What's the simplest way to define a confounding variable?

A confounder is a third variable that's associated with both what you're studying (the exposure) and the result you're measuring (the outcome), making it look like there's a direct relationship between the two when the real cause might be something else entirely. The fire-stations-and-fire-damage example is the classic: population size confounds the association, not any causal link between stations and damage.

Can a randomised controlled trial eliminate all confounding?

Randomisation distributes both known and unknown confounders across groups by chance, which is why RCTs are considered the strongest design for causal questions. But they're not immune — poor randomisation, small sample sizes, and dropout patterns can reintroduce imbalance. And many questions in plant-medicine research simply can't be answered with an RCT for practical or ethical reasons, so observational designs remain important.

What does 'statistically controlling for a confounder' actually mean?

When researchers say they 'controlled for' something — say, age or tobacco use — they've used statistical methods (usually regression) to estimate what the relationship between exposure and outcome would look like if everyone in the study had the same value for that variable. It's a mathematical adjustment, not a physical one. It only works for confounders that were actually measured, and only as well as those measurements are accurate.

Why is confounding such a common problem in cannabinoid and plant-medicine research specifically?

A few reasons. First, randomised trials in this space are difficult — regulatory hurdles, supply constraints, and blinding challenges all apply. Second, the populations who use these products tend to share characteristics (lifestyle, health attitudes, concurrent substance use) that cluster together as potential confounders. Third, the compounds themselves are often studied in real-world settings where multiple variables change simultaneously, making clean causal attribution hard.

Is residual confounding the same as a study being wrong?

Not exactly. Residual confounding means the statistical adjustment was incomplete — usually because a confounder was measured imprecisely, or because relevant confounders weren't measured at all. The study's conclusions might still be directionally correct, but the size of the effect could be over- or underestimated. It's a reason to treat the precise numbers with caution, not necessarily to throw out the finding altogether.

Confounding Variables Explained

By Hannah Bui · 30 June 2026 · 7 min read

a woman looking through a microscope into jars

The statistic that sounds airtight but isn't

About sixty per cent of the studies I read in a given week have a confounding problem sitting in plain sight. Not hidden , plain sight. The authors know it's there; they'll mention it in the limitations section, usually in one diplomatic sentence before moving on. Readers, understandably, don't always make it that far.

So let's talk about confounding. Not the textbook definition (though we'll get there), but the practical question: when you read a study about a cannabinoid, a terpene, a peptide , anything in the plant-medicine space , how do you know whether the relationship you're looking at is real, or whether some third variable is doing all the actual work?

What a confounding variable actually is

A confounder is a variable that's associated with both the exposure you're studying and the outcome you're measuring, but isn't on the causal path between them. It muddles the picture. Classic textbook example: towns with more fire stations tend to have more fire damage. Does that mean fire stations cause damage? No; population size is the confounder. Bigger towns have more fires and more fire stations. Fire stations aren't the problem.

That's obvious once you see it. In biomedical research, confounders are rarely that obvious. They're buried in recruitment criteria, or they're socioeconomic factors the researchers couldn't measure, or they're biological variables no one thought to record.

A real, verifiable example from cannabis research: early observational studies on cannabis use and various health outcomes were heavily confounded by tobacco co-use. Many people who smoked cannabis also smoked tobacco, and for a long time studies didn't separate these groups well. The apparent associations kept partially reflecting tobacco's effects, not just cannabis. The National Academies of Sciences, Engineering, and Medicine flagged this specifically in their 2017 report on cannabis and health as a pervasive methodological limitation across decades of literature. That's not a gotcha against the field, it's just what confounding looks like in practice.

Why this is especially sticky in plant-medicine research

The plant-medicine space has a particular confounding problem, and I'd argue it's underappreciated even by people who read carefully. Here's the short version: the populations who use these products are not randomly selected. They have things in common, health concerns, attitudes towards conventional medicine, lifestyle factors, income levels, where they live. That's a cluster of potential confounders right there, before anyone has even designed a study.

Take something like bioavailability research on cannabinoids. Studies comparing oral to inhaled delivery are trying to isolate the route of administration. But if the people choosing inhalation are systematically different from the people choosing oral; different health backgrounds, different usage patterns, different concurrent substance use, then any difference in outcome might reflect who they are, not how they dosed. You'd need randomisation to control for that, and genuine randomisation in this kind of research is difficult.

Similarly, research into the so-called entourage effect, the hypothesis that cannabinoids and terpenes interact in ways that affect overall activity; faces a confounding challenge every time researchers compare full-spectrum preparations to isolates. The preparations differ in multiple compounds simultaneously. If you see a difference in outcome, you don't automatically know which compound drove it. It's not confounding in the strict variable-association sense, but it's the same underlying epistemological problem: you can't cleanly attribute cause when you've changed several things at once.

The three questions to ask every time

I've developed a habit, honestly, more of a reflex at this point, of asking three things the moment a study result sounds convincing:

1. Who are these people, and how were they selected? Convenience samples (people who volunteered, who already use a product, who responded to an ad) are riddled with selection bias, which is confounding's close cousin. Look at the recruitment section. If it says "participants were recruited via social media," pump the brakes.

2. What else is different between the groups? In a comparison between two groups; users vs non-users, high-dose vs low-dose, ask what else varies systematically. Age, sex, other medications, diet, exercise levels, socioeconomic status. Did the study measure these? Did it adjust for them statistically? "We controlled for age and sex" is a start, but it's rarely the whole story.

