📊 statistics

Applying the Weibull Shape Parameter test for signal detection in pharmacovigilance using the R package WSPsignal

This paper introduces the R package WSPsignal, a unified open-source tool that facilitates the application of the Weibull shape parameter test for pharmacovigilance signal detection by offering flexible frequentist and Bayesian frameworks, simulation-based tuning for optimal test specifications, and practical demonstrations on both large and small datasets.

Original authors: Julia Dyck, Odile Sauzet

Published 2026-06-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Julia Dyck, Odile Sauzet

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a safety inspector for a new brand of cars. You know the car works well in the factory (clinical trials), but once thousands of people start driving them on real roads, you need to watch out for weird noises or sudden breakdowns that only happen after a few months of use. In the world of medicine, these "weird noises" are called Adverse Drug Reactions (ADRs).

This paper introduces a new digital tool, a software package called WSPsignal, designed to help safety inspectors spot these problems faster and more accurately using a method called the Weibull Shape Parameter (WSP) test.

Here is a breakdown of how it works, using simple analogies:

1. The Core Problem: The "Constant" vs. The "Spiky" Risk

When you take a new medicine, there is a risk of side effects.

The "Constant" Scenario: Imagine the risk of a side effect is like a flat, calm lake. The chance of something happening is the same on Day 1 as it is on Day 365. If the risk is flat, the medicine is likely safe regarding that specific side effect.
The "Spiky" Scenario: Imagine the risk is like a mountain range. Maybe the risk is low at first, then suddenly spikes in the 5th month, and drops off later. This "spike" suggests the drug is actually causing the problem, but only after the body has been exposed to it for a while.

The WSP test is a mathematical detector that looks at patient data to see if the risk is a flat lake or a spiky mountain. If it finds a spike, it raises a "signal" (an alert) for doctors to investigate further.

2. The New Tool: WSPsignal

Before this paper, doing this math was like trying to bake a complex cake using five different cookbooks, scattered recipes, and custom-made mixing bowls. It was hard for regular researchers to do.

The authors built WSPsignal, which is like a smart, all-in-one kitchen appliance.

It puts all the necessary recipes (statistical tests) into one box.
It has a "Default" button for quick results.
It has a "Tuning" knob that lets you customize the settings based on how much data you have (a small sample vs. a huge dataset).

3. Two Ways to Cook: Frequentist vs. Bayesian

The paper explains that the tool can cook the cake in two different styles, depending on the size of your party (sample size):

The "Frequentist" Style (The Big Party):
- When to use: When you have a massive amount of data (like 20,000 patient records).
- How it works: It looks strictly at the numbers in front of it. It asks, "Is the pattern in this huge pile of data statistically different from a flat line?"
- The Paper's Example: They tested this on a large dataset of muscle pain cases. The tool found a "spike" in risk around the 5th month, raising a signal.
The "Bayesian" Style (The Small Party):
- When to use: When you have a smaller amount of data (like 1,000 records) or when you already know a little bit about the drug.
- How it works: It mixes the new data with "prior knowledge" (what we already suspect). It's like a detective who says, "I know this drug usually causes issues in the middle of the treatment, so I'll look harder for that specific pattern."
- The Paper's Example: They tested this on a smaller dataset. They had to "tune" the detective's prior knowledge (setting the expectation that the risk happens in the middle of the year). Once tuned, the tool again found a signal.

4. The "Tuning" Process

One of the paper's main points is that you shouldn't just use the "Default" settings blindly. The tool includes a Simulation Engine.

Think of this like a flight simulator for the software. Before you fly the real plane (analyze real patient data), you run thousands of fake flights (simulations) with different settings:

What if the risk happens early?
What if the risk happens late?
What if the data is noisy?

The software runs these simulations to figure out which "recipe" (combination of math models and settings) catches the most real problems while avoiding false alarms. The paper shows that by doing this tuning, the tool becomes much sharper at finding the truth.

Summary

The paper doesn't claim to have found a new drug or cured a disease. Instead, it presents a better, easier-to-use toolbox for statisticians and safety experts.

The Goal: To make it easier to spot when a medicine causes a side effect that only appears after a certain amount of time.
The Method: Using a "shape detector" (WSP) that looks for spikes in risk over time.
The Innovation: A single software package that handles both big data (using strict math) and small data (using educated guesses), and includes a simulator to help users find the perfect settings for their specific situation.

In short, WSPsignal is the new, user-friendly dashboard that helps medical safety teams drive their investigations more safely and efficiently.

