Decision and Causality 1. Necessity. Objectives. 2. Recommendations? 3. Options/Alternatives. 4. Consequences: Likelihood and Importance 5. Compare the

Decision and CausalityDecision and Causality

1.1. Necessity. Objectives.Necessity. Objectives.2.2. Recommendations?Recommendations?3.3. Options/Alternatives.Options/Alternatives.4.4. Consequences: Likelihood and Consequences: Likelihood and

Importance Importance 5.5. Compare the alternatives. Compare the alternatives. 6.6. Feasibility and contingency plans. Feasibility and contingency plans. 7.7. Cost of deciding.Cost of deciding.


Evaluate causal modelsEvaluate causal models

Based on Giere, Understanding Based on Giere, Understanding Scientific Reasoning, 4th ed, 1997Scientific Reasoning, 4th ed, 1997



Does saccharin cause cancer?Does saccharin cause cancer?




Experiment on rats fed on 5% Experiment on rats fed on 5% saccharin for 2 generations.saccharin for 2 generations.





Significance: probability of this Significance: probability of this deviation from expected value by deviation from expected value by chance alone.chance alone.



Yes, but in rats and in high doses. Yes, but in rats and in high doses. However, could have small effect in However, could have small effect in humans.humans.

Modelling the Modelling the experimentexperiment

Real

Pop (U)

Hyp.

All C (X)

Hyp..

No C (K)

152Random

Saccharin

F(E)=7/78

No saccharin

F(E)=1/74

Random

Random

Hypothetical sample

1st Generation

P=7,5%

Hypothetical sample


Real

Pop (U)

Hyp.

All C (X)

Hyp..

No C (K)

183Random

Saccharin

F(E)=14/94

No saccharin

F(E)=0/89

Random

Random

Hypothetical sample

2nd Generation

P=0,3%

Hypothetical sample


Randomized experimental designRandomized experimental design, RED, RED

Take random sample from real Take random sample from real population.population.

Split and randomly assign cause to a Split and randomly assign cause to a group.group.

This group functions as a sample of This group functions as a sample of the hypothetical population X the hypothetical population X (eXperimental)(eXperimental)

The other group functions as a sample The other group functions as a sample of the hypothetical population K of the hypothetical population K («Kontrol»)(«Kontrol»)


Can we do this with people?...Can we do this with people?...



In 1747, James Lind carried out a In 1747, James Lind carried out a controlled experiment to discover a controlled experiment to discover a

cure for scurvy.cure for scurvy.

(from http://en.wikipedia.org/wiki/Design_of_experiments)(from http://en.wikipedia.org/wiki/Design_of_experiments)


Lind selected 12 men from the ship, Lind selected 12 men from the ship, all suffering from scurvy, and divided all suffering from scurvy, and divided them into six pairs, giving each group them into six pairs, giving each group different additions to their basic diet different additions to their basic diet for a period of two weeks. The for a period of two weeks. The treatments were all remedies that had treatments were all remedies that had been proposed at one time or another. been proposed at one time or another.


1.1. cider cider

2.2. elixir vitriolelixir vitriol

3.3. seawaterseawater

4.4. garlic, mustard and horseradishgarlic, mustard and horseradish

5.5. vinegar vinegar

6.6. two oranges and one lemon every two oranges and one lemon every day.day.



It depends on the issues. Ethical It depends on the issues. Ethical concerns may prevent this approach concerns may prevent this approach (cannot force people to smoke, for (cannot force people to smoke, for example)example)


For human testing it is also important For human testing it is also important to use blinded or double blinded to use blinded or double blinded experiments:experiments:

Blinded: The subjects do not know to Blinded: The subjects do not know to which group they belongwhich group they belong

Double-blind: Neither the subjects nor Double-blind: Neither the subjects nor the evaluators know to which group the evaluators know to which group each subject belongs. each subject belongs.


Double-blind studies are least Double-blind studies are least susceptible to bias (the experimenter susceptible to bias (the experimenter wants some result, placebo effect, etc)wants some result, placebo effect, etc)


One alternative: One alternative:

Prospective study.Prospective study.

Select individuals based on the Select individuals based on the presence or absence of the possible presence or absence of the possible cause (e.g. smokers and non-smokers)cause (e.g. smokers and non-smokers)

Wait, and check for the correlation of Wait, and check for the correlation of the effect with the possible cause.the effect with the possible cause.