3. Is this randomised? A randomised controlled trial (RCT) is specifically designed to distribute confounders evenly across groups, including ones the researchers never thought of. That's the beauty of randomisation. Observational studies, cohort studies, and case-control studies cannot do this, which is why causal claims from those designs need to be read with extra scepticism, even when they're methodologically sound by their own standards.

Residual confounding: when adjustment isn't enough

Even well-designed studies that statistically adjust for confounders can have what's called residual confounding. This happens when the adjustment was imperfect; the confounder was measured crudely, or there are confounders the researchers didn't even know to measure.

Consider pharmacokinetics research on compounds like cannabidiol. Studies frequently adjust for body weight, since it influences distribution. But body composition (fat mass vs lean mass) matters too, because cannabinoids are highly lipophilic, they're attracted to fatty tissue. If a study only captures weight and not body composition, the adjustment is incomplete. The confounder is still partially in the data.

This is especially relevant for anything involving first-pass metabolism, where individual variation in liver enzyme activity (particularly CYP450 enzyme expression) is a real confounder for oral cannabinoid studies, and it's extremely hard to measure at scale. Most studies don't try.

A brief note on what confounding is not

Confounding gets conflated with a few things it isn't, and the confusion is worth clearing up.

It's not the same as measurement error. If a study measures someone's cannabis use inaccurately (self-report bias, recall bias), that's a validity problem, not confounding per se, though the two can interact messily.

It's not the same as reverse causation. If a study finds an association between variable A and outcome B, it might be that B is actually causing A. That's a causal-direction problem. Confounding is about a third variable creating a spurious association, not about the arrow pointing the wrong way.

And it's not evidence that a study is worthless. A study can have significant confounding and still be the best available evidence on a question. The goal isn't to dismiss studies that have limitations; it's to weight them appropriately.

How researchers try to address it

Beyond randomisation, researchers use several approaches: multivariable regression (adjusting for multiple confounders simultaneously), propensity score matching (creating comparison groups that are balanced on measured characteristics), instrumental variable analysis, and sensitivity analyses that test how robust findings are to different assumptions.

None of these is magic. Each has its own assumptions and limitations. Propensity score matching only balances measured confounders. Sensitivity analyses tell you about a range of scenarios but can't confirm which scenario is real. Understanding which method a study used; and what its limitations are, is part of reading it properly.

For anyone who wants to go deeper on the statistical side, the STROBE guidelines (Strengthening the Reporting of Observational Studies in Epidemiology) are a freely available checklist of what good observational research should report. Looking up whether a study followed STROBE is a fast way to assess methodological seriousness. The guidelines are published and maintained by an international collaboration and available through journals including PLOS Medicine.

I'll admit I got a bit obsessed with STROBE a couple of years back, spent a rainy weekend in Hobart going through a stack of older phytomedicine papers against the checklist, my cattle dog Pip doing absolutely nothing helpful from the couch. The hit rate for full compliance was lower than I'd have liked, though honestly not surprising given how the field has matured over time.

The bottom line for readers of plant-medicine research

The plant-medicine space; cannabinoids, phytomedicine, peptides, psychedelics, attracts both genuinely rigorous science and a fair amount of research that moves faster than its methods justify. Confounding is one of the main reasons results that look clean in a press release look considerably messier in the actual paper.

Read the methods section. Read the limitations. Ask what else varied between groups. Ask whether the design could support a causal claim at all. And treat any single study, however well-conducted, as one data point, not a final answer.

That's not cynicism. That's just how evidence accumulates.

Sources

, Hannah Bui, Evidence & Research-Literacy Writer

Common questions

What's the simplest way to define a confounding variable?: A confounder is a third variable that's associated with both what you're studying (the exposure) and the result you're measuring (the outcome), making it look like there's a direct relationship between the two when the real cause might be something else entirely. The fire-stations-and-fire-damage example is the classic: population size confounds the association, not any causal link between stations and damage.
Can a randomised controlled trial eliminate all confounding?: Randomisation distributes both known and unknown confounders across groups by chance, which is why RCTs are considered the strongest design for causal questions. But they're not immune — poor randomisation, small sample sizes, and dropout patterns can reintroduce imbalance. And many questions in plant-medicine research simply can't be answered with an RCT for practical or ethical reasons, so observational designs remain important.
What does 'statistically controlling for a confounder' actually mean?: When researchers say they 'controlled for' something — say, age or tobacco use — they've used statistical methods (usually regression) to estimate what the relationship between exposure and outcome would look like if everyone in the study had the same value for that variable. It's a mathematical adjustment, not a physical one. It only works for confounders that were actually measured, and only as well as those measurements are accurate.
Why is confounding such a common problem in cannabinoid and plant-medicine research specifically?: A few reasons. First, randomised trials in this space are difficult — regulatory hurdles, supply constraints, and blinding challenges all apply. Second, the populations who use these products tend to share characteristics (lifestyle, health attitudes, concurrent substance use) that cluster together as potential confounders. Third, the compounds themselves are often studied in real-world settings where multiple variables change simultaneously, making clean causal attribution hard.
Is residual confounding the same as a study being wrong?: Not exactly. Residual confounding means the statistical adjustment was incomplete — usually because a confounder was measured imprecisely, or because relevant confounders weren't measured at all. The study's conclusions might still be directionally correct, but the size of the effect could be over- or underestimated. It's a reason to treat the precise numbers with caution, not necessarily to throw out the finding altogether.