Does Lead Exposure Really Kill Five Million People Per Year? (Probably, Yes)

Two years ago we did a deep dive into the evidence on lead exposure and children’s cognition, ultimately convincing ourselves that the evidence is causal, and what the size of the effect is. But a recent paper by Larsen and Sánchez-Triana found that most (three quarters) of the burden of lead exposure is due to effects on cardiovascular disease, rather than cognitive damage to children. They estimate 5.5 million people died in 2019 due to cardiovascular disease attributable to lead exposure. That’s more people than died of HIV/AIDS and malaria combined. Just how reliable are those estimates?

When it comes to causal evidence, we’re on thinner ground with cardiovascular disease than children’s cognition. We counted seven quasi-experimental, causal studies of the effects of lead on test scores, and there have since been four more published. For cardiovascular disease or elderly mortality I find just two. Hollingsworth and Rudik use exposure to leaded petrol at race car tracks as a natural experiment to estimate causal effects on adult mortality. Fletcher and Noghanibehambari use the distance to the closest lead manufacturer as an exogenous predictor of the amount of lead used in city water piping. Despite the small number of direct studies, we can increase our confidence that effects are causal by relying on our understanding of the biological mechanism, and from experimental studies with mice.

How do the magnitude of the causal estimates compare to the observational estimates used in estimating the number of deaths? First, the (observational) relationship between lead and cardiovascular disease deaths used by Larsen & Sánchez-Triana is non-linear. At the average blood lead level (BLL) in low- and middle-income countries, the relative risk ratio is around 1.5, with a 95 percent confidence interval from 1.1 to 1.9. So people with 5 μg/dl blood lead are 50 percent more likely to die. That’s a big effect! But is it consistent with the quasi-experimental evidence?

Lead exposure is correlated with increased risk of death

Image

Note: This figure is reproduced from Larsen and Sánchez-Triana (2023)

We need to do a few calculations to get comparable numbers. Hollingsworth and Rudik estimate effects of exposure to race tracks on child blood lead levels and separately on adult mortality. Their estimates on child blood are whether a child is above a 10 μg/dl threshold. This elevated blood lead rate increases by 17 percent, from an overall mean of 0.92 percent, with a standard deviation of 1.076 percent. So that’s a relatively small increase in the share of children with elevated levels, from 0.92 to 1.08 percent. To convert that to a change in mean BLL for comparison with the observational dose-response curve, we need some assumptions. If we assume a typical skewed log-normal distribution with a mean of 2.3 and log standard deviation of 0.8, then the 17 percent increase in children above the binary threshold is equivalent to a 0.1 μg/dl increase in mean blood levels. On mortality they find an effect due to cardiovascular disease of 37 deaths per 100,000 people, from an overall mean of 1,845 deaths. If we assume a linear slope between BLL and deaths, then moving from zero up to the average 5 μg/dl in LMICs is:

5 / 0.1 x 37 = 1,850 deaths, or a relative risk ratio of 2.0. That’s slightly larger than the 1.5 ratio in the observational estimates. It also undercounts the effect on total all-cause mortality, which is more than twice as large (91 deaths per 100,000 people).

In the second quasi-experimental paper, Fletcher and Noghanibehambari estimate that early life exposure to lead pipes causes a roughly 1 percent reduction in age at death. They don’t however have data on blood lead levels, so we can’t directly compare here with the observational dose-response curve. They do see a similar pattern whereby the quasi-experimental instrumental variable estimates are larger than the simple observational estimate. We can't, though, distinguish how much of this difference is because (a) the observational estimate was biased downwards, or (b) the instrumental variable estimates identify the effect only on a sub-population of people—the compilers—for whom effects may be larger than average.

The bottom line though is that the 5.5 million figure is big, but not implausible. And the limited causal evidence doesn’t contradict it. A 2021 editorial in the Journal of the American Heart Association stated that “the burden of proof has been met.” I have reasonably high confidence in that conclusion, but… I’d still like to see one or two more quasi-experimental causal studies, particularly from low- or middle-income countries.