Distribution ID Plot (Arbitrary Censoring)

On this page
  1. Purpose
  2. Key Capabilities
  3. Statistical Methods
  4. Input Fields & Options
  5. Output Descriptions
  6. Worked Example

Purpose

Distribution ID Plot (Arbitrary Censoring) helps you choose a life distribution when exact failure times are not always known. It accepts any mix of exact failures, right-censored units (still working when last seen), left-censored units (already failed at the first check) and interval-censored units (failed sometime between two inspections). It fits six distributions and draws a probability plot for each, ranked by how straight the plot is. Use it for inspection or audit data, where a failure is only found at the next check.

Key Capabilities

  • Data entered as a Start column and an End column; the pattern of blank and filled cells sets the type of censoring on each row.
  • Exact, right-, left- and interval-censored rows in the same data set.
  • Fits Weibull, Lognormal, Exponential, Normal, Logistic and Loglogistic in one run.
  • Ranks the distributions by the correlation coefficient of their probability plots.
  • One probability plot panel per distribution, with the fitted line and its confidence bounds.

Statistical Methods

Each row is read from its Start and End values: Start = End is an exact failure at that time; Start < End is an interval-censored failure somewhere between the two; End blank is right-censored at Start (the unit was still working then); Start blank is left-censored at End (the unit had already failed by then). Every row needs at least one value, and Start must not be greater than End.

Each distribution is fitted by maximum likelihood. An exact failure contributes the density f(t), a right-censored unit the survival probability 1 − F(Start), a left-censored unit F(End), and an interval-censored unit F(End) − F(Start), the probability of failing within the interval. The nonparametric estimate of the cumulative probability of failure for the plot points is the Turnbull estimate, an iterative method that spreads each interval-censored failure over the times where it could have happened.

As in the right-censoring tool, each panel transforms the axes so that its distribution plots as a straight line (log time for Weibull, Lognormal and Loglogistic), and the correlation coefficient is the Pearson correlation between the plotted x values and the transformed probabilities. Values close to 1 mean a straighter plot. With grouped inspection data there are fewer distinct plot points than units, so the correlation rests on fewer points; look at the panels as well as the ranking, and use engineering knowledge about the failure mechanism. Each panel also shows two-sided confidence bounds (dotted) at the Confidence level, as Minitab draws them: for each percent p, the percentile t_p of the fitted distribution with its variance by the delta method from the inverse Hessian of the log-likelihood, the interval formed on ln(t_p) for Weibull, Lognormal, Exponential and Loglogistic and on t_p for Normal and Logistic. Points well outside the bounds point to a poor fit.

Input Fields & Options

Field

Description

Start column

Required. Last known good time, or the exact failure time. Leave Start blank on a row to mark it left-censored.

End column

Required. First known failed time, or the exact failure time. Leave End blank on a row to mark it right-censored. Start = End means an exact failure; Start < End means interval-censored. Must be a different column from Start.

Confidence level (%)

Default 95; 50 to 99.99. Sets the confidence bounds drawn around each fitted line.

distribution-id-plot-arbitrary-censoring-inputs


Figure 1 — Options for the example: inspection data in a Start column and an End column.

Output Descriptions

  • Correlation ranking — each distribution and its correlation coefficient, from best to worst.
  • Messages — any distribution that could not be fitted, with the reason.
  • Distribution ID Plot for Start-End — one probability plot panel per distribution in ranking order, titled with its correlation (r), showing the plot points, the fitted line and its confidence bounds.
  • PDF export via Download.

distribution-id-plot-arbitrary-censoring-output


Figure 2 — Correlation ranking and probability plots for 50 windings checked every 10 hours: Loglogistic 0.990, Lognormal 0.989 and Weibull 0.987 are practically tied.

Worked Example

This example uses the 50 turbine engine windings tested at 80 °C from the Distribution ID Plot (Right Censoring) page, recorded as they would be if the windings were only checked at inspections every 10 hours from 30 hours onwards. The 4 windings found failed at the first inspection are left-censored (Start blank, End = 30). The 33 found failed at a later inspection are interval-censored, for example Start = 50 and End = 60 for a winding that was working at 50 hours and failed at 60. The 13 removed before failure are right-censored at their removal times (End blank). The engineer selects Start as the Start column and End as the End column.

The ranking is Loglogistic 0.990, Lognormal 0.989, Weibull 0.987, Exponential 0.980, Normal 0.975 and Logistic 0.972. Each panel has only five points, at 30, 40, 50, 60 and 79 hours: the times where the Turnbull estimate places failures (the one failure found between 70 and 80 hours is placed at 79 hours, after the last right-censored winding inside that interval). With so few points the coefficients are all high and close together, and the top three cannot be separated by the data alone. The fitted Weibull has a scale of 73.3 hours and a shape of 2.17, close to the 73.3 hours and 2.32 found from the exact times, so grouping the failures into inspection intervals has changed the fitted model only a little.

For the engineer, Loglogistic, Lognormal and Weibull all describe the data reasonably; Weibull remains a sensible choice for wear-out failures because its shape shows the trend of the failure rate. The next step is a Distribution Overview Plot (Arbitrary Censoring) with the chosen distribution, to see the fitted survival and hazard curves.