Tolerance Interval (Nonnormal Distribution)

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

Purpose

A tolerance interval gives the range that, with a stated confidence, contains at least a stated percentage of all the units in the population, for example: with 95% confidence, at least 95% of bulbs last longer than 1,093 hours. Tolerance Interval (Nonnormal Distribution) does this for data that are not normally distributed, such as lifetimes, strengths or impurity levels. It fits a chosen distribution and also gives a nonparametric interval that needs no distribution at all, so you can compare the two. Use it to set or check specification limits, warranty periods or release limits from a sample.

Key Capabilities

  • Eight distributions: Weibull, Lognormal, Gamma, Exponential, Smallest extreme value, Largest extreme value, Logistic and Loglogistic, or Assume none for the nonparametric interval only.
  • Nonparametric (distribution-free) interval from the ordered data, with its achieved confidence, shown in every case.
  • Two-sided interval, lower bound or upper bound.
  • Anderson-Darling goodness-of-fit test and the fitted parameters, to check the distribution chosen.
  • Several columns at once, each with its own interval and histogram with the tolerance limits marked.

Statistical Methods

Nonparametric method: the limits are order statistics of the sample. For a two-sided interval the tool uses (x₍ᵣ₎, x₍ₙ₋ᵣ₊₁₎), with r the largest value for which the achieved confidence, P(Beta(n − 2r + 1, 2r) > P), is still at least the confidence level, where P is the minimum proportion of the population. For a one-sided bound it uses x₍ₖ₎ (lower) or x₍ₙ₋ₖ₊₁₎ (upper), with k chosen the same way from the binomial distribution. Because only whole order statistics can be used, the achieved confidence is usually above the level asked for; with a small sample it can be below it, even with the minimum and maximum, and the table then shows the lower value.

Distribution methods, with parameters from maximum likelihood: Lognormal uses the normal tolerance interval of the log data, transformed back; Gamma uses the normal interval of the cube root of the data (Wilson-Hilferty); Exponential uses exact chi-square limits for one-sided bounds and the Fernandez method for two-sided; Weibull (on the log scale), Smallest and Largest extreme value use Lawless's exact method; Logistic and Loglogistic use the Bain and Engelhardt large-sample method. A two-sided interval is built from one-sided limits at α/2 and (1 + P)/2. Lognormal, Gamma, Exponential, Weibull and Loglogistic need positive data.

A distribution interval is only as good as the fit: check that the Anderson-Darling p-value is not small (above 0.05, say) before relying on it. When no distribution fits, use the nonparametric interval, which needs a larger sample to reach the same confidence. The achieved confidence applies only to the nonparametric method.

Input Fields & Options

Field

Description

Sample data

Required. One or more numeric columns; each gets its own interval.

Distribution

Weibull (default), Lognormal, Gamma, Exponential, Smallest extreme value, Largest extreme value, Logistic, Loglogistic, or Assume none, which shows only the nonparametric interval.

Confidence level

Default 95; 50 to 99.99.

Minimum percentage of population in interval

Default 95; 50 to 99.99. The proportion of the population the interval must contain.

Tolerance Interval

Two-sided (default), Lower bound or Upper bound.

Tolerance Interval Plot

Ticked by default: a histogram of each column with the tolerance limits.

Data Options: Subset (optional)

Exclude matching rows (default) or Include only matching rows, by row numbers (e.g. 3,7,12-15) and/or a condition.

tolerance-nonnormal-distribution-inputs


Figure 1 — Options for the example: Hours, Lognormal, 95% confidence, at least 95% of the population, lower bound.

Output Descriptions

  • Settings — distribution, confidence level and percent of population in interval.
  • Statistics — Variable, N, Mean and StDev for each column.
  • Tolerance table — for each column the interval or bound from the chosen distribution's method, the Nonparametric Method and its Achieved Confidence (with Assume none: Lower Bound, Upper Bound and Achieved Confidence).
  • Tolerance Interval for <column> — histogram with the limits drawn as dashed lines (the distribution method's limits when a distribution is chosen).
  • Distribution Parameters — the fitted parameters, e.g. Location and Scale, or Shape and Scale.
  • Goodness-of-Fit Test (Anderson-Darling) — AD statistic and p-value for the distribution chosen.
  • PDF export via Download.

tolerance-nonnormal-distribution-output


Figure 2 — Lognormal lower bound of 1,093.40 hours and nonparametric bound of 1,070.70 hours (achieved confidence 96.29%); the lognormal fits (AD 0.484, p = 0.224).

Worked Example

A quality engineer tests 100 light bulbs to failure and wants to state a burn time that, with 95% confidence, at least 95% of all bulbs will exceed. Lifetimes are often skewed, so the engineer fits a lognormal distribution. Sample data = Hours, Distribution = Lognormal, Confidence level = 95, Minimum percentage of population in interval = 95, Tolerance Interval = Lower bound.

The 100 bulbs have a mean of 1,248 hours and a standard deviation of 84.1 hours. The lognormal fits well (location 7.1271, scale 0.0671; AD 0.484, p = 0.224). The lognormal lower bound is 1,093.40 hours: with 95% confidence, at least 95% of bulbs last longer than that. The nonparametric bound, the second-shortest life in the sample, is 1,070.70 hours with an achieved confidence of 96.29%. It is lower because it uses no model, but it supports the same order of claim.

The choice of distribution matters. A Weibull does not fit these bulbs (AD 1.558, p < 0.010) and would give a lower bound of only 1,017.99 hours; a Gamma fits (AD 0.442, p > 0.250) and gives 1,091.16 hours, close to the lognormal. The engineer can state about 1,090 hours with confidence, and should not use the Weibull figure. For a two-sided interval with the lognormal, 95% of bulbs fall between 1,070.95 and 1,447.79 hours (nonparametric: 1,044.33 to 1,510.35, the shortest and longest lives).