Slice by dimensions most likely to reveal the pattern: platform (iOS/Android/web), geography, user tenure, activity tier (power/casual/new), acquisition channel, and feature usage.

Start with the dimension that carries the most explanatory power for the specific metric. A metric drop that only appears in new users points to onboarding. One limited to mobile Android points to an app bug. One in a specific geography points to a local competitor or incident. Segmentation should be hypothesis-driven, not exhaustive - enumerate likely causes, then pick the most diagnostic slices first.