Danielle Belardo, MD Profile picture
Cardiologist obsessed with prevention & nutrition science

Nov 23, 14 tweets

My brilliant med student asked me to explain correlation, causation, confounding &collider bias. I used the following ex… so sharing here in case anyone finds it helpful!

PS -I have learned much from @dnunan79 @Catalogofbias - a great resource for EBM. hopefully he approves😅

A statistical association between 2 variables implies that knowing the value of 1 variable, provides information about the value of the other. It does not necessarily imply that 1 causes the other. Correlation ≠ causation.

To claim that an association represents a causal effect, we need to first rule out 2 possible issues that lead to a non-causal association:

Confounding
Collider bias

Images are:
@Catalogofbias

Confounding occurs when an exposure & an outcome share a common cause.

Failure to control for the confounder makes it appear that there is an assoc btwn the exposure & the outcome, whereas in fact both are caused by the confounder & are not related to each other at all.

For ex: a study found that there is a strong assoc btwn ice cream consumption & crime. Does this mean eating ice cream increases the crime rate?

Not so fast! The relationship is confounded by the time of year (summer!) But it can *appear* that ice cream & crime will be related in an analysis that doesn’t control for summer.

Collider bias occurs when an exposure & outcome share a common effect (the collider). In this case, a distorted association between the exposure & the outcome is produced when we control for the collider.

For ex: let’s say you follow me on social media for only 1 of 2 reasons 1) you are interested in science/medicine or 2) you like eating avocados.

Next we do a study evaluating whether there is a relationship btwn being interested in science/medicine & liking avocados. We analyze 1000 of my social media followers & find an assoc btwn liking science/medicine & liking avocados. This seems plausible, right?

People who are interested in science may know how healthy avocados are.
But we repeat the analysis in a sample of 5000 individuals in the general pop & find no assoc btwn interest in science/medicine & liking avocados.

In this ex, both liking science/medicine & liking avocados are both independent reasons for following me on social media — which is the “collider” (I do enjoy both 😂)

When you stratify/sample based on the collider, you induce a false association from being interested in science/medicine to liking avocados.

Only if we can rule out these biases can we begin to evaluate if association may imply causation. The end

Wow! I am so glad so many of you have found this thread helpful! And happy to report it got the stamp of approval from @dnunan79

If you’re looking to learn more about critical appraisal & EBM (which to me, is a lifelong journey!) follow @dnunan79 , &check out @Catalogofbias !

@dnunan79 @Catalogofbias Another shoutout to @dnunan79 — when I started my journey in learning critical appraisal -he had recommended a few books that have been amazing resources! Learning is a lifelong process, does not finish once medical training is over! 🙂

Share this Scrolly Tale with your friends.

A Scrolly Tale is a new way to read Twitter threads with a more visually immersive experience.
Discover more beautiful Scrolly Tales like this.

Keep scrolling