1 of 40

Lecture 18: Cognitive Biases

Jacob Steinhardt

Stat 157, Spring 2022

2 of 40

Cognitive Biases

Any tendency of our brain that does not �reliably lead to correct conclusions.
Cognitive biases often don’t “feel” incorrect, �so it is useful to explicitly be aware of them.�
We will see that many forecasting �techniques “counter” common cognitive �biases.

3 of 40

Warm-up Question

Split into two groups. (Each sees different question)

4 of 40

Warm-up Question: Group 1

An oil spill happens off the coast of California. There are 2,000 gulls that are in danger of being caught in the spill. You are governor of California and can pay a non-profit to relocate the gulls out of harm’s way. How much are you willing to pay to achieve this?

5 of 40

Warm-up Question: Group 2

An oil spill happens off the coast of California. There are 200,000 gulls that are in danger of being caught in the spill. You are governor of California and can pay a non-profit to relocate the gulls out of harm’s way. How much are you willing to pay to achieve this?

6 of 40

Warm-up Question: Discussion

7 of 40

Bias: Scope Neglect

Our intuitions often do not take into account �the actual magnitude of numbers involved �(beyond “not many” vs. “a lot”).�
Fixes:

Fermi estimates
Sanity checking (why $X and not $5X?)

8 of 40

Forecasting Exercise

Split into two groups. (Each sees different question)

9 of 40

Forecasting Exercise: Group 1

What is the probability that the Bay Area has an earthquake (magnitude >= 4) in the next month?

10 of 40

Forecasting Exercise: Group 2

What is the probability that the Bay Area has an earthquake (magnitude >= 4) in the next year?

11 of 40

Forecasting Exercise: Discussion

12 of 40

Forecasting Exercise: Discussion

13 of 40

Bias Category: Extension Neglect

Generalization of scope neglect
Another instance: disjunction fallacy

“Will California impose a lockdown” vs. �“Will any U.S. state impose a lockdown”?�

Antidotes:

Make forecasts across different time horizons and check� if they are consistent

Or select one horizon that you have best intuition about

Decomposing the problem

14 of 40

Extension Neglect and Base Rates

Sometimes answer implicitly depends on a quantity�
Example: “What is the probability that I have Covid, given symptoms” depends not just on likelihood of symptoms but on the base rate (prior probability) of Covid.�
This is one reason why base rates and reference classes are so useful—they remind us to attend to relevant quantities that we might otherwise neglect.

15 of 40

Some More Prediction Exercises

Give 80% CIs for each of the following:

What is the freezing point of vodka?
How long is the pregnancy of an African bush elephant?
How long does it take Mars to orbit the sun?
What year was George Washington elected president?

16 of 40

Some More Prediction Exercises

Give 80% CIs for each of the following:

What is the freezing point of vodka? -16^oF / -27^oC
How long is the pregnancy of an African bush elephant? 22 months
How long does it take Mars to orbit the sun? 687 days
What year was George Washington elected president? 1789

17 of 40

Anchoring Bias

Anchoring: We tend to have too much “inertia” around our first idea or estimate.

Example: if you forecast with method A, then �forecast again with different method B, the �answer to B is often magically close to A �(because we fudge numbers to make it work).

Other manifestations:

We forget to revisit assumptions
We are too slow to revise hypotheses given new info.

18 of 40

Confirmation Bias

Confirmation bias: We reject info that contradicts �our beliefs and accept info that reinforces them.

We also over-rely on a single information source, �or on sources with similar biases.�
See also: “consumption” part of information �lifecycle

19 of 40

Anchoring and Confirmation: Antidotes

Sanity checking: Are there assumptions I made that I forgot to check / account for?
Reversal test: if I saw A then B, rather than B then A, would the answer be different?

Or: if I don’t want to increase X, how do I feel about decreasing X?

Discussing with other people
Have a wide variety of analogies / modeling tools / etc. so that you don’t anchor too much on any single one

Relatedly, consider problem from many different angles

20 of 40

Misha Yagudin��Another idea is allowing yourself to wear "multiple hats."

Say for "Alexander Lukashenko to remain president of Belarus on January 31st, 2021?" I was

- "busy super who makes appropriate reference class forecast and moves on",

- "a Russian with in-depth ground-on knowledge of the protest",

- "a senior Belarusian politician trying to outplay the situation,"

- "a senior Russian politician trying to outplay the West."

21 of 40

Availability, Substitution, and Exposure

Three related failure modes:

Availability bias: the tendency to rely on the data that are most salient, rather than systematically collecting all relevant evidence.
Substitution heuristic: our tendency to answer a simpler question instead of the question actually in front of us.
Exposure effect: the tendency to believe things just because they have been repeated many times.

22 of 40

Availability Bias

Availability bias: relying on most salient data rather than using all relevant evidence.

Example: “What is the probability of a military or terrorist strike on U.S. soil in the next 10 years?”

You might give different answers based on whether you are old enough to remember September 11th.
Corrective: base rate from Pearl Harbor (1941), US Capitol Shooting (1954), World Trade Center Bombing (1993), and September 11 (2001), which gives ~once per 20 years.

Remedies for availability bias:

Reference class forecasting / base rates
Reading history books

23 of 40

Availability Bias Hard Mode: Personal Experience

We tend to overrate data from our personal experience, and from our immediate social group / subculture.
Can mess up Fermi estimates (“how many people own a credit card?”)
From humanitarian perspective, can lead to misguided policies / aid

�Antidotes:

Talk to people from other backgrounds (and other social classes!)
Read opinion polls, surveys, etc.
Cultivate a diverse group of friends with different experiences / viewpoints

24 of 40

Substitution Heuristic

Substitution heuristic: tendency to answer a simpler question instead of the question actually in front of us.

Example: “What will the Metaculus community forecast of the question ‘Omicron variant deadlier than Delta’ be on Dec 8?”

I mostly thought about the probability that Omicron was deadlier than Delta
I should have spent my time modeling the Metaculus community dynamics

Antidote: sanity-checking. Or explicitly ask “am I using the substitution heuristic?”

Or: “are there important considerations in how the question will resolve that I haven’t accounted for?”

25 of 40

Exposure Effect

Exposure effect: the tendency to believe things �just because they have been repeated many times.

E.g.: if many news articles say something, I end up �believing it, even if they don’t provide evidence.

Antidotes:

Don’t trust anyone 😨
General statistical street-fighting / remember news copies from same source
Be skeptical of numbers / claims unless source is given / work is shown
Good information diet

The scariest bias of them all!

26 of 40