Clinical researchers often want the adjusted effect of one variable on an outcome: how much serum albumin changes sepsis mortality once age, kidney-injury stage, and liver markers are held fixed. Answering this normally means pooling every patient's full record in one place. Often the variable of interest sits somewhere else, such as a separate laboratory system, an implanted sensor, or a partner hospital. That site can only push a few bits per patient toward the analyst and can never hear back. The usual response is to quantize the variable crudely and live with the lost precision. This note argues for the opposite view. The same collinearity that makes a risk factor hard to isolate also makes it cheap to transmit, because the analyst's server can already partly predict the remote variable from the confounders it holds.
The mechanism comes from distributed source coding, the study of compressing one signal when the receiver holds a correlated one. Its central trick is binning: the sender transmits only which "bin" its value falls in, where each bin groups many widely separated values. The receiver then picks the member of that bin most consistent with what it already knows. In our setting the remote site sends a dithered modulo index of the exposure, meaning its value wrapped around a small range with shared random noise added. The server unwraps it using a prediction built from the confounders and, in a soft decoding step, the outcome itself. We claim that this makes the cost, in bits per patient, of keeping the adjusted odds ratio's confidence interval nearly as tight as with pooled data independent of how collinear the exposure is. Quantizers that ignore the server's data pay an extra half-log of the variance inflation factor on every patient.
Core Claim
Let $X_j$ be the remote exposure, $Z$ the server's confounders, and $Y$ the binary outcome. The precision of the adjusted log-odds coefficient $\theta_j$ depends only on the part of $X_j$ that $Z$ cannot explain, whose variance is $\sigma^2_{j|Z} = \sigma_j^2(1-R_j^2)$. Here $\sigma_j^2$ is the marginal variance of $X_j$ and $R_j^2$ is the fraction of that variance explained by $Z$. Quantization adds noise of variance $D$ to $X_j$, and $D$ competes with this residual signal. Standard high-resolution quantization theory then gives the two cases below, where $k$ is bits per patient and $c$ is a constant that includes the scalar space-filling loss:
$$D_{\text{no side info}} \approx c\,\sigma_j^2\,2^{-2k}, \qquad D_{\text{binning}} \approx c\,\sigma_{j|Z}^2\,2^{-2k}.$$
The measurement-error inflation that matters, $D/\sigma^2_{j|Z}$, therefore equals $c\,2^{-2k}\,\mathrm{VIF}_j$ without side information and $c\,2^{-2k}$ with binning. The variance inflation factor is $\mathrm{VIF}_j = 1/(1-R_j^2)$. Binning thus saves $\tfrac12\log_2 \mathrm{VIF}_j$ bits per patient. More to the point, its efficiency relative to pooled analysis stops depending on collinearity. This is a claim about inference on one coefficient, including its bias, standard error, and interval coverage. Reconstruction error of the raw data is not the target, and that is what separates this proposal from the source-coding work it borrows from.
The claim only holds if one practical problem is solved. Under a one-way link the server cannot report a wrong unwrap. Occasional large decoding errors act like gross measurement outliers, and they bias odds ratios toward or away from the null. Our second ingredient is therefore decoding that is aware of the downstream task. The server does not pick the single most likely bin member. It keeps the few nearest candidates and fits the logistic model by expectation-maximization over them. This uses both $p(x_j \mid z)$ and $p(y \mid x_j, z; \theta)$, so the outcome helps resolve ambiguous bins. Because the dither comes from shared randomness, the remaining error is independent of $X_j$, and a standard errors-in-variables correction then applies exactly.
Relation to Prior Work
Three papers supply the method and show where it stops. Learned Wyner-Ziv Compressors Recover Binning trains neural compressors on a one-dimensional Gaussian source with receiver side information. It finds that they discover periodic binning on their own and sit about 1.53 dB from the theoretical bound, which is the scalar space-filling loss. That tells us scalar binning already captures nearly all of the gain, and it absorbs the constant $c$ above. It measures only reconstruction error of synthetic data, whereas we measure coefficient inference on clinical tables. Wyner-Ziv Estimators introduces the Rotated Modulo Quantizer (RMQ), which sends modulo values that the server unwraps with a correlated vector, and uses shared random thresholds when the distance to the side information is unknown. We reuse its modulo-and-unwrap core and its correlated-sampling idea. RMQ assumes side information in the same space as the source and targets mean estimation. Our side information is a different set of variables that predicts the exposure through a learned regression, and our target is one adjusted coefficient. Universal interactive Gaussian quantization with side information shows that a one-way, layered polar-lattice scheme reaches the Wyner-Ziv bound when the decoder knows the noise level. We borrow its layered structure for robustness, but we need no Gaussianity: clinical labs are skewed and mixed-type.
