The Invisible Third Person in Every Relationship You've Ever Studied

The Problem First

You're a community medicine resident. Your thesis: "Association between mobile phone usage and brain tumours." You collect data from 400 adults. You find it. OR = 2.1, p = 0.008. Mobile phone users have twice the odds of brain tumours.

You present at the departmental seminar. You're ready for the applause.

Your professor doesn't applaud. She asks:

"Who uses mobile phones the most? Urban professionals. Who has better access to diagnostic imaging — MRI, CT? Urban professionals. Who gets diagnosed with brain tumours more often because they actually GET scanned? Urban professionals. Is it possible that mobile phones didn't CAUSE the tumours — the tumours were just detected more often in the same population that happens to use phones?"

You stare at her.

"Your mobile phone variable and your brain tumour variable are both associated with a third variable — urbanisation and healthcare access. That third variable is pulling both strings. You're watching two puppets dance and concluding they're in love."

That third variable — the puppeteer — is a confounder. And it just destroyed your thesis.

Before the Term — What's Actually Happening?

You see two things move together. A and B are correlated. Your brain screams: A causes B.

But there are actually THREE possible explanations for why A and B move together:

Explanation 1: A → B          (A causes B) Explanation 2: A ← B          (B causes A — reverse causation) Explanation 3: A ← C → B      (C causes BOTH — confounding)

Confounding is Explanation 3. Some THIRD variable — C — is independently associated with BOTH your exposure (A) AND your outcome (B), creating a phantom association between A and B that has nothing to do with A actually affecting B.

The analogy: Ice cream sales and drowning deaths are correlated. More ice cream sold → more drownings. Does ice cream cause drowning?

No. Summer causes both. Hot weather → people buy ice cream. Hot weather → people go swimming → some drown. Ice cream and drowning are connected through a third variable (temperature/season), not through each other.

Every first-year student laughs at the ice cream example. Then they go write a thesis full of the exact same error, just dressed in medical terminology.

Word Surgery: "Confounder"

"Confound"

Root: Latin confundere = con- (together) + fundere (to pour) → "to pour together" / "to mix together" / "to throw into confusion"

Original meaning: To MIX things so thoroughly that you can't separate them. To CONFUSE by mingling.

16th century English usage: "To confound" = "to defeat by mixing up" / "to destroy by confusion"

• "God confounded their language" (Tower of Babel — He MIXED their languages so they couldn't understand each other)
• "The evidence confounded the jury" (MIXED UP the jury's thinking)

→ In statistics: A confounder MIXES the effect of the exposure with its own effect, so you can't tell which one is causing the outcome. The exposure's effect and the confounder's effect are POURED TOGETHER, inseparable.

→ Aha: "Confound" = "pour together." A confounder POURS its own effect INTO the exposure-outcome relationship, contaminating it. You think you're measuring the effect of A on B. You're actually measuring the effect of A MIXED WITH the effect of C. The mixture is what "confounding" means — literally, a pouring together of effects you needed to keep separate.

"Confounding Variable" / "Confounder"

Literal meaning: "The variable that pours its effect together with the exposure's effect, making them inseparable"

→ The puppeteer analogy: You're watching Puppet A (exposure) and Puppet B (outcome) dance together. You conclude A makes B dance. But Confounder C is above the stage, pulling BOTH strings. Remove C's strings, and A and B stop dancing together. The association was C's effect, poured into the A-B relationship.

Naming Family

The Three Conditions — When Is Something a Confounder?

A variable C is a confounder of the A → B relationship if and only if ALL THREE of these are true:

Condition 1: C is associated with the exposure (A)

The confounder must be linked to the exposure. If smoking (C) is a confounder of the coffee (A) → cancer (B) relationship, it's because smokers are more likely to drink coffee. If smokers and non-smokers drank coffee equally, smoking couldn't confound.

Condition 2: C is independently associated with the outcome (B)

The confounder must ALSO cause (or be associated with) the outcome, independent of the exposure. Smoking causes cancer regardless of coffee consumption.

