Is a poly-Weibull model always bathtub-shaped?

No. Its shape depends on the component parameters; the worked example has an increasing hazard.

Is it a mixture model?

Not in the usual weighted-mixture sense. Its defining construction adds Weibull hazards.

Can identical shape parameters cause difficulty?

Yes. Their terms merge algebraically, so the individual scale contributions cannot be separated from the overall survival curve alone.

AI Glossary / Statistics and probability / Advanced

Poly-Weibull distribution

Definition

The poly-Weibull distribution is a continuous time-to-event distribution constructed by adding several Weibull-type hazard functions. With positive parameters lambda_j and nu_j, its survival function is S(t) = exp[-sum_j lambda_j t^nu_j] for nonnegative time. It can describe the earliest event among independent Weibull latent event times, or serve as a flexible model for an overall hazard. It is different from a mixture of Weibull distributions: adding hazards and averaging survival distributions are distinct constructions.

What it means

A single Weibull hazard is decreasing, constant or increasing, depending on its shape parameter. Several additive components can represent more complicated overall patterns. For example, an early decreasing component and a later increasing component can produce a bathtub-shaped hazard under suitable parameters. The components need not automatically correspond to identified real-world causes. If causes are unobserved, different parameter combinations may explain the overall data similarly. A useful fitted curve therefore does not, by itself, establish a physical decomposition of the event process.

How it works

Choose the number of components and state the parameterization. For component j, the cumulative hazard is lambda_j t^nu_j and the hazard is nu_j lambda_j t^(nu_j-1). Add the hazards, then obtain survival by exponentiating the negative total cumulative hazard. The density is the total hazard times survival. Fit parameters using a likelihood or Bayesian model that correctly handles censoring and other features of observation. Check whether components are distinguishable and whether predictions depend strongly on the number of components or prior assumptions. Extrapolation beyond observed follow-up deserves particular scrutiny because different fitted hazards can diverge substantially outside the data.

Three practical examples

Reliability research: model time to the first of several independent failure mechanisms with Weibull latent lifetimes.
Survival modeling: represent a changing overall event hazard when one Weibull component is too restrictive.
Simulation: generate event times under a specified additive-hazard structure to test how an analysis behaves when the underlying hazard is more complex than a single monotonic curve.

Think of it this way

Think of several clocks running toward different possible events, with the first event ending observation of the overall waiting time. Independent Weibull clocks give one interpretation of the model. The observed stopping time alone may not reveal which clock rang, so fitting several mathematical components does not necessarily identify the actual causes.

Worked example

Take an original two-component model with cumulative hazards H_1(t) = 0.1t and H_2(t) = 0.02t^2, using a specified common time unit. The first component has Weibull shape 1 and the second shape 2. Total survival is S(t) = exp[-0.1t - 0.02t^2], and total hazard is h(t) = 0.1 + 0.04t. At t = 5, the cumulative hazard is 1, survival is exp(-1), approximately 0.3679, and hazard is 0.3 per time unit. The density there is 0.3 exp(-1), approximately 0.1104 per time unit. The probability of an event between times 5 and 6 conditional on survival to 5 is 1 - S(6)/S(5) = 1 - exp(-0.32), approximately 0.2739. This is not equal to the instantaneous hazard 0.3. The example illustrates the distribution's arithmetic without claiming that its two mathematical components identify observable causes.

Illustrative case study

In a fictional case, Vale Reliability fits an overall failure-time model to equipment records. A single Weibull curve misses a change in hazard shape, so analysts investigate a two-component poly-Weibull model. They include right-censored equipment still operating at the end of follow-up and compare predictions with simpler alternatives. Because failure causes were inconsistently recorded, they describe the components as modeling devices rather than assigning each one to a specific mechanism. Their report highlights uncertainty in long-term extrapolation.

Common misconceptions and mistakes

A sum of hazards is not a weighted mixture of survival curves. A hazard is an instantaneous rate conditional on survival, not an interval probability. Independent latent event times provide a construction, but their independence is generally not something that overall first-event observations alone establish. If component shapes are equal, their cumulative hazards combine into a single Weibull term with summed lambda parameters, making the separate contributions unidentifiable from overall times. Even distinct shapes may be difficult to separate in limited data. More components do not automatically improve useful prediction.

Business use

Businesses evaluating reliability forecasts should ask which features are supported by observed follow-up and which rely on extrapolation assumptions. Retain censored records appropriately rather than treating operating equipment as if its lifetime were fully observed. Compare results across plausible component counts and parameter choices, and assess forecasts on held-out observations where possible. If a decision depends on preventing one particular failure cause, an overall fitted hazard may be insufficient: cause-specific information and a suitable causal or competing-risk analysis are needed. A flexible curve is useful only when its uncertainty and interpretation are communicated.

Check your understanding

Two components have the same shape nu and coefficients lambda_1 and lambda_2. What is their combined cumulative hazard?

Answer: (lambda_1 + lambda_2)t^nu, equivalent to one Weibull cumulative-hazard term.

Related terms

Weibull distribution
Survival function
Censoring
Mixture distribution
hazard function
competing risks
identifiability

Learn next

Learn how hazards, densities and survival probabilities differ, then study censoring likelihoods and why flexible survival models require careful assessment of identifiability and extrapolation.

Censoring

Sources

Demiris, Lunn and Sharples, Survival extrapolation using the poly-Weibull model: https://pmc.ncbi.nlm.nih.gov/articles/PMC4456429/ Timme, Modeling Multimodal Failure Effects of Complex Systems Using Polyweibull Distribution: https://scholar.afit.edu/etd/1906/

Previous: Policy value Next: Pólya urn model