Causation Flashcards
Controlled Experimental Design
-Propose causal structure in form of causal hypothesis
-Decide what would happen if particular variables were controlled for
- Compare observed result to the expected result
If causal hypothesis is true certain observations should be found
Issues:
-Other things that could have contributed to the DV
Randomized Experimental Design
-Through random assignment we can reduce the probability that other variables bias the effect we are interested in
(Could be 100s of factors and you don’t even need to know them)
Physical v. Statistical control
Physical-Directly manipulating or regulating the environment
Statistical-Using statistical techniques to account for the influence of confounding variables
(Can be done through regression analysis)
Causal Models
Ex: Rain causes the presence of mud
-Relationship is asymmetric because mud does not cause rain
Observational Models
Ex: Having observed rain will give us information about what we will observe concerning mud
-Not asymmetric
(Rain tells us about mud but the presence of mud also tells us about whether or not it has rained)
Statistical Models
-Similar to observational models
-Specifies the mathematical relationship between the variables as well as the probability distribution of the variables
How do we get from Cause to association back to cause
DAG/d-separation
Parts of a DAG
-Vertices/Nodes- Letters or variables
-Edges- The lines
Types of Confounds-Fork
-Also called common cause bias
(Some variable (z) is a common cause of both x and y)
-X and Y are independent conditional on Z
-By conditioning on Z backdoor is shut
Types of Confounds-Chain
-X influences Z which influences Y
-If you condition on Z the path from X to Y is blocked (Don’t want to do that)
Types of Confounds-Collider
-In a collider, there is no association between X and Y unless you condition on Z
-Conditioning on Z opens up the path and allows non-causal information to flow
(Not good)
Types of Confounds-Descendant
-Conditioning on a descendent of a variable is like conditioning on the variable itself but weaker
d-separation
-Algorithm developed by Judea Pearl
-Criterion for deciding if X is independent of Y, given Z (criterion for finding a backdoor/what to condition on)
Steps:
1. List all paths connecting X and Y
2. Classify each path by whether it is open or closed (Path is open if it does not contain a collider)
3. Classify each path by whether it is a backdoor (A backdoor path has an arrow entering X)
4. If there are any backdoor paths, decide which variables to condition on to close it
Why is this important?
-D-separation translates between language of causation and language of probability distributions
-In short, this is how we go from cause to association and back to cause
Simpson’s paradox
-Statistical phenomenon where a trend that exists between two variables is reversed when data is aggregated or disaggregated into subgroups
