Behavioral Risk Factor Surveillance System · 2024 · n = 414,633

How to predict how fat someone is

Every year the CDC phones roughly 400,000 adults and asks, among much else, their height and weight. Describe someone below and this model reports what that survey would expect them to weigh — and how little it really knows.

All US adults
This profile
0%25%50%75%100%

What moves the needle

Holding the rest of the profile fixed, this is the range of obesity probability across each characteristic’s options. The dot marks the current selection.

Read this before you believe any of it

Heights and weights here are self-reported, collected over the phone. People shave pounds and add inches. Measured data from NHANES puts adult obesity near 40%; this survey lands around 33%. The gap is the lying, and it is not evenly distributed — heavier respondents understate more.

This model is weak at the individual level, by design and by nature. Against a flat national baseline it reduces prediction error by only about 5%. Compare the two ribbons above: for most profiles they barely differ. Demographics shape populations far more than they determine any one person.

A profile is not a prognosis. Nothing here describes a person — it describes the average of everyone who filled in the same boxes.

Correlation only. Exercise shows the largest single swing, but this is one phone call, not a study over time. People who are already heavy exercise less, and the survey cannot tell you which came first. The same caution applies to income, employment and housing.

Smoking appears to lower obesity. It does — nicotine suppresses appetite and raises metabolic rate. It is a documented effect and a catastrophic weight-loss strategy.

BMI itself is a blunt instrument. It cannot separate muscle from fat, and its cutoffs were calibrated on largely European populations, which is one reason health risk begins at lower BMI values in South and East Asian groups than the standard thresholds imply.

Method. Weighted multinomial logistic regression on the 2024 BRFSS public-use file, fitted with the survey weight _LLCPWT, with sex interacted against race, age, marital status and number of children. Refusals are kept as their own category rather than dropped. Held-out calibration is within a percentage point on all four outcomes.