Why associations aren't always causal
A study finds that patients who receive Treatment A have better outcomes than those who receive Treatment B. Case closed?
Not so fast. What if the patients who got Treatment A were healthier to begin with? What if the sicker patients were more likely to drop out? What if the researchers only measured outcomes they expected to be different?
Bias and confounding are the two main reasons an observed association might not reflect a true causal relationship.
Learn to spot them, and you'll never be fooled by a headline again.
Bias is a systematic error that distorts results in one direction. Here are the big three.
The patients in your study aren't representative of the population you're trying to study.
The data you collect is systematically inaccurate or incomplete.
Patients who drop out of a study are systematically different from those who stay.
Bias can make an ineffective treatment look effective, or a harmful treatment look safe. And unlike random error, bias doesn't get better with a larger sample size.
A confounder is a third variable that distorts the relationship between your exposure and outcome.
The confounder is associated with both the exposure and the outcome, creating a spurious association.
Early studies found that coffee drinkers had higher rates of lung cancer. Did coffee cause cancer?
No. Coffee drinkers were more likely to smoke. Smoking was the confounder. It was associated with both coffee drinking (exposure) and lung cancer (outcome).
Once you account for smoking, the coffee-cancer association disappears.
RCTs solve confounding through randomization. In observational studies, you must identify and adjust for confounders. If you miss one, your results may be wrong.
How to identify bias and confounding when reading a paper.
"How do I know if there's selection bias?"
Look at Table 1. Compare the study population to typical patients. Are they older, sicker, from a single center? Ask: would these results apply to my patients?
"How do I know if there's information bias?"
Ask how data were collected. Prospective measurement is more reliable than retrospective chart review. Objective measures (lab values) are more reliable than subjective ones (pain scores recalled months later).
"How do I know if confounding is a problem?"
Look at baseline characteristics. If the treatment groups differ in important ways (age, comorbidities), those differences could explain the outcome. Did the authors adjust for them?
"The authors say they 'controlled for' confounders. Is that enough?"
Not always. You can only control for confounders you measured. Unmeasured confounding (e.g., socioeconomic status, health-seeking behavior) can still bias results. Statistical adjustment doesn't fix design flaws.
In observational studies, patients who receive different treatments are often different in ways that affect outcomes. No amount of statistical adjustment can fully overcome this.
This is why we do RCTs when we can.
Read each scenario and identify the type of bias or confounding.
Module 1 · Lesson 2 complete
When reading any observational study, ask: What are the potential sources of bias? What confounders might explain this association?