Model Selection for Weibull Series Systems: When Simpler Models Suffice

When can you safely use a simpler model for a series system? I ran extensive simulation studies with likelihood ratio tests to get a quantitative answer.

The Problem

In series system reliability, you estimate component parameters from masked failure data. For Weibull components, that means estimating (2m) parameters: shape (k_j) and scale (\lambda_j) for each of (m) components.

But what if the components have similar failure characteristics? A reduced model with homogeneous shape parameters uses only (m+1) parameters (one common (k) plus (m) scales). This roughly halves the parameter count and has a nice property: the system itself becomes Weibull-distributed.

The question is when this simplification is justified.

Key Findings

Robustness of the Reduced Model

For well-designed series systems (components with similar failure characteristics), the result is striking:

The reduced homogeneous-shape model cannot be rejected even with sample sizes approaching 30,000, far larger than anything typically available in practice.

With realistic sample sizes (50 to 500), the likelihood ratio test shows no evidence against the reduced model when components truly have similar shapes. This is strong justification for using the simpler model.

Sharp Boundaries

The paper pins down exactly how much heterogeneity it takes to trigger rejection:

Even modest deviations in a single component's shape parameter provide evidence against the reduced model. The boundaries are clean.

Practical Guidance

Use the reduced model when:

• Components come from similar manufacturing processes
• Historical data suggests similar wear-out patterns
• Sample sizes are moderate ((n < 500))
• You need a quick reliability assessment

Use the full model when:

• Components have fundamentally different failure modes (infant mortality vs wear-out)
• Large samples are available ((n > 1000))
• Precise component-level inference is critical
• Preliminary studies suggest model inadequacy

Connection to Related Work

This paper fits into a broader program on masked failure data:

The pieces address different aspects of the same problem:

1. Estimation (thesis): How to estimate parameters from masked data
2. Efficiency (expo-masked-fim): Closed-form variance bounds
3. Model selection (this paper): When simpler models suffice
4. Robustness (mdrelax): What happens when assumptions break

Technical Approach

Likelihood Ratio Test

The test compares:

• Full model (H_1): Each component has its own shape (k_j)
• Reduced model (H_0): All components share shape (k)

Test statistic:
$$\Lambda = -2\left[\ell(\hat{\theta}_0) - \ell(\hat{\theta}_1)\right]$$

Under (H_0), asymptotically (\Lambda \sim \chi^2_{m-1}).

Simulation Design

• System: 5-component series with Weibull lifetimes
• Masking: Candidate set width 2 (C1-C2-C3 conditions)
• Censoring: 30% right-censoring
• Sample sizes: 50, 100, 250, 500, 1000, 5000, 30000
• Shape deviations: Perturbing single component from 0.25 to 3.0

Well-Designed Systems

A well-designed series system has component MTTFs within a factor of 2 to 3, shape parameters within 20-30% of each other, and no single dominant weak point. This concept matters: the reduced model works for well-designed systems but fails when components have fundamentally different reliability profiles.

Why This Matters

The bias-variance tradeoff is real in reliability engineering. Simpler models have lower variance but potential bias. Complex models reduce bias but increase variance.

For the small samples typical in reliability testing, the reduced model's variance advantage often outweighs any bias from mild shape heterogeneity. This paper provides the quantitative boundaries for that decision.

Resources

• Paper: Model Selection for Reliability Estimation in Series Systems
• Master's Thesis: Reliability Estimation in Series Systems
• Related Paper: Closed-Form Fisher Information
• R Package: wei.series.md.c1.c2.c3