Abstract
This paper presents a critical introduction to the complex statistical models used in $^{14}$C dating, with a focus on estimating transit times between archaeological layers. While Bayesian methods are the industry standard due to their ability to handle non-regular models and provide intuitive credible sets, they rely on technical assumptions—priors—that can significantly bias scientific conclusions. Through toy model simulations and an analysis of Iron Age data from Tel Rehov, this study demonstrates that standard Bayesian procedures can introduce artificial signals. Ultimately, the current radiocarbon data from Tel Rehov is found to be insufficient to resolve the historical debate regarding the United Monarchy of David and Solomon.
Introduction: The Challenge of Irregular Data
In archaeology, establishing precise chronologies for soil layers (strata) is fundamental. For decades, researchers have favored Bayesian analysis, implemented in software like OxCal or BCal, to process radiocarbon ($^{14}$C) dates. The primary appeal of the Bayesian paradigm is its capacity to incorporate known archaeological constraints—such as the sequential nature of layers—into the statistical model.
However, radiocarbon data is "irregular." Standard frequentist confidence intervals are difficult to construct because the models do not follow typical regularity conditions. While Bayesian methods handle these constraints elegantly, they require the specification of "priors"—mathematical assumptions about the parameters before data is observed. Often treated as neutral "black boxes," these priors can exert a hidden, powerful influence on final dates, potentially creating "evidence" where none exists in the raw data.
Methodological Framework
The study employs a multi-faceted approach to evaluate these statistical tools:
- Simulated Toy Models: Artificial datasets with known "true" dates were used to compare the performance of various Bayesian priors against Frequentist Maximum Likelihood Estimators (MLE).
- Robust Estimation: To mitigate the impact of outliers (e.g., intrusive samples from different layers), Huber’s M-estimators were utilized to "pull" extreme data points toward the center.
- Bootstrap and PCA: Non-parametric and parametric bootstrapping, combined with Principal Component Analysis (PCA), allowed for the visualization of the "cloud" of possible chronological scenarios, providing a more transparent view of uncertainty.
- Tel Rehov Dataset: The analysis focused on a real-world dataset of 32 samples (86 measurements) from Tel Rehov, a site central to the "High vs. Low Chronology" debate in Biblical archaeology.
The "Prior Trap" and Hidden Bias
A central finding is that seemingly "non-informative" uniform priors are often anything but neutral. In models of consecutive layers, the prior density includes a factor inversely proportional to the duration of the layer. Mathematically, this strongly favors very short durations for archaeological periods, regardless of the actual radiocarbon evidence.
Furthermore, in a Bayesian framework, "remote" data points—those far from a transition boundary—can still shift the estimated transition time. This counter-intuitive property means that samples which should be irrelevant to a specific boundary can still influence its dating.
Results: Simulations and the Tel Rehov Debate
Simulations of event sequences revealed that standard Bayesian procedures can be strictly biased when events are not evenly spaced. These models tend to "smooth" results, which can obscure the true historical signal.
When applied to the Tel Rehov data, the bootstrap analysis produced a wide "cloud" of possible dates encompassing both the "High Chronology" (supporting the feasibility of the United Monarchy in the 10th century BCE) and the "Low Chronology" (pushing developments into the 9th century). Specifically, the transition between key strata was estimated to lie anywhere between 948 and 907 BCE. PCA results further suggested that the start and end of certain strata can be analyzed almost independently, contradicting the tightly coupled results often produced by standard Bayesian software.
Conclusion
The study concludes that the radiocarbon data currently available from Tel Rehov is "underdetermined" and cannot resolve the debate over the United Monarchy. While Bayesian methods remain a powerful tool, their results must be interpreted with caution. There is a pressing need for more robust priors and a more transparent use of frequentist tools, like the bootstrap, to complement Bayesian findings and ensure that technical assumptions do not inadvertently dictate historical conclusions.
Summary updated