CUD Diagnoses and Blood Clots: Who Is in the Cohort?

Photo by Vitaly Gariev via Pexels.

The exposure in the Mu, Cheah and Lai paper is not cannabis. It is a diagnosis: cannabis use disorder, a code a clinician has to enter in a record. The paper, Cannabis use disorder and risk of venous thromboembolism, appeared in the Journal of Cannabis Research with a listing dated Oct 19, 2026.

Here is where I stand. What I have in hand is the journal listing, not the full text. So this piece gives no effect sizes, sample counts, database names or follow-up periods. I won't guess at them, and you shouldn't trust anyone who quotes them without opening the paper.

My thesis is narrower and more useful. A diagnosis-coded cohort will get read as cannabis causes clots. Such a cohort tends to select for heavy, high-comorbidity patients, which limits what an ordinary dispensary customer can take from it. Our VA coverage showed how soft CUD codes can be, and that is why this angle matters. What follows is a way to read exposure definitions, so you can test this paper and the next one yourself.

The exposure is a diagnosis, not cannabis use

A person who uses cannabis and a person with a CUD code are not the same group. The code needs a clinician, a visit where cannabis comes up, and a decision to record it. Plenty of people use cannabis for years without any of those three things happening.

The design matters too. A retrospective cohort starts from existing records, sorts people by exposure, and looks back at who developed the outcome. It can show an association in a defined group. It cannot show cause, because nobody assigned the exposure and the groups differed before the clock started.

The questions below have answers only in the full text, so check them there:

• Which database did the cohort come from, and which years does it cover?
• How many people were in it?
• Which ICD-10 F12 codes defined CUD, and did that take one encounter or several?
• Do the authors write association or causation, and what does their limitations section concede?

I can't answer those from a listing, and a headline can't answer them either. If the authors say association and name confounding as a limit, the headline has already overshot the paper. Read the limitations paragraph before the abstract. Authors usually state the ceiling of their own design more carefully than the press does.

Who a CUD code selects for, and who it leaves out

Photo by Tima Miroshnichenko via Pexels.

Think about how someone ends up with a CUD code. They see a clinician, or get screened inside a health system, or arrive for some other problem and cannabis is noted along the way. That is a filtered group. My read is that the filter leans toward heavier use and more medical complexity. That is inference, not something I can cite to this paper.

I also have no validation study of F12 codes in hand to point to, so I'm keeping the claim qualitative. What I can say is that a code records what a clinician noticed and chose to write down, not how much anyone consumed.

Now consider who is absent:

• Occasional users who never raise cannabis with a clinician.
• Daily users who feel fine, skip checkups or see no reason to mention it.
• People who use cannabis with no recorded diagnosis, which is most people who use it.

The paper's table of baseline characteristics is where this shows up. Look at which comorbidities the CUD group carried compared with the rest: other substance use disorders, psychiatric diagnoses, obesity, cancer, recent surgery, hospitalization. Then look at how the authors adjusted for them. Statistical adjustment can balance recorded variables. It cannot rebuild a group that was different in unrecorded ways.

The typical dispensary customer, someone buying flower or edibles on a weekend with no clinical diagnosis, is mostly outside this population. That doesn't make the study irrelevant. It means it speaks directly to people who look like its cohort, and only indirectly to everyone else.

What our VA coverage showed about soft CUD codes

We looked at this problem from another side in VA Cannabis Misuse Diagnoses: Why the Data Isn't So Simple. That piece examined how cannabis-related diagnoses are applied in VA records and why they are messier than they look. I'm not repeating figures from it here, since I'd want to re-check each against the published article before quoting.

The connection to this study is straightforward. If a CUD code can be applied loosely, the exposed group is blurrier than the label suggests. Some coded patients may meet the clinical definition of the disorder. Others may carry the code for reasons that have more to do with workflow, screening habits or how a visit was documented.

Three things should stay separate here:

• Documented: our earlier reporting on soft VA codes.
• Inferred: the same softness may exist in whatever data this paper used.
• Open: whether that softness changes the result at all.

