Allen Downey | Probably Overthinking It | Talks at Google

TL;DR

Berkson's Paradox, a form of sampling bias, can lead to counterintuitive results and misinterpretation of data, as demonstrated by various examples.

Transcript

[MUSIC PLAYING] SPEAKER 1: Welcome to Google. Today, I'm excited to introduce Allen Downey to Google. He's a prolific author. He's a professor and an amazing guy. Allen worked here at Google from 2009 to 2010. And I see some people out in the audience who probably remember him from that time. He helped us analyze network performance, and we've been... Read More

Key Insights

• ⚾ Berkson's Paradox arises from sampling bias, where selecting based on an effect distorts the correlation among causes of the effect.
• 📶 Understanding causal diagrams and considering the direction and strength of causal relationships is essential in analyzing Berkson's Paradox.
• 🎮 Controlling for confounders is helpful, but controlling for colliders can induce bias, as demonstrated by the Everest paradox.
• ℹ️ It is important to be skeptical of counterintuitive research findings and consider the sources of information to avoid falling into statistical traps.

Questions & Answers

Q: What is Berkson's Paradox?

Berkson's Paradox refers to a sampling bias that occurs when two causes of an effect are negatively correlated, but selecting based on the effect creates a positive correlation due to the sampling process.

Q: How does the low birthweight paradox demonstrate Berkson's Paradox?

Babies of smokers tend to have lower birthweights and higher mortality rates. However, when only low birthweight babies are selected, babies of smokers have lower mortality rates due to other causes of low birthweight being more severe.

Q: How does the obesity paradox exemplify Berkson's Paradox?

Obese individuals have higher mortality rates, but when only individuals with kidney disease are selected, obese individuals have longer lifespans. This is due to other factors causing kidney disease and the bias in the selection process.

Q: Can sampling bias in research studies lead to misleading results?

Yes, if not careful, sampling bias can lead to misleading results and misinterpretation of data. It is crucial to consider the sampling process and potential biases when drawing conclusions from studies.

Summary & Key Takeaways

• Allen Downey discusses Berkson's Paradox and its implications in his book "Probably Overthinking It."

• Berkson's Paradox occurs when two causes of an effect are negatively correlated, but selecting based on the effect distorts the correlation to appear positive.

• Examples of Berkson's Paradox include the low birthweight paradox, obesity paradox, and twin paradox.