Reply to Moffitt et al.: Causal inference from observational data remains difficult

As children, we are largely “stuck” with our parents, neighborhoods, and schools. Such forced environments are gradually left behind as we age, and our intelligence quotient (IQ) and “non-cognitive factors” (NCFs) (interests, personality, temperament, etc.) influence which new environments we are exposed to. Some of these are more cognitively demanding than others, causing IQs to change. If these same NCFs affect risks of early-onset cannabis use/dependence, this generates a correlation between IQ change and cannabis use.

I suggested (1) that this mechanism would bias the estimates by Meier et al. (2) and show up as differing IQ trends across socioeconomic status (SES): research indicates that low SES is associated with higher levels of, and sensitivity to, such risk factors, as well as with less cognitively demanding home environments relative to school. This postulated SES/risk factor interaction now seems wrong: reported estimates are unchanged when estimated on mid-SES data, and IQ trends are similar across SES groups (3). However, with the exception of self-control, shown to be uncorrelated with IQ in the comments by Moffitt et al. (3) to Daly (4), neither NCFs in general nor my points on method are discussed in the response.

The Dunedin group’s method compares average outcomes across five groups defined by different cannabis-use patterns. This identifies an effect of cannabis only if these groups would have had the same IQ change in the absence of cannabis use. As noted (1), information on “background traits and early environment” across their five exposure groups could have indicated the strength of selection processes into the five exposure groups, but this and the other suggested analyses are ignored by Moffitt et al. (3). The common trend assumption required for inference is violated if there is selection on variables with causal effect on future environments and IQ, which seems plausible: Poulton has stated that the cannabis-dependent 21-y olds in the cohort “had a long history of anti-social behavior, going right back to when they were three years old[...] When stuff doesn’t work out right they just resort to violence” (5).

The evidence I cited (1) on causal effects of environment on IQ is also ignored, although Moffitt et al. (3) cite an earlier Dunedin study that failed to find simple environmental IQ determinants in childhood (6). Interestingly, this earlier work pointed to indications of “rebounding, level maintaining” processes related to “genotype-environment effects,” i.e., the effects at the heart of my argument: the Flynn-Dickens model of IQ finds that such effects can mask strong environmental effects within cohort-data, while causing IQs to change and IQ heritability to increase as we gradually age out of forced environments.

Finally, challenges of causal inference from observational data (7) are underplayed. Robustness of effect estimates to controls for causal back channels unrelated to neurotoxicity, simultaneous inclusion of multiple potential confounders, and changes to their statistical model remain undiscussed. Oddly, Moffitt et al. (3) claim that observational studies test actual effects while randomized clinical trials (RCTs) merely identify potential effects. A contrary view would be that RCTs have shown that even multiple, large-scale observational studies can be misleading, e.g., indicating health-protective effects of health-damaging therapies such as hormone-replacement therapy, β-carotene, and vitamin E (8).