reasoning with uncertainty. often, we want to reason from observable information to unobservable...

31
Reasoning with Uncertainty

Post on 21-Dec-2015

213 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Reasoning with Uncertainty

Page 2: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Reasoning with Uncertainty

• Often, we want to reason from observable information to unobservable information

• We want to calculate how our prior beliefs change given new available evidence

• Bayes rule tells us how to optimally reason with uncertainty. Do people reason like Bayes rule?

Page 3: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Bayes Rule

Prior probability

Evidence

Posterior Probability

Bayes rule tells us how the available evidence should alter our belief in something being true

Page 4: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Difficulties in Reasoning with Uncertainty

• Problems reasoning with probabilities– frequencies are easier to understand

• Problems understanding Conditional probability– doctors need to calculate the probability of disease

given the observed symptoms: P( disease | symptoms )

– Sometimes P( symptoms | disease ) is used incorrectly when reasoning about the likelihood of a disease

– Why is this wrong?

Page 5: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

• To get P( disease | symptom ), you need to know about P( symptom | disease ) and also the base rate -- prevalence of the disease before you have seen patient

• More intuitive example:– what is the probability of being tall given you are

player in the NBA? – what is the probability of being a player in the NBA

given that you are tall?

P( NBA player | tall ) ≠ P( tall | NBA player )

The base rate is important

Page 6: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Reasoning with base rates

• Suppose there is a disease that affects 1 out of 100 people

• There is a diagnostic test with the following properties:– If the person has the disease, the test will be positive

98% of the time– if the person does not have the disease, the test will

be positive 1% of the time

• A person tests positive, what is the probability that this person has the disease?– Frequent answer = .98– Correct answer ≈ .50

Page 7: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Are we really that bad in judging probabilities?

According to some researchers (e.g., Gigerenzer), it matters how the information is presented and processed. Processing frequencies is more intuitive than probabilities (even it leads to the same outcome).

Page 8: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

A counting heuristic (in tree form)

10,000 people

100 have disease 9,900 do not

98 test positive 2 test negative 99 test positive 9801 test negative

P( disease | test positive ) = 98 / ( 98 + 99 ) ≈ .50

Page 9: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

The same thing in words ...

• Let’s take 10,000 people.

• On average, 100 out of 10,000 actually have the disease and 98 of those will test positive (98% true positive rate)

• Among the 9,900 who do not have the disease, the test will falsely identify 1% as having it. 1% of 9,900 = 99

• On average, out of 10,000 people:98 test positive and they have the disease99 test positive and they do not have the disease.

• Therefore, a positive test outcome implies a 98/(98+99)≈50% chance of having the disease

Page 10: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Change the example

• What now if the disease affects only1 out of 10,000 people?

• Assume same diagnosticity of test (98% true positive rate, 1% false positive rate)

• A person tests positive, what now is the probability that this person has the disease?

Page 11: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

A counting heuristic (in tree form)

1,000,000 people

100 have disease 999,900 do not have the disease

98 test positive 2 test negative 9999 test positive 989901 test negative

P( disease | test positive ) = 98 / ( 98 + 9999 ) = .0097 (smaller than 1%)

Page 12: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Bayes Rule

• The previous example essentially is a simple way to apply Bayes rule:

| ( )( | )

| ( ) | ( )

P positive disease P diseaseP disease positive

P positive disease P disease P positive not disease P not disease

| .98P positive disease

| .01P positive not disease

.0001P disease

| .0097P disease positive

Page 13: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Normative Model

• Bayes rule tells you how you should reason with probabilities – it is a prescriptive (i.e., normative) model

• But do people reason like Bayes?In certain circumstances, the base rates are neglected base rate neglect

Page 14: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

The Taxi Problem: version 1• A witness sees a crime involving a taxi in Carborough.

The witness says that the taxi is blue. It is known from previous research that witnesses are correct 80% of the time when making such statements.

• What is the probability that a blue taxi was involved in the crime?

Page 15: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

The Taxi Problem: version 2• A witness sees a crime involving a taxi in Carborough.

The witness says that the taxi is blue. It is known from previous research that witnesses are correct 80% of the time when making such statements.

• The police also know that 15% of the taxis in Carborough are blue, the other 85% being green.

• What is the probability that a blue taxi was involved in the crime?

