Description
Test whether events in one system's history are becoming more or less frequent, and if so fit a recurrence-rate model to predict future counts.
For a single system, service or account, the gaps between successive events are only a sample of independent times if the event rate is not changing. Plots come first: cumulative events against age, gaps against event number, and each gap against the previous one. Formal tests follow: the Military Handbook test (for a power-law trend) and the Laplace test (for a loglinear trend) are powerful against a constant-rate Poisson process but can mislead if gaps simply have a non-exponential distribution; the Lewis-Robinson test corrects Laplace by the gaps' coefficient of variation. With no trend and independent gaps, ordinary distribution methods apply to the gaps. With a trend, a nonhomogeneous Poisson process with a power or loglinear rate is fitted by maximum likelihood and used to predict events in a future window.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Plot cumulative event count against age, gaps against event sequence, and gaps against lagged gaps.
- Apply the Military Handbook, Laplace and Lewis-Robinson tests for trend; prefer Lewis-Robinson when gaps may be non-exponential.
- If there is no trend and gaps look independent, treat the gaps as a sample and fit a single distribution.
- If a trend exists, fit a power-law or loglinear nonhomogeneous Poisson process by maximum likelihood and obtain intervals for its parameters.
- Predict the expected number of events in a future interval, with bootstrap bounds.
- For many systems, use the mean cumulative function instead unless they plausibly share one rate.
Sources
- Statistical Methods for Reliability Data (Meeker & Escobar, 1998)
Also known as
- Laplace test
- Military Handbook test
- MIL-HDBK-189 test
- Lewis-Robinson test
- Power-law NHPP
- Crow-AMSAA model
- Renewal process check
- Trend tests and nonhomogeneous Poisson process models for recurrence data from a single system
Categories
- Analytics & Behaviour
- Statistics & Modelling