This is a classic example of selection bias in an observational (= non-randomized) study. There is no mention in the abstract of whether other health variables

This is a classic example of selection bias in an observational (= non-randomized) study. There is no mention in the abstract of whether other health variables were statistically controlled for. And since it's a longitudinal study over a long time span, you can pretty much bet that it'd be insufficient anyways. You would also expect to see exactly this pattern in the data, with light drinking correlating strongly with otherwise-good health and non-drinking correlating mildly with otherwise-bad health. Parenthetically, the same effect turns up with elective Covid vaccinations. At-risk populations are much more likely to opt in, so vaccinated groups often have worse hospitalization and death rates than unvaccinated ones. Disingenuous observational studies are then used to justify anti-vax nonsense.