Condition 3: C is NOT on the causal pathway between A and B

This is the one students forget. If A → C → B (A causes C, which causes B), then C is a mediator, not a confounder. Adjusting for a mediator removes the real effect. Adjusting for a confounder reveals the real effect. Getting this wrong is catastrophic.

The diagram:

CONFOUNDER:        C ↙ ↘ A    B      (C causes both. A-B link is spurious.) → ADJUST for C to remove fake association. MEDIATOR:        A → C → B   (C is HOW A causes B. Part of the causal chain.) → DO NOT adjust for C. You'll erase the real effect. COLLIDER:        A → C ← B   (Both A and B cause C.) → DO NOT adjust for C. You'll CREATE a fake association.

The Coffee-Smoking-Cancer Example

The question: Does coffee cause lung cancer?

The data: Coffee drinkers have higher lung cancer rates.

The confounder: Smoking.

• Smokers drink more coffee (Condition 1: C associated with A ✓)
• Smoking causes lung cancer (Condition 2: C associated with B ✓)
• Smoking is NOT caused by coffee (Condition 3: C is not on the pathway ✓)

After adjusting for smoking: The coffee-cancer association disappears. The relationship was entirely driven by smoking. Coffee was innocent. The confounder was guilty.

Now compare — the aspirin-bleeding-stroke example:

The question: Does aspirin prevent stroke?

The proposed "confounder": Bleeding.

• Aspirin causes bleeding (Condition 1 ✓)
• Bleeding doesn't cause stroke... wait. Let's re-examine.

Actually: Aspirin → antiplatelet effect → reduced clotting → prevents stroke. But aspirin → antiplatelet effect → reduced clotting → increased bleeding. Bleeding is a SIDE EFFECT, not a confounder. It's on a SEPARATE causal pathway. Adjusting for it wouldn't remove confounding — it would create a nonsensical analysis.

The point: Not every third variable is a confounder. The causal structure matters. Drawing the DAG (Directed Acyclic Graph) is how you figure out what to adjust for and what to leave alone.

Who Named This? — The History of Seeing Invisible Third Variables

The Concept Before the Name

The idea that a hidden third variable can create a fake association is ancient:

John Graunt (1662) — One of the first to notice that death rates varied by neighbourhood. But he recognised that wealthier neighbourhoods had lower death rates not because wealth CURED disease, but because wealthy people had better nutrition, sanitation, and healthcare. Wealth was confounding the neighbourhood-mortality relationship.

Pierre-Simon Laplace (1812) — Noted that the association between the phase of the moon and hospital admissions could be explained by the fact that more people walked outside during full moons (better visibility), leading to more accidents. The moon didn't cause illness. It caused walking. Walking caused admissions.

The Statistical Formalisation

Ronald Fisher (1926) — In The Design of Experiments, Fisher identified the problem of "lurking variables" that could explain observed associations. His SOLUTION was randomisation: if you randomly assign treatments, known AND unknown confounders are distributed equally between groups, neutralising their effect.

Fisher's genius insight: You don't need to IDENTIFY every confounder. You don't need to MEASURE every confounder. You just need to RANDOMISE, and probability ensures they balance out. This is why the RCT is the gold standard — not because it measures confounders, but because it neutralises ALL of them simultaneously, including ones you didn't know existed.

Austin Bradford Hill (1965) — In his landmark paper on causal criteria (the "Bradford Hill criteria"), Hill explicitly identified confounding as a primary alternative explanation for observed associations. His criterion of "experiment" (does removing the exposure remove the outcome?) was partly an anti-confounding test.

The Word Itself in Statistics

The term "confounding" in its statistical sense was formalised in the mid-20th century by epidemiologists working on smoking and lung cancer:

Jerome Cornfield (1959) — Wrote the definitive early paper on confounding in the smoking-cancer debate. The tobacco industry argued that a "constitutional factor" (genetic predisposition to both smoking and cancer) could be confounding the association. Cornfield proved mathematically that for a confounder to fully explain the observed smoking-cancer association (RR ≈ 10), the confounder would need to be at least 10 times more common in smokers than non-smokers. No known variable met this criterion. Confounding was mathematically ruled out as a sufficient explanation.

