Did Netflix's Price Increase Cause Churn?

1. The Research Question

Did the January 2025 US price increase cause churn, or was any observed lift driven by seasonality, content gaps, competitor activity, or macro conditions? Answering this cleanly requires separating the price change from everything else happening in the same calendar window.

The pipeline uses a stacked Difference-in-Differences event studyon billing-derived price-exposure timing across the Feb 1 – Apr 30 2025 rollout window. Treatment timing is assigned from each member's first observed new-price charge, with a fallback imputation for members who churned before receiving one.

2. Why Staggered DiD?

A naive before-vs-after comparison confounds the price increase with anything else happening at the same time. A US-vs-non-US design removes the before/after problem but introduces a worse one: other countries have different catalogs, economies, and subscriber mixes, making them a poor counterfactual.

Because the rollout hit different members' billing cycles at different times, members exposed in week 1 can be compared against not-yet-exposed members in later weeks — all subject to the same national shocks. Seasonality, competitor events, and macro conditions cancel out by construction.

3. Event Study

Weekly voluntary-cancel effect relative to the price-exposure week. Pre-period coefficients are flat and near zero (consistent with parallel trends); the sharpest negative estimate lands on the exposure week itself before attenuating and reverting toward zero.

4. Static ATT & Parallel-Trends Test

Collapsing the post-period into a single average treatment effect on the treated (ATT), and formally testing whether the pre-period leads are jointly zero:

Relative to the 0.105% baseline weekly churn rate, the point estimate is a −1.57× baseline shift, though the interval includes zero at the 5% level.

We fail to reject the null of parallel pre-trends (p = 0.104) — the pre-period leads are jointly indistinguishable from zero, which supports the core identifying assumption.

5. Interpretation

Members exposed to the new price showed lower voluntary-cancel rates on average relative to not-yet-treated controls. Two mechanisms are consistent with the data:

Because the confidence interval crosses zero, the data are consistent with a null causal effect. If a true rollout-assignment table (intent-to-treat) exists, it should replace billing-derived timing to remove this selection concern.

6. Analysis Pipeline

7. Cohort Composition

Treated members by billing cohort and plan. Cohorts after Mar 17 had fewer than 20 treated or control members and were excluded from estimation, leaving 8 of 13 stacks.