📊 statistics

A modified score function for monotone likelihood in promotion time cure rate models

This paper proposes a modified score function based on Firth's bias-reduction method to resolve the monotone likelihood problem in promotion time cure rate models, thereby ensuring finite and stable parameter estimates where standard maximum likelihood estimation fails.

Original authors: Stephany Lima de Oliveira, Frederico Machado Almeida

Published 2026-08-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Stephany Lima de Oliveira, Frederico Machado Almeida

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

In the world of medical research, scientists often study how long patients survive after a diagnosis, looking for patterns that explain why some people recover while others do not. A common challenge in these studies is that many patients are still alive when the study ends, or they drop out before the event being tracked happens. This creates "censored" data, where the full story is missing. To make sense of this, researchers use special mathematical models that can account for a group of people who are essentially "cured"—individuals who will never experience the event, such as a cancer returning, no matter how long they are watched. One of the most useful tools for this is called the promotion time model. It imagines that a disease returns only after a certain number of hidden, competing causes have activated, and if none of these causes ever activate, the patient remains healthy forever. However, when researchers try to calculate the numbers behind these models, they sometimes hit a wall. If the data is unbalanced—for instance, if a specific risk factor is present in almost every patient who gets sick but absent in almost everyone who stays healthy—the standard math breaks down. The calculations try to run toward infinity, producing results that are impossible to interpret and leaving doctors without clear answers about which factors truly matter.

A team of researchers from Brazil set out to fix this mathematical breakdown, specifically for the promotion time model, which had previously been overlooked in this context. They developed a new way of calculating the model's numbers that acts like a stabilizer. Instead of letting the math run wild when the data is unbalanced, their method gently pulls the estimates back to a finite, sensible place. They tested this new approach by creating thousands of fake patient datasets on a computer, mimicking real-world scenarios where data is messy and unbalanced. In these simulations, the standard method often failed to produce a clear answer, or it produced answers that were wildly inaccurate and unreliable. In contrast, their new method consistently found stable numbers, even when the data was heavily skewed. While the new method did not make the estimates perfectly precise when the imbalance was extreme, it prevented the calculations from collapsing entirely, allowing researchers to see trends that would otherwise remain hidden.

To see if this approach worked in the real world, the team applied it to a collection of medical records from 215 patients treated for melanoma, a serious form of skin cancer, between 1995 and 2012. The goal was to understand which factors predicted whether the cancer would spread to other parts of the body. One specific factor, the presence of mitosis (a sign of rapid cell division), presented a perfect storm for the mathematical problem: every single patient in the study who had this sign eventually developed metastasis, while none of the patients without it did. When the researchers used the standard calculation method, the result for this factor was a massive, meaningless number with a huge margin of error, effectively telling them the factor was unimportant. This was a false conclusion caused by the mathematical instability. When they switched to their new, stabilized method, the result changed dramatically. The estimate became a reasonable, finite number, and the statistical evidence showed that mitosis was indeed a significant predictor of the cancer spreading. In one version of their model, the factor became statistically significant, confirming what doctors had long suspected but could not prove with the old tools.

The study also looked at how the balance of the data affected the results. They found that the more uneven the groups were, the harder it was to get precise numbers, even with the new method. However, the new approach was robust enough to handle the imbalance without failing completely. By applying this correction, the researchers were able to produce a clearer picture of the disease, identifying that patients with mitotic activity faced a higher risk of metastasis. This finding aligns with what is known in oncology, but the study proves that without this specific mathematical adjustment, the data itself could have led researchers to ignore a critical warning sign. The work demonstrates that by refining how we crunch the numbers, we can extract reliable truths from messy, imperfect data, ensuring that the stories told by patient records are accurate and actionable.

Technical Summary: A Modified Score Function for Monotone Likelihood in Promotion Time Cure Rate Models

Problem Statement
Survival analysis involving a cure fraction, specifically within the Promotion Time Cure Model (PTCM), faces a critical statistical challenge known as Monotone Likelihood (ML). ML occurs when the likelihood function increases monotonically along certain directions in the parameter space, preventing the existence of finite maximum likelihood estimates (MLEs). This phenomenon typically arises in scenarios with a high proportion of censored observations or, more critically, when binary covariates are severely unbalanced (e.g., all observed failures occur within a single category of a predictor). In such cases, standard MLEs diverge to $\pm\infty$ , rendering point estimates, confidence intervals, and hypothesis tests invalid. While ML has been extensively studied in standard survival models and generalized linear models, the literature contains a significant gap regarding its occurrence and resolution within the context of PTCMs. To date, no comprehensive methodological solution has been proposed for this specific class of models.

Methodology
To address this gap, the authors propose a modified score function based on Firth's bias-reduction method. The core of the approach involves penalizing the likelihood function to ensure finite parameter estimates even in the presence of separation or ML.

