When Code Gets Cheap — Day-level two-node DiD

Spec — day-level TWFE DiD (paper-aligned)

log_apps_gdc = α_g + γ_d + δ_c + Σ_h β_h · holiday_dummy_dc + β_M22 · (treated_c × post_May22_d) + β_Nov24 · (treated_c × post_Nov24_d) + ε_gdc

Panel unit: (genre × day_in_cycle × cycle)
day_in_cycle ∈ [0, 365): days from Mar 1 of cycle year
post_May22 = 1 if day ≥ 82 (= May 22)
post_Nov24 = 1 if day ≥ 268 (= Nov 24, second launch)

γ_d = day-of-year fixed effect (365 levels) — absorbs all calendar-aligned daily seasonality
Floating holidays added as dummies (year-specific): Christmas, Christmas Eve, NYE, NY, Easter, Eid al-Fitr, Diwali, Cyber Monday, Thanksgiving, CNY (next year)
Cluster: genre. Window: day_in_cycle ∈ [28, 330]. Cycles: 2023 / 2024 / 2025.

Event-study (weekly bins, day-level FE) — 3 datasets

Coefficients on (treated × event_week_bin), aggregated to 7-day bins for plot tractability, but underlying regression has day-of-year FE. Reference bin = 11 (mid-May, week before May 22). 橙线 = May 22 cutoff (event_bin 12); 蓝线 = Nov 24 cutoff (event_bin 38). Edge bins 0-3 和 48-51 trimmed.

Two-node DiD coefficients (day-level)

Dataset	Contrast	β_M22	p	β_{Nov24_inc}	p	M22 %	Nov24 inc %	Total post-Nov24	Verdict

核心 read

三个 dataset 在 day-level 收敛到 +43%. iOS V1 +43.2%, iOS V2 +43.3%, Android +43.5% — total post-Nov-24 effect 几乎相同。这是 3 个独立 sampling 抓的数据集，cross-platform/cross-sampling-design 都看到一致的 entry surge magnitude。

iOS V1 跟 paper +48% 接近. 我的 v45 spec 略小于 paper (43% vs 48%)，因为： (1) Cycle window 不同：我用 Mar-Feb (含 May 22)，paper 用 Jun-Apr； (2) Holiday FE 控制更激进； (3) Outcome 不同：我 log(1+entries)，paper raw level。方向、量级一致。

Day-of-year FE 已加入. 365 levels of day-in-cycle FE 自动吸收 Christmas, NYE, Thanksgiving (fixed-date) 季节性。额外加 10 个 floating-holiday dummies (CNY, Easter, Eid, Diwali, Cyber Monday) 处理 lunar / 浮动日期。 Week 43 Christmas 现象（之前 weekly chart 显示）在 day-level + holiday FE 后被吸收，event-study 更平滑。

Placebo identification. Day-level 比 weekly 更 noisy，placebo 也显示一些 drift (iOS V1/V2 placebo M22 +7-8% 显著)。 Android M22 placebo null ✓, Nov24_inc placebo +10% drift。 Treated effect 远超 placebo，但要承认还有 unmodeled secular trend。