Dissertation — Causal Inference and Decision-Making Under Uncertainty

Thesis

Reliable decision-making in uncertain environments requires a causal pipeline, not a predictive pipeline: decisions should be anchored to explicit interventions, identified under transparent assumptions, estimated with robustness diagnostics, and optimized against utility/regret rather than raw statistical significance.

1) Methods: from causal question to deployable policy

The study progressed through six units that mapped the full causal stack:
- Foundations: potential outcomes, counterfactual logic, SCMs, and DAG semantics to prevent category errors between prediction and intervention.
- Identification: backdoor/frontdoor logic, quasi-experimental designs, and assumption stress-testing to establish when causal claims are defensible.
- Estimation: weighting/matching/doubly robust approaches with overlap and sensitivity diagnostics to reduce model dependence.
- Sequential decisions: adaptive experimentation and policy learning under time-varying uncertainty.
- Decision theory: utility, regret, and value-of-information framing to connect evidence to action quality.
- Capstone synthesis: operational decision protocol with monitoring and rollback criteria.

Methodologically, the dominant principle is design before estimation. A weaker estimator with strong identification generally outperforms a sophisticated estimator attached to a vague intervention.

2) Evidence and findings

A. Conceptual finding

The most important practical move is tightening the mapping:
action set → estimand → identification assumptions → estimation diagnostics → utility decision rule.
Any break in this chain reduces trustworthiness and deployability.

B. Operational finding

Causal analysis is most decision-useful when outputs include:
1. Effect interval (not just point estimate),
2. Assumption fragility ranking,
3. Segment heterogeneity (where policy works/fails),
4. Explicit stop/rollback conditions.

C. Governance finding

Uncertainty communication is a core deliverable. Decision stakeholders can handle uncertainty when framed as expected utility and regret tradeoff; they fail when uncertainty is hidden behind model complexity.

3) Limitations

• Heavy emphasis on applied synthesis over formal proofs.
• Limited domain-specific simulation in longitudinal high-interference settings.
• External validity remains contingent on stable treatment implementation and measurement reliability.

4) Decision framework for real systems

A practical framework emerging from this curriculum:
1. Frame intervention clearly (what is changed, for whom, when).
2. Select estimand tied to policy objective (ATE/CATE/policy value).
3. Defend identification with falsifiable assumptions.
4. Estimate with robustness bundle (overlap, sensitivity, placebo, stability).
5. Optimize decision with utility/regret + VOI.
6. Deploy with monitoring contract (drift checks, rollback triggers, re-identification cadence).

This framework turns causal inference from an academic post-hoc exercise into an ongoing operational discipline.

5) Roadmap

Near-term

• Build reusable causal decision brief template for internal product/policy choices.
• Add automated assumption-audit checklist to experiment review.

Mid-term

• Integrate dynamic treatment regime methods for multi-step intervention policies.
• Standardize policy value dashboards with uncertainty-aware alerts.

Long-term

• Establish organization-level causal governance: every material decision includes explicit intervention framing, identification argument, and rollback policy.

Conclusion

The central lesson is straightforward: uncertainty is unavoidable, but avoidable errors come from causal ambiguity. Teams that treat causality as the backbone of decision architecture make fewer high-cost mistakes, adapt faster under drift, and communicate risk with greater integrity. In practice, better causal discipline is better strategic judgment.