Cornfield's inequality remains one of the most elegant arguments in epidemiology: it quantified how strong a confounder would NEED to be, and showed that no plausible confounder was strong enough.

How Confounding Distorts — The Full Mechanics

Positive Confounding (Inflates the Association)

Example: The association between alcohol and oesophageal cancer.

• True RR (after adjustment): 3.0
• Crude (unadjusted) RR: 5.5

Why the inflation? Smoking is a positive confounder:

• Drinkers smoke more (Condition 1 ✓)
• Smoking causes oesophageal cancer (Condition 2 ✓)
• Smoking is not caused by drinking (Condition 3 ✓... mostly)

Smoking's cancer-causing effect was ADDED to alcohol's effect in the crude analysis. The crude RR = alcohol's real effect + smoking's confounding effect. After removing smoking's contribution → RR drops from 5.5 to 3.0.

Negative Confounding (Hides the Association)

Example: The association between physical exercise and coronary heart disease (CHD).

Crude data might show a WEAK association between exercise and reduced CHD. But after adjusting for age:

• Crude RR: 0.85 (15% reduction — modest)
• Adjusted RR: 0.55 (45% reduction — substantial)

Why? Age is a negative confounder:

• Older people exercise LESS (Condition 1 ✓)
• Older people have MORE CHD (Condition 2 ✓)
• Age is not caused by exercise (Condition 3 ✓)

In the crude analysis, the exercise group is younger (because young people exercise more). Younger people naturally have less CHD. But the MIXING of "younger + exercising" meant the crude analysis UNDERESTIMATED exercise's benefit — some of the apparent benefit was actually just youth. Wait... that would make the crude RR LOWER (more protective), not higher.

Let me reconsider. Actually, negative confounding works the OTHER way here:

• Exercise group is younger → lower CHD (makes exercise look MORE protective in crude)
• But if we're seeing crude RR = 0.85 and adjusted RR = 0.55, age must be masking the effect differently

Correct example of negative confounding:

Coffee and Parkinson's disease. Crude OR ≈ 0.8 (modest protection). After adjusting for smoking: OR ≈ 0.5 (strong protection).

• Smokers drink more coffee (Condition 1 ✓)
• Smoking independently reduces Parkinson's risk (Condition 2 ✓ — one of the few protective effects of smoking)
• Smoking is not on the coffee → Parkinson's pathway (Condition 3 ✓)

In the crude analysis, coffee's protective effect was DILUTED because coffee-drinkers also smoked, and smoking has its OWN protective effect. After removing smoking's effect → coffee's TRUE protective effect emerges as stronger. The confounder was HIDING part of the real association.

Confounding That Reverses the Association (Simpson's Paradox)

The most terrifying form. The confounder doesn't just inflate or hide — it FLIPS the direction.

The famous Berkeley admissions example (1973):

• Overall: 44% of male applicants admitted, 35% of female applicants admitted. Gender bias? Men favoured?
• By department: In MOST departments, women were admitted at EQUAL or HIGHER rates than men.

The confounder: Department choice. Women disproportionately applied to MORE COMPETITIVE departments (lower acceptance rates). Men disproportionately applied to LESS COMPETITIVE departments (higher acceptance rates). When you aggregated across departments, it LOOKED like gender bias against women. Department by department, there was no bias — or slight bias FAVOURING women.

The aggregated data said the opposite of the truth because the confounder (department competitiveness) was distributed differently between men and women.

In medicine: A drug might appear harmful overall but beneficial within every subgroup — if sicker patients (who have worse outcomes regardless) are more likely to receive the drug. The crude association (drug → worse outcomes) is the REVERSE of the true effect (drug → better outcomes within each severity group). Severity confounds the drug-outcome relationship so powerfully that it flips the direction.

