[thaumasiotes]

https://www.gla.ac.uk/media/Media_774797_smxx.pdf

> We perform the first meta-analysis of the effect of lead on crime by pooling 529 estimates from 24 studies. We find evidence of publication bias across a range of tests.

> When we restrict our analysis to only high-quality studies that address endogeneity the estimated mean effect size is close to zero.

> When we use the full sample, the mean effect size is a partial correlation coefficient of 0.11, over ten times larger than the high-quality sample.

Interestingly, I met someone on HN once who insisted that effect sizes are always smaller in larger trials because that's just the nature of reality. This does not say anything good about scientists' conceptual understanding of what they're doing.