Confounding Bias

The defining graphical structure of confounding is a fork: an unobserved variable Z causally influences both the input X and the output Y. This creates a non-causal statistical association between X and Y that vanishes when conditioning on Z.

• Example: In RF fingerprinting, a specific power amplifier non-linearity (X) may appear highly predictive of a device ID (Y). However, a common cause like ambient temperature (Z) affects both the amplifier's behavior and the transmitter's oscillator drift, creating a spurious correlation.
• Key Insight: The model learns a correlation that is not invariant. When temperature changes in deployment, the learned association breaks, causing silent performance degradation.

Confounding inflates or deflates the perceived importance of a feature, misleading interpretability methods like SHAP or permutation feature importance. A confounded feature may appear highly significant, diverting engineering attention from the true causal drivers of system behavior.

• Mechanism: The model uses the confounded feature as a proxy for the unobserved confounder. The feature's importance score reflects the confounder's influence, not the feature's direct causal effect.
• RF Consequence: An engineer might waste resources optimizing a power amplifier parameter that SHAP flags as critical, when the true root cause of performance variance is an unmonitored power supply ripple affecting multiple components simultaneously.

A classic manifestation of confounding where a trend appears in several different groups of data but disappears or reverses when these groups are aggregated. This occurs when a confounding variable defines the group membership.

• Example: A cognitive radio model might show that increasing gain improves signal detection accuracy in each individual frequency band. However, when data from all bands is pooled, the overall trend shows accuracy decreasing with gain. The confounder is band-specific noise floor, which dictates both the optimal gain setting and the baseline detection difficulty.
• Diagnostic: Always disaggregate performance metrics by known operational modes, frequency bands, or environmental conditions to check for reversals.

From purely observational RF data, it is statistically impossible to distinguish a confounded relationship from a direct causal one without external knowledge or experimental intervention. The same joint probability distribution P(X, Y) can be generated by multiple causal graphs.

• Implication: A model trained on passively collected spectrum data cannot learn causal relationships. It learns associations that are a mixture of direct effects, reverse causation, and confounding.
• Mitigation: This necessitates randomized controlled experiments (e.g., intentionally varying transmission power while holding other factors constant) or strong, physics-based causal assumptions encoded via do-calculus to isolate the true effect of a signal parameter on receiver performance.

A closely related pitfall where conditioning on a common effect of two variables creates a spurious association between them. This is distinct from confounding but equally dangerous in model training.

• Example: A signal is successfully decoded only if it has both high SNR (X) and low interference (Y). If a model is trained only on successfully decoded signals (conditioning on the collider), it will learn a negative correlation between SNR and interference. In the real world, these are independent. The model incorrectly learns that high SNR implies low interference.
• RF Context: Training exclusively on high-quality, successfully demodulated signals introduces collider bias, making the model brittle when deployed on marginal or corrupted signals.

The primary graphical method for identifying a sufficient set of variables to condition on to remove confounding bias. A set of variables S satisfies the backdoor criterion relative to (X, Y) if no node in S is a descendant of X and S blocks every path between X and Y that has an arrow pointing into X.

• Application: In a digital pre-distortion (DPD) system, to estimate the true causal effect of a pre-distorter coefficient (X) on adjacent channel leakage ratio (Y), one must condition on the power amplifier's temperature and age, as these are common causes of both the chosen coefficient and the amplifier's non-linearity.
• Formula: P(Y | do(X)) = Σ_z P(Y | X, Z=z) P(Z=z). This adjustment provides a causally valid estimate from observational data if the correct backdoor set Z is measured.

Confounding bias occurs when a spurious correlation is learned between a signal feature and a label due to an unobserved confounder that causally drives both. In RF machine learning, this manifests when the model latches onto an environmental artifact—such as channel condition, hardware temperature, or transmitter location—rather than the true signal characteristic of interest. The result is a model that performs well in training but fails catastrophically when the confounder distribution shifts in deployment.

A common confounding trap in automatic modulation classification: a dataset is collected where high-order QAM signals are always recorded at high signal-to-noise ratio (SNR) in a lab, while BPSK signals are captured at low SNR in the field. The model learns to classify based on noise floor characteristics rather than modulation-specific features like the cyclostationary signature or cumulant patterns. When deployed, a high-SNR BPSK signal is misclassified as QAM because the confounder—SNR—was the true driver of the learned decision boundary.

The confounding relationship follows a specific directed acyclic graph (DAG) structure:

• Confounder (Z): An unobserved variable, such as receiver gain setting or ambient interference level
• Input Feature (X): The IQ sample or spectral representation fed to the model
• Output Label (Y): The classification target, such as modulation scheme or emitter identity

Z causally influences both X and Y, creating a backdoor path that induces a non-causal correlation. Without intervention, the model exploits this path rather than learning the true X→Y relationship. This is distinct from selection bias or covariate shift, which involve distributional mismatches rather than causal distortion.

Identifying confounding bias in RF models requires deliberate diagnostic approaches:

• Stratified evaluation: Test model performance within each confounder stratum (e.g., per-SNR bucket) to reveal hidden dependencies
• Feature attribution auditing: Apply SHAP or Integrated Gradients to verify that high-importance features align with known physical mechanisms, not environmental artifacts
• Adversarial confounder perturbation: Synthetically vary suspected confounders while holding the signal constant to observe prediction instability
• Causal discovery algorithms: Use constraint-based methods like the PC algorithm on model activations to infer potential confounding structures from observational data

Addressing confounding bias requires both data and architectural interventions:

• Data balancing: Ensure confounder distributions are uniform across all classes during collection, preventing the model from using confounders as shortcuts
• Domain randomization: In simulation-based training, deliberately vary channel impairments, noise profiles, and hardware parameters to force the model to learn invariant features
• Causal representation learning: Architect models with disentangled representations that separate confounder-driven variation from signal-intrinsic features
• Instrumental variable techniques: When a valid instrument exists—a variable that influences X but is independent of the confounder—use two-stage methods to recover the true causal effect

Confounding bias is the causal mechanism behind spurious feature learning. In RF fingerprinting, a model may appear to identify a transmitter by its unique hardware impairments but is actually keying off the carrier frequency offset induced by a shared local oscillator configuration in the collection setup. This is a confounded relationship: the collection geometry (Z) influences both the measured signal (X) and the transmitter label (Y). The learned feature is not a true emitter fingerprint but an artifact of the measurement process. Distinguishing genuine physical-layer signatures from confounded artifacts is the central challenge of robust RFML deployment.