2. Field data characteristics

Reliability data collected from the field rarely arrive as clean, complete lifetimes. A repairable pump monitored over three years may still be running when the observation window closes; a batch of bearings may only be inspected after a failure threshold is crossed; units from different production lines may behave as if drawn from different populations; and some failure records may simply be lost. Before any Bayesian model is specified, the analyst must recognize which of these four mechanisms — censoring, truncation, heterogeneity, and missingness — is present, because each one changes the likelihood function and therefore the posterior.

Censoring: partial information about event times

Censoring occurs when we know an event happened (or did not happen) within an interval, but not the exact time. The three forms most relevant to field reliability are:

  • Right censoring — the unit was still operating at the end of observation. We know T > tc. This is the dominant pattern in warranty and fleet data.
  • Left censoring — failure occurred before the first inspection, so T < tc. Common when failures are only discovered during scheduled maintenance.
  • Interval censoring — failure occurred between two inspection times, tL < T ≤ tR. Typical for periodic condition monitoring.

The likelihood contribution for a right-censored observation at time tc is the survival function S(tc | θ), not the density f(tc | θ). For an interval-censored observation it is S(tL | θ) − S(tR | θ). Mixing these up is one of the most common and most damaging errors in reliability modeling: treating a censored unit as a failure biases the estimated mean life downward, sometimes severely.

Practical check: For every record, ask "do I know the exact failure time, or only a bound?" If the answer is a bound, the record is censored and must enter the likelihood through S(·) or an interval probability.

Truncation: the observation window itself

Truncation differs from censoring in a subtle but important way. With censoring, the unit is observed and we know it survived past a point. With truncation, units whose event times fall outside a window are never observed at all. Two cases dominate:

  • Left truncation (delayed entry) — a unit only enters the study after surviving to time t0. Its likelihood is conditioned on survival to entry: f(t | θ) / S(t0 | θ).
  • Right truncation — only failures occurring before a cutoff are recorded, so long-lived units are systematically absent.

Ignoring left truncation inflates apparent reliability because the early, high-risk period is invisible. In field data, left truncation appears whenever components are installed before the monitoring program begins, or when only units that reached a certain age are eligible for a study.

Heterogeneity: one model does not fit all

Field populations are rarely homogeneous. Units may come from different suppliers, operating environments, duty cycles, or vintages. If this variation is ignored, the fitted lifetime distribution often shows an artificially decreasing hazard — a classic symptom of unmodeled heterogeneity (sometimes called a "mixture" or "frailty" effect).

Bayesian analysis handles heterogeneity naturally through hierarchical structure. Instead of a single parameter θ, we let each group j have its own θj, drawn from a common population distribution:

θ_j ~ p(θ | φ)        # group-level parameters
φ ~ p(φ)               # hyperprior
t_ij ~ f(t | θ_j)      # observations within group j

The hyperprior p(φ) pools information across groups, which is exactly what we want when some groups have few failures. A group with two failures borrows strength from the population; a group with two hundred failures is barely affected. This partial pooling is the Bayesian answer to the bias–variance trade-off in reliability estimation.

Missingness: when records are incomplete

Missing data in reliability datasets usually takes one of three forms:

  • Missing covariate values — e.g., operating temperature not logged for some units.
  • Missing failure mode — the unit failed, but the cause was not recorded.
  • Missing records entirely — units that were installed but never entered into the database.

The critical question is why the data are missing. If missingness is independent of the failure process (MCAR — missing completely at random), complete-case analysis is unbiased but inefficient. If missingness depends on observed covariates (MAR — missing at random), multiple imputation or a Bayesian joint model is appropriate. If missingness depends on the unobserved failure time itself (MNAR — missing not at random), the missingness mechanism must be modeled explicitly, or estimates will be biased.

Bayesian advantage: In a Bayesian framework, missing values can be treated as unknown parameters and sampled jointly with θ. This yields valid inference under MAR without discarding records, and it propagates imputation uncertainty into the posterior automatically.

How the four mechanisms interact

In real field datasets, these mechanisms rarely appear alone. A warranty database might contain right-censored units, left-truncated entries (components installed before the warranty period), heterogeneous suppliers, and missing failure modes — all at once. The figure below shows how each mechanism maps onto the likelihood and what the analyst must decide.

Field Data Mechanisms → Likelihood Contributions Censoring T > t_c or t_L < T ≤ t_R Truncation T > t_0 (delayed entry) Heterogeneity θ_j ~ p(θ | φ) Missingness MCAR / MAR / MNAR Likelihood term S(t_c | θ) S(t_L|θ) − S(t_R|θ) Likelihood term f(t | θ) / S(t_0 | θ) (conditioned on entry) Hierarchical model θ_j ~ p(θ | φ) φ ~ p(φ) Joint sampling t_miss ~ p(t | θ, obs) sample with θ Posterior: p(θ, φ, t_miss | data) ∝ L(data | θ, φ, t_miss) · p(θ | φ) · p(φ) All four mechanisms enter through the likelihood and the prior hierarchy Analyst decision checklist: 1. Is each record an exact time, a bound, or an interval? → choose f(t) or S(t) or interval probability. 2. Did every unit have the same chance of entering the dataset? → if not, condition on entry time. 3. Are there identifiable subgroups? → add a hierarchy; do not pool blindly.

What this means for the Bayesian workflow

The practical consequence is that the likelihood function is not a single formula but a composition. Each record contributes a term determined by its observation type, and the prior structure must reflect any grouping. A minimal but realistic specification for a heterogeneous, right-censored fleet dataset looks like this:

# Hierarchical Weibull model with right censoring (conceptual)
for j in groups:                     # e.g., supplier or site
    shape[j]  ~ LogNormal(0, 1)
    scale[j]  ~ LogNormal(mu_phi, sigma_phi)
    mu_phi    ~ Normal(0, 5)
    sigma_phi ~ HalfNormal(1)

for i in observations:
    if failed[i]:
        t[i] ~ Weibull(shape[group[i]], scale[group[i]])
    else:
        target += log(Surv_Weibull(t[i], shape[group[i]], scale[group[i]]))

Two things are worth noting. First, the censored branch uses the survival function, not the density. Second, the group-level parameters share hyperpriors, so information flows between groups. If left truncation were also present, each density term would be divided by S(t0,i | θgroup[i]). If covariates had missing values, those would be declared as parameters and sampled alongside the rest.

Common failure symptom: If a fitted Weibull shape parameter comes out below 1 (decreasing hazard) for a mechanical component that should wear out, suspect unmodeled heterogeneity or left truncation before concluding the physics is wrong. The data may be telling you about the sampling, not the failure mechanism.

Summary

Censoring, truncation, heterogeneity, and missingness are not data-cleaning nuisances to be patched before "real" analysis begins — they are structural features that define the correct likelihood and prior. Right and interval censoring change the observation term from a density to a survival or interval probability. Left truncation requires conditioning on entry. Heterogeneity calls for a hierarchical prior that partially pools information across groups. Missingness, when MAR, can be handled by treating unknown values as parameters and sampling them jointly. Recognizing which mechanisms are present, and encoding them explicitly, is the prerequisite for every Bayesian reliability model that follows.

← Previous 📖 Contents Next →