Causality

Artificial Intelligence Online Tutorials

Consider the following example. This is an artificial example, but is designed to show how we cannot predict how interventions will work given only observational data. That is, determining conditional probabilities will not predict what happens when there are interventions.

There is much debate about whether marijuana is a "gateway drug": taking marijuana leads to people take hard drugs. One may believe this is because taking marijuana is positively correlated with taking hard drugs. Consider the following belief network that provides a model of how marijuana could be correlated with taking hard drugs:

First turn on monitoring, and monitor each of the variables. Note the probability of each variable.

First, look at the right two variables:

• What happens to "Takes Hard Drugs" when "Takes Marijuana" is observed to be true? What happens when it is observed to be false?
• What happens to "Takes Marijuana" when "Takes Hard Drugs" is observed to be true? What happens to "Takes Marijuana" when "Takes Hard Drugs" is observed to be true?
• What do you think about the hypothesis that taking marijuana leads to taking hard drugs?
• What do you expect to happen if someone is forced to take marijuana? Are they more likely or less likely to take hard drugs? What if they were forced to not take marijuana?

The left most variable forces the Marijuana variable to have a different value.

• Suppose you force someone to not take marijuana, what do you expect to happen to the value of "Takes Hard Drugs"? Try setting "Force Marijuana" to "ForceTrue". What happens? Is it what you expected?
• What if you force them to not take marijuana? Try setting "Force Marijuana" to "ForceFalse". Does it do what you expect?

Before you look at the solution below try to explain why it is acting like this. What is going on here?

• What is going on here?