Needlessly hostile article. Confusing to call him Bird when there's already a communist Kevin Bird in this area. IB would be better, which also happens to be a Danish male first name. :)
"We use family-level fixed effects when estimating the bivariate association between birth order and these outcomes, meaning we are comparing siblings within the same family."
It cannot be a between family confounding factor, as you are suggesting.
Shared environment cannot differ between siblings. By definition, shared environment is whatever causal forces that make family members more similar that aren't genetic.
B_o (no maternal age) = B_o (maternal age controlled for) + Cov(Order + family + sex, maternal age) * Cov(Maternal age, education)
Indeed, Cov(order, ma) should be positive, and Cov(ma, education) should be positive. So B_o(no ma) > B_o (ma controlled for). This is what we see. AFAIK, fixed effects is not a replacement for controlling for individual gene scores.
As for article hostility, I tried not to use any ad hominems, I don't know how it is hostile other than I disagree with him. It would be nicer if I did not say he "strawmanned Plomin", but I really believe this, so what else can I say? I would be nicest if I did not publish a criticism of his post, but I believe this criticism is accurate.
You don't need to add variables for any parental values that are the same for siblings because they are already adjusted for using the fixed family term. Thus, there is no need to control for parental IQ/EA/PGS. But parental age at birth will vary between siblings, and thus can be entered into the model.
I thought about it, instead of dummy variable, fixed effects can be conceptualized as sib mean, right (this should be equivalent to the sibs all have the same binary dummy var ticked, right?)? So the model just looks like predicting individual EA with regression, throwing in an individual sex, sib mean (family fixed effect), maternal age at birth, and birth order effect.
OVB: Short equals long plus {the effect(s) of omitted times the regression(s) of omitted on included}
effect of omitted is positive (Ma on EA).
effect of regression omitted on included means how the Sibmean + Birthorder + Sex term goes up with Ma. This should be positive; higher Ma means higher sibmean and higher birthorder.
Thus, short is greater than long. Which is what we see, short is birth order effect with no Ma in the model, this is less negative than the long version.
If I'm right about what fixed effects means in this paper, it's not a substitute for observing genetics.
Edit: From "Mostly Harmless Econometrics", the dummy variable approach is the right representation, and this is apparently equivalent to prediction EA-family mean EA using all the variables minus their family means.
I have a slight disagreement with this: "Parents basically do not influence IQ at all, except through genes." I think it's worth qualifying this statement. It shows up in adoption studies within modern, Western countries, which produces a selection bias. The adoption process filters out a significant amount of maladaptive parental traits and ensures that the kids are planned.
If I recall correctly, black children who are adopted have higher IQs than ones who are not. This is despite the fact that the black kids put up for adoption likely have worse overall genetic and early environment factors than even average black children have.
Essentially, filtering out the worst environments does make a big difference, and I think it's worth mentioning. (Also, parental definitely matters when you're takking wbout global variation, though I doubt you'd disagree with this.)
Paternal age affecting rates of autism got me wondering (not really related to this specific article). We know that autism is inheritable, right? Could it be that autistic men are just less likely to find a partner at a younger age, partly because of social issues, partly because they hyperfocus more on their own interests rather than hunting for a mate. This could influence our view of how much paternal age really does affect rates of autism, since autistic men might just be more likely to have children later in life, with the same rate of heritability as autism always has. Could of course also be a combination of both, but older fathers might not have to worry so much about getting autistic children as the statistics might make them believe.
wanna hear a joke? how do U redpill people?
U cant because of thier mutationbal load hahahahahHAhaha
Needlessly hostile article. Confusing to call him Bird when there's already a communist Kevin Bird in this area. IB would be better, which also happens to be a Danish male first name. :)
https://academic.oup.com/pnasnexus/article/1/2/pgac051/6604844?login=false#381345265
"We use family-level fixed effects when estimating the bivariate association between birth order and these outcomes, meaning we are comparing siblings within the same family."
It cannot be a between family confounding factor, as you are suggesting.
Shared environment cannot differ between siblings. By definition, shared environment is whatever causal forces that make family members more similar that aren't genetic.
