SUBMISSION: Capstone 1 — Quasi-Experimentation with Difference-in-Differences (DiD) on a Product Launch

# SUBMISSION: Capstone 1 — Quasi-Experimentation with Difference-in-Differences (DiD) on a Product Launch > 🎯 **Capstone Portfolio Deliverable.** Randomization was impossible — the CEO launched city-wise. This capstone proves you can still do credible causal inference. ## 🏢 Business Scenario & Executive Prompt **RideWynk** (ride-hailing, 12 cities) launched *"Priority Pickup"* in 4 cities (Mumbai, Delhi, Bengaluru, Hyderabad) in July; the other 8 cities launch in October. Leadership claims *"+18% bookings in launch cities — the feature works!"* But July is monsoon season, and launch cities are the largest metros. The board asks: *"What did the feature itself cause — and should we roll it to all cities?"* No randomization is possible retroactively. ## 📋 Source Data & Input Schemas **`ridewynk_city_panel.csv`** — city × week panel, Jan–Sep 2026, 12 cities × 39 weeks = 468 rows: | column | dtype | description | |---|---|---| | `city` | str | city code | | `week` | date | week start (Monday) | | `treated` | 0/1 | 1 for the 4 launch cities (all weeks) | | `post` | 0/1 | 1 for weeks ≥ 2026-07-01 | | `bookings` | int | weekly completed bookings | | `rainfall_mm` | float | weekly rainfall (confounder: monsoon) | | `active_drivers` | int | supply proxy | Stakes: all-city rollout costs ~₹2.1Cr in driver incentives. A wrong call burns the quarter's growth budget. ## 🎯 Required Deliverables 1. **`did_analysis.ipynb`** — trends → DiD estimate → placebos → recommendation. 2. **`did_estimates.csv`** — `specification, estimate, se, ci_lo, ci_hi` (naive diff, DiD, DiD + rainfall + drivers, event-study avg post). 3. **`did_memo.md`** — causal estimate with CI, parallel-trends verdict, placebo results, rollout recommendation. 4. **`event_study.png`** — leads-and-lags coefficients with CIs (pre-period should straddle 0). ## 🛠️ Step-by-Step Implementation Guide 1. **Show the naive claim failing.** Plot launch vs non-launch booking trends Jan–Sep. The +18% post gap is visible — and so is the monsoon divergence. Compute the naive post-only difference. 2. **DiD estimator.** `Y = α + γ·treated + λ·post + δ·(treated×post) + ε`; δ is the causal estimate under parallel trends. Then add `rainfall_mm + active_drivers` and watch δ move — sign of confounding absorbed. 3. **Defend parallel trends.** Event-study spec: interact `treated` with each pre-period week dummy; plot leads — they must be ≈0 (joint F-test reported). This is the load-bearing assumption; one paragraph on *why it could still fail* (Ashenfelter-dip-style selection: biggest cities chosen first). 4. **Placebos.** Fake treatment date (April); fake treatment group (2 control cities). Both δ̂ should be ≈0. If not, say so and downgrade the recommendation. 5. **Recommend.** Roll out / hold / run a randomized holdout in 2 cities — with the rupee math. ## 💻 Production Starter Scaffold ```python import numpy as np, pandas as pd import statsmodels.formula.api as smf import matplotlib.pyplot as plt rng = np.random.default_rng(31) cities = ["MUM","DEL","BLR","HYD","CHN","KOL","PUN","AMD","JAI","LKO","KOCHI","IND"] treated_cities = {"MUM","DEL","BLR","HYD"} weeks = pd.date_range("2026-01-05", "2026-09-28", freq="W-MON") rows=[] for ci, c in enumerate(cities): base = 8000 + ci*1500 for w in weeks: monsoon = 1 + 0.35*np.exp(-((w.month-7)**2)/1.5) * (c in treated_cities) post = int(w >= "2026-07-01"); tr = int(c in treated_cities) y = base*(1 + 0.02*(w.weekofyear)) * monsoon + tr*post*900 + rng.normal(0, 400) rows.append([c, w, tr, post, int(y), round(float(40*monsoon + rng.normal(0,8)),1), int(base/4 + rng.normal(0,150))]) df = pd.DataFrame(rows, columns=["city","week","treated","post","bookings","rainfall_mm","active_drivers"]) m1 = smf.ols("bookings ~ treated*post", data=df).fit(cov_type="cluster", cov_kwds={"groups": df.city}) m2 = smf.ols("bookings ~ treated*post + rainfall_mm + active_drivers", data=df).fit( cov_type="cluster", cov_kwds={"groups": df.city}) print(m1.summary().tables[1]); print(m2.summary().tables[1]) # Event study: replace post with week dummies x treated, plot leads/lags (see notebook). ``` ## 🧪 Acceptance Criteria & Verification Checks - [ ] Naive post-only gap (+18%-ish) computed and explicitly rejected with the monsoon confounder shown. - [ ] DiD δ̂ reported for 3+ specs with cluster-robust (city) SEs; rainfall-adjusted δ̂ differs from naive. - [ ] `event_study.png` shows pre-period leads straddling 0 + joint F-test p-value printed. - [ ] Two placebos (fake date, fake group) with ≈0 estimates documented. - [ ] Memo gives a costed rollout recommendation conditional on the placebo verdicts. --- ## 📊 Evaluation Rubric (100 points — passing ≥ 80) | # | Criterion | Points | What earns full marks | |---|---|---|---| | 1 | Statistical correctness | 25 | Right distributions/tests/CIs; assumptions checked, not assumed; math is reproducible | | 2 | Code quality & reproducibili