Minimax Bounds for Distributed Logistic Regression constrains the framing rather than supplying the method. It proves that when each node holds a whole sample and sends $k$ bits one-way with no side information, parameter error is at least of order $d^2/(kn\sigma^2)$ for sub-Gaussian data, and degrades to $2^{-k}$ scaling under finite second moments. Our setting splits each patient's variables across sites and gives the server side information, so that bound does not apply directly. Its sensitivity to heavy tails is a warning we test explicitly, since skewed labs like bilirubin could break a modulo scheme. On the application side, An Interpretable ICU Mortality Risk Calculator and the SA-AKI MIMIC study provide centralized L1 and multivariate logistic pipelines with reported odds ratios. These serve as our pooled-data reference, and neither considers any communication constraint.
The secret-key literature (Secret-Key Generation Using Correlated Sources and Channels, Secret Key Agreement: General Capacity and Second-Order Asymptotics) shows that one-way protocols must over-provision their messages against the receiver's worst-case uncertainty. That is precisely the cost of choosing a modulo range without knowing $\sigma_{j|Z}$. We measure this cost directly rather than assume interaction away.
Research Program
We will build a one-way protocol with three parts. At the remote site, a shared-seed dither and a modulo scalar quantizer produce $k$ bits per patient, with an optional coarse layer carrying the top bits so that robustness can be tuned. Continuous skewed labs are transformed monotonically (log or rank-Gaussian) before binning, and ordinal variables are binned on their integer grid. At the server, a conditional density model $p(x_j\mid z)$ is fit on a small paired calibration set; gradient boosting with a quantile or mixture head is the default. Candidate lists are formed from the bin structure, and the EM-based logistic fit runs over them with sandwich standard errors. As the analysis part, we will turn the heuristic in the Core Claim into a statement for the Gaussian-design logistic case. It will bound the asymptotic relative efficiency of $\hat\theta_j$ as a function of $k$, $R_j^2$, and wrap probability, and will say when the collinearity-independence breaks. The expected break is when the calibration set is too small to estimate $\sigma_{j|Z}$, or when tails make wrap errors frequent. We will also study the case where the modulo range is chosen from a misestimated $\sigma_{j|Z}$, which is the open "unknown correlation" problem that Jha and Tyagi leave. Our remedy is a two-layer message whose coarse layer caps the damage.
Baselines
The naive reference is the centralized, uncompressed multivariate logistic regression on pooled records. It is the ceiling every method is measured against. On the SA-AKI cohort it reports an albumin odds ratio of 0.606, bilirubin 1.055, AKI Stage 2 1.470, and a combined AUC of 0.796 (SA-AKI MIMIC study). On the larger ICU cohort, L1 logistic regression reports AUC 0.80 (ICU risk calculator). We will rerun both pipelines ourselves so that the cohort definitions, imputation, and covariate sets match exactly across all methods.
The "do nothing clever" communication baseline is a uniform scalar quantizer on the exposure's marginal range with $k$ bits. The server fits logistic regression on the reconstructed values with no side information. This is the regime analyzed by Minimax Bounds for Distributed Logistic Regression, for which only lower bounds, not empirical numbers, are reported.
The strongest competitor we plan is side-information-free quantization plus a classical measurement-error correction. The site sends a dithered uniform $k$-bit quantization, and the server applies regression calibration, replacing $X_j$ with $E[X_j \mid \text{quantized value}, Z]$ estimated on the calibration set. This is the approach a biostatistician would reach for, and it already uses $Z$ at decoding time. What it lacks is binning at encoding time.
The strongest existing binning method is RMQ from Wyner-Ziv Estimators, adapted so that its side information is the server's prediction $\hat x_j(z)$ and followed by hard unwrapping and ordinary logistic regression. That paper reports "orders of magnitude lower MSE" than side-information-free quantizers on synthetic vectors up to $d=1024$, but no reported result was found on Lacuna for RMQ on logistic coefficient estimation or on clinical data.
Two ablations isolate our ingredients. The first swaps EM-over-candidates for hard unwrapping, which tests task-aware decoding. The second removes the outcome from the decoding posterior, which tests whether $Y$ helps resolve bins.
All methods receive the same bit budget per patient, the same calibration set, the same density model for $p(x_j\mid z)$ where they use one, and the same hyperparameter search budget. The modulo range and dither scale are tuned on calibration data only. We count as beating a baseline only this: reaching 95% relative efficiency against pooled analysis (ratio of standard-error squares) with interval coverage within one point of nominal, at fewer bits per patient than the baseline needs. The gap must also grow with $\mathrm{VIF}_j$ roughly as $\tfrac12\log_2\mathrm{VIF}_j$.
Why not just…?
