Meta-analysis

A meta-analysis takes a set of separately published studies on the same question and combines their results into a single, pooled estimate, usually weighted by each study's sample size or precision.

What it buys you

The Cochrane Handbook (chapter 10, Deeks and colleagues) defines it as "the statistical combination of results from two or more separate studies", and in its inverse-variance method each study's weight is one over the square of its standard error, so larger, more precise studies count for more. The main advantage is sample size. Veale et al. (2015), the reference this site treats as standard, pooled studies in which a health professional measured at least 50 men using a standard procedure, reaching up to 15,521 men - far more than any single constituent study could gather on its own, which narrows the uncertainty around the pooled mean considerably.

The one caveat that matters

A meta-analysis does not fix the studies it pools. If the underlying studies used different measurement protocols, recruited different populations, or carry their own biases, the pooled estimate inherits all of that. Combining many imperfect studies produces a more precise estimate of something, but it does not upgrade any one of them into a better-designed study, and it cannot correct a systematic error that runs through several of them in the same direction.

This is why pooling is not a substitute for reading what went into the pool. When the constituent studies disagree with each other by more than chance would predict, that disagreement is called heterogeneity - Veale et al. report the greatest variability between studies for stretched flaccid length - and what it does to a pooled mean is worth understanding before trusting a single combined figure at face value. The individual study estimates that get pooled are usually shown first as a forest plot, and reading one is a useful skill before trusting the single pooled line beneath it.

A meta-analysis also does not manufacture new measurements. It reorganises and weights existing ones, which means its precision is bounded by how the source studies were conducted in the first place - a well-run meta-analysis of poorly measured studies is still a precise estimate of a poorly measured quantity, not a correction of the underlying problem. That distinction is easy to lose once a single pooled figure starts circulating on its own, detached from the studies that produced it.

Where to read the whole paper

This post is a definition, not a summary of Veale 2015 itself. For the full walkthrough of what that paper set out to do, what it pooled and what it reports, see Veale et al. 2015, explained in full.

A meta-analysis of measured data is also a different kind of thing entirely from a rating produced by Rate Cock, which pools nothing statistically and instead judges a single photograph on its own terms. An AI model scoring an image has no pooled dataset of physical measurements behind its output either - what it is actually doing is pattern recognition on one picture, not meta-analysis of anything. A ranked gallery like Penis Rater or a verdict from a human judge are likewise not statistical pooling exercises, whatever the word "average" ends up meaning in either context.