>It cannot be a between family confounding factor, as you are suggesting.
I could be wrong here, but per my understanding, family fixed effects just means using a dummy variable to denote family. This will capture the the family phenotype mean, roughly. https://towardsdatascience.com/fixed-effect-regression-simply-explained-ab690bd885cf
We get 2 models, education =~ Sex*B_S + Family_vector * Family_weights_vector + Order * B_o and education =~ Sex*B_S + Family_vector * Family_weights_vector + Order * B_o + Maternal_age * B_m.
They are plotting Order * B_o in these models. For some reason it's not linear when they don't have maternal age. I don't know why this is.
B_o gets more negative when they throw in maternal age. They don't show us what happens to the family weights.
From the OVB formula, (https://www.masteringmetrics.com/wp-content/uploads/2020/07/lny20n08MRU_R2.pdf)
B_o (no maternal age) = B_o (maternal age controlled for) + Cov(Order + family + sex, maternal age) * Cov(Maternal age, education)
Indeed, Cov(order, ma) should be positive, and Cov(ma, education) should be positive. So B_o(no ma) > B_o (ma controlled for). This is what we see. AFAIK, fixed effects is not a replacement for controlling for individual gene scores.
As for article hostility, I tried not to use any ad hominems, I don't know how it is hostile other than I disagree with him. It would be nicer if I did not say he "strawmanned Plomin", but I really believe this, so what else can I say? I would be nicest if I did not publish a criticism of his post, but I believe this criticism is accurate.
You don't need to add variables for any parental values that are the same for siblings because they are already adjusted for using the fixed family term. Thus, there is no need to control for parental IQ/EA/PGS. But parental age at birth will vary between siblings, and thus can be entered into the model.
I thought about it, instead of dummy variable, fixed effects can be conceptualized as sib mean, right (this should be equivalent to the sibs all have the same binary dummy var ticked, right?)? So the model just looks like predicting individual EA with regression, throwing in an individual sex, sib mean (family fixed effect), maternal age at birth, and birth order effect.
OVB: Short equals long plus {the effect(s) of omitted times the regression(s) of omitted on included}
effect of omitted is positive (Ma on EA).
effect of regression omitted on included means how the Sibmean + Birthorder + Sex term goes up with Ma. This should be positive; higher Ma means higher sibmean and higher birthorder.
Thus, short is greater than long. Which is what we see, short is birth order effect with no Ma in the model, this is less negative than the long version.
If I'm right about what fixed effects means in this paper, it's not a substitute for observing genetics.
Edit: From "Mostly Harmless Econometrics", the dummy variable approach is the right representation, and this is apparently equivalent to prediction EA-family mean EA using all the variables minus their family means.
This comes out to a formula you can find here https://chat.openai.com/share/05cb70c9-9379-432e-a0b9-986ba3455257
Seems this will indeed give a positive OVB correction term per GPT simulation.
I have a slight disagreement with this: "Parents basically do not influence IQ at all, except through genes." I think it's worth qualifying this statement. It shows up in adoption studies within modern, Western countries, which produces a selection bias. The adoption process filters out a significant amount of maladaptive parental traits and ensures that the kids are planned.
If I recall correctly, black children who are adopted have higher IQs than ones who are not. This is despite the fact that the black kids put up for adoption likely have worse overall genetic and early environment factors than even average black children have.
Essentially, filtering out the worst environments does make a big difference, and I think it's worth mentioning. (Also, parental definitely matters when you're takking wbout global variation, though I doubt you'd disagree with this.)
Paternal age affecting rates of autism got me wondering (not really related to this specific article). We know that autism is inheritable, right? Could it be that autistic men are just less likely to find a partner at a younger age, partly because of social issues, partly because they hyperfocus more on their own interests rather than hunting for a mate. This could influence our view of how much paternal age really does affect rates of autism, since autistic men might just be more likely to have children later in life, with the same rate of heritability as autism always has. Could of course also be a combination of both, but older fathers might not have to worry so much about getting autistic children as the statistics might make them believe.