Causal Inference Theory and Its Business Applications in Ctrip Train Ticket Operations

Background Ctrip, as a travel platform, needs to understand the causal impact of various strategies on conversion and revenue, controlling for complex, unobservable confounding factors. Five typical causal questions arise: product feature evaluation, virtual product value assessment, precise marketing, evaluation without AB tests, and external environment impact.

The three main solution approaches are: (1) designing proper AB experiments, (2) observational causal inference, and (3) combining machine‑learning algorithms with data/experiments to construct counterfactual reasoning. All three share the core idea of causal inference.

2. Basic Ideas and Theoretical Framework

2.1 Basic Idea – Causality differs from correlation; causality implies that A leads to B, not merely that they co‑occur. The goal of causal inference is to separate true causal effects from mere associations.

2.2 Frameworks – Two major frameworks are presented:

Rubin's Potential Outcome model focuses on finding suitable control groups (e.g., matching, synthetic control) to estimate the unobservable counterfactual.

Pearl's Causal Graph model uses directed graphs to represent variable relationships and compute conditional distributions for bias elimination.

Both frameworks aim to estimate the effect of an intervention when confounding variables exist, with Rubin emphasizing average treatment effect and Pearl emphasizing distributional changes.

3. Practical Cases

3.1 User Operations – UPLIFT Model: Identifies strategy‑sensitive users to reduce SMS costs and improve ROI. Modeling methods include S‑Learner, T‑Learner, X‑Learner; evaluation uses QINI curves. Results show a 10% top‑scoring segment yields a 0.011 incremental increase.

3.2 Virtual Value Assessment – Propensity Score Matching (PSM): Matches users in enterprise WeChat environments with similar users outside to evaluate incremental value. Multiple matching specifications (A, B, C) are discussed, highlighting sample‑selection bias and the trade‑off between accuracy and sample size.

3.3 Experiment Design – Synthetic Control Method (SCM): Constructs a weighted combination of multiple cities to form a synthetic control for a target city, enabling policy impact evaluation without a perfect control group.

3.4 Policy Intervention – Regression Discontinuity Design (RDD): Evaluates the effect of changing a WeChat public‑account reminder from strong to weak on 3‑day and 7‑day conversion rates, showing significant declines after the change.

4. Summary of Causal Inference Usage

4.1 When to Use – When perfect random experiments are infeasible, observational data can be used to estimate causal effects, isolating the impact of the factor of interest.

4.2 Scenario Identification – Four typical scenarios are outlined: (1) non‑experimental strategy effect evaluation, (2) heterogeneous treatment effect analysis in experiments, (3) sensitive‑user identification, and (4) causal impact analysis of metrics, each with recommended methods such as PSM, SCM, causal forests, uplift models, and double machine learning.

These guidelines help practitioners select appropriate causal inference techniques and apply them to real‑world business problems.