Methods of Controlling Confounding

At the Design Stage (BEFORE collecting data)

At the Analysis Stage (AFTER collecting data)

Why Randomisation Is King

Every analytical method has the same fatal flaw: it can only control for confounders you MEASURED.

If there's a confounder you didn't think of — didn't measure — didn't include in your model — it's still confounding. And you don't know it.

Randomisation controls for confounders you measured AND confounders you didn't measure AND confounders you don't even know exist. It does this by the simple act of flipping a coin. Coin flips don't care about confounders.

This is why an RCT with 200 patients often provides stronger evidence than an observational study with 200,000 patients. The observational study has more data but unresolved confounding. The RCT has less data but no confounding (in expectation). Quantity of data doesn't compensate for quality of design.

The Confounder vs The Imposters — Critical Distinctions

Confounder vs Mediator

This is the distinction that destroys careers when confused.

Confounder: A → B association exists BECAUSE of C. Remove C → the association disappears. C was creating a fake link.

Mediator: A → B association exists THROUGH C. Remove C → you've removed the actual mechanism. You've accidentally proven the drug "doesn't work" by removing the pathway through which it works.

Clinical example:

Statin → LDL reduction → reduced MI.

LDL reduction is a MEDIATOR. It's HOW statins prevent MI. If you "adjust for LDL" in your regression, statins will appear to have NO effect on MI. You haven't removed confounding. You've removed the causal mechanism. You've statistically proven that statins don't work — which is absurd.

The mistake: Adjusting for mediators is one of the most common analytical errors in medical research. It's called overadjustment bias. You think you're being thorough by "controlling for everything." You're actually erasing real effects.

How to tell the difference:

Does C come BEFORE A in time?     → Possible confounder Does C come BETWEEN A and B?      → Possible mediator Does C come AFTER both A and B?   → Possible collider — DON'T adjust

Confounder vs Effect Modifier (Interaction)

Confounder: C creates a FAKE association (or distorts a real one). After adjusting for C, the TRUE association is revealed. The adjusted association is ONE NUMBER — it doesn't vary by C.

Effect modifier: C changes the SIZE of the real association. The TRUE association IS DIFFERENT at different levels of C. There is no single "adjusted" number — you need to report the effect separately for each level of C.

Clinical example:

Aspirin prevents MI in men (RR = 0.56) but NOT in women (RR = 0.99). Sex is an EFFECT MODIFIER — the real effect of aspirin genuinely differs by sex. You should NOT "adjust for sex" and report one number. You should STRATIFY by sex and report two numbers.

The test: If the stratum-specific estimates are SIMILAR → confounding (report one adjusted estimate). If they are DIFFERENT → effect modification (report separate estimates for each stratum and test the interaction).

Confounder vs Collider

This is the subtlest and most recently appreciated distinction.

Confounder: C → A and C → B. C is a COMMON CAUSE. Adjusting for C removes bias.

Collider: A → C and B → C. C is a COMMON EFFECT. Adjusting for C CREATES bias.

The classic collider example — the "obesity paradox":

Among hospitalised patients with heart failure, obese patients appear to survive LONGER than non-obese patients. Obesity is "protective"?

The explanation: Hospitalisation is a COLLIDER. Both obesity (A) and cardiac risk factors (B) increase the probability of hospitalisation (C). If you study ONLY hospitalised patients (conditioning on C), you create a spurious negative association between obesity and cardiac risk. Among hospitalised patients, obese patients are there partly BECAUSE of their obesity — they might have fewer other risk factors. Non-obese patients who are hospitalised must have MORE severe cardiac risk factors (otherwise they wouldn't be hospitalised). So within the hospital, obesity APPEARS protective because you've conditioned on the collider.

In the general population, obesity is NOT protective. The "paradox" is entirely an artefact of conditioning on a collider.

