Causal Inference

tags : Machine Learning, Statistics

FAQ §

“correlation does not implies causation” meme §

• “correlation does not implies causation” : Yes, commonly said
• But causation also does not imply correlation too. There can be hidden variables.

What are associations? §

• correlation is a limited measure of association
  • variables can be associated but have no correlation
• association are bi-directional: associations between variables run in both direction

Basics §

What is Causal Interference §

• What happens when some intervention is done
• Can only be done if we have a causal model

Causal Prediction §

• Different from normal prediction
• Predicting the effect
  • Being able to predict the consequences of an intervention.
  • Eg. Movement of the trees and wind are statically associated. But nothing in the data tells you that wind causes the trees to move.
• What if I do this?

Causal Imputation §

• Knowing the cause
  • Being able to construct some unobserved counterfactual outcome
• What if I had done something else?

Related topics to Causal Inference §

• Description (of population)
  • The sample is caused by things
• Design (of research project)
  • things need to be drawn w a causal logic which will help us design/calculate around them
  • thinking about why sample differs from the population

Models §

DAGs §

• Example
  • Here A influences the treatment X
  • Here B influences the outcome Y
  • Here C influences both X and Y (Confound, we’d want to control it)
• We can ask multiple questions to this model
• DAGs are intuition pumps: get head out of data, into science
• Gives you a strategy for which control variables you need to play with

GOLEMS §

statistical models to produce scientific insight

• They require additional scientific (causal) models
• The reasons are not found in the data, but rather in the causes of the data
  • i.e We should not try to infer the reason in the data.
• No causes in(the data), No causes out(from the data)

Decision Tree/Flowchart §

• We use it for selecting an appropriate statistical procedure.

• But this kind of using a decision tree to select the statistical procedure is not much helpful for a research scientist because this quite limiting
• Each of these procedures can have bayesian and frequestist version of them
• Mostly useful in industrial testing