Unpublished Article on Dumb Minor Crémieux Drama

I wrote this one ages ago and buried it deep in the archives because I’m not a fan of making enemies and I quite like this community. I also wanted to make a more comprehensive review of his work, following up on a few other cases in which others criticized Crém’s sourcing and claims.

But now it’s all buried in my bottomless morass of Twitter bookmarks and it’s currently safe to criticise him, so here it is.

Lead and Crime

Rates of violent crime have dropped precipitously in many developed nations since 1990 without any single clear cause. Most common explanations are, frankly, astoundingly bad, such as “America legalized abortion” or “America increased incarceration rates” — all relying on the all-too-common assumption that America is the only country on Earth (something only partly true).

Indeed, not only was this decline in violent crime observed worldwide, certain nations actually decreased abortion access and incarceration rates during that same period, yet saw similar reductions in violence!

This leaves the global phaseout of lead additives from consumer products — and subsequent fall in blood lead levels — as one of the few remaining serious explanations (which I have seen). If the research is to be believed, not only is lead exposure uniquely pernicious as there is no safe exposure level (unlike most other toxins, which are considered acceptable below a certain concentration), but it also carries a host of secondary effects such as decreased IQ, heightened aggression, and an increased propensity to commit violent crime.

At least, that’s what you’ve probably heard.

Crémieux says that after correcting for publication bias, the correlation between blood lead levels and crime is zero, citing this study and attaching this graph:

And by adding the qualifier that the correlation is brought to 0 when using “any commonly used correction” (an inclusion which later struck me as somewhat load-bearing), he appears to be referring to the data in Table 4, Panel A, showing what is essentially a null effect for lead on PCCs (partial correlation coefficients) after adjusting for bias.

Pretty convincing, huh? I mean, I was convinced by it.

But…

I still would be convinced, had I not happened upon this reply thread by one of the paper’s authors, pointing out “for some reason he cut out the bottom of that chart”, showing “when we use elasticities rather than PCCs […] there is a clear effect”¹.

Here’s the full, uncropped image from the paper.

Moreover, he points out that his study actually ran a metra-regression analysis to adjust for heterogeneity in between studies, ultimately concluding there is an effect between lead and crime which remains after adjusting for publication bias (in both partial correlation and elasticity).

Many thanks to Higney for his patience in answering my questions!

Update

Crémieux has semi-retracted his tweet on lead and crime, saying he “doesn’t trust the lead author”, who “has made it clear they’re surly, reactive, and untrustworthy”, along with preferring “using their harder-to-vet estimates, which reads suspiciously”.

I read every single tweet made by Higney since Crémieux’s post went up and the only other ones in which he refers to Crém are one in which he calls him “not quite correct” and another in which he calls him a “bad faith actor”.

Of course, they might have had a hostile conversation in private, so I asked Higney directly, and:

To Crém’s credit, I’m not exactly clear on how the original author reached his adjusted figures either, and I don’t know if they’re any good. However, I believe this to be a moot point: the issue at hand is that Crémieux appears to treat data as trustworthy when it helps him make the case for his beliefs, yet maligns the very same research as “untrustworthy” and “suspicious” when it no longer does so.

So, in summary:

• Crémieux cites a study which concludes the exact opposite of what he claims

• Crémieux crops out data that contradicts his claims

• Crémieux cherry-picks data to support his claims, adding language like “with any commonly used correction” in what (I assume) is an attempt to justify this

• When confronted on the aforementioned acts by the study’s author, Crémieux claims he is “surly, reactive, and untrustworthy” and calls his work “suspicious”.
  (You can judge for yourself if their interactions warrant this characterization.)

1. TL;DR of both is PCC = correlation as measured in SDs, while elasticity = correlation as measured in real-world variables.