Comment by tofof | Hacker News Reader

Comment by tofof | Hacker News Reader

Parent Full thread

> How does prenatal test affect someone after they've gone through puberty

The test doesn't affect them. Hormone levels during development are the primary way the body controls development, so even small alterations in hormone levels in utero would be expected to have large physiological consequences.

> Is that not a far reaching statement to associate the two

Luckily, the field of statistics gives us ways to reason about whether or not to believe an observed effect.

In this case[1], the researchers used data on 100% of Norwegian births from 1967-1978 (n=728,842) including 13,800 twins. From these large n's the researchers are able to get the statistical power to measure these effects at the level of precision they did. The p-values for all the reported effects are presented in the appendix [2] and ranged from <.001 to .054 for the outcomes reported to be different.

It's worth noting that many of these differences had already been observed in multiple previous studies. The difference in marriage rates, for example, already appeared in a study analyzing records of 18th and 19th century (1734–1888) Finns [3].

So, in summary, while these statements are interesting, they're in no way an overreach, given the amount of data analyzed.

1: https://www.pnas.org/content/early/2019/03/14/1812786116

2: https://www.pnas.org/content/pnas/suppl/2019/03/14/181278611...

3: https://www.pnas.org/content/104/26/10915?ijkey=2c55ba770f14...

Being statistically illiterate, could I get an explanation of of "The p-values for all the reported effects are presented in the appendix [2] and ranged from <.001 to .054 for the outcomes reported to be different." I looked at the link, but I'm still not sure what the table means, or how I could explain what this means to someone else who doesn't know stats.

I'd like to share this with her, but neither of us know stats so the significance is kinda lost.

Sure.

P-value is the universal* way of expressing statistical likelihood. It corresponds to a percentage: p=.05 just means 5%, and p=.001 means 0.1%, etc.

It's often inaccurately explained as the likelihood of getting our results through chance alone. That's wrong for reasons that are technically important, but not in a way that really inhibits understanding of the strength of results that have small p-values.

* It has flaws, and a growing number of researchers believe it should not have the prominent importance currently placed on it.

---- Stop reading here if you're already satisfied. ----

We want to measure if there's a difference between two groups. So, we take measurements of a portion of group A, and measurements of a portion of group B.

Mathematically, we assume that what we've actually done is sampled from the same population both times. If that's true, our data sets should be quite similar to one another, but of course there will be some difference just due to noise.

So, we compute mathematically the chance of seeing a difference at least as large as the one we see between our data sets, if that assumption is true.

We decide beforehand on a small error rate. If not otherwise stated, 5% (p=.05) is the universal standard.

If we find that the likelihood of observing a same-or-larger difference between the populations is smaller than that already-small 5% chance, we REJECT our assumption that the samples came from the same population, and conclude that the populations must actually be different.

In other words, we conclude that we measured an actual difference that contrasts two mathematically-separate populations.

> Luckily, the field of statistics gives us ways to reason about whether or not to believe an observed effect.

What you mean is correlation, right? Effect would imply causation.

No, I mean effect, i.e. causation.

In this case* the direction of the correlation is 100% knowable: economic and education outcomes in the 1990s cannot possibly have altered whether or not you had a male twin in the 1970s, unless you can reverse time's arrow.

*The OP was specifically asking about marriage rate and education outcomes. The paper also examined fertility. It's at least remotely plausible that there's some 3rd/confounding biological mechanism that actually the cause of both the presence of a male twin and the increased fertility of the female twin. In that case I would agree that 'correlation' is the more appropriate choice of term, but note that it's still correct in that case to assign 'effect' to both of the observed properties as they'd be downstream of the confounder.

Reply on news.ycombinator.com