Page 16: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Base Rate Neglect: The Taxi Problem

• Failure to take prior probabilities (i.e., base rates) into account

• In the taxi story, the addition of:

“The police also know that 15% of the taxis in Carborough are blue, the other 85% being green.”

has little influence on rated probability

Page 17: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Base Rate Neglect (2)

• Kahneman & Tversky (1973).

group A: 70 engineers and 30 lawyers

group B: 30 engineers and 70 lawyers

• What is probability of picking an engineer in group A and B? Subjects can do this …

Page 18: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Provide some evidence …

• “Jack is a 45 year-old man. He is married and has four children. He is generally conservative, careful, and ambitious. He shows no interest in political and social issues and spends most of his free time on his many hobbies, which include home carpentry, sailing, and mathematical puzzles”

• What now is probability Jack is an engineer?

• Estimates for both group A and group B was P = .9

Page 19: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Heuristics and Biases

• Tversky & Kahneman propose that people often do not follow rules of probability

• Instead, decision making may be based on heuristics

• Lower cognitive load but may lead to systematic errors and biases

• Example heuristics– representativeness– availability

Page 20: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

All the families having exactly six children in a particular city were surveyed. In 72 of the families, the exact order of the births of boys and girls was: G B G B B G

What is your estimate of the number of families surveyed in which the exact order of births was: B G B B B B

Answer: a) < 72 b) 72 c) >72

Page 21: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

The sequence “G B G B B G” is seen as

A) more representative of all possible birth sequences.

B) better reflecting the random process of B/G

Representativeness Heuristic

Page 22: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

A coin is flipped. What is a more likely sequence?A) H T H T T HB) H H H H H H

A) #H = 3 and #T = 3 (in some order)B) #H = 6

Gambler’s fallacy: wins are perceived to be more likely after a string of losses

Page 23: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs
Page 24: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Does the “hot hand” phenomenon exist?

Most basketball coaches/players/fans refer to players having a “Hot hand” or being in a “Hot zone” and show “Streaky shooting”

However, making a shot after just making three shots is pretty much as likely as after just missing three shots

not much statistical evidence that basketball players switch between a state of “hot hand” and “cold hand”

(Gilovich, Vallone, & Tversky, 1985)

Page 25: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Availability Heuristic

• Are there more words in the English language that begin with the letter V or that have V as their third letter?

• What about the letter R, K, L, and N?

(Tversky & Kahneman, 1973)

Page 26: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.

Rate the likelihood that the following statements about Linda are true:

a) Linda is active in the feminist movement

b) Linda is a bank teller

c) Linda is a bank teller and is active in the feminist movement

Rating C as more likely than B and A is a Conjunction Fallacy

Page 27: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

What to make of these results?

• One interpretation of Tversky & Kahneman’s findings: – people do not use proper probabilistic reasoning– people use arbitrary mechanisms/ heuristics with no

apparent rationale

• However, Gigerenzer and Todd show in their “Fast and Frugal Heuristics” research program that heuristics can often be very effective

Page 28: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Which city has a larger population?A) San DiegoB) San Antonio

• 66% accuracy with University of Chicago undergraduates. However, 100% accuracy with German students.

• San Diego was recognized as American cities by 78% of German students. San Antonio: 4%

With lack of information, use recognition heuristic

(Goldstein & Gigerenzer, 2002)

Page 29: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

How to pick a stock

Problem: what stocks to invest in?

Solution 1—“optimizing”:– Gather lots of info about many companies– Process with sophisticated tools and choose

Solution 2—the recognition heuristic:– Purchase stocks from recognized companies

(slide from Peter Todd)

Page 30: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

“Paying for the name…….”

(slide from Peter Todd)

Page 31: Reasoning with Uncertainty. Often, we want to reason from observable information to unobservable information We want to calculate how our prior beliefs

Picking Stocks with Recognition Heuristic

• Borges et al. (1999) – can “ignorance” beat the stock market?

• 180 German lay-people recognition of German stocks• 6 month return on DAX 30: Dec 1996 – Jun 1997

• Note: this result has not replicated in other studies (e.g., Boyd, 2001; Rakow, 2002) -- don’t rush to use this heuristic on your own money!

Market Index Recognition Rate > 90%

Recognition Rate < 10%

+34% +47% +13%