The rule: Before adjusting for ANY variable, draw the DAG. If it's a confounder (common cause) → adjust. If it's a collider (common effect) → DO NOT adjust. If it's a mediator (on the pathway) → DO NOT adjust (unless you specifically want to decompose direct vs indirect effects).

Branch-by-Branch — Where Confounders Destroy Research

General Medicine

The scenario: Observational study: Vitamin D levels are inversely associated with cardiovascular mortality. Low vitamin D → more heart deaths. OR = 1.8, p < 0.001.

The confounder avalanche:

• Sick people go outside less → less sun → less vitamin D → reverse causation/confounding by illness
• Obese people sequester vitamin D in fat → lower serum levels. Obesity also causes CVD → confounding by obesity
• Lower socioeconomic status → worse diet → lower vitamin D. Lower SES also → worse CVD outcomes → confounding by SES
• Older people have less vitamin D. Older people also die more from CVD → confounding by age

The result: Multiple large RCTs of vitamin D supplementation (VITAL, ViDA, D-Health) showed NO cardiovascular benefit. The observational association was entirely confounded. Vitamin D was a MARKER of health (healthy people go outside, exercise, eat well), not a CAUSE of health.

The lesson: Every observational association between a "healthy" behaviour/biomarker and a good outcome should be assumed confounded until an RCT proves otherwise. People who take vitamins, eat organic, do yoga, and meditate also tend to be wealthy, educated, non-smoking, and health-conscious. The specific vitamin/behaviour is a PASSENGER in a vehicle of general healthiness.

Surgery

The scenario: Retrospective study: Patients who received robotic surgery had lower complication rates than those who received open surgery. Adjusted OR = 0.65, p = 0.01.

The confounders the adjustment missed:

• Surgeon selection bias: The best, most experienced surgeons adopted robotic surgery first. They would have had lower complications with ANY technique.
• Patient selection bias: Surgeons chose robotic for "easier" cases (lower BMI, fewer adhesions, earlier stage). The hard cases got open surgery.
• Hospital volume: Robotic surgery concentrated at high-volume centres. High-volume centres have better outcomes regardless of technique.
• Era effect: Robotic surgery was adopted later. Later era = better anaesthesia, better antibiotics, better perioperative care. The robot gets credit for improvements in everything else.

The residual confounding: Even after "adjusting for" age, BMI, ASA score, and stage, the measured variables don't capture surgeon skill, case complexity, or institutional quality. Residual confounding — the confounding left over after adjustment — is likely substantial. The OR = 0.65 may reflect surgeon quality and patient selection, not robotic superiority.

Paediatrics

The scenario: Cohort study: Breastfed children have higher IQ scores at age 7. Mean difference = 4.2 points, p < 0.001.

The confounder you can never fully adjust for: maternal intelligence and socioeconomic status.

• Higher maternal IQ → more likely to breastfeed (follows health recommendations)
• Higher maternal IQ → children genetically predisposed to higher IQ
• Higher SES → more likely to breastfeed (time, support, information)
• Higher SES → children raised in more stimulating environments → higher IQ

The PROBIT trial (Belarus): A cluster-RCT that RANDOMISED breastfeeding PROMOTION (not breastfeeding itself — you can't randomise that). At 16 years: no significant IQ difference. The observational association was largely confounded by maternal IQ and SES.

The lesson: The breastfeeding-IQ association is one of the most replicated and most confounded findings in paediatric epidemiology. Every observational study finds it. The one quasi-experimental study doesn't. The confounder (maternal intelligence) is so tightly bound to the exposure (breastfeeding) that no amount of statistical adjustment can fully separate them.

Obstetrics

The scenario: Observational data shows that elective caesarean section at 39 weeks has lower neonatal morbidity than vaginal delivery. Adjusted OR = 0.7.

