Description
Models how often events happen (errors, support calls, crashes) and how the risk changes over time.
Counts of events per unit of time or exposure, such as errors per thousand sessions or support contacts per week, follow Poisson distributions, which give proper intervals and tests for rates. When the time until an event matters (time to first error, time to abandonment), the exponential distribution assumes a constant rate while the Weibull allows the rate to rise or fall over time, revealing whether problems cluster early (onboarding) or emerge later (wear-out). Bruce and Bruce introduce these as tools for estimating failure rates.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Define the event and the exposure (sessions, users, days).
- Estimate the rate and a Poisson confidence interval.
- Compare rates across versions with a Poisson or negative binomial regression.
- For time-to-event data, fit exponential and Weibull models and check the shape parameter.
- Use the shape to tell whether risk is concentrated early or late.
Sources
- Practical Statistics for Data Scientists (Bruce & Bruce)
Also known as
- Poisson rate estimation
- Weibull failure analysis
Categories
- Analytics & Behaviour
- Statistics & Modelling