Hacker News Reader

> That's about enough to conclude all this doesn't mean much.

You list a common confounder, then skip over the section that says 5/6 studies that have none of these confounders find a decline. That's a fairly important followup.

> However, Auger identify six (of their original seventy) studies that they say are extremely well done and suffer from none of these potential confounders. They say five of the six still show declining sperm counts, and believe that the effect might be real (they say there is no evidence it is global rather than limited to these five regions but I think if an effect happens in five regions, and there is only one region where it is not happening, it is fair to wonder whether it represents a broader trend). I interpret their conclusions as very cautiously pro-decline-hypothesis (Fisch remains against).

> Edit: One important thing the article doesn't talk about is how sperm is actually counted. Obviously researchers don't count all 150 to 300 million sperms per sample; they probably count a small (or very small) sub-sample and extrapolate

When you've got hundreds of millions of something distributed essentially randomly in some space, counting a subsample and extrapolating is fine.

But the method needs to be consistent over time for comparisons to be meaningful.

If for example, they did rough extrapolations in the 50s, and then in the 90s they were able to do an exact and exhaustive count, it's a problem.

IDK if that's what happened, but we would want to know.

Consistent in certain aspects, not all.

> If for example, they did rough extrapolations in the 50s, and then in the 90s they were able to do an exact and exhaustive count, it's a problem.

That's not a change that's happened, smaller volumes are still analysed, but no that's not necessarily a problem.

At such scales, "small" samples can be totally fine and have an extremely high chance of matching a complete analysis. That's why polling and sampling works.

The problem shrinks even more when you're looking at averages over groups. Smaller samples are noisy, not biased. Averaging over many doesn't lower bias but it does lower the noise.

if the incoming data is not random, but is in fact skewed by factors relevant to the data we're looking for, there's very little value in it

Looking at people who donate sperm effectively is a random sample of men unless there's a causal link between sperm count and the desire to be a donor.

Sampling only people who self-select themselves into one specific group (sperm donors) definitely isn’t a random sample. There are obviously correlations between income, geographic area and age with the desire to become a donor.

Don't sperm banks collect all of that relevant information? I would assume they collect at least age, income, and education. Probably some health history as well.

There can easily be a causal link. It’s a self selected group specifically on sexual reproductive strategy. I don’t need to be able to explain the relationship in detail; it’s enough that it’s plausible and that it hasn’t already been studied.

You also need to live somewhere where there are sperm banks. Lots of countries don't have them.

When you couple the decline in sperm count with testosterone levels which have plummeted, doesn’t that seem concerning?