The confounders:

• Indication bias: Elective CS at 39 weeks is planned. These pregnancies are SCREENED — anomalies, complications, and high-risk features are detected and managed. Vaginal delivery includes unscreened, unplanned, and complicated deliveries.
• Socioeconomic selection: Women who choose elective CS at 39 weeks tend to be in private healthcare, better nourished, better monitored, and deliver at tertiary centres with NICUs.
• Time-of-day effect: Elective CS happens during daytime with full staffing. Emergency vaginal deliveries happen at 3 AM with skeleton staff.

The confounder you can't measure: The "optimised pregnancy" — a planned 39-week CS represents the endpoint of a carefully managed pregnancy. The vaginal delivery comparison group includes everything from perfect births to unattended emergencies. You're comparing a curated population to an uncurated one and attributing the difference to the mode of delivery.

Psychiatry

The scenario: Observational study: Cannabis use in adolescence is associated with schizophrenia. OR = 2.1.

The confounding minefield:

• Shared genetic risk: The same genes that predispose to schizophrenia may also predispose to cannabis use (novelty-seeking, dopamine sensitivity). This is not classical confounding — it's confounding by shared genetic architecture.
• Prodromal self-medication: Schizophrenia has a prodromal phase (subtle symptoms years before diagnosis). Adolescents in the prodrome may use cannabis to self-medicate anxiety, social discomfort, or perceptual disturbances. This is reverse causation, not forward causation.
• Social environment: Both cannabis use and schizophrenia risk are associated with urban upbringing, social deprivation, childhood trauma, and social isolation.

The Mendelian randomisation evidence: Studies using genetic variants associated with cannabis use as instruments (to bypass environmental confounding) have shown SOME evidence of a causal effect — but weaker than the observational OR of 2.1 suggests. The truth is probably: cannabis has a modest causal effect on schizophrenia risk, substantially inflated by confounding and reverse causation in observational studies.

The policy problem: If the real causal effect is OR = 1.3 (not 2.1), the public health messaging changes dramatically. "Cannabis doubles your risk of schizophrenia" becomes "Cannabis increases your risk by about 30%." Both are concerning. But the first sounds like a near-certainty. The second sounds like one risk factor among many. The confounder inflated the fear.

Community Medicine / PSM

The scenario: National survey: Households with clean cooking fuel (LPG/electricity) have lower child mortality than households with biomass fuel (wood/dung). Adjusted RR = 0.6.

The confounders that "adjustment" can't fully capture:

• Clean fuel → higher income → better nutrition, sanitation, healthcare, education → lower mortality. The fuel is a MARKER of socioeconomic status, not necessarily the cause of reduced mortality.
• Clean fuel → urban residence → closer to hospitals → faster emergency care
• Clean fuel → maternal education → better child-rearing practices

The adjustment problem: You can adjust for income, education, and urban/rural. But "income" is a crude variable — the difference between ₹15,000/month and ₹50,000/month is enormous, yet both might fall in the same income bracket in your survey. Residual confounding within crude categories is the silent killer of observational epidemiology.

The policy implication: Clean fuel DOES reduce indoor air pollution, which DOES reduce respiratory infections, which DOES reduce child mortality. There IS a real causal pathway. But the MAGNITUDE — RR = 0.6 — is almost certainly inflated by confounding. The real attributable effect of switching fuel (holding everything else constant) is probably smaller. If policy allocates resources assuming RR = 0.6 and the true effect is RR = 0.85, the cost-effectiveness calculations are wrong.

Orthopaedics

The scenario: Registry study: Patients who received physical therapy after ACL reconstruction had better functional outcomes than those who didn't. Adjusted OR = 2.3.

The confounder: motivation.

• Patients who attend physiotherapy are motivated. Motivated patients also do home exercises, follow restrictions, maintain healthy weight, and have better psychological coping.
• Patients who DON'T attend physiotherapy may be depressed, non-compliant, have transport barriers, or have other injuries limiting participation.

Motivation is a confounder you CANNOT measure and CANNOT adjust for. No registry, no database, no questionnaire captures "how much this person actually cares about their recovery." It confounds virtually every physiotherapy outcome study.

