Description
Plot censored time-to-event estimates on axes that turn a candidate distribution into a straight line, to choose a distribution, read rough parameters and spot mixed populations.
Before fitting a parametric time-to-event model, check which shape suits the data. Probability plotting takes the nonparametric estimate of the fraction failed (Kaplan-Meier or Turnbull) and transforms the axes so that a chosen distribution plots as a straight line: for the Weibull, log time against log(-log survival); for the lognormal, log time against the normal quantile. Points are placed at the midpoint of each jump so that a correct fitted line runs through them, and simultaneous confidence bands show whether a straight line could fit. Hazard plotting is the companion: accumulate 1 over the number at risk at each event to get Nelson's cumulative hazard and plot it on the same linearising scales; the slope of a Weibull hazard plot estimates the shape. A sharp bend often reveals two failure modes or subpopulations, which is frequently the most useful finding.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Evaluation, Post-launch
- Data
- Quantitative
- Effort
- Hours
Steps
- Compute the nonparametric estimate suited to the censoring: Kaplan-Meier for right-censored data, Turnbull for interval-censored data, or Nelson's cumulative hazard from reverse ranks.
- Choose candidate distributions (exponential, Weibull, lognormal, loglogistic) and transform axes so each would plot as a straight line.
- Plot the estimates at midpoint-of-jump positions and add simultaneous confidence bands rather than pointwise ones.
- Judge whether a straight line can pass between the bands for each candidate; expect waviness in small samples (at n = 10 even correct models often look poor), and simulate samples of the same size if unsure.
- Investigate bends by splitting the data by plausible sources (platform, cohort, cause of exit) and plotting each part.
- Read rough parameters from the line (a Weibull slope near 1 means a constant hazard, below 1 a falling hazard, above 1 a rising one), then fit by maximum likelihood and overlay the fit for presentation.
Sources
- Statistical Methods for Reliability Data (Meeker & Escobar, 1998)
- Statistical Methods for Survival Data Analysis (Lee & Wang, 2003)
Also known as
- Weibull plot
- Lognormal probability plot
- Probability paper
- Cumulative hazard plot
- Nelson hazard plot
- Weibull hazard plot
- Probability plotting of censored life data with simultaneous confidence bands
Categories
- Analytics & Behaviour
- Statistics & Modelling