Choice of baseline matters

Looks like folks are back to arguing about whether there's a J-curve in alcohol consumption. The J-curve plots out the relationship between all-cause mortality and drinking. Non-drinkers are at the left-hand upwards tip of the J, light-to-moderate drinkers are in the dip, then heavy drinkers are in the upwards tilt at the right hand side.

And the ballpark numbers I keep in my head on this are from Di Castelnuovo and Donati's metastudy from 2006 that has light drinkers (about a drink a day) with a relative risk of about 0.84 as compared to non-drinkers when former drinkers are excluded.

The Lancet has a piece up that the press are covering as showing no J-curve.

But they start their curve with a reference category of light drinkers: people consuming a small amount of alcohol per week. The J-curve normally starts with non-drinkers as a reference category. If you pull the non-drinkers out of any J-curve, then you'll have a hockey-stick instead: no downward tilt followed by a sharp upwards lift.

And, as Chris Snowdon points out, they do have the standard form buried over in an appendix. Page 31 of the appendix. I've copied that figure below. The one on the right is the all-cause mortality one. Once you put non-drinkers back in, you have a J-curve again. If you also have ex-drinkers, you have a sharper J-curve. The ex-drinkers are the even higher risk folks at the far left.

The reference category are people consuming from 0 to 5 standard drinks per week - so a bit under a drink a day. The category that always has the lowest all-source mortality because they're drinking a bit less than a standard drink per day. And the risk from drinking about 4 standard drinks per day (300 grams per week) is the same as the risk among never-drinkers - with much higher risk beyond that.

The same as we've known since Donati 2006.

Am I missing something? The newspapers are yelling about how the study means there's no J-curve.