Causal effect of ranking position on clicks

Causal Inference
Recommendations
AIPW
Causal ML
A notebook-driven causal inference project estimating whether top-3 recommendation exposure increases click probability in MIND impression logs.

Decision Question

Does placing a news item in the top 3 recommendation positions cause more clicks, or are top-ranked items simply more relevant and therefore more likely to be clicked anyway?

Causal Setup

  • Treatment is whether the item appears in the top 3 recommendation positions.
  • Outcome is click on the displayed item.
  • Adjustment uses user history, item metadata, slate size, time context, and item exposure features.

Methods

  • Propensity modeling and IPW
  • Doubly robust / AIPW estimation
  • Heterogeneous effects
  • Policy simulation
  • ML nuisance models and EconML extensions

Decision Takeaway

The project demonstrates the full applied causal workflow, including confounding checks, adjustment, doubly robust estimation, heterogeneity, policy simulation, and sensitivity analysis for a ranking decision.

Selected Figures

02 Estimator Comparison

03 Category Heterogeneous Effects

04 Policy Simulation

Notebook Sequence

  1. 01 - MIND Rank Position EDA
  2. 02 - Propensity Modeling And IPW
  3. 03 - Doubly Robust Estimation
  4. 04 - Heterogeneous Treatment Effects
  5. 05 - Policy Simulation
  6. 06 - Sensitivity And Limitations
  7. 07 - ML Nuisance Models With LightGBM And XGBoost
  8. 08 - EconML Causal ML Estimators

Generated Artifacts

Limitations

These notebook-driven causal analyses should be read with their identification assumptions, support diagnostics, measurement choices, and sensitivity checks in view.