Definition
Competing risks arise when one event prevents another event of interest from occurring under the chosen outcome definition. For a first-resolution analysis, an exchange and a refund can be competing outcomes if each unit receives only one initial resolution. Cumulative incidence describes the probability of a particular event by a given time while accounting for the competing events.
Why It Matters
- Counting only refunds can hide a shift toward exchanges, while combining every resolution can hide different operational workloads.
- Within a Commerce Intelligence OS, the event definition should match the decision: planning first-resolution demand differs from tracking every later customer action.
How It Works
- Choose the unit, starting event, horizon, and mutually exclusive first outcomes. Define tie-breaking for ambiguous timestamps and separate incomplete follow-up.
- Preserve time and event type for each unit. Do not treat a competing first event as an ordinary loss to follow-up when estimating real-world event probability.
- Use an appropriate cumulative-incidence estimator, such as Aalen-Johansen, and document assumptions about censoring. One minus Kaplan-Meier with competing events censored generally overestimates event probability.
- If customers can exchange and later obtain a refund, use a richer multi-state or repeated-event analysis for that lifecycle question. First-resolution probabilities do not describe every eventual resolution.
Ecommerce Example
Context: Illustrative example: an apparel merchant follows delivered units until their first refund, first exchange, or reporting cutoff.
Recommended move: Estimate separate first-refund and first-exchange cumulative incidence curves. Keep a later refund of an exchanged item in a linked lifecycle record.
Why it matters: Operations can distinguish the initial resolution mix without declaring exchanged customers permanently unable to refund. This is an illustrative analysis design, not an iKawn customer finding.
iKawn Framework
Define
The iKawn ontology framework names resolution states and permitted transitions.
Observe
Link each initial outcome to the purchased unit and its event time.
Estimate
Present outcome-specific probabilities with their observation assumptions.
Apply
Connect first-resolution demand with service planning while retaining later events.
Concise Summary
Competing-risk analysis estimates mutually exclusive event probabilities over time. Define the first outcome carefully and model later transitions separately when needed.