liveBook · Manning

Post-treatment period

As explained in Causal Inference for Data Science

The post-treatment period is the time interval after an event or treatment during which the observed outcome is compared with its counterfactual.

A post-treatment period is the time interval after event time T in which a model evaluates the counterfactual outcome: what would have happened without the event. It usually contains multiple timepoints and follows a pretreatment period used for training. section 11.3.2

Choose its endpoints to balance the amount of evidence against changes in the causal context:

  • Starting earlier provides more observations for evaluating impact, but later observations may reflect delayed effects or a changed context. section 11.3.6
  • Starting later can reduce contamination from anticipation effects, but leaves fewer data points for measuring the event’s impact. section 11.3.6
  • Use domain knowledge when setting endpoints, and consider predictive performance and the stability of relationships between outcomes and controls. section 11.3.6

Movie-release impact

Suppose a new movie is released at time T. Older movies may lose viewership after the release, but viewers may also anticipate it and watch older movies several days beforehand. To estimate the no-release counterfactual, exclude those influenced days from training and evaluate the observed-versus-counterfactual difference across the post-treatment period. section 11.3.6

How Causal Inference for Data Science covers it

How deep it goes
Defines the interval and explains endpoint tradeoffs involving anticipation, delayed effects, sample size, prediction accuracy, and context change.
Examples
Uses a movie release, observed-versus-counterfactual plots, and CausalImpact analyses over whole periods or single event times.
What it assumes
Assumes familiarity with predictive models, control time series, counterfactual outcomes, and the distinction between pretreatment and post-treatment data.
What you can do after
Operate CausalImpact with selected pretreatment and post-treatment ranges to estimate impacts across a period or at one event time.
How it connects
Connects synthetic controls with pretreatment periods, counterfactuals, predictive models, anticipation effects, delayed effects, and causal effects.
Cover of Causal Inference for Data Science
liveBookmanning.com
Read Causal Inference for Data Science on liveBook Get the book at manning.com

This explainer was generated by Manning's knowledge graph from the text of Causal Inference for Data Science. The section links point to the parts of the book it draws on.