Introduction
Estimating parameters for time-to-event data is a critical task in fields ranging from medical survival analysis to hydrological modeling. The log-logistic distribution is frequently employed for its ability to model heavy-tailed data. However, real-world observations are often left-truncated, meaning data points below a certain threshold are unobserved. Traditional Maximum Likelihood Estimation (MLE) often encounters numerical instability or poor convergence in these scenarios. This paper presents a robust Bayesian framework for the left-truncated log-logistic (LTLL) distribution, demonstrating superior stability and uncertainty quantification compared to traditional methods, especially when data is sparse or heavily truncated.
Methodology
The authors define the LTLL probability density function by adjusting the standard log-logistic PDF for truncation at $x_L$. To ensure parameter positivity and incorporate prior knowledge, independent Gamma distributions are assigned as priors for the scale ($\alpha$) and shape ($\beta$) parameters.
Given the complexity of the resulting posterior distribution, the Metropolis-Hastings algorithm—a Markov Chain Monte Carlo (MCMC) method—is utilized to sample the parameter space. This approach allows for the construction of full posterior distributions, providing a more comprehensive view of parameter uncertainty than point estimates. This is particularly useful for visualizing joint credible regions, which represent the relationship between $\alpha$ and $\beta$.
Experimental Evaluation
The framework was validated using synthetic datasets (sample sizes $N=50$ to $1000$) and three real-world datasets: bladder cancer remission times and historical precipitation records from Berlin, Toronto, and Dallas-Fort Worth.
Moment Sensitivity
A key finding of the study is the extreme sensitivity of the LTLL distribution's moments—mean, variance, skewness, and kurtosis—to the shape parameter $\beta$. Accurate estimation of $\beta$ is therefore paramount for reliable predictions in weather forecasting or clinical outcomes, as small changes in the parameter can lead to significantly different expected values and variances.
Key Results and Discussion
The Bayesian approach consistently outperformed MLE in several key areas:
- Superior Stability: In small-sample simulations ($N=50$), Bayesian estimators achieved lower variance and Root Mean Squared Error (RMSE) than MLE.
- Tighter Uncertainty: Bayesian 95% credible intervals were consistently narrower than MLE confidence intervals, offering more precise parameter localization.
- Real-World Robustness: In the Dallas-Fort Worth precipitation analysis, the Bayesian method remained stable under high truncation levels where MLE estimates began to fluctuate significantly.
The Gamma priors effectively regularize the estimation process, providing a stabilizing influence when the likelihood surface is flat or irregular due to heavy truncation.
Conclusion
This research establishes Bayesian inference as a superior alternative to MLE for left-truncated log-logistic distributions, particularly in data-sparse environments. While MCMC sampling introduces higher computational costs, the gains in stability and precision are significant for critical applications in medicine and hydrology. Future work may explore hierarchical models to account for group-level variations and the extension of this framework to cases where the truncation point itself is unknown.
Summary updated