In reliability engineering, a series system fails when any of its \(m\) components fails. When the system is observed in the field, we often know when it failed but not which component caused the failure. Instead, we observe a candidate set — a subset of components that plausibly contains the true cause. This is masked failure data.
The statistical challenge is to estimate component-level lifetime parameters from system-level observations where the failure cause is partially or fully masked and the failure time may be censored (right, left, or interval).
This vignette describes a family of R packages that address this problem at different levels of generality, from foundational distribution algebra through closed-form maximum likelihood estimation to fully general numerical inference.
The packages form a layered dependency graph. Solid arrows indicate
Imports dependencies; the dashed arrow indicates a
Suggests dependency used for cross-validation testing.
algebraic.dist ─────────────────────────┐
│ │
▼ │
algebraic.mle │
│ │
├──────────────┐ │
▼ ▼ ▼
compositional.mle likelihood.model generics
│ │ │
┌────┘ └────┐ │
▼ ▼ │
flexhaz maskedcauses
│ │
▼ │
serieshaz │
│ │
└──────┐ ┌──────┘
▼ ▼
maskedhazAll packages are available on r-universe. Packages marked with [CRAN] below are published on CRAN; the remainder are targeting CRAN submission.
Four packages provide the algebraic and inferential foundation that the reliability-specific packages build on.
algebraic.dist [CRAN] — An algebra over probability distributions. Compose, sample, and auto-simplify distributions (normal, exponential, multivariate normal, empirical) using S3 generics. Provides the distribution abstraction used throughout the ecosystem.
algebraic.mle [CRAN] — An algebra over maximum
likelihood estimators. Unified interface for mle,
mle_numerical, mle_boot, and
rmap_mle objects with delta-method and bootstrap inference.
The MLE result class used by fitting functions across the ecosystem.
likelihood.model [CRAN] — A Fisherian likelihood
framework. Defines the core generics — loglik,
score, hess_loglik, fim,
observed_info, rdata, assumptions
— that all likelihood models in this ecosystem implement.
Contribution-based models support heterogeneous observation types
(exact, right-censored, left-censored, interval-censored) in a single
likelihood.
compositional.mle [CRAN] — Composable MLE solvers as
first-class functions. Sequential chaining (%>>%),
parallel racing (%|%), and random restarts turn solver
construction into an algebra. Useful for building robust optimization
strategies for the multi-modal likelihoods that arise in masked data
problems.
flexhaz — Define a lifetime distribution by
specifying its hazard function \(h(t;
\boldsymbol{\theta})\) and get the survival function, density,
CDF, quantiles, random sampling, and MLE automatically. The
dfr_dist class (“dynamic failure rate distribution”) is the
building block for component-level models.
Built-in constructors include dfr_exponential(),
dfr_weibull(), dfr_gompertz(), and
dfr_loglogistic(), but any user-defined hazard function
works. flexhaz implements the likelihood_model interface
from likelihood.model, so every dfr_dist object can be used
directly with loglik(), score(),
fit(), and friends.
serieshaz — Compose dfr_dist components
from flexhaz into a series system. For a series system of \(m\) components, the system hazard is the
sum of component hazards:
\[h_{\text{sys}}(t) = \sum_{j=1}^{m} h_j(t; \boldsymbol{\theta}_j)\]
and the system survival is the product of component survivals:
\[S_{\text{sys}}(t) = \prod_{j=1}^{m} S_j(t; \boldsymbol{\theta}_j).\]
The dfr_dist_series class inherits all flexhaz methods
(density, survival, quantiles, sampling, MLE) and adds a parameter
layout system that maps a flat parameter vector to per-component slices
— enabling standard optimizers to work on the full system
likelihood.
maskedcauses (this package) provides closed-form maximum likelihood estimation for masked series system data under the standard masking conditions (C1, C2, C3). Three likelihood models form a natural nesting chain:
| Model | Parameters | Constructor |
|---|---|---|
| Exponential series | \(m\) rates: \((\lambda_1, \ldots, \lambda_m)\) | exp_series_md_c1_c2_c3() |
| Homogeneous Weibull series | \(m + 1\): \((k, \beta_1, \ldots, \beta_m)\) | wei_series_homogeneous_md_c1_c2_c3() |
| Heterogeneous Weibull series | \(2m\): \((k_1, \beta_1, \ldots, k_m, \beta_m)\) | wei_series_md_c1_c2_c3() |
Each nesting level is testable via likelihood ratio test (see the
vignette("model_selection") for a complete treatment).
The key advantage of maskedcauses is analytical tractability: the exponential model has closed-form log-likelihood, score, and Hessian for all four observation types (exact, right, left, interval). The Weibull models have analytical expressions for exact and right-censored observations, with efficient closed-form or numerical expressions for left and interval censoring.
maskedcauses also provides the masked data
infrastructure used across the ecosystem: composable
observation functors (observe_right_censor,
observe_left_censor, observe_periodic,
observe_mixture), candidate set generation under C1/C2/C3
conditions, and the series_md S3 class hierarchy.
maskedhaz — When component lifetimes follow distributions beyond exponential and Weibull — Gompertz, log-logistic, or arbitrary user-defined hazard functions — maskedhaz provides MLE for the masked series system likelihood using numerical methods.
maskedhaz builds on serieshaz for system construction and
maskedcauses for masked data infrastructure. It uses numerical
integration (stats::integrate) for left and
interval-censored contributions and numerical differentiation
(numDeriv::jacobian) for the Hessian.
The trade-off is clear: maskedhaz handles any component lifetime distribution that flexhaz can represent, at the cost of speed and the availability of analytical derivatives. For exponential and Weibull components, maskedcauses is the faster and more precise choice.
Cross-validation tests in maskedhaz verify that its numerical results match the analytical results from maskedcauses for the exponential special case, providing confidence in both implementations.
| You need… | Use |
|---|---|
| Component lifetimes are exponential | maskedcauses::exp_series_md_c1_c2_c3() |
| Component lifetimes are Weibull (shared shape) | maskedcauses::wei_series_homogeneous_md_c1_c2_c3() |
| Component lifetimes are Weibull (different shapes) | maskedcauses::wei_series_md_c1_c2_c3() |
| Model selection across the Weibull chain | maskedcauses — see
vignette("model_selection") |
| Component lifetimes are Gompertz, log-logistic, or custom | maskedhaz with flexhaz components |
| Series system distribution (no masking) | serieshaz with flexhaz components |
| Single-component DFR distribution | flexhaz |
| MLE algebra (delta method, bootstrap) | algebraic.mle |
| Composing optimization strategies | compositional.mle |