In twenty-eight real families, the top-scoring child was actually the tallest in seven
This particular test was not a model. Real brothers and sisters who had already grown up, scored as if they had been embryos.
Most of what is claimed for embryo selection comes out of computer models, including the headline gains further down this page. One piece of this work did not, and it is the piece to start with: a test of the height score against real families whose children had already grown up.
Researchers took 28 real families with between 3 and 20 grown children each, average about ten. Every child's height was known. The team scored each of them with a polygenic height score, then asked a simple question: if these children had been embryos and you had picked the one with the best score, would you have got the tallest?
The authors put it in one sentence: These mean values are accompanied by wide prediction intervals, and indeed, in large nuclear families, the majority of children top-scoring for height are not the tallest.
The rest of the paper is modelling, and its headline figure for what selection buys, with five embryos and current scores, is about 2.5 cm of height or about 2.5 IQ points. That is a simulated average. The spread around it is the part that matters.
Using today's score, the modelled 95% prediction interval for a child's adult height is about plus or minus 9 to 10 cm. Even if a future score explained the entire inherited component of height, which no score does, the interval would still be about plus or minus 5 cm. For IQ, even under those very generous assumptions, the interval is about plus or minus 13 to 19 points. Based on parents alone it is plus or minus 24 to 27 points. The authors' own worked example, on assumptions they describe as more favourable than reality, is a predicted IQ of 106 plus or minus 19, meaning 88 to 125, with selection, against 100 plus or minus 27, meaning 73 to 127, without it.
The same authors state what that spread means for a couple selecting for intelligence: The future child has a non-negligible probability (≈0.25, assuming a normal distribution) to have an IQ below the population average.
That is roughly a one in four chance, after selecting.
And selecting for more than one thing shrinks each of them. The paper shows mathematically that if you select for several traits at once, the average gain in each trait falls as the number of traits rises, and that in the limiting case, where two traits are maximally anti-correlated, the gain per trait completely vanishes.
Every company sells a panel, not a single test. So the single-condition figures on the previous page are ceilings for what a panel can deliver per condition, not estimates of it.
"What is the range around the average gain, not just the average?"
"If I am selecting for several conditions at once, what happens to the benefit for each one?"