Is this just a rounded continuous Weibull value?

The type I form can arise from flooring an appropriately parameterized continuous Weibull variable; other rounding conventions or discrete forms need different formulas.

Can it model counts as well as lifetimes?

Yes, but the interpretation and estimation setup must match the application.

Does beta = 1 give a Poisson distribution?

No. It gives a geometric distribution under the stated zero-based convention.

AI Glossary / Statistics and probability / Advanced

Discrete Weibull distribution

Definition

The discrete Weibull distribution is a family of distributions for nonnegative integer outcomes, often used for discrete lifetimes or count data. In the common type I form, its probability mass at integer k is q^(k^beta) - q^((k+1)^beta), with 0 < q < 1 and beta > 0. It is related to the continuous Weibull through a discretization of survival probabilities. The type and parameter convention should be stated because several discrete Weibull formulations exist.

What it means

The distribution offers a flexible way to describe counts or event times recorded in whole intervals. Unlike a Poisson model, it does not impose equality of mean and variance, and appropriate parameter combinations can represent different dispersion patterns. In a lifetime interpretation, beta controls whether the conditional event probability rises, falls or stays constant across intervals. These mathematical features make it a candidate model, not evidence that a particular process follows it. Parameter values and fit must be estimated and checked for the application.

How it works

For the type I convention used here, P(X >= k) = q^(k^beta) and P(X <= k) = 1 - q^((k+1)^beta) for nonnegative integers k. Adjacent survival differences give the probability masses, which sum to one by telescoping. The discrete hazard, P(X = k | X >= k), equals 1 - q^[(k+1)^beta - k^beta]. It increases with k when beta exceeds one, decreases when beta is below one, and is constant when beta equals one. At beta = 1, the distribution becomes a geometric distribution on zero, one, two and so on. Fit the model with a method appropriate to complete, censored or covariate-dependent observations.

Three practical examples

Reliability research: model lifetimes recorded as completed operating intervals when a discrete event-time representation is appropriate.
Count regression: consider a discrete Weibull response model when dispersion relative to Poisson varies with conditions.
Simulation: generate integer outcomes under specified parameter settings to study a planning rule's behavior before fitting real observations.

Think of it this way

Think of a ladder of survival probabilities, one step for each integer interval. The probability of an event at a particular step is the drop between neighboring levels. The parameters determine the shape of the ladder, but the ladder still needs to match the process being studied.

Worked example

Take a hypothetical type I distribution with q = 0.5 and beta = 2. The probability of X = 0 is 1 - 0.5 = 0.5. For X = 1 it is 0.5 - 0.5^4 = 0.4375. For X = 2 it is 0.5^4 - 0.5^9 = 0.060546875. The remaining probability of X >= 3 is 0.5^9 = 0.001953125. These four quantities sum exactly to one. Conditional on reaching interval one, the event probability there is 0.4375/0.5 = 0.875; conditional on reaching interval two, it is 0.060546875/0.0625 = 0.96875. Thus the conditional hazard rises even though the unconditional probability mass at two is smaller than at one. The calculation illustrates the distinction between mass and hazard, not a fitted equipment-failure model.

Illustrative case study

Fictional case: Spruce Reliability records device lifetimes in inspection intervals and considers a discrete Weibull model. Analysts define whether zero denotes failure during the first interval or an actual zero-duration event, and record devices still operating at the end of observation as censored. They compare predicted and observed survival patterns and investigate differences across device types. This invented case illustrates careful model specification and validation. It does not establish that this distribution is superior to competing lifetime models or that interval-recorded observations reveal exact continuous failure times.

Common misconceptions and mistakes

Do not apply a continuous Weibull density directly as if it were a discrete probability mass. Avoid mixing support beginning at zero with a shifted version beginning at one. Parameter q is not the mean, and beta is not a generic variance parameter. An increasing hazard does not mean the sequence of probability masses must increase. Treating censored observations as exact failure times distorts estimation. Flexibility in dispersion does not guarantee a good fit, and excess zeros, dependence, mixtures or changing conditions may require a different model or additional structure.

Business use

Businesses may encounter discrete Weibull models in reliability analysis and specialized count forecasting. The important decision is whether the support, interval definition and modeled conditional behavior match the data. Require parameterization details, uncertainty estimates, diagnostic checks and comparisons with plausible alternatives. For operational planning, translate probabilities into relevant failure or demand events instead of relying on parameter names alone. Forecasts far beyond observed intervals can be sensitive to the assumed tail, so evaluate those implications explicitly before using the model for replacement or capacity decisions.

Check your understanding

Why is the hazard at interval two large while P(X = 2) is only about 0.0605?

Answer: The hazard conditions on the small group that has reached interval two. Its denominator is P(X >= 2) = 0.0625. Most of that remaining probability mass ends at two, although most of the original population ended earlier.

Related terms

Weibull distribution
Discrete distribution
Survival function
Censoring
Probability mass function (pmf)
geometric distribution
discrete hazard
count regression
overdispersion
underdispersion
probability mass function

Learn next

derive masses by subtracting successive survival values and verify normalization. Compare beta values below, equal to and above one while keeping the support convention fixed. Then study likelihood contributions for censored data and assess whether a count-regression or lifetime interpretation fits the actual question.

Sources

NIST Dataplot, Discrete Weibull probability mass function: https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/diwpdf.htm A Simple and Adaptive Dispersion Regression Model for Count Data: https://arxiv.org/abs/1511.00634 Discrete Weibull Generalized Additive Model: An Application to Count Fertility Data: https://academic.oup.com/jrsssc/article/68/3/565/7058404

Previous: Discrete wavelet transform Next: Discretionary access control