Decide when correlation is not enough and where counterfactual thinking is required. Test your grasp of confounding, identification, and targeting for real interventions.
What is the primary goal that distinguishes causal inference from pure prediction?
Estimating the effect of an action or treatment on outcomes
Detecting anomalies in streaming telemetry
Maximizing overall predictive accuracy on historical data
Compressing the dataset without losing variance
Which technique is most directly designed to remove confounding when estimating treatment effects from observational data?
K-means clustering
Propensity score methods (matching, weighting, or stratification)
Early stopping in gradient boosting
Principal component analysis
Randomized experiments help causal identification primarily because randomization ensures ______.
All covariates are perfectly measured
Errors follow a normal distribution
Treatment assignment is independent of potential outcomes
Outliers are eliminated from analysis
A model that predicts churn risk well may still be poor for causal decisions if it ______.
Uses regularization to avoid overfitting
Outputs calibrated probabilities
Scores are monotonic with risk
Relies on features affected by the treatment itself
Backdoor adjustment in causal graphs addresses bias introduced by ______.
Common causes of treatment and outcome
Large sample sizes increasing variance
Measurement noise in the outcome only
Random missing values in features
In uplift modeling, the target concept is best described as ______.
Average outcome under historical policy
Individual treatment effect (difference in outcomes with vs. without treatment)
Model residuals after regularization
Probability of class membership regardless of treatment
When using observational data, which assumption is typically required for unbiased causal estimates after conditioning on covariates?
Linear relationships between all variables
Perfectly balanced class labels
No unmeasured confounding (conditional ignorability)
Homoskedastic errors in the outcome model
Why can purely predictive feature importance be misleading for causal decisions?
Models cannot rank features without permutation tests
Correlated features always have zero importance
Causal features must be binary indicators
A predictor may correlate with the outcome without changing it under intervention
Instrumental variable methods require an instrument that affects treatment but ______.
Has no direct path to the outcome other than through treatment
Eliminates the need for any covariates
Is measured with substantial error
Perfectly predicts the outcome on its own
For policy targeting, a model is most useful if it ranks customers by ______.
Similarity to past customers
Expected incremental benefit from the action
Baseline risk regardless of action
Likelihood of responding under any circumstance
Starter
You’re getting familiar with interventions and counterfactuals—keep practicing on simple diagrams.
Solid
Strong grasp of identification and targeting—try applying uplift ideas to your own experiments.
Expert!
Master-level: you can separate prediction from causation and choose designs that stand up to scrutiny.
