The question
Two questions, really. First: are there natural types of people when it comes to substance use, or is everyone spread along a continuum? Second: among the things this survey records, which are most associated with having had a period of serious depression in the past year?
The first is a grouping problem: nobody tells the method what the groups are, and it has to find them. The second is a prediction problem: we know who reported depression, and we ask which characteristics would have let us guess.
The survey, and why its design matters
NSDUH asks tens of thousands of people every year about alcohol, marijuana, cigarettes, their health, their income and their mental health. Three years pooled gives 134,897 adults, which is large enough to look at small subgroups.
The survey is not a simple lottery. It deliberately interviews more young people than their share of the population, because that is the only way to say anything precise about them. To compensate, everyone carries a weight: a number saying how many people in the country they stand for. A young adult might count for fewer people than an older one.
Get this wrong and your answer describes the people in the room rather than the country. Unweighted, 11.29% of respondents had a depressive episode. Weighted, 8.44%. The first is a fact about the survey; the second is an estimate about America. This site reports both, every time.
What we found
Depression is strongly patterned by age
Roughly one adult in six aged 18 to 25 reported a depressive episode in the past year. Among adults 65 and over it is closer to one in thirty-four. Women reported about 1.6 times the rate of men. How people rate their own health tracks depression more closely than any of the substance measures.
There are four recognisable groups
The first question was whether there are natural "types" of people when it comes to substance use. Grouping respondents on their use and circumstances produces four profiles. Because the labels do the explaining, here is exactly what each one means.
Very high co-use of alcohol, marijuana and cigarettes together. Alcohol 83.6%, marijuana 80.1%, cigarettes 48.5%. Depression 19.6%. n = 26,510.
Low on every substance. The youngest group, and the lowest income. Alcohol 21.4%, marijuana 5.2%, cigarettes 1.5%. Depression 11.1%. n = 37,171.
Drink, but almost nothing else. The oldest and highest-income group, and the healthiest. Alcohol 79.9%, marijuana 4.1%, cigarettes 1.5%. Depression 6.2%. n = 45,853.
Low alcohol and marijuana, but the highest cigarette use and the poorest self-rated health. Alcohol 16.7%, marijuana 3.8%, cigarettes 20.7%. Depression 12.2%. n = 25,363.
The dividing line is not drinking. It is how many things travel together. The largest group drinks heavily as a group, at about 80%, and has the lowest depression rate of the four. The group that stands out is the one combining alcohol, marijuana and cigarettes, whose depression rate is more than three times as high.
A "social drinker" in this data is someone who drinks and does very little else, is older, has higher income and rates their health well. A "light user" is younger, low on everything, and lower income. A "heavy multi-substance user" is the polysubstance group: high on all three at once. Those are descriptions of a profile, not diagnoses, and nobody in the survey was asked to place themselves in a category.
Worth saying plainly: this method always returns four groups if you ask it for four. We checked that the same groups come back when the analysis is repeated on random parts of the data, and they do. That means the pattern is reproducible. It does not prove these are real categories of people rather than convenient lines drawn through a continuum.
A model can rank people, but a naive cut-off is useless
The prediction model is genuinely informative: given two people, one depressed and one not, it ranks them correctly about 0.7406 of the time. But asked to label individuals using the standard rule, it declares almost everyone not depressed, and is 88.7% accurate for doing so.
This is the most useful thing on the site. Because only about one adult in nine had an episode, a model that says "no" to everybody is right most of the time. High accuracy can mean a model has learned nothing. Lower the bar for flagging someone and it finds far more of the people who were actually depressed, at the cost of flagging more people overall. Which is right depends on what a flag leads to.
What this cannot tell you
Everyone was asked about substance use and depression at the same moment, so nothing here shows what came first. Marijuana use and depression travel together; whether one leads to the other, or something else drives both, is outside what this survey can answer.
The measures are coarse. "Used in the past month" counts one drink and daily heavy use identically. Everything is self-reported, and people under-report behaviour they expect to be judged for. Adolescents are not included at all: they are asked a different set of questions and were excluded from this analysis.
Most of what drives depression is simply not in this dataset. Genetics, life events, social support, treatment history: none are recorded here. That is why a model built on these seven variables ranks people only moderately well, and why it would be a mistake to read it as a diagnostic tool.