Why not just quantize locally and correct for measurement error at the server? This is the real threat. At high bit rates, regression calibration may recover almost everything, because once $D \ll \sigma^2_{j|Z}$ it hardly matters how the error arose. The theory predicts that its bit cost exceeds binning's by $\tfrac12\log_2 \mathrm{VIF}_j$ whatever the correction, because correction cannot restore residual information the quantizer discarded. For a variable with $R_j^2 = 0.9$, that is about 1.7 bits per patient. The synthetic sweep over $R_j^2$ settles this directly. If the measured gap is under half a bit at realistic collinearity, the proposal reduces to recommending regression calibration, and we will report exactly that.
Why not just screen variables at the remote site and send only the significant ones? The remote site holds neither the confounders nor, in our partition, the outcome, so it has nothing to screen against. Even where it could screen, marginal screening is exactly what multivariate isolation exists to avoid. In the SA-AKI cohort, single markers reach AUC 0.674 (BUN) and 0.665 (AKI stage) against 0.796 for the adjusted model (SA-AKI MIMIC study).
Why not just allow two-way communication, as in federated averaging? Interaction does beat one-way protocols in theory (Secret-Key Generation Using Compound Sources and One-Way Public Communication). But the regimes that motivate this work, such as low-power implants, broadcast-only sensors, and data-diode hospital links, cannot receive messages. We do not claim to beat interactive methods. Instead we report the gap to an idealized interactive oracle, one that knows $\sigma_{j|Z}$ exactly, to show what the one-way constraint costs.
Why not just train an end-to-end neural compressor on the task loss? Learned Wyner-Ziv Compressors Recover Binning suggests a learned scalar compressor would rediscover modulo binning and land within the same space-filling constant. It would add training instability without guaranteed error independence, and that independence is what makes our correction exact. We include a small learned variant on the synthetic benchmark to confirm that it matches but does not beat the hand-built quantizer.
Proposed Experiments
The first experiment is synthetic. We generate logistic data in which the exposure is a linear-plus-noise function of 5 to 20 confounders, with $R_j^2$ swept from 0 to 0.98. We cover Gaussian, sub-exponential, and log-normal exposures, the last mirroring the heavy-tail regime where Barnes and Özgür predict collapse. $k$ ranges from 1 to 8 bits and $n$ from $10^3$ to $10^5$. We report bias of $\hat\theta_j$, 95% interval coverage, relative efficiency against pooled data, and the bits per patient needed to reach 95% efficiency. The decisive check is the slope: required bits should be flat in $\mathrm{VIF}_j$ for our method and rise as $\tfrac12\log_2\mathrm{VIF}_j$ for the calibration baseline. If both are flat, the core claim is false.
The second experiment is semi-synthetic, using MIMIC-III v1.4. We reconstruct the SA-AKI cohort of 4,868 patients (SA-AKI MIMIC study) and simulate a vertical split. The server keeps age, AKI stage, AST, antibiotic delay, and mortality. A remote laboratory holds one exposure at a time: albumin (moderate collinearity), BUN (high collinearity with kidney-injury stage), or total bilirubin (heavily skewed, strongly tied to AST). Twenty percent of patients serve as the paired calibration set, and results are averaged over resampled splits. The decisive clinical check has two parts. The first is whether the protocol reproduces the pooled adjusted odds ratio and its interval, with the interval overlapping the pooled interval and its width within 5%, at 2 to 3 bits per patient where uniform quantization needs noticeably more. The second is whether bilirubin, with its skewed tail, shows any wrap-induced bias once the EM decoder is removed. We repeat the analysis on the 36,330-record ICU cohort (ICU risk calculator), where the target is recovering the same sparse set of selected predictors and an AUC near 0.80.
The third experiment tests robustness to unknown correlation. We misestimate $\sigma_{j|Z}$ by factors of 0.5 to 2, shrink the calibration set to 1%, and shift the conditional distribution between calibration and deployment by training on one admission era and deploying on another. We report how much the coarse layer recovers and at what added bit cost. That result tells us whether the one-way tax is a few tenths of a bit or large enough to make the scheme impractical.
References
- Learned Wyner-Ziv Compressors Recover Binning
- Wyner-Ziv Estimators: Efficient Distributed Mean Estimation with Side-Information
- Universal interactive Gaussian quantization with side information
- Minimax Bounds for Distributed Logistic Regression
- An Interpretable Intensive Care Unit Mortality Risk Calculator
- A study on the factors influencing mortality risk in sepsis-induced acute kidney injury based on analysis of the MIMIC database
- Secret-Key Generation using Correlated Sources and Channels
- Secret key agreement: General capacity and second-order asymptotics
- Secret-Key Generation Using Compound Sources and One-Way Public Communication