When Silence Is Informative: Treating Channel-Inversion Outages as a Designed Censoring Mechanism

Wireless sensors on industrial machines often use truncated channel inversion. The transmitter raises its power to cancel out fading, the random weakening of the radio signal, but it stops transmitting entirely whenever the channel is too weak to be worth the power. The engineering rationale is sound: inverting a deep fade would need unbounded power. The side effect is that a device can fall silent for long stretches. Fleet-monitoring pipelines then record "lost contact" as right-censoring, the survival-analysis convention for "we stopped observing this unit before it failed," and assume that losing contact says nothing about failure. That assumption breaks exactly when it matters. A machine that is degrading vibrates harder, runs hotter, and drains its battery, and each of these also degrades the radio link. The power-control rule then preferentially silences the units closest to failure. Every downstream survival model, whether Cox or not, is fit to a risk set that has been quietly cleaned of its riskiest members.

This note argues that this censoring is not an unknown nuisance. It is produced by a rule the engineer wrote down. Under Rayleigh fading, the standard model in which the channel power gain is exponentially distributed, the probability of an outage is a closed-form function of the device's mean channel gain and the chosen threshold. Because the exponential distribution is memoryless, the mean gain can be estimated without bias from the channel reports a device sends while it is transmitting. The server can therefore compute physics-derived censoring weights and correct a non-proportional-hazards survival model for outage-induced informative censoring. The same calculation also exposes a third cost of the truncation threshold, alongside power and interference: the statistical information lost from the failure process.

Core Claim

The proposal makes three linked claims, each checkable. First, when channel degradation is coupled to the failure process, treating loss-of-contact as non-informative censoring biases survival estimates. It also manufactures hazard ratios that drift over time, because degraded units leave the risk set early. This is the same depletion mechanism that makes marginal hazard ratios decline under unobserved frailty in Neural Frailty Machine. An analyst would therefore see apparent non-proportional hazards that are an artifact of the radio, not of the machine.

Second, the censoring mechanism can be identified from physical-layer quantities. Write device $k$'s channel power gain in slot $s$ as $h_{k,s}$, exponential with slowly varying mean $\bar\gamma_k(t)$, and write $h_{\text{th}}$ for the truncation threshold. A slot is silent with probability

$$p_k(t) = 1 - e^{-h_{\text{th}}/\bar\gamma_k(t)}.$$

If the operator declares a device lost after $L$ consecutive silent slots, the per-slot probability that a censoring run starts is approximately $(1-p_k)p_k^{L}$. The quantity $\bar\gamma_k(t)$ is recoverable from transmitted slots because $E[h - h_{\text{th}} \mid h \ge h_{\text{th}}] = \bar\gamma_k(t)$ under the exponential law. The server can therefore build inverse-probability-of-censoring weights, which upweight observed units to stand in for similar units that were censored. It builds them from a two-parameter physical model rather than a black-box classifier.

Third, the threshold $h_{\text{th}}$ should be chosen with the variance of these weights in view. A high threshold saves power but inflates the weights and shrinks the effective sample size.

Relation to Prior Work

The truncation literature supplies the mechanism and treats outages as exogenous. ElSawy et al. model the threshold as a trade-off between received signal quality and outage probability. Ali et al. show truncation acts as a decentralized interference limiter. Zhong et al. show it stays energy-efficient with stale channel estimates. None of them asks what an outage reveals about the transmitter's own state, and that is the question here.

The over-the-air learning papers are the closest systems work. AirFL-Mem and Digital versus Analog Transmissions treat truncation as dropping gradient coordinates. They fix the dropping with long-term memory or study it under convex losses on MNIST. In both, the channel is assumed independent of the data. The present proposal is the opposite case, where the channel is correlated with the label.

On the survival side, When Does IPCW Help? motivates the work. It shows inverse-probability-of-censoring weighting matters for relative effects mainly under differentially informative censoring, and that weight truncation is needed to control variance. It also constrains the design, because any weighting scheme must report effective sample size. A Federated Cox Model with Non-Proportional Hazards and MTLR supply the downstream model. Both are discrete-time models whose per-interval covariate weights allow crossing hazards, and their loss is a sum over intervals that accepts per-observation weights. Gao et al. supply restricted mean survival time, the area under the survival curve up to a horizon, as an estimand that stays meaningful when hazards cross. Jiang et al. show that hazard ratios under non-proportionality are not transitive. This is why the evaluation targets survival curves and restricted mean survival time rather than a single hazard ratio.

Research Program

The work has three parts.

The first part is a bias characterization. Couple $\bar\gamma_k(t)$ to the latent failure process through a single strength parameter $\alpha$, with $\alpha = 0$ meaning independence. Then derive and simulate how the naive Kaplan–Meier curve, the naive Cox hazard ratio between device classes, and a discrete-time non-proportional model drift as $\alpha$ grows. The specific prediction is that the naive hazard ratio declines over time even when the true conditional hazards are proportional, mirroring the frailty depletion effect.

The second part is the estimator. Each device appends its measured gain to every packet it sends, which costs a few bits. From these reports the server fits a smoothed $\bar\gamma_k(t)$ using the memoryless excess-over-threshold estimator together with the observed transmission rate. It then computes the censoring survival function from the closed form above and fits an MTLR-style discrete-time model with stabilized, truncated inverse-probability-of-censoring weights. A doubly robust variant also adds silence-run length as a time-varying covariate. This hedges against misspecifying the channel model, which matters for Rician fading or shadowing, where memorylessness fails. In that case the estimator falls back to a parametric fit of the truncated distribution.

