How Deceptive Studies Use Multiple Regression To Reach Misleading Conclusions

The presence of firearms is not associated with increased mortality after controlling for the presence of ammunition

Image made by me with MidJourney

Five days ago, Plos One published a study titled Body mass index and all-cause mortality in a 21^st century U.S. population: A National Health Interview Survey analysis. It contained some rather surprising conclusions:

This is from the “conclusion” portion of the abstract: The risk of all-cause mortality was elevated by 21–108% among participants with BMI ≥30. BMI may not necessarily increase mortality independently of other risk factors in adults, especially older adults, with overweight BMI. Further studies incorporating weight history, body composition, and morbidity outcomes are needed to fully characterize BMI-mortality associations.

And the following is from the conclusion at the end of the article: In conclusion, our findings suggest that BMI in the overweight range is generally not associated with increased risk of all-cause mortality. Our study suggests that BMI may not necessarily increase mortality independently of other risk factors in those with BMI of 25.0–29.9 and in older adults with BMI of 25.0–34.9.

Well, that certainly sounds groundbreaking: being overweight doesn’t increase the risk of death until you’re extremely obese. What’s this about “independently of other risk factors” though?

Here’s a quote from the “methods” section of the abstract: We estimated risk of all-cause mortality using multivariable Cox proportional hazards regression, adjusting for covariates, accounting for the survey design, and performing subgroup analyses to reduce analytic bias.

Okay, so they adjusted for some covariates. That’s another way of saying they controlled for some other variables that could affect a person’s likelihood of dying, in order to isolate the effects of body weight. That’s a totally normal element of studies like these. Let’s see which other variables they controlled for.

Oh. Oh dear.