(a “time delay” correlation…)(a “time delay” correlation…)

Model of Prospective Model of Prospective StudyStudy

Framingham study:Framingham study:

In 1950, selected 3074 men and 3433 In 1950, selected 3074 men and 3433 women at random, ages 30-59.women at random, ages 30-59.

Examined every 2 years for 20 years.Examined every 2 years for 20 years.Coronary Heart Disease (CHD) at ages 40-49:Coronary Heart Disease (CHD) at ages 40-49:

Men: 29%Men: 29%

Women: 14%Women: 14%



Controlling for other factorsControlling for other factors

Coronary Heart Disease (CHD) at ages 40-49:Coronary Heart Disease (CHD) at ages 40-49:

Men: Men: Smoking: 22%Smoking: 22% Non-Non-S:11%S:11%

Women: Women: Smoking: 7%Smoking: 7% Non-S:6%Non-S:6%





Coffee drinkers also had significantly Coffee drinkers also had significantly more CHD than non-drinkers.more CHD than non-drinkers.

Could it be correlation with tobacco?Could it be correlation with tobacco?






Could it be correlation with tobacco?Could it be correlation with tobacco?






Nonsmokers had no difference in CHD Nonsmokers had no difference in CHD as a function of coffee drinking.as a function of coffee drinking.


Farmingham

Men aged 30-39

Smokers Nonsmokers

Random

X fx(CHD)=22%

K fk(CHD)=11%

NonrandomAll C No C

Frequency of E


A A prospectiveprospective study (or experiment) study (or experiment) examines the correlation between two examines the correlation between two factors, but the possible cause is factors, but the possible cause is chosen before the effect is evident.chosen before the effect is evident.

There may be effects from other There may be effects from other factors, but these can be accounted factors, but these can be accounted for, and a prospective study can be for, and a prospective study can be quite conclusivequite conclusive


Example: Example: 1960s, National Cancer Institute (USA)1960s, National Cancer Institute (USA)

37,000 smokers and 37,000 nonsmokers37,000 smokers and 37,000 nonsmokers

After 3 years smokers hadAfter 3 years smokers had

Double death rateDouble death rate

Double death rate from heart diseaseDouble death rate from heart disease

Nine times death rate from lung cancerNine times death rate from lung cancer

Correlated with time, amount, inhalationCorrelated with time, amount, inhalation

Decreased death rate for former smokersDecreased death rate for former smokers


A different approach: A different approach:

Breast cancer and contraceptionBreast cancer and contraception

In the 1980s, UK researchers In the 1980s, UK researchers questioned women who had breast questioned women who had breast cancer and were younger than 36 cancer and were younger than 36 years old. 755 responded.years old. 755 responded.

For each of these women researchers For each of these women researchers selected one woman at random with selected one woman at random with no breast cancer.no breast cancer.


A different approach:A different approach:

Each woman was interviewed about Each woman was interviewed about children, marriage, cohabitation, oral children, marriage, cohabitation, oral contraceptives, etc.contraceptives, etc.


Results: Women using oral Results: Women using oral contraceptives for more than 4 yearscontraceptives for more than 4 years


A A retrospectiveretrospective study: study:

Selects sample based on the effect Selects sample based on the effect and tries to reason backwards towards and tries to reason backwards towards cause.cause.

Most susceptible to bias. In this case:Most susceptible to bias. In this case:

ResponseResponse

SurveillanceSurveillance

RecallRecall

InterviewInterview


Response bias:Response bias:

Only some women agreed to Only some women agreed to participate, and this may not be a participate, and this may not be a random samplerandom sample

Surveillance:Surveillance:

Women using contraceptives go to the Women using contraceptives go to the doctor more oftendoctor more often


Recall:Recall:

Subjects may not remember past Subjects may not remember past events accurately, or may have a events accurately, or may have a biased memory.biased memory.

Interview:Interview:

Interviewers knewInterviewers knew


Results: Women using oral Results: Women using oral contraceptives for more than 4 yearscontraceptives for more than 4 years

Model, Retrospective Model, Retrospective StudyStudy

Women UK

Cancer No cancer

Match?

