This question evaluates a data scientist's proficiency in causal inference for experiments with noncompliance and interference, covering estimands such as intent-to-treat and local average treatment effects, instrumental-variable and cluster-randomization strategies, exposure mapping for spillovers, and conducting sensitivity analyses.
You ran an A/B test that assigned some drivers to receive surge recommendations. About 30% of assigned (treated) drivers ignored the suggestion (noncompliance). Surge in one zone can also affect neighboring zones (interference/spillovers). The outcome is completed trips per driver (e.g., per shift or time block).
To make treatment well-defined for both treated and control drivers, assume you can generate a "ghost" recommendation for control drivers by running the recommendation algorithm offline to record the counterfactual target zone they would have been encouraged to move to if treated.
Propose an analysis plan to estimate the causal effect of following a surge recommendation on completed trips that:
Login required