The only solution: Randomise. But randomising patients to "no physiotherapy" after ACL reconstruction is ethically questionable. So the confounding persists — and every observational physiotherapy study carries it.

The 6 Ways Not Knowing Confounders Destroys You

1. You believe every observational association is causal

"Red meat causes cancer." "Screen time causes depression." "Vitamin D prevents everything." These observational associations are confounded to varying degrees. Without understanding confounding, you accept them as facts and make clinical recommendations based on phantom relationships.

2. You adjust for the wrong variables

You throw every variable into your regression "to be thorough." Some are mediators — you've removed the real effect. Some are colliders — you've created new bias. Overadjustment is as dangerous as underadjustment, and more common because it FEELS rigorous.

3. You can't evaluate observational studies in journal club

Every observational study claims to have "adjusted for confounders." Your job is to ask: Which ones did they miss? Is residual confounding plausible? How strong would an unmeasured confounder need to be to explain the result? Without this skill, you take adjusted estimates at face value.

4. You can't explain why RCTs are the gold standard

"RCTs are better because they're randomised" is circular. The REASON randomisation is powerful is that it eliminates confounding — including confounders you didn't measure, didn't think of, and don't know exist. If you don't understand confounding, you can't explain why the design hierarchy exists.

5. You misinterpret null RCT results after positive observational studies

Vitamin D: Observational studies → protective. RCTs → null. Hormone replacement therapy: Observational studies → cardioprotective. RCT (WHI) → harmful. Without understanding confounding, these reversals seem like "science can't make up its mind." With understanding, they make perfect sense: the observational signal was confounded, the RCT removed the confounding, and the truth was different.

6. You confuse confounders with effect modifiers in clinical practice

If age is a CONFOUNDER: the treatment works the same at every age, but age was distorting the crude estimate. Report one adjusted estimate.

If age is an EFFECT MODIFIER: the treatment works differently at different ages. Report separate estimates. Prescribe differently for different ages.

Treating a confounder as an effect modifier means you unnecessarily stratify. Treating an effect modifier as a confounder means you miss that the drug helps some patients and harms others. Both errors change clinical decisions.

The DAG — Your Confounder Detection Tool

Word Surgery: DAG

DAG = Directed Acyclic Graph

• Directed = arrows point in ONE direction (cause → effect)
• Acyclic = no loops (A can't cause B which causes A)
• Graph = a diagram of nodes (variables) and edges (arrows)

→ "A one-way, no-loops diagram of what causes what."

How to Use a DAG

Before ANY observational analysis:

1. Draw every variable you think is relevant (exposure, outcome, potential confounders, mediators, colliders)
2. Draw arrows showing what you believe causes what
3. Identify confounders: Common causes of both exposure and outcome (arrows pointing FROM the variable TO both exposure and outcome)
4. Identify mediators: Variables on the causal pathway (exposure → variable → outcome)
5. Identify colliders: Variables caused BY both exposure and outcome (arrows pointing TO the variable FROM both)
6. Adjust for confounders. Leave mediators and colliders alone.

The DAG is the single most important analytical tool in observational epidemiology. It forces you to make your causal assumptions explicit BEFORE running the analysis. Without a DAG, you're adjusting blind — hoping you picked the right variables, with no systematic way to check.

The One Thing to Remember

A confounder is a third variable that creates a fake connection — or distorts a real one — between the thing you're studying and the outcome you're measuring. It does this by being associated with BOTH, giving the illusion that one causes the other when actually the confounder is driving both.

The resident who reads "smoking is associated with pancreatic cancer, OR = 2.8, adjusted for age and sex" and asks "What about alcohol? What about diet? What about SES? What about BMI? How strong would an unmeasured confounder need to be to explain this?" — that resident understands confounding.

The resident who reads "OR = 2.8, adjusted" and accepts it as causal truth — that resident is watching puppets dance and believing they're in love.

The puppeteer is always there. Your job is to look up.