Staggered Rollouts

By the end of this lesson, you should be able to: recognise a staggered rollout, explain how the standard two-way fixed effects regression compares groups, show why it goes wrong when effects change over time, and choose an estimator that doesn't.

Rollouts in waves

Many changes don't launch everywhere at once. A feature goes to a few countries first, then more. A policy starts in some cities, then spreads. This is a staggered rollout, and it looks ideal for difference-in-differences: at any point, some units have the change and others don't yet.

Facebook's early spread is a real example. It launched at Harvard in February 2004 and reached hundreds of US colleges over the next two years, each on its own date. Researchers used those staggered arrival dates to estimate Facebook's effect on students' mental health.

The regression everyone runs

The standard approach is a two-way fixed effects (TWFE) regression:

Y_{it} = \alpha_i + \gamma_t + \tau \cdot \text{Treated}_{it} + \varepsilon_{it}

$\alpha_i$ is a separate baseline for each unit (lesson 13), $\gamma_t$ a separate effect for each period, and $\text{Treated}_{it}$ is 1 once unit $i$ has the change. The coefficient $\tau$ is reported as "the effect".

It looks like a natural generalisation of DiD. When the effect is the same for every unit and every period after launch, it is. When the effect changes over time, it isn't.