SAMPLE_SIZE)?
The design of the study is that some people randomly (at least the hope is it was random enough) get more year of education. You should instead have something like:
edu = np.random.normal(0, 1, SAMPLE_SIZE)
Because the added year of education aren't from "g" but from an external froce: the government.
The "g * G_TO_EDU_EFFECT" represents that smarter (higher g) people get more education. How much education you get isn't purely random, it's also dependent on how smart you are.
(I don't think this has any particular effect on the result, though; you could set G_TO_EDU_EFFECT to 0, and I think that would still yield the basic observations that even though education improves the outcome, 1) education's impact on IQ is perfectly negatively g-loaded, and 2) the outcome-ability correlations are perfectly positively g-loaded.
But we're talking about a study and situation where education would not be correlation to g! We're talking about the effects of increasing people's Education alone. By having G causally related to Education, you're just making a model where the outcome-Edu correlation isn't necessarily causal, but just a correlation
Setting G_TO_EDU_EFFECT = 0 actually kills most of the correlation between outcome and education though.
And with the remaining correlation being weak enough that the simplicity and linearity of the assumtions may be the cause. For example, the simplicity of the model made it so the least g-loaded still have a very big 0.2 correlation with outcome. I don't know whether this is true in real life. (Don't have the time right now to check tbh).
For example, if at line 23 I write "factors[:,0] = 0" it will make it so the first ability is just a random normally distributed thing, right? like something with 0 G-loading? Yet it is still correlated to outcome. So if we add a random item into an IQ test (like the lenghth of your nostril hairs) it'll still have a 0.08 correlation with outcome? This looks like what your model says.
I think you are misunderstanding my view and my claims.
I absolutely agree that the correlation between education and outcomes is not just due to education improving outcomes, but is also confounded by g (and that this confounding shouldn't be considered a benefit of education); that is precisely *why* I put in the "g * G_TO_EDU_EFFECT" term, to represent this confounding. As such, I also agree that the correlation between education and outcomes in my model is not just due to education improving outcomes.
"Setting G_TO_EDU_EFFECT = 0 actually kills most of the correlation between outcome and education though.
And with the remaining correlation being weak enough that the simplicity and linearity of the assumtions may be the cause."
It is correct that setting G_TO_EDU_EFFECT = 0 reduces the correlation between the outcome and education due to removing the confounded fragment.
However, it is incorrect to think that the remaining correlation between education and outcome is due to confounding. You can read it off in the code; on line 20, education influences the ability factors, and on line 22, the ability factors influence overall ability, and on line 23, overall ability influences the outcomes. So you've got a causal chain: education -> ability factors -> overall ability ("iq") -> outcome. This causal chain implies that education must improve the outcome.
"For example, if at line 23 I write "factors[:,0] = 0" it will make it so the first ability is just a random normally distributed thing, right? like something with 0 G-loading?"
By line 23, the factors have already asserted their effect, so that is a bad place to put it. Instead I recommend between line 14 and 15 to set G_LOADINGS[0] = 0, and between line 15 and 16 to set EDU_LOADINGS[0] = 0. This would make the first ability a random normally distributed thing.
"So if we add a random item into an IQ test (like the lenghth of your nostril hairs) it'll still have a 0.08 correlation with outcome? This looks like what your model says."
You seem to be misunderstanding my model. Partly this is my fault because it's a simplified model, so while I believe different abilities have different degrees of effect on outcomes, these differences have not been included in the model. So if you want a modified model that also describes variation in the relationship between different abilities and outcomes of interest, then feel free to request that and I will make it.
But let's just have a natural-language explanation of what I would make:
Different abilities differ in their effect on outcomes based on what real-world tasks people encounter, how much people's real-world tasks depend on the abilities, and how much people vary in their abilities.
So for instance, people who know how to program are likely to work as programmers, and their various programming skills like debugging ability etc. vary a lot and influence what sorts of tasks they can solve and how quickly they can solve them, and in effect influences what jobs they get and how much money they make. This is _heavily_ influenced by g, but it's one of the _mechanisms_ by which g acts; if there was a non-g way for a programmer to improve their skills, then that would also improve their outcomes.
As another example, consider ability to play the 1992 video game "Night Trap". It's not useful for anything; there aren't even progaming contests you could play it in to win prizes AFAIK, so "Night Trap" ability won't improve your outcomes, even if it is g-loaded (as it likely is, just like any other cognitive ability). However, since it is g-loaded, it will likely still correlate with your outcomes, because g improves your outcomes.
However, consider something like hard work. Conscientiousness doesn't correlate with g, but conscientiousness still contributes to positive outcomes, because it makes you make safer and more productive decisions, even if they aren't smarter per se. So conscientiousness would be an example of a non-g-loaded variable which improves your outcomes.
On the other hand, yes, there are things like the length of your nostril hairs, which is neither g-loaded nor improves your outcomes. Probably most variables are like this. I didn't include such variables in my simulation for efficiency reasons, but if you would like me to, I can add them in another simulation.
Possible way out: Education may improve real world outcomes that are not associated with g or intelligence, but are often associated with non-cognitive skills, for example socialization and empathy, which can impact physical health and longevity.
I think that is the opposite of a way out? Emil's post semi-successfully argues that education does not do much to improve real-world outcomes. However, I'm not sure I buy his argument that education does not improve intelligence. So the paradox is that education does improve intelligence but does not improve outcomes, whereas your proposal gives a mechanism by which education could improve outcomes without improving intelligence.
I do not know if he is referring to fluid intelligence or crystalline intelligence, where former is g-strong while the latter is education-strong (or grit-strong, in reference to Freddie DeBoer).
The first study Emil looked at found education to improve both.
Physics nitpicking here: having an absolute zero doesn't imply a meaningful or well-defined multiplication operation. In fact, the example of the Kelvin scale is good for illustrating this: as it is a little-known fact that negative Kelvin degrees are not only theoretically possible but have also been obtainable in the lab for several decades now - and negative Kelvin degrees are actually hotter than any finite positive-Kelvin temperature. So clearly -10K is not -1 hotter than 10K. As it happens, temperature just isn't a concept with a natural preferred scale - e.g., there is no objective advantage of using Kelvin over Rankine.
Just the blanket term education implies a "good" education which is a joke. How many bad teachers have people had who were pretty much worthless or even destructive to teaching anything worthwhile? On the other hand I believe that constantly reading on one's own, especially good books actually does increase IQ. A person who reads constantly will undoubtedly be quicker on the draw mentally to his peers who do not. I mean think about it. No brainer. Reading is to the mind what exercise is to the body. Case and point is the life story of Dr. Ben Carson the famous brain surgeon turned Department Of Housing Secretary. Carson hailed from the poorest of slums of Detroit with no father. How did he succeed? His mother made him take out any two books from the library he wanted every week and do a book report on them. Engendering a life long love of reading early is the hugest factor in increasing mental acuity.