Model Formulation: The study utilizes the PTCM, which models the time to an event as arising from a latent number of competing causes ( $M$ ), typically following a Poisson distribution with mean $\theta$ . The survival function for the population is defined as $S_p(t|\psi) = \exp(-\theta F(t|\vartheta))$ , where $F$ is the cumulative distribution function for susceptible subjects. The model incorporates covariates via an accelerated failure time (AFT) framework, considering both Weibull and Log-Logistic distributions for the underlying survival times.
Modified Score Function: The standard score function $U(\psi) = \partial \ell(\psi) / \partial \psi$ $U (ψ) = \partial ℓ (ψ) / \partial ψ$ is modified by adding an adjustment term $A(\psi)$ $A (ψ)$ derived from the first-order asymptotic bias of the MLE. The modified score equation is $U^*(\psi) = U(\psi) + A(\psi) = 0$ $U^{*} (ψ) = U (ψ) + A (ψ) = 0$ .
- The adjustment term is defined as $A(\psi) = -I(\psi)b(\psi)$ , where $I(\psi)$ is the Fisher information matrix and $b(\psi)$ is the vector of first-order bias terms.
- Equivalently, this corresponds to maximizing a penalized log-likelihood $\ell^*(\psi) = \ell(\psi) + \frac{1}{2}\log|I(\psi)|$ , which is equivalent to penalizing the likelihood with Jeffreys' invariant prior in exponential families.
Estimation: The penalized MLE, $\hat{\psi}^*$ , is obtained by solving the modified likelihood equation. The authors note that while the procedure has a Bayesian interpretation (posterior mode under Jeffreys' prior), the estimation remains entirely frequentist.

Key Contributions

Novel Application: This is the first study, to the authors' knowledge, to explicitly address the ML problem within the context of Promotion Time Cure Models.
Methodological Development: The paper derives the specific modified score function and the associated penalized likelihood for PTCMs, adapting Firth's correction to this specific biological and statistical framework.
Comprehensive Evaluation: The study provides a rigorous assessment of the proposed method through extensive Monte Carlo simulations and a real-world application, filling a void in the literature where such solutions were previously nonexistent for PTCMs.

Results

Simulation Studies:
- The simulations evaluated the finite-sample performance of the modified estimators under scenarios with varying degrees of covariate imbalance (inducing ML) and sample sizes ( $n=50$ to $1000$).
- Standard MLE Failure: In scenarios with severe imbalance (e.g., a 6:994 success-to-failure ratio), standard MLEs exhibited substantial relative bias (up to -22% for the binary covariate coefficient), inflated empirical standard deviations, and coverage rates significantly below the nominal 95% level.
- Performance of Modified Estimators: The Firth-corrected estimators successfully produced finite values for all parameters, even in cases where standard MLEs diverged. While some bias and coverage issues persisted under extreme imbalance (particularly for small sample sizes), the performance metrics (bias, RMSE, coverage) improved markedly as the sample size increased and as the covariate balance improved.
- Balancing Experiment: A specific experiment varying the proportion of successes in the binary covariate demonstrated that even modest improvements in balance (e.g., moving from 0.6% to 2.5% successes) drastically reduced the incidence of ML and improved estimation precision.
Real Data Application (Melanoma Dataset):
- The method was applied to a dataset of 215 melanoma patients. A key feature of this data was "perfect separation" regarding the mitotic activity covariate: all patients with mitotic activity developed metastasis, while none without it did.
- Standard Approach: Using the standard score function, the coefficient for mitotic activity diverged (estimate $\approx -16$ with SE $\approx 506$ ), resulting in a $p$ -value near 1 and a failure to reject the null hypothesis. The intercept also showed extreme instability.
- Modified Approach: The Firth-corrected method yielded finite, stable estimates. In the PTM-Log-Logistic model, the mitotic activity factor became statistically significant ( $p=0.044$ ), with a reasonable coefficient estimate ($-2.285$) and standard error ($1.134$).
- Model Fit: Information criteria (AIC/BIC) favored the PTM-Log-Logistic model. Quantile residual plots indicated that the penalization improved numerical stability without materially altering the distributional behavior of the residuals or the overall model fit.

Significance and Claims
The paper claims that the proposed modified score function provides a robust and reliable framework for inference in Promotion Time Cure Models when data exhibit monotone likelihood. The primary significance lies in enabling the estimation of finite parameters and valid statistical inference in situations where standard maximum likelihood estimation fails completely.

The authors modestly conclude that while the method ensures finite estimates, the precision of these estimates (particularly for the covariate inducing the imbalance) remains dependent on the degree of covariate balance and sample size. However, the application to the melanoma dataset demonstrates a practical advantage: the methodology successfully identified the mitotic factor as a statistically significant prognostic marker, a conclusion that was obscured by the ML problem under standard estimation. This aligns with existing oncology literature, validating the method's utility in real-world medical research where cure fractions and separation issues are common.

Technical Summary: A Modified Score Function for Monotone Likelihood in Promotion Time Cure Rate Models

More like this