The third part is threshold design. Add the variance of the weighted restricted-mean-survival-time estimator as a third axis to the power–interference trade-off, and compute the Pareto frontier over $h_{\text{th}}$ and the timeout $L$.

Baselines

The naive reference point treats every loss-of-contact as non-informative censoring and fits the same discrete-time non-proportional model with uniform weights. This is current practice, and any method that fails to beat it under strong coupling has no reason to exist.

The oracle uses the same model, trained on the same units with the channel truncation removed from the observation process, so every unit is followed until its true failure or administrative end of study. It is the ceiling.

For real-data context only, the reported performance of the downstream model class under ideal communication is useful. A Federated Cox Model with Non-Proportional Hazards reports a C-index of about 66.7 on METABRIC for its non-proportional network, against 63.5 for the linear proportional model. MTLR reports a 12-month Brier score of 0.158 on the Alberta registry. Our oracle on METABRIC should land near the first figure. These are not results under informative truncation; I found no reported result on Lacuna for any survival estimator under wireless-induced informative censoring.

The strongest real competitor is generic inverse-probability-of-censoring weighting in the style of When Does IPCW Help?. A flexible censoring model, a gradient-boosted discrete-time hazard, is fit on everything the server observes, including the reported channel gains and transmission rates. It receives the same inputs, the same weight-truncation and stabilization rules, and the same tuning budget, with hyperparameters selected by the same validation split as our method. Beating it means lower absolute bias in restricted mean survival time at the chosen horizon, and coverage closer to nominal, at matched effective sample size. The advantage should be clearest when few lost devices are available to fit the censoring model.

The key ablation swaps our physical censoring model for that learned model while holding the weighting and survival model fixed. This isolates whether the physics, rather than merely having access to channel information, does the work.

A last baseline is store-and-forward. Devices buffer data locally during outages and upload it once the link recovers. Whenever a device survives its outage this removes censoring, but whenever the device fails during the outage the buffer is lost.

Why not just…?

Why not just use standard IPCW with a learned censoring model? This is the real threat. If the learned model is given the same channel features, it may match the physical model in large, well-mixed fleets. The ablation above settles it. We expect the physical model to win when lost-device events are rare, because it has two parameters rather than hundreds. We also expect it to win under extrapolation, for example a change of threshold between training and deployment, because the closed form transfers and a learned model does not. If the learned model matches on both counts, the honest conclusion is that the contribution reduces to "log the channel gain and include it in the censoring model." That is still useful, and the bias characterization still stands.

Why not just buffer data and upload after the outage? Store-and-forward fixes survivors but not failures during outages, which are exactly the informative cases. It also delays alarms in maintenance. It is included as a baseline. We will report the fraction of failures that occur mid-outage as a function of $\alpha$, because that fraction determines how much residual bias store-and-forward leaves.

Why not just lower the threshold or use full inversion so outages vanish? Full inversion needs unbounded expected power under Rayleigh fading. Lowering the threshold raises average power and network interference, as ElSawy et al. model. The threshold-design study makes this explicit by reporting how much power must be spent to push naive bias below a tolerance, compared with how much our correction needs.

Why not just add silence as a covariate, or model it as a competing risk? Including silence-run length as a covariate in a model like Neural Frailty Machine improves prediction but does not by itself restore an unbiased population survival curve. Once a unit is censored, its later hazard is still unobserved. Silence is therefore used inside the doubly robust variant, and the covariate-only version is reported separately.

Proposed Experiments

No public dataset pairs wireless telemetry with failure times, so the evaluation is semi-synthetic, with real event processes and simulated radio. In the fully synthetic arm, discrete-time event times come from the EventTimesSampler in pydts. The truth is set to be crossing-hazard and non-proportional, so the oracle survival curves are known exactly. In the semi-synthetic arm, covariates and event or censoring times come from SUPPORT, METABRIC, GBSG and, for scale, KKBOX. Each unit is assigned a simulated Rayleigh channel whose mean gain declines toward its true event time with strength $\alpha$. Truncation with timeout $L$ is then applied on top of the dataset's own censoring.

The sweep covers $\alpha$, $h_{\text{th}}$, $L$, and the fading law (Rayleigh versus Rician and log-normal shadowing). Metrics are absolute bias and confidence-interval coverage of restricted mean survival time at a fixed horizon, the drift over time of the naive hazard ratio between two device classes, the censoring-adjusted integrated Brier score, C-index, effective sample size, and average transmit power.

Three checks are decisive. At $\alpha = 0$, all estimators must agree with the oracle. If they do not, the pipeline is wrong. As $\alpha$ grows, the naive estimator's restricted-mean-survival-time bias should grow monotonically, and its hazard ratio should drift even though the simulated conditional hazards are proportional. Finally, the physics-weighted estimator should stay within the oracle's confidence interval, and it should beat learned censoring weights when lost-device counts are small and when the threshold changes between training and test. Under Rician fading we expect the closed form to degrade and the doubly robust variant to recover. Reporting where the physical model breaks is part of the result.