Technical Summary: Applying the Weibull Shape Parameter Test for Signal Detection in Pharmacovigilance using the R package WSPsignal

Problem Statement
Post-marketing pharmacovigilance relies on identifying potential adverse drug reactions (ADRs) from routinely collected health data. While clinical trials often involve limited populations, real-world usage may reveal risks only apparent after market approval. Detecting these signals is crucial for maintaining safety profiles, potentially leading to label changes or market withdrawals. Existing signal detection methods include disproportionality approaches, temporal pattern analysis, and exposure models. However, time-to-event (TTE) based methods, specifically the Weibull Shape Parameter (WSP) tests, which leverage temporal information to assess whether the hazard of an adverse event changes over time since drug initiation, have been difficult to apply. Historically, implementing WSP variants required substantial custom coding and scattered resources, limiting their accessibility to applied researchers.

Methodology
The paper introduces the WSPsignal R package, an open-source interface that consolidates the family of WSP signal detection tests. The methodology is grounded in the assumption that a drug-related ADR results in a non-constant hazard function over time, whereas a lack of temporal dependence implies a constant hazard.

The framework supports two estimation approaches:

Frequentist: Utilizes Maximum Likelihood Estimation (MLE) to fit Weibull-type models.
Bayesian: Utilizes posterior sampling (via Stan) with specified prior distributions (lognormal, gamma, fixed-scale variants) for model parameters.

The package implements three specific TTE distributions:

Weibull (W): Standard single-shape parameter model.
Double Weibull (dW): Estimates two Weibull models (one for the full data, one censored at the midpoint of the observation period).
Power Generalized Weibull (pgW): A more flexible model with additional shape parameters.

Signal Detection Logic:

Frequentist: A signal is raised if the shape parameter(s) statistically differ from 1 (indicating non-constant hazard) at a specified confidence level ( $1-\alpha$ ).
Bayesian: Decisions are based on the comparison between the posterior credibility interval (CI) and a Region of Practical Equivalence (ROPE) around the null value of 1. The package offers three sensitivity options to combine interim results from multiple parameters into a binary signal/no-signal output.

A key feature of the package is its simulation-based tuning capability. Users can simulate data scenarios reflecting specific characteristics (sample size, background rate, ADR rate, observation duration) to rank test specifications based on the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC). This allows practitioners to identify the optimal distribution, confidence/credibility level, and prior specifications for their specific data context.

Key Contributions

Software Development: The primary contribution is the WSPsignal package, which unifies previously fragmented WSP methodologies into a single, reproducible, and open-source tool.
Workflow Standardization: The paper provides a step-by-step workflow for applying these tests, including data exploration, default testing, and simulation-based tuning.
Demonstration of Flexibility: The authors illustrate the package's utility through two synthetic datasets:
- Large-sample setting (~20,000 observations): A frequentist approach using a Double Weibull model with a 97% confidence level was shown to be effective, aligning with previous recommendations.
- Small-sample setting (~1,000 observations): A Bayesian approach was employed. The authors demonstrated the necessity of tuning priors and credibility levels, finding that a Power Generalized Weibull model with a lognormal prior, 60% credibility level, and sensitivity option 3 yielded the best performance for this specific scenario.
Performance Evaluation: The paper details how to evaluate test performance regarding convergence, execution time, effective sample sizes (for Bayesian methods), and statistical metrics (AUC, FPR, TPR).

Results
The application of the package to the synthetic datasets confirmed that:

In large samples, default specifications (e.g., dW with 97% confidence) perform well, but tuning can confirm optimality.
In small samples, the choice of prior belief and distribution is critical. The simulation revealed that the dW model with a lognormal prior suffered from convergence issues and low effective sample sizes in the small-sample scenario, leading to its exclusion from the final analysis in favor of the pgW model.
The performance of the tests is sensitive to the timing of the ADR occurrence; tests performed better when ADRs occurred at the beginning or end of the study period in the large-sample scenario, and towards the middle in the small-sample Bayesian scenario.
The tuning process successfully identified specific configurations (e.g., specific credibility levels and sensitivity options) that maximized the AUC for the given data characteristics.

Significance
The paper claims that by integrating all WSP test variants into a single, transparent, and reproducible package, the technical barrier for applied researchers in statistics, pharmacovigilance, and related fields is lowered. The authors state that this facilitates broader adoption, critical review, and further methodological development. The package enables practitioners to move beyond "one-size-fits-all" approaches by allowing them to perform simulation-based tuning to identify the optimal test for a given data scenario, thereby improving the reliability of signal detection in post-marketing surveillance.