所属模块：
M1 · 数学与统计基础 (Mathematics & Statistics Fundamentals)| 专题分类：因果推断 (Causal Inference)| 难度等级：Medium
一、核心一句话结论 (One-Sentence Summary)
比较处理组与对照组的’前后变化之差’；关键假设是平行趋势。
Difference-in-Differences estimates causal effects by comparing the pre-post changes in a treatment group against changes in an untreated control group over the same time horizon: $Delta_{text{DID}} = (bar{Y}{T, 2} – bar{Y}}) – (bar{Y{C, 2} – bar{Y})$.
二、核心考点要义 (Key Insights)
- 📌 平行趋势不可直接检验，需前趋势图/事件研究法
- 📌 交错处理需用现代 DID 估计量（Callaway-Sant’Anna）
English Insights:
– Core Mechanism: First difference subtracts baseline time-invariant user heterogeneity; second difference subtracts macroeconomic time trends.
– Parallel Trends Assumption: In the absence of treatment, the average outcome of treatment and control groups would have followed parallel trajectories over time.
– Regression Formulation: $Y_{it} = beta_0 + beta_1 text{Treated}i + beta_2 text{Post}_t + delta (text{Treated}_i times text{Post}_t) + epsilon$, where $delta$ is the causal DID effect.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$hattau_{DID}=(bar Y_{T,post}-bar Y_{T,pre})-(bar Y_{C,post}-bar Y_{C,pre})$$
DID 的核心思想是用对照组的变化扣除时间趋势：处理组的前后变化 = 处理效应 + 时间趋势，对照组的前后变化 = 时间趋势，两者相减即得处理效应。这比简单前后比较（会混入时间趋势）和简单组间比较（会混入组间固有差异）都更强。关键识别假设是平行趋势（parallel trends）：若无处理，处理组与对照组的结果本会沿平行路径演变。注意这是反事实假设，不可直接检验（因为处理组的反事实路径不可观测），只能间接支持。DID 可写成双向固定效应回归：Y_it=α+β·Treat_i·Post_t+γ_i+δ_t+ε_it，其中 γ_i、δ_t 为个体与时间固定效应，β 即 DID 估计量。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Expectations table across groups and time periods: (1) Control Pre: $E[Y_{C, 1}] = beta_0$. (2) Control Post: $E[Y_{C, 2}] = beta_0 + beta_2$. Change in control: $Delta_C = beta_2$ (time trend). (3) Treatment Pre: $E[Y_{T, 1}] = beta_0 + beta_1$. (4) Treatment Post: $E[Y_{T, 2}] = beta_0 + beta_1 + beta_2 + delta$. Change in treatment: $Delta_T = beta_2 + delta$. Taking the difference of differences: $Delta_{text{DID}} = Delta_T – Delta_C = (beta_2 + delta) – beta_2 = delta$. The estimate isolates the true causal impact $delta$ while controlling for baseline group differences $beta_1$ and macro shocks $beta_2$.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
实践要点：① 平行趋势的诊断——用事件研究法（event study）画处理前多期的动态效应，若处理前的系数接近 0 且无趋势，则支持平行趋势假设；若处理前已存在趋势差异，DID 失效。其他方法包括：用不同的对照组做安慰剂检验、做合成控制（synthetic control）构造更可信的反事实。② 交错处理（staggered adoption）的陷阱——当不同个体在不同时间接受处理时，用传统双向固定效应（TWFE）回归会因’已处理组被当作对照组’而产生负权重问题（Goodman-Bacon 分解证明 β 是多个 2×2 DID 的加权和，部分权重为负），可能给出方向相反的结论。解法是使用现代估计量：Callaway & Sant’Anna (2021) 的组-时间 ATT、Sun & Abraham (2021) 的交互加权、或 Borusyak et al. 的插补法。③ 其他威胁——预期效应（处理前已反应）、组成变化（样本进出）、其他同期政策冲击，需用安慰剂检验与稳健性分析排除。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
DID is widely used for policy evaluation, natural experiments, and geo-level marketing campaigns (e.g. Launching TV ads in California while using Texas as control). The Parallel Trends assumption cannot be directly proven for the post-treatment period; scientists validate it empirically using Event Study plots (lead-lag regressions showing estimated coefficients $delta_t = 0$ for all pre-treatment periods $t < 0$).
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 用 TWFE 处理交错处理（负权重问题）
- ⚠️ 只看处理前一期就断言平行趋势（需多期事件研究）
English Pitfalls:
– Applying standard DID when parallel trends fail pre-treatment (pre-existing divergence violates the identifying assumption).
– Staggered adoption bias: When different units receive treatment at different times, standard two-way fixed effects regressions produce biased estimates (addressed by Callaway-Sant’Anna estimators).
六、高频深度面试追问与预测 (Follow-Up Questions)
- 如何检验平行趋势？
- How do event study regressions formally test the pre-treatment parallel trends assumption?
- 交错 DID 的负权重问题是什么？
- Why does standard two-way fixed effects (TWFE) fail in staggered rollout designs with heterogeneous treatment effects over time?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
因果推断框架：潜在结果模型、倾向评分匹配与双重差分(Causal Inference: Potential Outcomes, PSM & DiD) - 🗺️ 知识图谱模块：
数据科学与因果实验导图
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。