“Scott Alexander” writes,

When they caught up with these kids at age 25, the intervention group was found to have an odds ratio of around 0.6 to 0.7 of having developed various psychiatric disorders the study was testing for, including antisocial personality disorder, ADHD, depression, or anxiety. They had odds ratios around 0.7 of developing drug and alcohol abuse problems by various measures. They reported less risky sexual behavior, less domestic abuse, and fewer violent crimes. All of this was significant at the p < 0.05 level, and some of it was significant at much higher levels like p = 0.001 or below. Subgroup analysis found the data were very similar when you restricted the analysis to various subgroups like boys, girls, whites, blacks, highest-risk, lowest-risk, and by study site (it was a multi-site study)

This was a randomized, controlled study of a group of many interventions. “Scott” goes on to point out a number of caveats. The group of interventions was expensive. A lot of other indicators, including employment rates, did not improve. We do not know whether the results came from one or two of the interventions, or from the combination of all of them.

Still, it looks as though something managed to defeat the null hypothesis. As a controlled trial, it gets over the hurdle of confusing correlation with causality. As a study of long-term outcomes, it gets over the hurdle of fade-out. The results are numerically significant, not just statistically significant. The only remaining hurdle is replicability. My guess is, given the complexity of all those interventions, that the replicability hurdle will be a challenge.