Long-term causal effects of recommendation exposure

Causal Inference
Longitudinal Data
Marginal Structural Models
G-Computation
A sequential causal project estimating whether short-term recommendation exposure affects longer-term engagement and retention-style outcomes.
Published

April 28, 2026

Decision Question

How do sequential recommendation exposures affect future engagement or retention?

Causal Setup

  • Treatment: short-term recommendation exposure across time.
  • Outcome: future engagement and retention-style behavior.
  • Challenge: time-varying confounding affected by prior exposure.

Methods

  • Long-term outcome definition
  • Time-dependent propensity weights
  • Marginal structural models
  • G-computation
  • Doubly robust and heterogeneous effect analysis

Portfolio Takeaway

The project is designed to show that durable value often requires sequential estimands and diagnostics, not just same-session click lift.

Selected Figures

01 Estimator Comparison

02 Weight Balance

04 Secondary Outcomes

Notebook Sequence

The links below open rendered HTML versions of the notebooks. The source .ipynb files are kept in the matching folder under notebooks/projects/.

  1. Notebook 01: KuaiRec Sequence EDA for Long-Term Causal Effects
  2. Notebook 02: Defining the Long-Term Causal Estimand
  3. Notebook 03: Time-Varying Confounding and Propensity Weights
  4. Notebook 04: Marginal Structural Model for Long-Term Effects
  5. Notebook 05: G-Computation for Long-Term Effects
  6. Notebook 06: Doubly Robust Heterogeneous Effects
  7. Notebook 07: Sensitivity, Final Report, and Portfolio Artifacts

Generated Artifacts

Limitations

These are notebook-driven causal analyses, not production guarantees. Each project should be read with its identification assumptions, support diagnostics, measurement choices, and sensitivity checks in view.