Does postselection bias equally distort positive and negative thrust measurements, or does it preferentially amplify one direction?

# Postselection Bias: Directional Asymmetry in Thrust Measurements

Postselection bias does not affect positive and negative thrust equally—it preferentially amplifies signals in one direction, though the mechanism depends on your specific measurement procedure.

The Core Asymmetry

The key insight comes from how selection thresholds interact with noise. When you select events based on a criterion like signal-to-noise ratio (SNR), you create an asymmetry: "sources that are nominally below threshold are more likely to be detected than sources above threshold are to be missed" [AG-2024.04-481]. This matters for thrust because a true zero signal corrupted by noise can fluctuate in either direction, but selection biases the *upward* fluctuations more than the downward ones.

More concretely, if your selection rule is "accept measurements above threshold T," then:

• A true zero-thrust system will occasionally produce noise that pushes it above T by chance, and you'll record it as a positive detection.
• A true small-positive thrust system will occasionally dip below T due to noise, and you'll miss it entirely.

The result is an overestimation of signals in whichever direction the threshold naturally favors—typically the positive direction if you're searching for a new effect.

Asymmetric Overestimation

This bias has a directional feature: amplitude parameters tend to be overestimated when their true value is nonzero, but show no false deviation when the signal is truly absent [AG-2025.01-657]. So for thrust measurements:

• If true thrust is zero: postselection introduces spurious positive detections (one-sided bias).
• If true thrust is small and positive: postselection overestimates it.
• If true thrust is genuinely absent: you'll see false positives, not false negatives.

This creates an important asymmetry: negative thrust measurements (if you're searching for them) won't be amplified the same way, because selection naturally favors the direction aligned with your detection threshold.

Why This Matters

For frontier propulsion claims, this directional bias is critical. If your analysis selects events above an SNR threshold, you'll preferentially record and overestimate positive signals while missing small negative ones—a form of one-sided systematic error that mimics a real effect even when none exists [AG-2024.04-481]. Properly accounting for selection effects (either in the likelihood or post-hoc, with care [AG-2024.05-183]) is essential to avoid false claims of asymmetric propulsion.