Is the hazard constant?

Not in the positive-growth form; it increases exponentially.

Can the hazard exceed one?

Yes. It is a rate with time units, not a probability.

What happens as the growth parameter approaches zero?

With the initial rate fixed, the distribution approaches the exponential model.

AI Glossary / Statistics and probability / Intermediate

Gompertz distribution

Definition

The Gompertz distribution is a continuous distribution for nonnegative event times whose hazard changes exponentially with time. In the increasing-hazard form used here, h(t) = b*exp(a*t), with b > 0 and a > 0. The hazard is the instantaneous event rate among cases that have not yet experienced the event. This distribution is distinct from a deterministic Gompertz growth curve, even though both carry the same name.

What it means

A constant-hazard model assumes that the instantaneous event rate stays the same as time passes. A Gompertz model allows that rate to increase by a constant multiplicative factor over equal time intervals. For example, a hypothetical component's failure rate might rise as it ages.

This describes a pattern in conditional rates, not a fixed schedule for each component. Individual event times remain random, and the model does not say that every item fails at its average or median lifetime.

Its relevance to AI is often as a statistical component of event-time prediction, reliability analysis, or simulation. Choosing the distribution is a substantive assumption that needs comparison with data and alternatives.

How it works

With time t at least zero, rate b and growth parameter a both positive, the cumulative hazard is H(t) = (b/a)[exp(a*t) - 1]. Survival is S(t) = exp[-H(t)], and the event probability by t is F(t) = 1 - S(t). The density is f(t) = h(t)*S(t).

The median solves S(t) = 0.5, giving t_median = log[1 + (a/b)*log(2)]/a. As a approaches zero with b fixed, the model approaches an exponential distribution with constant hazard b.

Parameter names differ across software. SciPy's positive-shape convention corresponds to c = b/a and scale = 1/a when location is zero. Some survival packages also allow negative a, which creates a nonzero probability of no finite event; that extension requires explicit interpretation and is outside the worked example.

Fitting must account for censoring and the observation process where relevant, rather than pretending unobserved event times equal study-end times.

Three practical examples

Equipment planning: evaluate whether increasing event rates provide a plausible description of component failures over the observed age range before estimating replacement needs.
Simulation: generate hypothetical event times under an explicitly stated increasing-hazard scenario to test how a planning system responds to aging assets.
AI-assisted reliability: compare a parametric event-time baseline with more flexible predictions, checking calibration and performance on later observations rather than assuming complexity improves accuracy.

Think of it this way

Think of a clock whose alarm becomes progressively more likely to ring in the next instant as time passes. The alarm has no predetermined ringing time. A Gompertz model describes how that conditional rate increases, while the survival curve describes the chance that the clock has remained silent up to a given time.

Worked example

Use an original hypothetical component model with time in months, b = 0.02 per month and a = 0.10 per month. At month ten, the hazard is 0.02*exp(1) = 0.05437 per month. The cumulative hazard is (0.02/0.10)[exp(1) - 1] = 0.34366.

Therefore S(10) = exp(-0.34366) = 0.70917. The probability of an event by month ten is approximately 29.08%, not 5.44%. The latter number comes from an instantaneous rate and cannot be substituted for a cumulative event probability.

The median event time is log[1 + 5*log(2)]/0.10 = 14.9643 months. Substituting that result back into the survival formula gives exactly 0.5 within numerical precision, providing an independent inverse check.

For an item that survives to month ten, the probability of an event before month eleven is 1 - S(11)/S(10) = 0.05557, about 5.56%. This differs slightly from multiplying the month-ten hazard by one month because the hazard changes during that interval. Direct exponential calculations verify all values. These selected parameters illustrate the formulas and are not fitted estimates or operational recommendations.

Illustrative case study

In a fictional packaging plant, engineers compare several event-time models for a replaceable component. A Gompertz model seems plausible because failures become more frequent among surviving older units. They include units still operating at the end of observation as censored records and account for differences in operating conditions.

The team checks predictions on a later group of units and investigates whether a maintenance-policy change alters the observed pattern. It uses the distribution as one input to planning, alongside downtime costs and inspection evidence. The fictional case does not establish that all components age exponentially or that a fitted curve determines an optimal replacement policy.

Common misconceptions and mistakes

Common mistakes include confusing the hazard with a probability, transferring parameter values between packages without checking their definitions, and equating a Gompertz lifetime distribution with a growth curve.

A monotonically increasing hazard cannot represent every failure pattern, such as a bathtub-shaped rate, without additional structure. A good fit over a short interval does not establish the shape of a distant tail.

Negative-shape extensions need special care: they are not simply an ordinary finite-lifetime distribution with the same positive-shape interpretation. Similarly, adding a location parameter changes the time origin and support; it should have a substantive rationale rather than being an invisible fitting convenience.

Business use

Use the distribution when its hazard pattern is defensible and the event definition, time origin, censoring, and population are clear. Report parameterization, time units, fit uncertainty, and supported prediction horizons. Business decisions should reflect uncertainty and the costs of different outcomes, not just a single predicted lifetime.

Check your understanding

If b = 0.02 per month and a = 0.10 per month, by what factor does the hazard change over five months?

Answer: It is multiplied by exp(0.5), approximately 1.649. This is a hazard multiplier, not a five-month event probability; survival probabilities require integrating the rate over time.

Related terms

Hazard rate
Survival function
Censoring
Reliability
cumulative hazard
exponential distribution
lifetime distribution
Gompertz growth curve

Learn next

Learn the relationship between hazards and survival probabilities, then study censored-data likelihoods and model comparison. Practice translating parameterizations before comparing software results.

Sources

Christopher Jackson, flexsurv Gompertz documentation: https://chjackson.github.io/flexsurv/reference/Gompertz.html — primary formulas and unrestricted-shape interpretation. SciPy, scipy.stats.gompertz: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.gompertz.html — standardized positive-shape density and scaling convention. SciPy, Gompertz distribution tutorial: https://scipy.github.io/devdocs/tutorial/stats/continuous_gompertz.html — distribution and quantile functions. The worked example uses original hypothetical parameters and independently checked calculations.

Previous: Gomory-Hu Tree Next: Goodman and Kruskal's gamma