The limit is real. The VA findings concern one health system. This paper may draw on a different source, such as commercial claims, an EHR network or a registry, and I can't tell you which from a listing. If it isn't VA data, our coverage is a caution, not evidence about this cohort. Blur in the exposure can also cut both ways. It might dilute a true effect or inflate one by pulling in sicker patients. Without the paper's data, I can't say which.

Five questions to ask before believing a cannabis-and-clots headline

Photo by Ron Lach via Pexels.

Here is a checklist you can reuse on any cannabis health headline, not just this one:

1. How was exposure defined? A diagnosis code, a survey answer, a urine test and a self-reported habit describe different people.
2. Who was the comparison group? No diagnosis, matched controls and people with other substance use disorders each answer a different question.
3. Was tobacco or nicotine adjusted for, and how was it measured? A diagnosis code, a cessation prescription and a self-report vary a lot in quality.
4. How was VTE defined? Deep vein thrombosis, pulmonary embolism or both, and was it confirmed by imaging or treatment, or only coded?
5. What are the absolute risks and confidence intervals? A large relative risk on a rare outcome can still be a small absolute number.

I can't give you this paper's answers. For each, open the full text and write down what you find. Pay particular attention to the gap between unadjusted and adjusted estimates. A big drop after adjustment tells you how much of the first number was the cohort's baggage, not the exposure.

Two more things to check. First, detection bias: people newly diagnosed with CUD are in contact with clinicians, so clots may simply be found more often. Look for a sensitivity analysis that excludes early events after diagnosis, and note whether the authors ran one or admit they didn't.

Second, residual confounding from tobacco. It is a standard caveat in cannabis research because many people use both, and records capture nicotine badly. See what the authors say about it, and how much weight they put on it.

The strongest counterargument, and where it holds

The best counterargument goes like this. Confounding exists in every observational study, yet a signal that survives adjustment can still be real. Researchers have also looked at cannabis and VTE in other groups, including trauma patients and people with HIV. Several designs pointing the same direction would carry more weight than one coded cohort.

I'll concede part of that. A soft code does not make a finding false. It makes the population harder to describe, and that is a different problem. If the paper reports a robust adjusted association with sensible sensitivity analyses, I would not wave it away because of coding concerns. I haven't verified what the trauma and HIV studies found, so I won't characterize them. Check them before relying on them in either direction.

What I won't do is say cannabis causes clots, or that it doesn't. The paper's own wording sets that limit, and a retrospective cohort rarely reaches causation.

What would settle it? Roughly three things:

• Use measured some way other than diagnosis codes, such as structured interviews or biomarkers.
• Dose and frequency data, so risk can be compared across how much people use and by what route.
• A design that handles tobacco and comorbidity directly, with prospective collection or at least careful matching.

Until something like that exists, the fair reading is a flag about a specific, heavily medicalized group. Whether it extends beyond that group is the open question.

What this means if you shop at a dispensary

The study describes people diagnosed with CUD. It can't be applied one-to-one to everyone who buys cannabis, and I'm keeping any health claim to what the paper supports. If the authors report an association, it is an association in that cohort.

There is still a practical step that doesn't depend on the paper. If you have a history of blood clots, take blood thinners or have surgery coming up, tell your clinician about your cannabis use and ask about it directly. This is general guidance, not medical advice. Clinicians can't weigh something they don't know about.

Remember also that tobacco mixed into joints or blunts is a separate exposure, and one studies often struggle to separate out. If you mix, the risk picture you are reading about may not describe what you are actually doing. Rules on cannabis vary by state, so check local law, and read the paper itself instead of relying on a summary, including this one.

The open question is whether the next cohort will measure actual use rather than a diagnosis. Until one does, treat headlines about clots as a prompt to read the methods section.

The exposure definition decides who a study speaks for. A diagnosis is a narrow doorway into a population: it lets in people who were seen, screened and coded, and it leaves the rest of the cannabis-using public outside, whatever the headline implies.

So what would a study need to measure for a dispensary customer to learn something about their own risk? My answer is actual use, meaning how often, how much and by which route, along with tobacco and the conditions that raise clot risk anyway. Until a study measures those, read the methods before the abstract.

Browse our seed collection.