Pooling three years of a large national survey gives enough respondents to split the population finely. It also makes it easy to forget that the survey is not a simple random sample. Both halves of that sentence turned out to matter here.
The data
NSDUH is run every year by the U.S. government and asks tens of thousands of people about substance use, mental health and their circumstances. It is a stratified multistage probability sample, not a simple random draw: some groups are deliberately over-sampled so that estimates for them are precise enough to publish.
That design is the reason the survey works, and it is also the thing an analysis has to respect. Every respondent carries a weight saying how many people they represent. Ignore the weights and you describe the respondents rather than the country.
The public-use files for 2022 to 2024 total about 1.1 GB, roughly 59,000 respondents and 2,650 variables each year. They are not republished here. Download them from SAMHSA.
What the design changes
Every descriptive figure on this site is computed twice from the same respondents: once ignoring the survey design, and once using it. The difference is not cosmetic.
Depression is the clearest case. Unweighted, 11.29% of respondents report a past-year depressive episode. Weighted to the population, it is 8.44%. The unweighted figure overstates it by about a third, because the survey deliberately over-samples young adults and young adults report depression at much higher rates.
What was found
Depression falls steeply with age. Around one in six adults aged 18 to 25 report a past-year episode; among those 65 and over it is closer to one in thirty-four. Women report roughly 1.6 times the rate of men, and self-rated health tracks it more strongly than any substance measure.
Substance use clusters into recognisable groups. Reducing seven measures to their main components and grouping respondents produces four types, one of which stands out: about 19.7% of the sample, defined by very high marijuana use (80.1%), and reporting depression at 19.6%.
A model can rank people by risk but not usefully label them. A penalised logistic regression reaches an AUC of 0.7406, which is a real signal. At the default cut-off it is also 88.7% accurate while flagging only 805 people as depressed out of 134,897. That is the most instructive result on the site, and it is explained on the analysis page.
Where to go next
The analysis
Descriptives both ways, PCA and clustering with a stability check, and the classifier in full.
Live