This is a good point. I don't think it should make that much of a difference given how young LessWrong is on average, but it can't hurt to try. 

My two problem are 1) finding SAT statistics for nationally representative samples, and not just seniors that take the SAT (the latter are obviously selected) is difficult, and 2) I’d need more detailed data than just the SAT averages—I'd have to adjust each person’s SAT z-score based on the year they took the test.

By this logic, if rationalists are selected based on IQ and not height, and the average rationalist height is +1.85 SD, then we’d have to assume that rationalists’ IQ is +9.25 SD (assuming an IQ-height correlation of 0.2), which is, of course, impossible.

For another example of why this logic doesn’t work, consider this: if you have a variable that is uncorrelated with IQ (r = 0), and rationalists are just slightly above average for that variable, then we'd be forced to conclude that rationalists are infinitely smart (or, if they're below average, infinitely dumb) depending on the direction of the deviation. This is clearly nonsensical.

For an explanation of why this logic doesn't work, see my reply to Unnamed's comment. And for the correct calculations, see my reply to faul_sname's comment.

Your argument assumes a uniform prior, but a Gaussian prior is more realistic in this case. In practice, IQ scores are distributed normally, so it's more likely that someone with a high SAT score comes from a more common IQ range than from a very high outlier. For example, say the median rationalist has an SAT score of +2 SD (chosen for ease of computation), and the SAT-IQ correlation is 0.80. The IQ most likely to produce an SAT of +2 SD is 137.5 (+2.5 SD). However, IQs of 137.5 are rare (99.4%-ile). While lower IQs are less likely to achieve such a high SAT score, there are more people in the lower IQ ranges, making it more probable that someone with a +2 SD SAT score falls into a lower IQ bracket.

This shift between the MLE (Maximum Likelihood Estimate) and MAP (Maximum A Posteriori) estimates is illustrated in the graph, where the MLE estimate would be +2.5 SD, but the MAP estimate, accounting for the Gaussian prior, is closer to +1.6 SD, as expected. (You may also be interested in my reply to faul_sname's comment.)

Eric Neyman is right. They are both valid! 

In general, if we have two vectors $X$ and $Y$ which are jointly normally distributed, we can write the joint mean $\mu$ and the joint covariance matrix $\Sigma$ as

The conditional distribution for $Y$ given $X$ is given by $Y|X\sim N\left({\mu }_{Y|X},K_{Y|X}\right)$,
defined by conditional mean 

${\mu }_{Y|X}={\mu }_Y+K_{YX}{K_{XX}}^{-1}(X-{\mu }_X)$

and conditional variance 

$K_{Y|X}=K_{YY}{K_{XX}}^{-1}{K}_{XY}$

Our conditional distribution for the IQ of the median rationalist, given their SAT score is $N(0+(0.8\cdot 1\cdot (2.42-0)),1-(0.8\ast 1\ast 0.8\left)\right)=N(1.94,0.36)$ 
(That's a mean of 129 and a standard deviation of 9 IQ points.)

Our conditional distribution for the IQ of the median rationalist, given their height is $N(0+(0.2\ast 1\ast (1.85-0)),1-(0.2\ast 1\ast 0.2\left)\right)=N(0.37,0.96)$
(That's a mean of 106 and a standard deviation of 14.7 IQ points.)

Our conditional distribution for the IQ of the median rationalist, given their SAT score and height is $N(0+\left(\left[\begin{array}{cc}0.8& 0.2\end{array}\right]{\left[\begin{array}{cc}1& 0.16\\ 0.16& 1\end{array}\right]}^{-1}\left[\begin{array}{c}2.42\\ 1.85\end{array}\right]\right),1-\left(\left[\begin{array}{cc}0.8& 0.2\end{array}\right]{\left[\begin{array}{cc}1& 0.16\\ 0.16& 1\end{array}\right]}^{-1}\left[\begin{array}{c}0.8\\ 0.2\end{array}\right]\right))=N(2.04,0.35)$(That's a mean of 131 and a standard deviation of 8.9 IQ points)

Unfortunately, since men are taller than women, and rationalists are mostly male, we can't use the height as-is when estimating the IQ of the median rationalist (maybe normalizing height within each sex would work?). 

You're missing the point. While I agree that we don't want to select too hard for personality traits, the bigger problem is that we're not able to robustly select for personality traits the way we're able to select for IQ. If you try to select for Extraversion, you may end up selecting for people particularly prone to social desirability bias. This isn't a Goodhart thing; the way our personality tests are currently constructed means that all the personality traits have fairly large correlations with social desirability, which is not what you want to select for. Also, the specific personality traits our tests measure don't seem real in the same way IQ is real (that's what testing for a common pathway model tells us).

The key distinction is that IQ demonstrates a robust common pathway structure - different cognitive tests correlate with each other because they're all tapping into a genuine underlying cognitive ability. In contrast, personality measures often fail common pathway tests, suggesting that the correlations between different personality indicators might arise from multiple distinct sources rather than a single underlying trait. This makes genetic selection for personality traits fundamentally different from selecting for IQ - not just in terms of optimal selection strength, but in terms of whether we can meaningfully select for the intended trait at all.

The problem isn't just about avoiding extreme personalities - it's about whether our measurement and selection tools can reliably target the personality constructs we actually care about, rather than accidentally selecting for measurement artifacts or superficial behavioral patterns that don't reflect genuine underlying traits.

A key consideration when selecting for latent mental traits is whether a common pathway model holds for the latent variable under selection. In an ideal common pathway model, all covariance between indicators is mediated by a single underlying construct.
When this model fails, selecting for one trait can lead to unintended consequences. For instance, attempting to select for Openness might not reliably increase open-mindedness or creativity. Instead, such selection could inadvertently target specific parts of whatever went into the measurement, like liberal political values, aesthetic preferences, or being the kind of person with an inflated view of yourself.
Unlike personality factors, which demonstrate mixed evidence for a coherent latent structure, IQ has been more consistently modeled using a common pathway approach.

TL;DR: Selecting for IQ good. Will get smarter children. Selecting for personality risky. Might get child that likes filling in the rightmost bubble on tests.

Sources:
https://psycnet.apa.org/record/2013-24385-001
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7839945/

It turns out that in adults, BMI is negatively correlated with height. So, if human heights have been increasing over time, we'd actually expect BMI to decrease over time. 

Since molecular squiggle maximizers and paperclip maximizers both result in a universe-shard that's a boring wasteland, despite the fact that they maximize different things, what's the practical difference between talking about molecular squiggle maximizers instead of paperclip maximizers?