The study appears to have a valid design, subject to some questions not answered in the Bloomberg coverage. I think the more reasonable question is what exactly

The study appears to have a valid design, subject to some questions not answered in the Bloomberg coverage. I think the more reasonable question is what exactly the design is measuring. Using the Ashley Madison data is basically what we would call a "reduced form estimate". The idea is that someone in the Ashley Madison data is more likely (or at least not less likely) than a comparable individual not in the Ashley Madison data to cheat on their spouse. They would like to measure whether or not people actually cheat, but they can't, so they measure this instead. This kind of reduced form estimate is a fairly common thing when you can't measure actual treatment compliance, but you can measure an "invitation to treat". We, in fact, know that Ashley Madison wasn't actually used for cheating because there were no (zero) real women on the site. Extensive analyses of the data leak suggest basically that men signed up and were faced with phony bots. But the idea should still be that the choice to sign up on Ashley Madison reveals an attempt to cheat. The other major challenge for the design is that the control group of CEOs need to have what we would think are baseline similar propensity to commit fraud etc in all respects except for the choice to cheat. So, at the very least I would want to do a matching, weighting, or propensity score design that accounts for differential company characteristics (sector, size, age, any characteristics we would a priori affect propensity to engage in financial malfeasance.) If what we learn is that people who run shady payday loan companies also cheat on their spouses, then this is maybe not interesting. But if what we learn is that among F2000 blue chip companies in similar sectors, cheating spouses are more likely to be cheating CEOs, this is more interesting, right? I agree with you that it's possible that the degree of difficulty / competence factor might be a confounder. Probably they could compensate for this by looking at something other that rate of conviction for malfeasance, rather by modelling types / degrees of malfeasance and degree of difficulty for getting caught. Finally, the casual inference here is impossible. So it's possible that they have a good design to answer a question they're not asking. I would assume the default assumption is not that cheating on one's spouse causes one to cheat on financial things, but rather that both are outcomes which flow from an underlying propensity for deceit. In this case, the main takeaway is that early indicators of deceit may allow us to head off or catch financial malfeasance. That's interesting, but maybe not the exact question they're asking. I'd be really interested in reading the final thing because I could see it going either way. I think it's an interesting proposal for a design. Would love to see a pre-print