How do null-test evidence ratios differ from standard Bayes factors in detecting absorbed systematics?

# Null-Test Evidence Ratios vs. Standard Bayes Factors in Detecting Absorbed Systematics

The key distinction is diagnostic scope: standard Bayes factors compare models globally, while null-test evidence ratios are specifically designed to catch when systematics from one component contaminate estimates of another [AG-2025.02-241].

The Problem Standard Bayes Factors Miss

When you fit a composite model—say, a foreground plus a signal in 21-cm astronomy—a poor model for the foreground can "absorb" into the signal component. The result: your overall fit looks good (standard Bayes factors might favor the composite model), but your recovered signal is biased and your foreground estimates are wrong [AG-2025.02-241]. This is the category II scenario: components interact in ways that trick conventional model comparison.

How BaNTER (Bayes Null Test Evidence Ratio) Works

The BaNTER framework [AG-2025.02-241] adds a validation layer by testing whether each component of a winning model can stand alone and predict data independently. Rather than just comparing composite models in aggregate, you compute evidence ratios that interrogate *whether the components are truly separable*.

Think of it this way: a standard Bayes factor asks "does this model fit the data?" A null-test evidence ratio asks "does this model fit the data *and* validate each piece independently?"

Why It Matters

In fields like 21-cm cosmology or gravitational-wave detection, systematics are often subtle. A standard Bayes factor might declare victory while hidden biases corrupt science-critical parameters. BaNTER catches this by refusing to declare a model winner until its sub-components prove robust [AG-2025.02-241]. This is especially important when minor instrumental or modeling errors can masquerade as real signals.

The broader literature on Bayes factors in new-physics searches [AG-2024.01-1092] and stochastic signal detection [AG-2024.12-298] reinforces that evidence ratios are powerful, but BaNTER's innovation is adding component-level validation to prevent silent absorption of systematics into favored models.