X fx(OC-22)=68%

K fk(OC-22)=69%

NonrandomAll E No E

Frequency of C

Evaluate the modelEvaluate the model

1.1. Model and PopulationModel and Population

2.2. Sample DataSample Data

3.3. Experimental designExperimental design

4.4. Random Sampling and biasRandom Sampling and bias

5.5. SignificanceSignificance

6.6. Summary and conclusionSummary and conclusion


Decision is related to causal models:Decision is related to causal models:

Because we need to understand the Because we need to understand the effects our decisions will cause, andeffects our decisions will cause, and

Because we need to decide which Because we need to decide which experiments to do to test causal experiments to do to test causal modelsmodels

Designing an experimentDesigning an experiment

1.1. Necessity. Objectives.Necessity. Objectives.2.2. Recommendations?Recommendations?3.3. Options/Alternatives.Options/Alternatives.4.4. Consequences: Likelihood and Consequences: Likelihood and

ImportanceImportance 5.5. Compare the alternatives. Compare the alternatives. 6.6. Feasibility and contingency plans. Feasibility and contingency plans. 7.7. Cost of deciding.Cost of deciding.


Examples:Examples:Second hand smoking causes cancer?Second hand smoking causes cancer?1.1. Necessity: important to determine effectNecessity: important to determine effect2.2. Recommendations? Similar to smoking?Recommendations? Similar to smoking?3.3. Options: RED, Prospective, retrospectiveOptions: RED, Prospective, retrospective4.4. Consequences: RED may be unethical, Consequences: RED may be unethical,

prospective takes too long…prospective takes too long…5.5. Compare: RetrospectiveCompare: Retrospective6.6. Feasibility: FeasibleFeasibility: Feasible7.7. Cost of deciding: …Cost of deciding: …


Examples:Examples:Does birth date affect personality (according to Does birth date affect personality (according to

astrology)?astrology)?

1.1. Necessity: not much…Necessity: not much…2.2. Recommendations? No…Recommendations? No…3.3. Options: RED, Prospective, retrospectiveOptions: RED, Prospective, retrospective4.4. Consequences: No bad consequences, RED Consequences: No bad consequences, RED

is most reliableis most reliable5.5. Compare: Best is double blind REDCompare: Best is double blind RED6.6. Feasibility: Double blind may not be Feasibility: Double blind may not be

feasible.feasible.7.7. Cost of deciding: …Cost of deciding: …


Examples:Examples:

Shawn Carlson, 1989Shawn Carlson, 1989http://psychicinvestigator.com/demo/AstroSkc.htmhttp://psychicinvestigator.com/demo/AstroSkc.htm

30 astrologers were asked to match 30 astrologers were asked to match 116 natal charts each to one of 3 116 natal charts each to one of 3 personality profiles (using the personality profiles (using the California Personality Inventory, California Personality Inventory, with the agreement of the with the agreement of the astrologers)astrologers)


Examples:Examples:


Examples:Examples:Does birth date affect personality Does birth date affect personality

(according to astrology)?(according to astrology)?

In this case we do not test the actual causal In this case we do not test the actual causal model of birth date and personality. But model of birth date and personality. But the absence of a correlation between the the absence of a correlation between the astrologers’ predictions and CPI shows astrologers’ predictions and CPI shows there is no causal relation between the there is no causal relation between the factors identified. factors identified. A causal relation A causal relation implies a correlation.implies a correlation.

ConsequencesConsequences

Consequences are an important part Consequences are an important part of any decision.of any decision.

Decisions under uncertainty: Decisions under uncertainty: consequences must be weighted consequences must be weighted with the probability.with the probability.


Example:Example:

0.3% probability of child having 0.3% probability of child having Down’s if mother over 35Down’s if mother over 35

0.5% probability of miscarriage from 0.5% probability of miscarriage from amniocentesis.amniocentesis.

Is it worth the risk? It depends on the Is it worth the risk? It depends on the utility values…utility values…


Example:Example:

Biofuel may reduce carbon emissions Biofuel may reduce carbon emissions by a small fraction (industrialized by a small fraction (industrialized agriculture demands lots of fuel).agriculture demands lots of fuel).

However, crops used for fuel will raise However, crops used for fuel will raise food prices globally.food prices globally.


Example:Example:

Biofuel may reduce carbon emissions Biofuel may reduce carbon emissions by a small fraction (industrialized by a small fraction (industrialized agriculture demands lots of fuel).agriculture demands lots of fuel).

However, crops used for fuel will raise However, crops used for fuel will raise food prices globally.food prices globally.

Lowering consumption could be an Lowering consumption could be an answer. But what is the cost of answer. But what is the cost of decreased economic growth?decreased economic growth?

Documents

Decision and Causality 1. Necessity. Objectives. 2. Recommendations? 3. Options/Alternatives. 4. Consequences: Likelihood and Importance 5. Compare the