Causal Inference to Avoid Design Fallacies

Illustrate Simpson's paradox using a lingerie dataset, showing how a style's perceived effectiveness reverses when data is stratified by a confounder (e.g., model's body type).

Draw a causal Directed Acyclic Graph (DAG) for lingerie style, model attributes, and photoshoot rating to identify confounders and colliders.

Apply the back-door criterion to identify the correct adjustment set for estimating the true causal effect of a lingerie style on photoshoot ratings.

Demonstrate Berkson's paradox by showing how conditioning on a collider (e.g., 'pieces selected for the final shoot') creates a spurious negative correlation between two desirable attributes (e.g., comfort and aesthetic appeal).

Solve a numerical exercise: given observational data exhibiting Simpson's paradox, apply an adjustment strategy to estimate the correct causal effect.